drivers/infiniband/ulp/isert/ib_isert.c

   1 /*******************************************************************************
   2  * This file contains iSCSI extentions for RDMA (iSER) Verbs
   3  *
   4  * (c) Copyright 2013 Datera, Inc.
   5  *
   6  * Nicholas A. Bellinger <nab@linux-iscsi.org>
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License as published by
  10  * the Free Software Foundation; either version 2 of the License, or
  11  * (at your option) any later version.
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  ****************************************************************************/
  18
  19 #include <linux/string.h>
  20 #include <linux/module.h>
  21 #include <linux/scatterlist.h>
  22 #include <linux/socket.h>
  23 #include <linux/in.h>
  24 #include <linux/in6.h>
  25 #include <rdma/ib_verbs.h>
  26 #include <rdma/rdma_cm.h>
  27 #include <target/target_core_base.h>
  28 #include <target/target_core_fabric.h>
  29 #include <target/iscsi/iscsi_transport.h>
  30 #include <linux/semaphore.h>
  31
  32 #include "isert_proto.h"
  33 #include "ib_isert.h"
  34
  35 #define ISERT_MAX_CONN          8
  36 #define ISER_MAX_RX_CQ_LEN      (ISERT_QP_MAX_RECV_DTOS * ISERT_MAX_CONN)
  37 #define ISER_MAX_TX_CQ_LEN      (ISERT_QP_MAX_REQ_DTOS  * ISERT_MAX_CONN)
  38 #define ISER_MAX_CQ_LEN         (ISER_MAX_RX_CQ_LEN + ISER_MAX_TX_CQ_LEN + \
  39                                  ISERT_MAX_CONN)
  40
  41 static int isert_debug_level;
  42 module_param_named(debug_level, isert_debug_level, int, 0644);
  43 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0 (default:0)");
  44
  45 static DEFINE_MUTEX(device_list_mutex);
  46 static LIST_HEAD(device_list);
  47 static struct workqueue_struct *isert_comp_wq;
  48 static struct workqueue_struct *isert_release_wq;
  49
  50 static void
  51 isert_unmap_cmd(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn);
  52 static int
  53 isert_map_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
  54                struct isert_rdma_wr *wr);
  55 static void
  56 isert_unreg_rdma(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn);
  57 static int
  58 isert_reg_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
  59                struct isert_rdma_wr *wr);
  60 static int
  61 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd);
  62 static int
  63 isert_rdma_post_recvl(struct isert_conn *isert_conn);
  64 static int
  65 isert_rdma_accept(struct isert_conn *isert_conn);
  66 struct rdma_cm_id *isert_setup_id(struct isert_np *isert_np);
  67
  68 static void isert_release_work(struct work_struct *work);
  69
  70 static inline bool
  71 isert_prot_cmd(struct isert_conn *conn, struct se_cmd *cmd)
  72 {
  73         return (conn->pi_support &&
  74                 cmd->prot_op != TARGET_PROT_NORMAL);
  75 }
  76
  77
  78 static void
  79 isert_qp_event_callback(struct ib_event *e, void *context)
  80 {
  81         struct isert_conn *isert_conn = context;
  82
  83         isert_err("%s (%d): conn %p\n",
  84                   ib_event_msg(e->event), e->event, isert_conn);
  85
  86         switch (e->event) {
  87         case IB_EVENT_COMM_EST:
  88                 rdma_notify(isert_conn->cm_id, IB_EVENT_COMM_EST);
  89                 break;
  90         case IB_EVENT_QP_LAST_WQE_REACHED:
  91                 isert_warn("Reached TX IB_EVENT_QP_LAST_WQE_REACHED\n");
  92                 break;
  93         default:
  94                 break;
  95         }
  96 }
  97
  98 static int
  99 isert_query_device(struct ib_device *ib_dev, struct ib_device_attr *devattr)
 100 {
 101         int ret;
 102
 103         ret = ib_query_device(ib_dev, devattr);
 104         if (ret) {
 105                 isert_err("ib_query_device() failed: %d\n", ret);
 106                 return ret;
 107         }
 108         isert_dbg("devattr->max_sge: %d\n", devattr->max_sge);
 109         isert_dbg("devattr->max_sge_rd: %d\n", devattr->max_sge_rd);
 110
 111         return 0;
 112 }
 113
 114 static struct isert_comp *
 115 isert_comp_get(struct isert_conn *isert_conn)
 116 {
 117         struct isert_device *device = isert_conn->device;
 118         struct isert_comp *comp;
 119         int i, min = 0;
 120
 121         mutex_lock(&device_list_mutex);
 122         for (i = 0; i < device->comps_used; i++)
 123                 if (device->comps[i].active_qps <
 124                     device->comps[min].active_qps)
 125                         min = i;
 126         comp = &device->comps[min];
 127         comp->active_qps++;
 128         mutex_unlock(&device_list_mutex);
 129
 130         isert_info("conn %p, using comp %p min_index: %d\n",
 131                    isert_conn, comp, min);
 132
 133         return comp;
 134 }
 135
 136 static void
 137 isert_comp_put(struct isert_comp *comp)
 138 {
 139         mutex_lock(&device_list_mutex);
 140         comp->active_qps--;
 141         mutex_unlock(&device_list_mutex);
 142 }
 143
 144 static struct ib_qp *
 145 isert_create_qp(struct isert_conn *isert_conn,
 146                 struct isert_comp *comp,
 147                 struct rdma_cm_id *cma_id)
 148 {
 149         struct isert_device *device = isert_conn->device;
 150         struct ib_qp_init_attr attr;
 151         int ret;
 152
 153         memset(&attr, 0, sizeof(struct ib_qp_init_attr));
 154         attr.event_handler = isert_qp_event_callback;
 155         attr.qp_context = isert_conn;
 156         attr.send_cq = comp->cq;
 157         attr.recv_cq = comp->cq;
 158         attr.cap.max_send_wr = ISERT_QP_MAX_REQ_DTOS;
 159         attr.cap.max_recv_wr = ISERT_QP_MAX_RECV_DTOS + 1;
 160         attr.cap.max_send_sge = device->dev_attr.max_sge;
 161         isert_conn->max_sge = min(device->dev_attr.max_sge,
 162                                   device->dev_attr.max_sge_rd);
 163         attr.cap.max_recv_sge = 1;
 164         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 165         attr.qp_type = IB_QPT_RC;
 166         if (device->pi_capable)
 167                 attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
 168
 169         ret = rdma_create_qp(cma_id, device->pd, &attr);
 170         if (ret) {
 171                 isert_err("rdma_create_qp failed for cma_id %d\n", ret);
 172                 return ERR_PTR(ret);
 173         }
 174
 175         return cma_id->qp;
 176 }
 177
 178 static int
 179 isert_conn_setup_qp(struct isert_conn *isert_conn, struct rdma_cm_id *cma_id)
 180 {
 181         struct isert_comp *comp;
 182         int ret;
 183
 184         comp = isert_comp_get(isert_conn);
 185         isert_conn->qp = isert_create_qp(isert_conn, comp, cma_id);
 186         if (IS_ERR(isert_conn->qp)) {
 187                 ret = PTR_ERR(isert_conn->qp);
 188                 goto err;
 189         }
 190
 191         return 0;
 192 err:
 193         isert_comp_put(comp);
 194         return ret;
 195 }
 196
 197 static void
 198 isert_cq_event_callback(struct ib_event *e, void *context)
 199 {
 200         isert_dbg("event: %d\n", e->event);
 201 }
 202
 203 static int
 204 isert_alloc_rx_descriptors(struct isert_conn *isert_conn)
 205 {
 206         struct isert_device *device = isert_conn->device;
 207         struct ib_device *ib_dev = device->ib_device;
 208         struct iser_rx_desc *rx_desc;
 209         struct ib_sge *rx_sg;
 210         u64 dma_addr;
 211         int i, j;
 212
 213         isert_conn->rx_descs = kzalloc(ISERT_QP_MAX_RECV_DTOS *
 214                                 sizeof(struct iser_rx_desc), GFP_KERNEL);
 215         if (!isert_conn->rx_descs)
 216                 goto fail;
 217
 218         rx_desc = isert_conn->rx_descs;
 219
 220         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 221                 dma_addr = ib_dma_map_single(ib_dev, (void *)rx_desc,
 222                                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 223                 if (ib_dma_mapping_error(ib_dev, dma_addr))
 224                         goto dma_map_fail;
 225
 226                 rx_desc->dma_addr = dma_addr;
 227
 228                 rx_sg = &rx_desc->rx_sg;
 229                 rx_sg->addr = rx_desc->dma_addr;
 230                 rx_sg->length = ISER_RX_PAYLOAD_SIZE;
 231                 rx_sg->lkey = device->pd->local_dma_lkey;
 232         }
 233
 234         return 0;
 235
 236 dma_map_fail:
 237         rx_desc = isert_conn->rx_descs;
 238         for (j = 0; j < i; j++, rx_desc++) {
 239                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 240                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 241         }
 242         kfree(isert_conn->rx_descs);
 243         isert_conn->rx_descs = NULL;
 244 fail:
 245         isert_err("conn %p failed to allocate rx descriptors\n", isert_conn);
 246
 247         return -ENOMEM;
 248 }
 249
 250 static void
 251 isert_free_rx_descriptors(struct isert_conn *isert_conn)
 252 {
 253         struct ib_device *ib_dev = isert_conn->device->ib_device;
 254         struct iser_rx_desc *rx_desc;
 255         int i;
 256
 257         if (!isert_conn->rx_descs)
 258                 return;
 259
 260         rx_desc = isert_conn->rx_descs;
 261         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 262                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 263                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 264         }
 265
 266         kfree(isert_conn->rx_descs);
 267         isert_conn->rx_descs = NULL;
 268 }
 269
 270 static void isert_cq_work(struct work_struct *);
 271 static void isert_cq_callback(struct ib_cq *, void *);
 272
 273 static void
 274 isert_free_comps(struct isert_device *device)
 275 {
 276         int i;
 277
 278         for (i = 0; i < device->comps_used; i++) {
 279                 struct isert_comp *comp = &device->comps[i];
 280
 281                 if (comp->cq) {
 282                         cancel_work_sync(&comp->work);
 283                         ib_destroy_cq(comp->cq);
 284                 }
 285         }
 286         kfree(device->comps);
 287 }
 288
 289 static int
 290 isert_alloc_comps(struct isert_device *device,
 291                   struct ib_device_attr *attr)
 292 {
 293         int i, max_cqe, ret = 0;
 294
 295         device->comps_used = min(ISERT_MAX_CQ, min_t(int, num_online_cpus(),
 296                                  device->ib_device->num_comp_vectors));
 297
 298         isert_info("Using %d CQs, %s supports %d vectors support "
 299                    "Fast registration %d pi_capable %d\n",
 300                    device->comps_used, device->ib_device->name,
 301                    device->ib_device->num_comp_vectors, device->use_fastreg,
 302                    device->pi_capable);
 303
 304         device->comps = kcalloc(device->comps_used, sizeof(struct isert_comp),
 305                                 GFP_KERNEL);
 306         if (!device->comps) {
 307                 isert_err("Unable to allocate completion contexts\n");
 308                 return -ENOMEM;
 309         }
 310
 311         max_cqe = min(ISER_MAX_CQ_LEN, attr->max_cqe);
 312
 313         for (i = 0; i < device->comps_used; i++) {
 314                 struct ib_cq_init_attr cq_attr = {};
 315                 struct isert_comp *comp = &device->comps[i];
 316
 317                 comp->device = device;
 318                 INIT_WORK(&comp->work, isert_cq_work);
 319                 cq_attr.cqe = max_cqe;
 320                 cq_attr.comp_vector = i;
 321                 comp->cq = ib_create_cq(device->ib_device,
 322                                         isert_cq_callback,
 323                                         isert_cq_event_callback,
 324                                         (void *)comp,
 325                                         &cq_attr);
 326                 if (IS_ERR(comp->cq)) {
 327                         isert_err("Unable to allocate cq\n");
 328                         ret = PTR_ERR(comp->cq);
 329                         comp->cq = NULL;
 330                         goto out_cq;
 331                 }
 332
 333                 ret = ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP);
 334                 if (ret)
 335                         goto out_cq;
 336         }
 337
 338         return 0;
 339 out_cq:
 340         isert_free_comps(device);
 341         return ret;
 342 }
 343
 344 static int
 345 isert_create_device_ib_res(struct isert_device *device)
 346 {
 347         struct ib_device_attr *dev_attr;
 348         int ret;
 349
 350         dev_attr = &device->dev_attr;
 351         ret = isert_query_device(device->ib_device, dev_attr);
 352         if (ret)
 353                 return ret;
 354
 355         /* asign function handlers */
 356         if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS &&
 357             dev_attr->device_cap_flags & IB_DEVICE_SIGNATURE_HANDOVER) {
 358                 device->use_fastreg = 1;
 359                 device->reg_rdma_mem = isert_reg_rdma;
 360                 device->unreg_rdma_mem = isert_unreg_rdma;
 361         } else {
 362                 device->use_fastreg = 0;
 363                 device->reg_rdma_mem = isert_map_rdma;
 364                 device->unreg_rdma_mem = isert_unmap_cmd;
 365         }
 366
 367         ret = isert_alloc_comps(device, dev_attr);
 368         if (ret)
 369                 return ret;
 370
 371         device->pd = ib_alloc_pd(device->ib_device);
 372         if (IS_ERR(device->pd)) {
 373                 ret = PTR_ERR(device->pd);
 374                 isert_err("failed to allocate pd, device %p, ret=%d\n",
 375                           device, ret);
 376                 goto out_cq;
 377         }
 378
 379         /* Check signature cap */
 380         device->pi_capable = dev_attr->device_cap_flags &
 381                              IB_DEVICE_SIGNATURE_HANDOVER ? true : false;
 382
 383         return 0;
 384
 385 out_cq:
 386         isert_free_comps(device);
 387         return ret;
 388 }
 389
 390 static void
 391 isert_free_device_ib_res(struct isert_device *device)
 392 {
 393         isert_info("device %p\n", device);
 394
 395         ib_dealloc_pd(device->pd);
 396         isert_free_comps(device);
 397 }
 398
 399 static void
 400 isert_device_put(struct isert_device *device)
 401 {
 402         mutex_lock(&device_list_mutex);
 403         device->refcount--;
 404         isert_info("device %p refcount %d\n", device, device->refcount);
 405         if (!device->refcount) {
 406                 isert_free_device_ib_res(device);
 407                 list_del(&device->dev_node);
 408                 kfree(device);
 409         }
 410         mutex_unlock(&device_list_mutex);
 411 }
 412
 413 static struct isert_device *
 414 isert_device_get(struct rdma_cm_id *cma_id)
 415 {
 416         struct isert_device *device;
 417         int ret;
 418
 419         mutex_lock(&device_list_mutex);
 420         list_for_each_entry(device, &device_list, dev_node) {
 421                 if (device->ib_device->node_guid == cma_id->device->node_guid) {
 422                         device->refcount++;
 423                         isert_info("Found iser device %p refcount %d\n",
 424                                    device, device->refcount);
 425                         mutex_unlock(&device_list_mutex);
 426                         return device;
 427                 }
 428         }
 429
 430         device = kzalloc(sizeof(struct isert_device), GFP_KERNEL);
 431         if (!device) {
 432                 mutex_unlock(&device_list_mutex);
 433                 return ERR_PTR(-ENOMEM);
 434         }
 435
 436         INIT_LIST_HEAD(&device->dev_node);
 437
 438         device->ib_device = cma_id->device;
 439         ret = isert_create_device_ib_res(device);
 440         if (ret) {
 441                 kfree(device);
 442                 mutex_unlock(&device_list_mutex);
 443                 return ERR_PTR(ret);
 444         }
 445
 446         device->refcount++;
 447         list_add_tail(&device->dev_node, &device_list);
 448         isert_info("Created a new iser device %p refcount %d\n",
 449                    device, device->refcount);
 450         mutex_unlock(&device_list_mutex);
 451
 452         return device;
 453 }
 454
 455 static void
 456 isert_conn_free_fastreg_pool(struct isert_conn *isert_conn)
 457 {
 458         struct fast_reg_descriptor *fr_desc, *tmp;
 459         int i = 0;
 460
 461         if (list_empty(&isert_conn->fr_pool))
 462                 return;
 463
 464         isert_info("Freeing conn %p fastreg pool", isert_conn);
 465
 466         list_for_each_entry_safe(fr_desc, tmp,
 467                                  &isert_conn->fr_pool, list) {
 468                 list_del(&fr_desc->list);
 469                 ib_dereg_mr(fr_desc->data_mr);
 470                 if (fr_desc->pi_ctx) {
 471                         ib_dereg_mr(fr_desc->pi_ctx->prot_mr);
 472                         ib_dereg_mr(fr_desc->pi_ctx->sig_mr);
 473                         kfree(fr_desc->pi_ctx);
 474                 }
 475                 kfree(fr_desc);
 476                 ++i;
 477         }
 478
 479         if (i < isert_conn->fr_pool_size)
 480                 isert_warn("Pool still has %d regions registered\n",
 481                         isert_conn->fr_pool_size - i);
 482 }
 483
 484 static int
 485 isert_create_pi_ctx(struct fast_reg_descriptor *desc,
 486                     struct ib_device *device,
 487                     struct ib_pd *pd)
 488 {
 489         struct pi_context *pi_ctx;
 490         int ret;
 491
 492         pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
 493         if (!pi_ctx) {
 494                 isert_err("Failed to allocate pi context\n");
 495                 return -ENOMEM;
 496         }
 497
 498         pi_ctx->prot_mr = ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG,
 499                                       ISCSI_ISER_SG_TABLESIZE);
 500         if (IS_ERR(pi_ctx->prot_mr)) {
 501                 isert_err("Failed to allocate prot frmr err=%ld\n",
 502                           PTR_ERR(pi_ctx->prot_mr));
 503                 ret = PTR_ERR(pi_ctx->prot_mr);
 504                 goto err_pi_ctx;
 505         }
 506         desc->ind |= ISERT_PROT_KEY_VALID;
 507
 508         pi_ctx->sig_mr = ib_alloc_mr(pd, IB_MR_TYPE_SIGNATURE, 2);
 509         if (IS_ERR(pi_ctx->sig_mr)) {
 510                 isert_err("Failed to allocate signature enabled mr err=%ld\n",
 511                           PTR_ERR(pi_ctx->sig_mr));
 512                 ret = PTR_ERR(pi_ctx->sig_mr);
 513                 goto err_prot_mr;
 514         }
 515
 516         desc->pi_ctx = pi_ctx;
 517         desc->ind |= ISERT_SIG_KEY_VALID;
 518         desc->ind &= ~ISERT_PROTECTED;
 519
 520         return 0;
 521
 522 err_prot_mr:
 523         ib_dereg_mr(pi_ctx->prot_mr);
 524 err_pi_ctx:
 525         kfree(pi_ctx);
 526
 527         return ret;
 528 }
 529
 530 static int
 531 isert_create_fr_desc(struct ib_device *ib_device, struct ib_pd *pd,
 532                      struct fast_reg_descriptor *fr_desc)
 533 {
 534         fr_desc->data_mr = ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG,
 535                                        ISCSI_ISER_SG_TABLESIZE);
 536         if (IS_ERR(fr_desc->data_mr)) {
 537                 isert_err("Failed to allocate data frmr err=%ld\n",
 538                           PTR_ERR(fr_desc->data_mr));
 539                 return PTR_ERR(fr_desc->data_mr);
 540         }
 541         fr_desc->ind |= ISERT_DATA_KEY_VALID;
 542
 543         isert_dbg("Created fr_desc %p\n", fr_desc);
 544
 545         return 0;
 546 }
 547
 548 static int
 549 isert_conn_create_fastreg_pool(struct isert_conn *isert_conn)
 550 {
 551         struct fast_reg_descriptor *fr_desc;
 552         struct isert_device *device = isert_conn->device;
 553         struct se_session *se_sess = isert_conn->conn->sess->se_sess;
 554         struct se_node_acl *se_nacl = se_sess->se_node_acl;
 555         int i, ret, tag_num;
 556         /*
 557          * Setup the number of FRMRs based upon the number of tags
 558          * available to session in iscsi_target_locate_portal().
 559          */
 560         tag_num = max_t(u32, ISCSIT_MIN_TAGS, se_nacl->queue_depth);
 561         tag_num = (tag_num * 2) + ISCSIT_EXTRA_TAGS;
 562
 563         isert_conn->fr_pool_size = 0;
 564         for (i = 0; i < tag_num; i++) {
 565                 fr_desc = kzalloc(sizeof(*fr_desc), GFP_KERNEL);
 566                 if (!fr_desc) {
 567                         isert_err("Failed to allocate fast_reg descriptor\n");
 568                         ret = -ENOMEM;
 569                         goto err;
 570                 }
 571
 572                 ret = isert_create_fr_desc(device->ib_device,
 573                                            device->pd, fr_desc);
 574                 if (ret) {
 575                         isert_err("Failed to create fastreg descriptor err=%d\n",
 576                                ret);
 577                         kfree(fr_desc);
 578                         goto err;
 579                 }
 580
 581                 list_add_tail(&fr_desc->list, &isert_conn->fr_pool);
 582                 isert_conn->fr_pool_size++;
 583         }
 584
 585         isert_dbg("Creating conn %p fastreg pool size=%d",
 586                  isert_conn, isert_conn->fr_pool_size);
 587
 588         return 0;
 589
 590 err:
 591         isert_conn_free_fastreg_pool(isert_conn);
 592         return ret;
 593 }
 594
 595 static void
 596 isert_init_conn(struct isert_conn *isert_conn)
 597 {
 598         isert_conn->state = ISER_CONN_INIT;
 599         INIT_LIST_HEAD(&isert_conn->node);
 600         init_completion(&isert_conn->login_comp);
 601         init_completion(&isert_conn->login_req_comp);
 602         init_completion(&isert_conn->wait);
 603         kref_init(&isert_conn->kref);
 604         mutex_init(&isert_conn->mutex);
 605         spin_lock_init(&isert_conn->pool_lock);
 606         INIT_LIST_HEAD(&isert_conn->fr_pool);
 607         INIT_WORK(&isert_conn->release_work, isert_release_work);
 608 }
 609
 610 static void
 611 isert_free_login_buf(struct isert_conn *isert_conn)
 612 {
 613         struct ib_device *ib_dev = isert_conn->device->ib_device;
 614
 615         ib_dma_unmap_single(ib_dev, isert_conn->login_rsp_dma,
 616                             ISER_RX_LOGIN_SIZE, DMA_TO_DEVICE);
 617         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 618                             ISCSI_DEF_MAX_RECV_SEG_LEN,
 619                             DMA_FROM_DEVICE);
 620         kfree(isert_conn->login_buf);
 621 }
 622
 623 static int
 624 isert_alloc_login_buf(struct isert_conn *isert_conn,
 625                       struct ib_device *ib_dev)
 626 {
 627         int ret;
 628
 629         isert_conn->login_buf = kzalloc(ISCSI_DEF_MAX_RECV_SEG_LEN +
 630                                         ISER_RX_LOGIN_SIZE, GFP_KERNEL);
 631         if (!isert_conn->login_buf) {
 632                 isert_err("Unable to allocate isert_conn->login_buf\n");
 633                 return -ENOMEM;
 634         }
 635
 636         isert_conn->login_req_buf = isert_conn->login_buf;
 637         isert_conn->login_rsp_buf = isert_conn->login_buf +
 638                                     ISCSI_DEF_MAX_RECV_SEG_LEN;
 639
 640         isert_dbg("Set login_buf: %p login_req_buf: %p login_rsp_buf: %p\n",
 641                  isert_conn->login_buf, isert_conn->login_req_buf,
 642                  isert_conn->login_rsp_buf);
 643
 644         isert_conn->login_req_dma = ib_dma_map_single(ib_dev,
 645                                 (void *)isert_conn->login_req_buf,
 646                                 ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_FROM_DEVICE);
 647
 648         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_req_dma);
 649         if (ret) {
 650                 isert_err("login_req_dma mapping error: %d\n", ret);
 651                 isert_conn->login_req_dma = 0;
 652                 goto out_login_buf;
 653         }
 654
 655         isert_conn->login_rsp_dma = ib_dma_map_single(ib_dev,
 656                                         (void *)isert_conn->login_rsp_buf,
 657                                         ISER_RX_LOGIN_SIZE, DMA_TO_DEVICE);
 658
 659         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_rsp_dma);
 660         if (ret) {
 661                 isert_err("login_rsp_dma mapping error: %d\n", ret);
 662                 isert_conn->login_rsp_dma = 0;
 663                 goto out_req_dma_map;
 664         }
 665
 666         return 0;
 667
 668 out_req_dma_map:
 669         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 670                             ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_FROM_DEVICE);
 671 out_login_buf:
 672         kfree(isert_conn->login_buf);
 673         return ret;
 674 }
 675
 676 static int
 677 isert_connect_request(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 678 {
 679         struct isert_np *isert_np = cma_id->context;
 680         struct iscsi_np *np = isert_np->np;
 681         struct isert_conn *isert_conn;
 682         struct isert_device *device;
 683         int ret = 0;
 684
 685         spin_lock_bh(&np->np_thread_lock);
 686         if (!np->enabled) {
 687                 spin_unlock_bh(&np->np_thread_lock);
 688                 isert_dbg("iscsi_np is not enabled, reject connect request\n");
 689                 return rdma_reject(cma_id, NULL, 0);
 690         }
 691         spin_unlock_bh(&np->np_thread_lock);
 692
 693         isert_dbg("cma_id: %p, portal: %p\n",
 694                  cma_id, cma_id->context);
 695
 696         isert_conn = kzalloc(sizeof(struct isert_conn), GFP_KERNEL);
 697         if (!isert_conn)
 698                 return -ENOMEM;
 699
 700         isert_init_conn(isert_conn);
 701         isert_conn->cm_id = cma_id;
 702
 703         ret = isert_alloc_login_buf(isert_conn, cma_id->device);
 704         if (ret)
 705                 goto out;
 706
 707         device = isert_device_get(cma_id);
 708         if (IS_ERR(device)) {
 709                 ret = PTR_ERR(device);
 710                 goto out_rsp_dma_map;
 711         }
 712         isert_conn->device = device;
 713
 714         /* Set max inflight RDMA READ requests */
 715         isert_conn->initiator_depth = min_t(u8,
 716                                 event->param.conn.initiator_depth,
 717                                 device->dev_attr.max_qp_init_rd_atom);
 718         isert_dbg("Using initiator_depth: %u\n", isert_conn->initiator_depth);
 719
 720         ret = isert_conn_setup_qp(isert_conn, cma_id);
 721         if (ret)
 722                 goto out_conn_dev;
 723
 724         ret = isert_rdma_post_recvl(isert_conn);
 725         if (ret)
 726                 goto out_conn_dev;
 727
 728         ret = isert_rdma_accept(isert_conn);
 729         if (ret)
 730                 goto out_conn_dev;
 731
 732         mutex_lock(&isert_np->mutex);
 733         list_add_tail(&isert_conn->node, &isert_np->accepted);
 734         mutex_unlock(&isert_np->mutex);
 735
 736         return 0;
 737
 738 out_conn_dev:
 739         isert_device_put(device);
 740 out_rsp_dma_map:
 741         isert_free_login_buf(isert_conn);
 742 out:
 743         kfree(isert_conn);
 744         rdma_reject(cma_id, NULL, 0);
 745         return ret;
 746 }
 747
 748 static void
 749 isert_connect_release(struct isert_conn *isert_conn)
 750 {
 751         struct isert_device *device = isert_conn->device;
 752
 753         isert_dbg("conn %p\n", isert_conn);
 754
 755         BUG_ON(!device);
 756
 757         if (device->use_fastreg)
 758                 isert_conn_free_fastreg_pool(isert_conn);
 759
 760         isert_free_rx_descriptors(isert_conn);
 761         if (isert_conn->cm_id)
 762                 rdma_destroy_id(isert_conn->cm_id);
 763
 764         if (isert_conn->qp) {
 765                 struct isert_comp *comp = isert_conn->qp->recv_cq->cq_context;
 766
 767                 isert_comp_put(comp);
 768                 ib_destroy_qp(isert_conn->qp);
 769         }
 770
 771         if (isert_conn->login_buf)
 772                 isert_free_login_buf(isert_conn);
 773
 774         isert_device_put(device);
 775
 776         kfree(isert_conn);
 777 }
 778
 779 static void
 780 isert_connected_handler(struct rdma_cm_id *cma_id)
 781 {
 782         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 783         struct isert_np *isert_np = cma_id->context;
 784
 785         isert_info("conn %p\n", isert_conn);
 786
 787         mutex_lock(&isert_conn->mutex);
 788         isert_conn->state = ISER_CONN_UP;
 789         kref_get(&isert_conn->kref);
 790         mutex_unlock(&isert_conn->mutex);
 791
 792         mutex_lock(&isert_np->mutex);
 793         list_move_tail(&isert_conn->node, &isert_np->pending);
 794         mutex_unlock(&isert_np->mutex);
 795
 796         isert_info("np %p: Allow accept_np to continue\n", isert_np);
 797         up(&isert_np->sem);
 798 }
 799
 800 static void
 801 isert_release_kref(struct kref *kref)
 802 {
 803         struct isert_conn *isert_conn = container_of(kref,
 804                                 struct isert_conn, kref);
 805
 806         isert_info("conn %p final kref %s/%d\n", isert_conn, current->comm,
 807                    current->pid);
 808
 809         isert_connect_release(isert_conn);
 810 }
 811
 812 static void
 813 isert_put_conn(struct isert_conn *isert_conn)
 814 {
 815         kref_put(&isert_conn->kref, isert_release_kref);
 816 }
 817
 818 /**
 819  * isert_conn_terminate() - Initiate connection termination
 820  * @isert_conn: isert connection struct
 821  *
 822  * Notes:
 823  * In case the connection state is BOUND, move state
 824  * to TEMINATING and start teardown sequence (rdma_disconnect).
 825  * In case the connection state is UP, complete flush as well.
 826  *
 827  * This routine must be called with mutex held. Thus it is
 828  * safe to call multiple times.
 829  */
 830 static void
 831 isert_conn_terminate(struct isert_conn *isert_conn)
 832 {
 833         int err;
 834
 835         switch (isert_conn->state) {
 836         case ISER_CONN_TERMINATING:
 837                 break;
 838         case ISER_CONN_UP:
 839         case ISER_CONN_BOUND:
 840         case ISER_CONN_FULL_FEATURE: /* FALLTHRU */
 841                 isert_info("Terminating conn %p state %d\n",
 842                            isert_conn, isert_conn->state);
 843                 isert_conn->state = ISER_CONN_TERMINATING;
 844                 err = rdma_disconnect(isert_conn->cm_id);
 845                 if (err)
 846                         isert_warn("Failed rdma_disconnect isert_conn %p\n",
 847                                    isert_conn);
 848                 break;
 849         default:
 850                 isert_warn("conn %p teminating in state %d\n",
 851                            isert_conn, isert_conn->state);
 852         }
 853 }
 854
 855 static int
 856 isert_np_cma_handler(struct isert_np *isert_np,
 857                      enum rdma_cm_event_type event)
 858 {
 859         isert_dbg("%s (%d): isert np %p\n",
 860                   rdma_event_msg(event), event, isert_np);
 861
 862         switch (event) {
 863         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 864                 isert_np->cm_id = NULL;
 865                 break;
 866         case RDMA_CM_EVENT_ADDR_CHANGE:
 867                 isert_np->cm_id = isert_setup_id(isert_np);
 868                 if (IS_ERR(isert_np->cm_id)) {
 869                         isert_err("isert np %p setup id failed: %ld\n",
 870                                   isert_np, PTR_ERR(isert_np->cm_id));
 871                         isert_np->cm_id = NULL;
 872                 }
 873                 break;
 874         default:
 875                 isert_err("isert np %p Unexpected event %d\n",
 876                           isert_np, event);
 877         }
 878
 879         return -1;
 880 }
 881
 882 static int
 883 isert_disconnected_handler(struct rdma_cm_id *cma_id,
 884                            enum rdma_cm_event_type event)
 885 {
 886         struct isert_np *isert_np = cma_id->context;
 887         struct isert_conn *isert_conn;
 888         bool terminating = false;
 889
 890         if (isert_np->cm_id == cma_id)
 891                 return isert_np_cma_handler(cma_id->context, event);
 892
 893         isert_conn = cma_id->qp->qp_context;
 894
 895         mutex_lock(&isert_conn->mutex);
 896         terminating = (isert_conn->state == ISER_CONN_TERMINATING);
 897         isert_conn_terminate(isert_conn);
 898         mutex_unlock(&isert_conn->mutex);
 899
 900         isert_info("conn %p completing wait\n", isert_conn);
 901         complete(&isert_conn->wait);
 902
 903         if (terminating)
 904                 goto out;
 905
 906         mutex_lock(&isert_np->mutex);
 907         if (!list_empty(&isert_conn->node)) {
 908                 list_del_init(&isert_conn->node);
 909                 isert_put_conn(isert_conn);
 910                 queue_work(isert_release_wq, &isert_conn->release_work);
 911         }
 912         mutex_unlock(&isert_np->mutex);
 913
 914 out:
 915         return 0;
 916 }
 917
 918 static int
 919 isert_connect_error(struct rdma_cm_id *cma_id)
 920 {
 921         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 922
 923         list_del_init(&isert_conn->node);
 924         isert_conn->cm_id = NULL;
 925         isert_put_conn(isert_conn);
 926
 927         return -1;
 928 }
 929
 930 static int
 931 isert_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 932 {
 933         int ret = 0;
 934
 935         isert_info("%s (%d): status %d id %p np %p\n",
 936                    rdma_event_msg(event->event), event->event,
 937                    event->status, cma_id, cma_id->context);
 938
 939         switch (event->event) {
 940         case RDMA_CM_EVENT_CONNECT_REQUEST:
 941                 ret = isert_connect_request(cma_id, event);
 942                 if (ret)
 943                         isert_err("failed handle connect request %d\n", ret);
 944                 break;
 945         case RDMA_CM_EVENT_ESTABLISHED:
 946                 isert_connected_handler(cma_id);
 947                 break;
 948         case RDMA_CM_EVENT_ADDR_CHANGE:    /* FALLTHRU */
 949         case RDMA_CM_EVENT_DISCONNECTED:   /* FALLTHRU */
 950         case RDMA_CM_EVENT_DEVICE_REMOVAL: /* FALLTHRU */
 951         case RDMA_CM_EVENT_TIMEWAIT_EXIT:  /* FALLTHRU */
 952                 ret = isert_disconnected_handler(cma_id, event->event);
 953                 break;
 954         case RDMA_CM_EVENT_REJECTED:       /* FALLTHRU */
 955         case RDMA_CM_EVENT_UNREACHABLE:    /* FALLTHRU */
 956         case RDMA_CM_EVENT_CONNECT_ERROR:
 957                 ret = isert_connect_error(cma_id);
 958                 break;
 959         default:
 960                 isert_err("Unhandled RDMA CMA event: %d\n", event->event);
 961                 break;
 962         }
 963
 964         return ret;
 965 }
 966
 967 static int
 968 isert_post_recvm(struct isert_conn *isert_conn, u32 count)
 969 {
 970         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 971         int i, ret;
 972         struct iser_rx_desc *rx_desc;
 973
 974         for (rx_wr = isert_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
 975                 rx_desc = &isert_conn->rx_descs[i];
 976                 rx_wr->wr_id = (uintptr_t)rx_desc;
 977                 rx_wr->sg_list = &rx_desc->rx_sg;
 978                 rx_wr->num_sge = 1;
 979                 rx_wr->next = rx_wr + 1;
 980         }
 981         rx_wr--;
 982         rx_wr->next = NULL; /* mark end of work requests list */
 983
 984         isert_conn->post_recv_buf_count += count;
 985         ret = ib_post_recv(isert_conn->qp, isert_conn->rx_wr,
 986                            &rx_wr_failed);
 987         if (ret) {
 988                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
 989                 isert_conn->post_recv_buf_count -= count;
 990         }
 991
 992         return ret;
 993 }
 994
 995 static int
 996 isert_post_recv(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc)
 997 {
 998         struct ib_recv_wr *rx_wr_failed, rx_wr;
 999         int ret;
1000
1001         rx_wr.wr_id = (uintptr_t)rx_desc;
1002         rx_wr.sg_list = &rx_desc->rx_sg;
1003         rx_wr.num_sge = 1;
1004         rx_wr.next = NULL;
1005
1006         isert_conn->post_recv_buf_count++;
1007         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_failed);
1008         if (ret) {
1009                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
1010                 isert_conn->post_recv_buf_count--;
1011         }
1012
1013         return ret;
1014 }
1015
1016 static int
1017 isert_post_send(struct isert_conn *isert_conn, struct iser_tx_desc *tx_desc)
1018 {
1019         struct ib_device *ib_dev = isert_conn->cm_id->device;
1020         struct ib_send_wr send_wr, *send_wr_failed;
1021         int ret;
1022
1023         ib_dma_sync_single_for_device(ib_dev, tx_desc->dma_addr,
1024                                       ISER_HEADERS_LEN, DMA_TO_DEVICE);
1025
1026         send_wr.next    = NULL;
1027         send_wr.wr_id   = (uintptr_t)tx_desc;
1028         send_wr.sg_list = tx_desc->tx_sg;
1029         send_wr.num_sge = tx_desc->num_sge;
1030         send_wr.opcode  = IB_WR_SEND;
1031         send_wr.send_flags = IB_SEND_SIGNALED;
1032
1033         ret = ib_post_send(isert_conn->qp, &send_wr, &send_wr_failed);
1034         if (ret)
1035                 isert_err("ib_post_send() failed, ret: %d\n", ret);
1036
1037         return ret;
1038 }
1039
1040 static void
1041 isert_create_send_desc(struct isert_conn *isert_conn,
1042                        struct isert_cmd *isert_cmd,
1043                        struct iser_tx_desc *tx_desc)
1044 {
1045         struct isert_device *device = isert_conn->device;
1046         struct ib_device *ib_dev = device->ib_device;
1047
1048         ib_dma_sync_single_for_cpu(ib_dev, tx_desc->dma_addr,
1049                                    ISER_HEADERS_LEN, DMA_TO_DEVICE);
1050
1051         memset(&tx_desc->iser_header, 0, sizeof(struct iser_hdr));
1052         tx_desc->iser_header.flags = ISER_VER;
1053
1054         tx_desc->num_sge = 1;
1055         tx_desc->isert_cmd = isert_cmd;
1056
1057         if (tx_desc->tx_sg[0].lkey != device->pd->local_dma_lkey) {
1058                 tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
1059                 isert_dbg("tx_desc %p lkey mismatch, fixing\n", tx_desc);
1060         }
1061 }
1062
1063 static int
1064 isert_init_tx_hdrs(struct isert_conn *isert_conn,
1065                    struct iser_tx_desc *tx_desc)
1066 {
1067         struct isert_device *device = isert_conn->device;
1068         struct ib_device *ib_dev = device->ib_device;
1069         u64 dma_addr;
1070
1071         dma_addr = ib_dma_map_single(ib_dev, (void *)tx_desc,
1072                         ISER_HEADERS_LEN, DMA_TO_DEVICE);
1073         if (ib_dma_mapping_error(ib_dev, dma_addr)) {
1074                 isert_err("ib_dma_mapping_error() failed\n");
1075                 return -ENOMEM;
1076         }
1077
1078         tx_desc->dma_addr = dma_addr;
1079         tx_desc->tx_sg[0].addr  = tx_desc->dma_addr;
1080         tx_desc->tx_sg[0].length = ISER_HEADERS_LEN;
1081         tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
1082
1083         isert_dbg("Setup tx_sg[0].addr: 0x%llx length: %u lkey: 0x%x\n",
1084                   tx_desc->tx_sg[0].addr, tx_desc->tx_sg[0].length,
1085                   tx_desc->tx_sg[0].lkey);
1086
1087         return 0;
1088 }
1089
1090 static void
1091 isert_init_send_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1092                    struct ib_send_wr *send_wr)
1093 {
1094         struct iser_tx_desc *tx_desc = &isert_cmd->tx_desc;
1095
1096         isert_cmd->rdma_wr.iser_ib_op = ISER_IB_SEND;
1097         send_wr->wr_id = (uintptr_t)&isert_cmd->tx_desc;
1098         send_wr->opcode = IB_WR_SEND;
1099         send_wr->sg_list = &tx_desc->tx_sg[0];
1100         send_wr->num_sge = isert_cmd->tx_desc.num_sge;
1101         send_wr->send_flags = IB_SEND_SIGNALED;
1102 }
1103
1104 static int
1105 isert_rdma_post_recvl(struct isert_conn *isert_conn)
1106 {
1107         struct ib_recv_wr rx_wr, *rx_wr_fail;
1108         struct ib_sge sge;
1109         int ret;
1110
1111         memset(&sge, 0, sizeof(struct ib_sge));
1112         sge.addr = isert_conn->login_req_dma;
1113         sge.length = ISER_RX_LOGIN_SIZE;
1114         sge.lkey = isert_conn->device->pd->local_dma_lkey;
1115
1116         isert_dbg("Setup sge: addr: %llx length: %d 0x%08x\n",
1117                 sge.addr, sge.length, sge.lkey);
1118
1119         memset(&rx_wr, 0, sizeof(struct ib_recv_wr));
1120         rx_wr.wr_id = (uintptr_t)isert_conn->login_req_buf;
1121         rx_wr.sg_list = &sge;
1122         rx_wr.num_sge = 1;
1123
1124         isert_conn->post_recv_buf_count++;
1125         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_fail);
1126         if (ret) {
1127                 isert_err("ib_post_recv() failed: %d\n", ret);
1128                 isert_conn->post_recv_buf_count--;
1129         }
1130
1131         return ret;
1132 }
1133
1134 static int
1135 isert_put_login_tx(struct iscsi_conn *conn, struct iscsi_login *login,
1136                    u32 length)
1137 {
1138         struct isert_conn *isert_conn = conn->context;
1139         struct isert_device *device = isert_conn->device;
1140         struct ib_device *ib_dev = device->ib_device;
1141         struct iser_tx_desc *tx_desc = &isert_conn->login_tx_desc;
1142         int ret;
1143
1144         isert_create_send_desc(isert_conn, NULL, tx_desc);
1145
1146         memcpy(&tx_desc->iscsi_header, &login->rsp[0],
1147                sizeof(struct iscsi_hdr));
1148
1149         isert_init_tx_hdrs(isert_conn, tx_desc);
1150
1151         if (length > 0) {
1152                 struct ib_sge *tx_dsg = &tx_desc->tx_sg[1];
1153
1154                 ib_dma_sync_single_for_cpu(ib_dev, isert_conn->login_rsp_dma,
1155                                            length, DMA_TO_DEVICE);
1156
1157                 memcpy(isert_conn->login_rsp_buf, login->rsp_buf, length);
1158
1159                 ib_dma_sync_single_for_device(ib_dev, isert_conn->login_rsp_dma,
1160                                               length, DMA_TO_DEVICE);
1161
1162                 tx_dsg->addr    = isert_conn->login_rsp_dma;
1163                 tx_dsg->length  = length;
1164                 tx_dsg->lkey    = isert_conn->device->pd->local_dma_lkey;
1165                 tx_desc->num_sge = 2;
1166         }
1167         if (!login->login_failed) {
1168                 if (login->login_complete) {
1169                         if (!conn->sess->sess_ops->SessionType &&
1170                             isert_conn->device->use_fastreg) {
1171                                 ret = isert_conn_create_fastreg_pool(isert_conn);
1172                                 if (ret) {
1173                                         isert_err("Conn: %p failed to create"
1174                                                " fastreg pool\n", isert_conn);
1175                                         return ret;
1176                                 }
1177                         }
1178
1179                         ret = isert_alloc_rx_descriptors(isert_conn);
1180                         if (ret)
1181                                 return ret;
1182
1183                         ret = isert_post_recvm(isert_conn,
1184                                                ISERT_QP_MAX_RECV_DTOS);
1185                         if (ret)
1186                                 return ret;
1187
1188                         /* Now we are in FULL_FEATURE phase */
1189                         mutex_lock(&isert_conn->mutex);
1190                         isert_conn->state = ISER_CONN_FULL_FEATURE;
1191                         mutex_unlock(&isert_conn->mutex);
1192                         goto post_send;
1193                 }
1194
1195                 ret = isert_rdma_post_recvl(isert_conn);
1196                 if (ret)
1197                         return ret;
1198         }
1199 post_send:
1200         ret = isert_post_send(isert_conn, tx_desc);
1201         if (ret)
1202                 return ret;
1203
1204         return 0;
1205 }
1206
1207 static void
1208 isert_rx_login_req(struct isert_conn *isert_conn)
1209 {
1210         struct iser_rx_desc *rx_desc = (void *)isert_conn->login_req_buf;
1211         int rx_buflen = isert_conn->login_req_len;
1212         struct iscsi_conn *conn = isert_conn->conn;
1213         struct iscsi_login *login = conn->conn_login;
1214         int size;
1215
1216         isert_info("conn %p\n", isert_conn);
1217
1218         WARN_ON_ONCE(!login);
1219
1220         if (login->first_request) {
1221                 struct iscsi_login_req *login_req =
1222                         (struct iscsi_login_req *)&rx_desc->iscsi_header;
1223                 /*
1224                  * Setup the initial iscsi_login values from the leading
1225                  * login request PDU.
1226                  */
1227                 login->leading_connection = (!login_req->tsih) ? 1 : 0;
1228                 login->current_stage =
1229                         (login_req->flags & ISCSI_FLAG_LOGIN_CURRENT_STAGE_MASK)
1230                          >> 2;
1231                 login->version_min      = login_req->min_version;
1232                 login->version_max      = login_req->max_version;
1233                 memcpy(login->isid, login_req->isid, 6);
1234                 login->cmd_sn           = be32_to_cpu(login_req->cmdsn);
1235                 login->init_task_tag    = login_req->itt;
1236                 login->initial_exp_statsn = be32_to_cpu(login_req->exp_statsn);
1237                 login->cid              = be16_to_cpu(login_req->cid);
1238                 login->tsih             = be16_to_cpu(login_req->tsih);
1239         }
1240
1241         memcpy(&login->req[0], (void *)&rx_desc->iscsi_header, ISCSI_HDR_LEN);
1242
1243         size = min(rx_buflen, MAX_KEY_VALUE_PAIRS);
1244         isert_dbg("Using login payload size: %d, rx_buflen: %d "
1245                   "MAX_KEY_VALUE_PAIRS: %d\n", size, rx_buflen,
1246                   MAX_KEY_VALUE_PAIRS);
1247         memcpy(login->req_buf, &rx_desc->data[0], size);
1248
1249         if (login->first_request) {
1250                 complete(&isert_conn->login_comp);
1251                 return;
1252         }
1253         schedule_delayed_work(&conn->login_work, 0);
1254 }
1255
1256 static struct iscsi_cmd
1257 *isert_allocate_cmd(struct iscsi_conn *conn, struct iser_rx_desc *rx_desc)
1258 {
1259         struct isert_conn *isert_conn = conn->context;
1260         struct isert_cmd *isert_cmd;
1261         struct iscsi_cmd *cmd;
1262
1263         cmd = iscsit_allocate_cmd(conn, TASK_INTERRUPTIBLE);
1264         if (!cmd) {
1265                 isert_err("Unable to allocate iscsi_cmd + isert_cmd\n");
1266                 return NULL;
1267         }
1268         isert_cmd = iscsit_priv_cmd(cmd);
1269         isert_cmd->conn = isert_conn;
1270         isert_cmd->iscsi_cmd = cmd;
1271         isert_cmd->rx_desc = rx_desc;
1272
1273         return cmd;
1274 }
1275
1276 static int
1277 isert_handle_scsi_cmd(struct isert_conn *isert_conn,
1278                       struct isert_cmd *isert_cmd, struct iscsi_cmd *cmd,
1279                       struct iser_rx_desc *rx_desc, unsigned char *buf)
1280 {
1281         struct iscsi_conn *conn = isert_conn->conn;
1282         struct iscsi_scsi_req *hdr = (struct iscsi_scsi_req *)buf;
1283         int imm_data, imm_data_len, unsol_data, sg_nents, rc;
1284         bool dump_payload = false;
1285         unsigned int data_len;
1286
1287         rc = iscsit_setup_scsi_cmd(conn, cmd, buf);
1288         if (rc < 0)
1289                 return rc;
1290
1291         imm_data = cmd->immediate_data;
1292         imm_data_len = cmd->first_burst_len;
1293         unsol_data = cmd->unsolicited_data;
1294         data_len = cmd->se_cmd.data_length;
1295
1296         if (imm_data && imm_data_len == data_len)
1297                 cmd->se_cmd.se_cmd_flags |= SCF_PASSTHROUGH_SG_TO_MEM_NOALLOC;
1298         rc = iscsit_process_scsi_cmd(conn, cmd, hdr);
1299         if (rc < 0) {
1300                 return 0;
1301         } else if (rc > 0) {
1302                 dump_payload = true;
1303                 goto sequence_cmd;
1304         }
1305
1306         if (!imm_data)
1307                 return 0;
1308
1309         if (imm_data_len != data_len) {
1310                 sg_nents = max(1UL, DIV_ROUND_UP(imm_data_len, PAGE_SIZE));
1311                 sg_copy_from_buffer(cmd->se_cmd.t_data_sg, sg_nents,
1312                                     &rx_desc->data[0], imm_data_len);
1313                 isert_dbg("Copy Immediate sg_nents: %u imm_data_len: %d\n",
1314                           sg_nents, imm_data_len);
1315         } else {
1316                 sg_init_table(&isert_cmd->sg, 1);
1317                 cmd->se_cmd.t_data_sg = &isert_cmd->sg;
1318                 cmd->se_cmd.t_data_nents = 1;
1319                 sg_set_buf(&isert_cmd->sg, &rx_desc->data[0], imm_data_len);
1320                 isert_dbg("Transfer Immediate imm_data_len: %d\n",
1321                           imm_data_len);
1322         }
1323
1324         cmd->write_data_done += imm_data_len;
1325
1326         if (cmd->write_data_done == cmd->se_cmd.data_length) {
1327                 spin_lock_bh(&cmd->istate_lock);
1328                 cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1329                 cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1330                 spin_unlock_bh(&cmd->istate_lock);
1331         }
1332
1333 sequence_cmd:
1334         rc = iscsit_sequence_cmd(conn, cmd, buf, hdr->cmdsn);
1335
1336         if (!rc && dump_payload == false && unsol_data)
1337                 iscsit_set_unsoliticed_dataout(cmd);
1338         else if (dump_payload && imm_data)
1339                 target_put_sess_cmd(&cmd->se_cmd);
1340
1341         return 0;
1342 }
1343
1344 static int
1345 isert_handle_iscsi_dataout(struct isert_conn *isert_conn,
1346                            struct iser_rx_desc *rx_desc, unsigned char *buf)
1347 {
1348         struct scatterlist *sg_start;
1349         struct iscsi_conn *conn = isert_conn->conn;
1350         struct iscsi_cmd *cmd = NULL;
1351         struct iscsi_data *hdr = (struct iscsi_data *)buf;
1352         u32 unsol_data_len = ntoh24(hdr->dlength);
1353         int rc, sg_nents, sg_off, page_off;
1354
1355         rc = iscsit_check_dataout_hdr(conn, buf, &cmd);
1356         if (rc < 0)
1357                 return rc;
1358         else if (!cmd)
1359                 return 0;
1360         /*
1361          * FIXME: Unexpected unsolicited_data out
1362          */
1363         if (!cmd->unsolicited_data) {
1364                 isert_err("Received unexpected solicited data payload\n");
1365                 dump_stack();
1366                 return -1;
1367         }
1368
1369         isert_dbg("Unsolicited DataOut unsol_data_len: %u, "
1370                   "write_data_done: %u, data_length: %u\n",
1371                   unsol_data_len,  cmd->write_data_done,
1372                   cmd->se_cmd.data_length);
1373
1374         sg_off = cmd->write_data_done / PAGE_SIZE;
1375         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
1376         sg_nents = max(1UL, DIV_ROUND_UP(unsol_data_len, PAGE_SIZE));
1377         page_off = cmd->write_data_done % PAGE_SIZE;
1378         /*
1379          * FIXME: Non page-aligned unsolicited_data out
1380          */
1381         if (page_off) {
1382                 isert_err("unexpected non-page aligned data payload\n");
1383                 dump_stack();
1384                 return -1;
1385         }
1386         isert_dbg("Copying DataOut: sg_start: %p, sg_off: %u "
1387                   "sg_nents: %u from %p %u\n", sg_start, sg_off,
1388                   sg_nents, &rx_desc->data[0], unsol_data_len);
1389
1390         sg_copy_from_buffer(sg_start, sg_nents, &rx_desc->data[0],
1391                             unsol_data_len);
1392
1393         rc = iscsit_check_dataout_payload(cmd, hdr, false);
1394         if (rc < 0)
1395                 return rc;
1396
1397         /*
1398          * multiple data-outs on the same command can arrive -
1399          * so post the buffer before hand
1400          */
1401         rc = isert_post_recv(isert_conn, rx_desc);
1402         if (rc) {
1403                 isert_err("ib_post_recv failed with %d\n", rc);
1404                 return rc;
1405         }
1406         return 0;
1407 }
1408
1409 static int
1410 isert_handle_nop_out(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1411                      struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1412                      unsigned char *buf)
1413 {
1414         struct iscsi_conn *conn = isert_conn->conn;
1415         struct iscsi_nopout *hdr = (struct iscsi_nopout *)buf;
1416         int rc;
1417
1418         rc = iscsit_setup_nop_out(conn, cmd, hdr);
1419         if (rc < 0)
1420                 return rc;
1421         /*
1422          * FIXME: Add support for NOPOUT payload using unsolicited RDMA payload
1423          */
1424
1425         return iscsit_process_nop_out(conn, cmd, hdr);
1426 }
1427
1428 static int
1429 isert_handle_text_cmd(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1430                       struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1431                       struct iscsi_text *hdr)
1432 {
1433         struct iscsi_conn *conn = isert_conn->conn;
1434         u32 payload_length = ntoh24(hdr->dlength);
1435         int rc;
1436         unsigned char *text_in = NULL;
1437
1438         rc = iscsit_setup_text_cmd(conn, cmd, hdr);
1439         if (rc < 0)
1440                 return rc;
1441
1442         if (payload_length) {
1443                 text_in = kzalloc(payload_length, GFP_KERNEL);
1444                 if (!text_in) {
1445                         isert_err("Unable to allocate text_in of payload_length: %u\n",
1446                                   payload_length);
1447                         return -ENOMEM;
1448                 }
1449         }
1450         cmd->text_in_ptr = text_in;
1451
1452         memcpy(cmd->text_in_ptr, &rx_desc->data[0], payload_length);
1453
1454         return iscsit_process_text_cmd(conn, cmd, hdr);
1455 }
1456
1457 static int
1458 isert_rx_opcode(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc,
1459                 uint32_t read_stag, uint64_t read_va,
1460                 uint32_t write_stag, uint64_t write_va)
1461 {
1462         struct iscsi_hdr *hdr = &rx_desc->iscsi_header;
1463         struct iscsi_conn *conn = isert_conn->conn;
1464         struct iscsi_cmd *cmd;
1465         struct isert_cmd *isert_cmd;
1466         int ret = -EINVAL;
1467         u8 opcode = (hdr->opcode & ISCSI_OPCODE_MASK);
1468
1469         if (conn->sess->sess_ops->SessionType &&
1470            (!(opcode & ISCSI_OP_TEXT) || !(opcode & ISCSI_OP_LOGOUT))) {
1471                 isert_err("Got illegal opcode: 0x%02x in SessionType=Discovery,"
1472                           " ignoring\n", opcode);
1473                 return 0;
1474         }
1475
1476         switch (opcode) {
1477         case ISCSI_OP_SCSI_CMD:
1478                 cmd = isert_allocate_cmd(conn, rx_desc);
1479                 if (!cmd)
1480                         break;
1481
1482                 isert_cmd = iscsit_priv_cmd(cmd);
1483                 isert_cmd->read_stag = read_stag;
1484                 isert_cmd->read_va = read_va;
1485                 isert_cmd->write_stag = write_stag;
1486                 isert_cmd->write_va = write_va;
1487
1488                 ret = isert_handle_scsi_cmd(isert_conn, isert_cmd, cmd,
1489                                         rx_desc, (unsigned char *)hdr);
1490                 break;
1491         case ISCSI_OP_NOOP_OUT:
1492                 cmd = isert_allocate_cmd(conn, rx_desc);
1493                 if (!cmd)
1494                         break;
1495
1496                 isert_cmd = iscsit_priv_cmd(cmd);
1497                 ret = isert_handle_nop_out(isert_conn, isert_cmd, cmd,
1498                                            rx_desc, (unsigned char *)hdr);
1499                 break;
1500         case ISCSI_OP_SCSI_DATA_OUT:
1501                 ret = isert_handle_iscsi_dataout(isert_conn, rx_desc,
1502                                                 (unsigned char *)hdr);
1503                 break;
1504         case ISCSI_OP_SCSI_TMFUNC:
1505                 cmd = isert_allocate_cmd(conn, rx_desc);
1506                 if (!cmd)
1507                         break;
1508
1509                 ret = iscsit_handle_task_mgt_cmd(conn, cmd,
1510                                                 (unsigned char *)hdr);
1511                 break;
1512         case ISCSI_OP_LOGOUT:
1513                 cmd = isert_allocate_cmd(conn, rx_desc);
1514                 if (!cmd)
1515                         break;
1516
1517                 ret = iscsit_handle_logout_cmd(conn, cmd, (unsigned char *)hdr);
1518                 break;
1519         case ISCSI_OP_TEXT:
1520                 if (be32_to_cpu(hdr->ttt) != 0xFFFFFFFF)
1521                         cmd = iscsit_find_cmd_from_itt(conn, hdr->itt);
1522                 else
1523                         cmd = isert_allocate_cmd(conn, rx_desc);
1524
1525                 if (!cmd)
1526                         break;
1527
1528                 isert_cmd = iscsit_priv_cmd(cmd);
1529                 ret = isert_handle_text_cmd(isert_conn, isert_cmd, cmd,
1530                                             rx_desc, (struct iscsi_text *)hdr);
1531                 break;
1532         default:
1533                 isert_err("Got unknown iSCSI OpCode: 0x%02x\n", opcode);
1534                 dump_stack();
1535                 break;
1536         }
1537
1538         return ret;
1539 }
1540
1541 static void
1542 isert_rx_do_work(struct iser_rx_desc *rx_desc, struct isert_conn *isert_conn)
1543 {
1544         struct iser_hdr *iser_hdr = &rx_desc->iser_header;
1545         uint64_t read_va = 0, write_va = 0;
1546         uint32_t read_stag = 0, write_stag = 0;
1547
1548         switch (iser_hdr->flags & 0xF0) {
1549         case ISCSI_CTRL:
1550                 if (iser_hdr->flags & ISER_RSV) {
1551                         read_stag = be32_to_cpu(iser_hdr->read_stag);
1552                         read_va = be64_to_cpu(iser_hdr->read_va);
1553                         isert_dbg("ISER_RSV: read_stag: 0x%x read_va: 0x%llx\n",
1554                                   read_stag, (unsigned long long)read_va);
1555                 }
1556                 if (iser_hdr->flags & ISER_WSV) {
1557                         write_stag = be32_to_cpu(iser_hdr->write_stag);
1558                         write_va = be64_to_cpu(iser_hdr->write_va);
1559                         isert_dbg("ISER_WSV: write_stag: 0x%x write_va: 0x%llx\n",
1560                                   write_stag, (unsigned long long)write_va);
1561                 }
1562
1563                 isert_dbg("ISER ISCSI_CTRL PDU\n");
1564                 break;
1565         case ISER_HELLO:
1566                 isert_err("iSER Hello message\n");
1567                 break;
1568         default:
1569                 isert_warn("Unknown iSER hdr flags: 0x%02x\n", iser_hdr->flags);
1570                 break;
1571         }
1572
1573         isert_rx_opcode(isert_conn, rx_desc,
1574                         read_stag, read_va, write_stag, write_va);
1575 }
1576
1577 static void
1578 isert_rcv_completion(struct iser_rx_desc *desc,
1579                      struct isert_conn *isert_conn,
1580                      u32 xfer_len)
1581 {
1582         struct ib_device *ib_dev = isert_conn->cm_id->device;
1583         struct iscsi_hdr *hdr;
1584         u64 rx_dma;
1585         int rx_buflen;
1586
1587         if ((char *)desc == isert_conn->login_req_buf) {
1588                 rx_dma = isert_conn->login_req_dma;
1589                 rx_buflen = ISER_RX_LOGIN_SIZE;
1590                 isert_dbg("login_buf: Using rx_dma: 0x%llx, rx_buflen: %d\n",
1591                          rx_dma, rx_buflen);
1592         } else {
1593                 rx_dma = desc->dma_addr;
1594                 rx_buflen = ISER_RX_PAYLOAD_SIZE;
1595                 isert_dbg("req_buf: Using rx_dma: 0x%llx, rx_buflen: %d\n",
1596                          rx_dma, rx_buflen);
1597         }
1598
1599         ib_dma_sync_single_for_cpu(ib_dev, rx_dma, rx_buflen, DMA_FROM_DEVICE);
1600
1601         hdr = &desc->iscsi_header;
1602         isert_dbg("iSCSI opcode: 0x%02x, ITT: 0x%08x, flags: 0x%02x dlen: %d\n",
1603                  hdr->opcode, hdr->itt, hdr->flags,
1604                  (int)(xfer_len - ISER_HEADERS_LEN));
1605
1606         if ((char *)desc == isert_conn->login_req_buf) {
1607                 isert_conn->login_req_len = xfer_len - ISER_HEADERS_LEN;
1608                 if (isert_conn->conn) {
1609                         struct iscsi_login *login = isert_conn->conn->conn_login;
1610
1611                         if (login && !login->first_request)
1612                                 isert_rx_login_req(isert_conn);
1613                 }
1614                 mutex_lock(&isert_conn->mutex);
1615                 complete(&isert_conn->login_req_comp);
1616                 mutex_unlock(&isert_conn->mutex);
1617         } else {
1618                 isert_rx_do_work(desc, isert_conn);
1619         }
1620
1621         ib_dma_sync_single_for_device(ib_dev, rx_dma, rx_buflen,
1622                                       DMA_FROM_DEVICE);
1623
1624         isert_conn->post_recv_buf_count--;
1625 }
1626
1627 static int
1628 isert_map_data_buf(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1629                    struct scatterlist *sg, u32 nents, u32 length, u32 offset,
1630                    enum iser_ib_op_code op, struct isert_data_buf *data)
1631 {
1632         struct ib_device *ib_dev = isert_conn->cm_id->device;
1633
1634         data->dma_dir = op == ISER_IB_RDMA_WRITE ?
1635                               DMA_TO_DEVICE : DMA_FROM_DEVICE;
1636
1637         data->len = length - offset;
1638         data->offset = offset;
1639         data->sg_off = data->offset / PAGE_SIZE;
1640
1641         data->sg = &sg[data->sg_off];
1642         data->nents = min_t(unsigned int, nents - data->sg_off,
1643                                           ISCSI_ISER_SG_TABLESIZE);
1644         data->len = min_t(unsigned int, data->len, ISCSI_ISER_SG_TABLESIZE *
1645                                         PAGE_SIZE);
1646
1647         data->dma_nents = ib_dma_map_sg(ib_dev, data->sg, data->nents,
1648                                         data->dma_dir);
1649         if (unlikely(!data->dma_nents)) {
1650                 isert_err("Cmd: unable to dma map SGs %p\n", sg);
1651                 return -EINVAL;
1652         }
1653
1654         isert_dbg("Mapped cmd: %p count: %u sg: %p sg_nents: %u rdma_len %d\n",
1655                   isert_cmd, data->dma_nents, data->sg, data->nents, data->len);
1656
1657         return 0;
1658 }
1659
1660 static void
1661 isert_unmap_data_buf(struct isert_conn *isert_conn, struct isert_data_buf *data)
1662 {
1663         struct ib_device *ib_dev = isert_conn->cm_id->device;
1664
1665         ib_dma_unmap_sg(ib_dev, data->sg, data->nents, data->dma_dir);
1666         memset(data, 0, sizeof(*data));
1667 }
1668
1669
1670
1671 static void
1672 isert_unmap_cmd(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn)
1673 {
1674         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1675
1676         isert_dbg("Cmd %p\n", isert_cmd);
1677
1678         if (wr->data.sg) {
1679                 isert_dbg("Cmd %p unmap_sg op\n", isert_cmd);
1680                 isert_unmap_data_buf(isert_conn, &wr->data);
1681         }
1682
1683         if (wr->rdma_wr) {
1684                 isert_dbg("Cmd %p free send_wr\n", isert_cmd);
1685                 kfree(wr->rdma_wr);
1686                 wr->rdma_wr = NULL;
1687         }
1688
1689         if (wr->ib_sge) {
1690                 isert_dbg("Cmd %p free ib_sge\n", isert_cmd);
1691                 kfree(wr->ib_sge);
1692                 wr->ib_sge = NULL;
1693         }
1694 }
1695
1696 static void
1697 isert_unreg_rdma(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn)
1698 {
1699         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1700
1701         isert_dbg("Cmd %p\n", isert_cmd);
1702
1703         if (wr->fr_desc) {
1704                 isert_dbg("Cmd %p free fr_desc %p\n", isert_cmd, wr->fr_desc);
1705                 if (wr->fr_desc->ind & ISERT_PROTECTED) {
1706                         isert_unmap_data_buf(isert_conn, &wr->prot);
1707                         wr->fr_desc->ind &= ~ISERT_PROTECTED;
1708                 }
1709                 spin_lock_bh(&isert_conn->pool_lock);
1710                 list_add_tail(&wr->fr_desc->list, &isert_conn->fr_pool);
1711                 spin_unlock_bh(&isert_conn->pool_lock);
1712                 wr->fr_desc = NULL;
1713         }
1714
1715         if (wr->data.sg) {
1716                 isert_dbg("Cmd %p unmap_sg op\n", isert_cmd);
1717                 isert_unmap_data_buf(isert_conn, &wr->data);
1718         }
1719
1720         wr->ib_sge = NULL;
1721         wr->rdma_wr = NULL;
1722 }
1723
1724 static void
1725 isert_put_cmd(struct isert_cmd *isert_cmd, bool comp_err)
1726 {
1727         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1728         struct isert_conn *isert_conn = isert_cmd->conn;
1729         struct iscsi_conn *conn = isert_conn->conn;
1730         struct isert_device *device = isert_conn->device;
1731         struct iscsi_text_rsp *hdr;
1732
1733         isert_dbg("Cmd %p\n", isert_cmd);
1734
1735         switch (cmd->iscsi_opcode) {
1736         case ISCSI_OP_SCSI_CMD:
1737                 spin_lock_bh(&conn->cmd_lock);
1738                 if (!list_empty(&cmd->i_conn_node))
1739                         list_del_init(&cmd->i_conn_node);
1740                 spin_unlock_bh(&conn->cmd_lock);
1741
1742                 if (cmd->data_direction == DMA_TO_DEVICE) {
1743                         iscsit_stop_dataout_timer(cmd);
1744                         /*
1745                          * Check for special case during comp_err where
1746                          * WRITE_PENDING has been handed off from core,
1747                          * but requires an extra target_put_sess_cmd()
1748                          * before transport_generic_free_cmd() below.
1749                          */
1750                         if (comp_err &&
1751                             cmd->se_cmd.t_state == TRANSPORT_WRITE_PENDING) {
1752                                 struct se_cmd *se_cmd = &cmd->se_cmd;
1753
1754                                 target_put_sess_cmd(se_cmd);
1755                         }
1756                 }
1757
1758                 device->unreg_rdma_mem(isert_cmd, isert_conn);
1759                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1760                 break;
1761         case ISCSI_OP_SCSI_TMFUNC:
1762                 spin_lock_bh(&conn->cmd_lock);
1763                 if (!list_empty(&cmd->i_conn_node))
1764                         list_del_init(&cmd->i_conn_node);
1765                 spin_unlock_bh(&conn->cmd_lock);
1766
1767                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1768                 break;
1769         case ISCSI_OP_REJECT:
1770         case ISCSI_OP_NOOP_OUT:
1771         case ISCSI_OP_TEXT:
1772                 hdr = (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
1773                 /* If the continue bit is on, keep the command alive */
1774                 if (hdr->flags & ISCSI_FLAG_TEXT_CONTINUE)
1775                         break;
1776
1777                 spin_lock_bh(&conn->cmd_lock);
1778                 if (!list_empty(&cmd->i_conn_node))
1779                         list_del_init(&cmd->i_conn_node);
1780                 spin_unlock_bh(&conn->cmd_lock);
1781
1782                 /*
1783                  * Handle special case for REJECT when iscsi_add_reject*() has
1784                  * overwritten the original iscsi_opcode assignment, and the
1785                  * associated cmd->se_cmd needs to be released.
1786                  */
1787                 if (cmd->se_cmd.se_tfo != NULL) {
1788                         isert_dbg("Calling transport_generic_free_cmd for 0x%02x\n",
1789                                  cmd->iscsi_opcode);
1790                         transport_generic_free_cmd(&cmd->se_cmd, 0);
1791                         break;
1792                 }
1793                 /*
1794                  * Fall-through
1795                  */
1796         default:
1797                 iscsit_release_cmd(cmd);
1798                 break;
1799         }
1800 }
1801
1802 static void
1803 isert_unmap_tx_desc(struct iser_tx_desc *tx_desc, struct ib_device *ib_dev)
1804 {
1805         if (tx_desc->dma_addr != 0) {
1806                 isert_dbg("unmap single for tx_desc->dma_addr\n");
1807                 ib_dma_unmap_single(ib_dev, tx_desc->dma_addr,
1808                                     ISER_HEADERS_LEN, DMA_TO_DEVICE);
1809                 tx_desc->dma_addr = 0;
1810         }
1811 }
1812
1813 static void
1814 isert_completion_put(struct iser_tx_desc *tx_desc, struct isert_cmd *isert_cmd,
1815                      struct ib_device *ib_dev, bool comp_err)
1816 {
1817         if (isert_cmd->pdu_buf_dma != 0) {
1818                 isert_dbg("unmap single for isert_cmd->pdu_buf_dma\n");
1819                 ib_dma_unmap_single(ib_dev, isert_cmd->pdu_buf_dma,
1820                                     isert_cmd->pdu_buf_len, DMA_TO_DEVICE);
1821                 isert_cmd->pdu_buf_dma = 0;
1822         }
1823
1824         isert_unmap_tx_desc(tx_desc, ib_dev);
1825         isert_put_cmd(isert_cmd, comp_err);
1826 }
1827
1828 static int
1829 isert_check_pi_status(struct se_cmd *se_cmd, struct ib_mr *sig_mr)
1830 {
1831         struct ib_mr_status mr_status;
1832         int ret;
1833
1834         ret = ib_check_mr_status(sig_mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
1835         if (ret) {
1836                 isert_err("ib_check_mr_status failed, ret %d\n", ret);
1837                 goto fail_mr_status;
1838         }
1839
1840         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1841                 u64 sec_offset_err;
1842                 u32 block_size = se_cmd->se_dev->dev_attrib.block_size + 8;
1843
1844                 switch (mr_status.sig_err.err_type) {
1845                 case IB_SIG_BAD_GUARD:
1846                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_GUARD_CHECK_FAILED;
1847                         break;
1848                 case IB_SIG_BAD_REFTAG:
1849                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_REF_TAG_CHECK_FAILED;
1850                         break;
1851                 case IB_SIG_BAD_APPTAG:
1852                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_APP_TAG_CHECK_FAILED;
1853                         break;
1854                 }
1855                 sec_offset_err = mr_status.sig_err.sig_err_offset;
1856                 do_div(sec_offset_err, block_size);
1857                 se_cmd->bad_sector = sec_offset_err + se_cmd->t_task_lba;
1858
1859                 isert_err("PI error found type %d at sector 0x%llx "
1860                           "expected 0x%x vs actual 0x%x\n",
1861                           mr_status.sig_err.err_type,
1862                           (unsigned long long)se_cmd->bad_sector,
1863                           mr_status.sig_err.expected,
1864                           mr_status.sig_err.actual);
1865                 ret = 1;
1866         }
1867
1868 fail_mr_status:
1869         return ret;
1870 }
1871
1872 static void
1873 isert_completion_rdma_write(struct iser_tx_desc *tx_desc,
1874                             struct isert_cmd *isert_cmd)
1875 {
1876         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1877         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1878         struct se_cmd *se_cmd = &cmd->se_cmd;
1879         struct isert_conn *isert_conn = isert_cmd->conn;
1880         struct isert_device *device = isert_conn->device;
1881         int ret = 0;
1882
1883         if (wr->fr_desc && wr->fr_desc->ind & ISERT_PROTECTED) {
1884                 ret = isert_check_pi_status(se_cmd,
1885                                             wr->fr_desc->pi_ctx->sig_mr);
1886                 wr->fr_desc->ind &= ~ISERT_PROTECTED;
1887         }
1888
1889         device->unreg_rdma_mem(isert_cmd, isert_conn);
1890         wr->rdma_wr_num = 0;
1891         if (ret)
1892                 transport_send_check_condition_and_sense(se_cmd,
1893                                                          se_cmd->pi_err, 0);
1894         else
1895                 isert_put_response(isert_conn->conn, cmd);
1896 }
1897
1898 static void
1899 isert_completion_rdma_read(struct iser_tx_desc *tx_desc,
1900                            struct isert_cmd *isert_cmd)
1901 {
1902         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1903         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1904         struct se_cmd *se_cmd = &cmd->se_cmd;
1905         struct isert_conn *isert_conn = isert_cmd->conn;
1906         struct isert_device *device = isert_conn->device;
1907         int ret = 0;
1908
1909         if (wr->fr_desc && wr->fr_desc->ind & ISERT_PROTECTED) {
1910                 ret = isert_check_pi_status(se_cmd,
1911                                             wr->fr_desc->pi_ctx->sig_mr);
1912                 wr->fr_desc->ind &= ~ISERT_PROTECTED;
1913         }
1914
1915         iscsit_stop_dataout_timer(cmd);
1916         device->unreg_rdma_mem(isert_cmd, isert_conn);
1917         cmd->write_data_done = wr->data.len;
1918         wr->rdma_wr_num = 0;
1919
1920         isert_dbg("Cmd: %p RDMA_READ comp calling execute_cmd\n", isert_cmd);
1921         spin_lock_bh(&cmd->istate_lock);
1922         cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1923         cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1924         spin_unlock_bh(&cmd->istate_lock);
1925
1926         if (ret) {
1927                 target_put_sess_cmd(se_cmd);
1928                 transport_send_check_condition_and_sense(se_cmd,
1929                                                          se_cmd->pi_err, 0);
1930         } else {
1931                 target_execute_cmd(se_cmd);
1932         }
1933 }
1934
1935 static void
1936 isert_do_control_comp(struct work_struct *work)
1937 {
1938         struct isert_cmd *isert_cmd = container_of(work,
1939                         struct isert_cmd, comp_work);
1940         struct isert_conn *isert_conn = isert_cmd->conn;
1941         struct ib_device *ib_dev = isert_conn->cm_id->device;
1942         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1943
1944         isert_dbg("Cmd %p i_state %d\n", isert_cmd, cmd->i_state);
1945
1946         switch (cmd->i_state) {
1947         case ISTATE_SEND_TASKMGTRSP:
1948                 iscsit_tmr_post_handler(cmd, cmd->conn);
1949         case ISTATE_SEND_REJECT:   /* FALLTHRU */
1950         case ISTATE_SEND_TEXTRSP:  /* FALLTHRU */
1951                 cmd->i_state = ISTATE_SENT_STATUS;
1952                 isert_completion_put(&isert_cmd->tx_desc, isert_cmd,
1953                                      ib_dev, false);
1954                 break;
1955         case ISTATE_SEND_LOGOUTRSP:
1956                 iscsit_logout_post_handler(cmd, cmd->conn);
1957                 break;
1958         default:
1959                 isert_err("Unknown i_state %d\n", cmd->i_state);
1960                 dump_stack();
1961                 break;
1962         }
1963 }
1964
1965 static void
1966 isert_response_completion(struct iser_tx_desc *tx_desc,
1967                           struct isert_cmd *isert_cmd,
1968                           struct isert_conn *isert_conn,
1969                           struct ib_device *ib_dev)
1970 {
1971         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1972
1973         if (cmd->i_state == ISTATE_SEND_TASKMGTRSP ||
1974             cmd->i_state == ISTATE_SEND_LOGOUTRSP ||
1975             cmd->i_state == ISTATE_SEND_REJECT ||
1976             cmd->i_state == ISTATE_SEND_TEXTRSP) {
1977                 isert_unmap_tx_desc(tx_desc, ib_dev);
1978
1979                 INIT_WORK(&isert_cmd->comp_work, isert_do_control_comp);
1980                 queue_work(isert_comp_wq, &isert_cmd->comp_work);
1981                 return;
1982         }
1983
1984         cmd->i_state = ISTATE_SENT_STATUS;
1985         isert_completion_put(tx_desc, isert_cmd, ib_dev, false);
1986 }
1987
1988 static void
1989 isert_snd_completion(struct iser_tx_desc *tx_desc,
1990                       struct isert_conn *isert_conn)
1991 {
1992         struct ib_device *ib_dev = isert_conn->cm_id->device;
1993         struct isert_cmd *isert_cmd = tx_desc->isert_cmd;
1994         struct isert_rdma_wr *wr;
1995
1996         if (!isert_cmd) {
1997                 isert_unmap_tx_desc(tx_desc, ib_dev);
1998                 return;
1999         }
2000         wr = &isert_cmd->rdma_wr;
2001
2002         isert_dbg("Cmd %p iser_ib_op %d\n", isert_cmd, wr->iser_ib_op);
2003
2004         switch (wr->iser_ib_op) {
2005         case ISER_IB_SEND:
2006                 isert_response_completion(tx_desc, isert_cmd,
2007                                           isert_conn, ib_dev);
2008                 break;
2009         case ISER_IB_RDMA_WRITE:
2010                 isert_completion_rdma_write(tx_desc, isert_cmd);
2011                 break;
2012         case ISER_IB_RDMA_READ:
2013                 isert_completion_rdma_read(tx_desc, isert_cmd);
2014                 break;
2015         default:
2016                 isert_err("Unknown wr->iser_ib_op: 0x%x\n", wr->iser_ib_op);
2017                 dump_stack();
2018                 break;
2019         }
2020 }
2021
2022 /**
2023  * is_isert_tx_desc() - Indicate if the completion wr_id
2024  *     is a TX descriptor or not.
2025  * @isert_conn: iser connection
2026  * @wr_id: completion WR identifier
2027  *
2028  * Since we cannot rely on wc opcode in FLUSH errors
2029  * we must work around it by checking if the wr_id address
2030  * falls in the iser connection rx_descs buffer. If so
2031  * it is an RX descriptor, otherwize it is a TX.
2032  */
2033 static inline bool
2034 is_isert_tx_desc(struct isert_conn *isert_conn, void *wr_id)
2035 {
2036         void *start = isert_conn->rx_descs;
2037         int len = ISERT_QP_MAX_RECV_DTOS * sizeof(*isert_conn->rx_descs);
2038
2039         if ((wr_id >= start && wr_id < start + len) ||
2040             (wr_id == isert_conn->login_req_buf))
2041                 return false;
2042
2043         return true;
2044 }
2045
2046 static void
2047 isert_cq_comp_err(struct isert_conn *isert_conn, struct ib_wc *wc)
2048 {
2049         if (wc->wr_id == ISER_BEACON_WRID) {
2050                 isert_info("conn %p completing wait_comp_err\n",
2051                            isert_conn);
2052                 complete(&isert_conn->wait_comp_err);
2053         } else if (is_isert_tx_desc(isert_conn, (void *)(uintptr_t)wc->wr_id)) {
2054                 struct ib_device *ib_dev = isert_conn->cm_id->device;
2055                 struct isert_cmd *isert_cmd;
2056                 struct iser_tx_desc *desc;
2057
2058                 desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
2059                 isert_cmd = desc->isert_cmd;
2060                 if (!isert_cmd)
2061                         isert_unmap_tx_desc(desc, ib_dev);
2062                 else
2063                         isert_completion_put(desc, isert_cmd, ib_dev, true);
2064         } else {
2065                 isert_conn->post_recv_buf_count--;
2066                 if (!isert_conn->post_recv_buf_count &&
2067                     isert_conn->state >= ISER_CONN_BOUND)
2068                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
2069         }
2070 }
2071
2072 static void
2073 isert_handle_wc(struct ib_wc *wc)
2074 {
2075         struct isert_conn *isert_conn;
2076         struct iser_tx_desc *tx_desc;
2077         struct iser_rx_desc *rx_desc;
2078
2079         isert_conn = wc->qp->qp_context;
2080         if (likely(wc->status == IB_WC_SUCCESS)) {
2081                 if (wc->opcode == IB_WC_RECV) {
2082                         rx_desc = (struct iser_rx_desc *)(uintptr_t)wc->wr_id;
2083                         isert_rcv_completion(rx_desc, isert_conn, wc->byte_len);
2084                 } else {
2085                         tx_desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
2086                         isert_snd_completion(tx_desc, isert_conn);
2087                 }
2088         } else {
2089                 if (wc->status != IB_WC_WR_FLUSH_ERR)
2090                         isert_err("%s (%d): wr id %llx vend_err %x\n",
2091                                   ib_wc_status_msg(wc->status), wc->status,
2092                                   wc->wr_id, wc->vendor_err);
2093                 else
2094                         isert_dbg("%s (%d): wr id %llx\n",
2095                                   ib_wc_status_msg(wc->status), wc->status,
2096                                   wc->wr_id);
2097
2098                 if (wc->wr_id != ISER_FASTREG_LI_WRID)
2099                         isert_cq_comp_err(isert_conn, wc);
2100         }
2101 }
2102
2103 static void
2104 isert_cq_work(struct work_struct *work)
2105 {
2106         enum { isert_poll_budget = 65536 };
2107         struct isert_comp *comp = container_of(work, struct isert_comp,
2108                                                work);
2109         struct ib_wc *const wcs = comp->wcs;
2110         int i, n, completed = 0;
2111
2112         while ((n = ib_poll_cq(comp->cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
2113                 for (i = 0; i < n; i++)
2114                         isert_handle_wc(&wcs[i]);
2115
2116                 completed += n;
2117                 if (completed >= isert_poll_budget)
2118                         break;
2119         }
2120
2121         ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP);
2122 }
2123
2124 static void
2125 isert_cq_callback(struct ib_cq *cq, void *context)
2126 {
2127         struct isert_comp *comp = context;
2128
2129         queue_work(isert_comp_wq, &comp->work);
2130 }
2131
2132 static int
2133 isert_post_response(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd)
2134 {
2135         struct ib_send_wr *wr_failed;
2136         int ret;
2137
2138         ret = isert_post_recv(isert_conn, isert_cmd->rx_desc);
2139         if (ret) {
2140                 isert_err("ib_post_recv failed with %d\n", ret);
2141                 return ret;
2142         }
2143
2144         ret = ib_post_send(isert_conn->qp, &isert_cmd->tx_desc.send_wr,
2145                            &wr_failed);
2146         if (ret) {
2147                 isert_err("ib_post_send failed with %d\n", ret);
2148                 return ret;
2149         }
2150         return ret;
2151 }
2152
2153 static int
2154 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2155 {
2156         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2157         struct isert_conn *isert_conn = conn->context;
2158         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2159         struct iscsi_scsi_rsp *hdr = (struct iscsi_scsi_rsp *)
2160                                 &isert_cmd->tx_desc.iscsi_header;
2161
2162         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2163         iscsit_build_rsp_pdu(cmd, conn, true, hdr);
2164         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2165         /*
2166          * Attach SENSE DATA payload to iSCSI Response PDU
2167          */
2168         if (cmd->se_cmd.sense_buffer &&
2169             ((cmd->se_cmd.se_cmd_flags & SCF_TRANSPORT_TASK_SENSE) ||
2170             (cmd->se_cmd.se_cmd_flags & SCF_EMULATED_TASK_SENSE))) {
2171                 struct isert_device *device = isert_conn->device;
2172                 struct ib_device *ib_dev = device->ib_device;
2173                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2174                 u32 padding, pdu_len;
2175
2176                 put_unaligned_be16(cmd->se_cmd.scsi_sense_length,
2177                                    cmd->sense_buffer);
2178                 cmd->se_cmd.scsi_sense_length += sizeof(__be16);
2179
2180                 padding = -(cmd->se_cmd.scsi_sense_length) & 3;
2181                 hton24(hdr->dlength, (u32)cmd->se_cmd.scsi_sense_length);
2182                 pdu_len = cmd->se_cmd.scsi_sense_length + padding;
2183
2184                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2185                                 (void *)cmd->sense_buffer, pdu_len,
2186                                 DMA_TO_DEVICE);
2187
2188                 isert_cmd->pdu_buf_len = pdu_len;
2189                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2190                 tx_dsg->length  = pdu_len;
2191                 tx_dsg->lkey    = device->pd->local_dma_lkey;
2192                 isert_cmd->tx_desc.num_sge = 2;
2193         }
2194
2195         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2196
2197         isert_dbg("Posting SCSI Response\n");
2198
2199         return isert_post_response(isert_conn, isert_cmd);
2200 }
2201
2202 static void
2203 isert_aborted_task(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2204 {
2205         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2206         struct isert_conn *isert_conn = conn->context;
2207         struct isert_device *device = isert_conn->device;
2208
2209         spin_lock_bh(&conn->cmd_lock);
2210         if (!list_empty(&cmd->i_conn_node))
2211                 list_del_init(&cmd->i_conn_node);
2212         spin_unlock_bh(&conn->cmd_lock);
2213
2214         if (cmd->data_direction == DMA_TO_DEVICE)
2215                 iscsit_stop_dataout_timer(cmd);
2216
2217         device->unreg_rdma_mem(isert_cmd, isert_conn);
2218 }
2219
2220 static enum target_prot_op
2221 isert_get_sup_prot_ops(struct iscsi_conn *conn)
2222 {
2223         struct isert_conn *isert_conn = conn->context;
2224         struct isert_device *device = isert_conn->device;
2225
2226         if (conn->tpg->tpg_attrib.t10_pi) {
2227                 if (device->pi_capable) {
2228                         isert_info("conn %p PI offload enabled\n", isert_conn);
2229                         isert_conn->pi_support = true;
2230                         return TARGET_PROT_ALL;
2231                 }
2232         }
2233
2234         isert_info("conn %p PI offload disabled\n", isert_conn);
2235         isert_conn->pi_support = false;
2236
2237         return TARGET_PROT_NORMAL;
2238 }
2239
2240 static int
2241 isert_put_nopin(struct iscsi_cmd *cmd, struct iscsi_conn *conn,
2242                 bool nopout_response)
2243 {
2244         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2245         struct isert_conn *isert_conn = conn->context;
2246         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2247
2248         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2249         iscsit_build_nopin_rsp(cmd, conn, (struct iscsi_nopin *)
2250                                &isert_cmd->tx_desc.iscsi_header,
2251                                nopout_response);
2252         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2253         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2254
2255         isert_dbg("conn %p Posting NOPIN Response\n", isert_conn);
2256
2257         return isert_post_response(isert_conn, isert_cmd);
2258 }
2259
2260 static int
2261 isert_put_logout_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2262 {
2263         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2264         struct isert_conn *isert_conn = conn->context;
2265         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2266
2267         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2268         iscsit_build_logout_rsp(cmd, conn, (struct iscsi_logout_rsp *)
2269                                 &isert_cmd->tx_desc.iscsi_header);
2270         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2271         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2272
2273         isert_dbg("conn %p Posting Logout Response\n", isert_conn);
2274
2275         return isert_post_response(isert_conn, isert_cmd);
2276 }
2277
2278 static int
2279 isert_put_tm_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2280 {
2281         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2282         struct isert_conn *isert_conn = conn->context;
2283         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2284
2285         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2286         iscsit_build_task_mgt_rsp(cmd, conn, (struct iscsi_tm_rsp *)
2287                                   &isert_cmd->tx_desc.iscsi_header);
2288         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2289         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2290
2291         isert_dbg("conn %p Posting Task Management Response\n", isert_conn);
2292
2293         return isert_post_response(isert_conn, isert_cmd);
2294 }
2295
2296 static int
2297 isert_put_reject(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2298 {
2299         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2300         struct isert_conn *isert_conn = conn->context;
2301         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2302         struct isert_device *device = isert_conn->device;
2303         struct ib_device *ib_dev = device->ib_device;
2304         struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2305         struct iscsi_reject *hdr =
2306                 (struct iscsi_reject *)&isert_cmd->tx_desc.iscsi_header;
2307
2308         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2309         iscsit_build_reject(cmd, conn, hdr);
2310         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2311
2312         hton24(hdr->dlength, ISCSI_HDR_LEN);
2313         isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2314                         (void *)cmd->buf_ptr, ISCSI_HDR_LEN,
2315                         DMA_TO_DEVICE);
2316         isert_cmd->pdu_buf_len = ISCSI_HDR_LEN;
2317         tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2318         tx_dsg->length  = ISCSI_HDR_LEN;
2319         tx_dsg->lkey    = device->pd->local_dma_lkey;
2320         isert_cmd->tx_desc.num_sge = 2;
2321
2322         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2323
2324         isert_dbg("conn %p Posting Reject\n", isert_conn);
2325
2326         return isert_post_response(isert_conn, isert_cmd);
2327 }
2328
2329 static int
2330 isert_put_text_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2331 {
2332         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2333         struct isert_conn *isert_conn = conn->context;
2334         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2335         struct iscsi_text_rsp *hdr =
2336                 (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
2337         u32 txt_rsp_len;
2338         int rc;
2339
2340         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2341         rc = iscsit_build_text_rsp(cmd, conn, hdr, ISCSI_INFINIBAND);
2342         if (rc < 0)
2343                 return rc;
2344
2345         txt_rsp_len = rc;
2346         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2347
2348         if (txt_rsp_len) {
2349                 struct isert_device *device = isert_conn->device;
2350                 struct ib_device *ib_dev = device->ib_device;
2351                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2352                 void *txt_rsp_buf = cmd->buf_ptr;
2353
2354                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2355                                 txt_rsp_buf, txt_rsp_len, DMA_TO_DEVICE);
2356
2357                 isert_cmd->pdu_buf_len = txt_rsp_len;
2358                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2359                 tx_dsg->length  = txt_rsp_len;
2360                 tx_dsg->lkey    = device->pd->local_dma_lkey;
2361                 isert_cmd->tx_desc.num_sge = 2;
2362         }
2363         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2364
2365         isert_dbg("conn %p Text Response\n", isert_conn);
2366
2367         return isert_post_response(isert_conn, isert_cmd);
2368 }
2369
2370 static int
2371 isert_build_rdma_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
2372                     struct ib_sge *ib_sge, struct ib_rdma_wr *rdma_wr,
2373                     u32 data_left, u32 offset)
2374 {
2375         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
2376         struct scatterlist *sg_start, *tmp_sg;
2377         struct isert_device *device = isert_conn->device;
2378         struct ib_device *ib_dev = device->ib_device;
2379         u32 sg_off, page_off;
2380         int i = 0, sg_nents;
2381
2382         sg_off = offset / PAGE_SIZE;
2383         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
2384         sg_nents = min(cmd->se_cmd.t_data_nents - sg_off, isert_conn->max_sge);
2385         page_off = offset % PAGE_SIZE;
2386
2387         rdma_wr->wr.sg_list = ib_sge;
2388         rdma_wr->wr.wr_id = (uintptr_t)&isert_cmd->tx_desc;
2389         /*
2390          * Perform mapping of TCM scatterlist memory ib_sge dma_addr.
2391          */
2392         for_each_sg(sg_start, tmp_sg, sg_nents, i) {
2393                 isert_dbg("RDMA from SGL dma_addr: 0x%llx dma_len: %u, "
2394                           "page_off: %u\n",
2395                           (unsigned long long)tmp_sg->dma_address,
2396                           tmp_sg->length, page_off);
2397
2398                 ib_sge->addr = ib_sg_dma_address(ib_dev, tmp_sg) + page_off;
2399                 ib_sge->length = min_t(u32, data_left,
2400                                 ib_sg_dma_len(ib_dev, tmp_sg) - page_off);
2401                 ib_sge->lkey = device->pd->local_dma_lkey;
2402
2403                 isert_dbg("RDMA ib_sge: addr: 0x%llx  length: %u lkey: %x\n",
2404                           ib_sge->addr, ib_sge->length, ib_sge->lkey);
2405                 page_off = 0;
2406                 data_left -= ib_sge->length;
2407                 if (!data_left)
2408                         break;
2409                 ib_sge++;
2410                 isert_dbg("Incrementing ib_sge pointer to %p\n", ib_sge);
2411         }
2412
2413         rdma_wr->wr.num_sge = ++i;
2414         isert_dbg("Set outgoing sg_list: %p num_sg: %u from TCM SGLs\n",
2415                   rdma_wr->wr.sg_list, rdma_wr->wr.num_sge);
2416
2417         return rdma_wr->wr.num_sge;
2418 }
2419
2420 static int
2421 isert_map_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2422                struct isert_rdma_wr *wr)
2423 {
2424         struct se_cmd *se_cmd = &cmd->se_cmd;
2425         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2426         struct isert_conn *isert_conn = conn->context;
2427         struct isert_data_buf *data = &wr->data;
2428         struct ib_rdma_wr *rdma_wr;
2429         struct ib_sge *ib_sge;
2430         u32 offset, data_len, data_left, rdma_write_max, va_offset = 0;
2431         int ret = 0, i, ib_sge_cnt;
2432
2433         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2434
2435         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2436         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2437                                  se_cmd->t_data_nents, se_cmd->data_length,
2438                                  offset, wr->iser_ib_op, &wr->data);
2439         if (ret)
2440                 return ret;
2441
2442         data_left = data->len;
2443         offset = data->offset;
2444
2445         ib_sge = kzalloc(sizeof(struct ib_sge) * data->nents, GFP_KERNEL);
2446         if (!ib_sge) {
2447                 isert_warn("Unable to allocate ib_sge\n");
2448                 ret = -ENOMEM;
2449                 goto unmap_cmd;
2450         }
2451         wr->ib_sge = ib_sge;
2452
2453         wr->rdma_wr_num = DIV_ROUND_UP(data->nents, isert_conn->max_sge);
2454         wr->rdma_wr = kzalloc(sizeof(struct ib_rdma_wr) * wr->rdma_wr_num,
2455                                 GFP_KERNEL);
2456         if (!wr->rdma_wr) {
2457                 isert_dbg("Unable to allocate wr->rdma_wr\n");
2458                 ret = -ENOMEM;
2459                 goto unmap_cmd;
2460         }
2461
2462         wr->isert_cmd = isert_cmd;
2463         rdma_write_max = isert_conn->max_sge * PAGE_SIZE;
2464
2465         for (i = 0; i < wr->rdma_wr_num; i++) {
2466                 rdma_wr = &isert_cmd->rdma_wr.rdma_wr[i];
2467                 data_len = min(data_left, rdma_write_max);
2468
2469                 rdma_wr->wr.send_flags = 0;
2470                 if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2471                         rdma_wr->wr.opcode = IB_WR_RDMA_WRITE;
2472                         rdma_wr->remote_addr = isert_cmd->read_va + offset;
2473                         rdma_wr->rkey = isert_cmd->read_stag;
2474                         if (i + 1 == wr->rdma_wr_num)
2475                                 rdma_wr->wr.next = &isert_cmd->tx_desc.send_wr;
2476                         else
2477                                 rdma_wr->wr.next = &wr->rdma_wr[i + 1].wr;
2478                 } else {
2479                         rdma_wr->wr.opcode = IB_WR_RDMA_READ;
2480                         rdma_wr->remote_addr = isert_cmd->write_va + va_offset;
2481                         rdma_wr->rkey = isert_cmd->write_stag;
2482                         if (i + 1 == wr->rdma_wr_num)
2483                                 rdma_wr->wr.send_flags = IB_SEND_SIGNALED;
2484                         else
2485                                 rdma_wr->wr.next = &wr->rdma_wr[i + 1].wr;
2486                 }
2487
2488                 ib_sge_cnt = isert_build_rdma_wr(isert_conn, isert_cmd, ib_sge,
2489                                         rdma_wr, data_len, offset);
2490                 ib_sge += ib_sge_cnt;
2491
2492                 offset += data_len;
2493                 va_offset += data_len;
2494                 data_left -= data_len;
2495         }
2496
2497         return 0;
2498 unmap_cmd:
2499         isert_unmap_data_buf(isert_conn, data);
2500
2501         return ret;
2502 }
2503
2504 static inline void
2505 isert_inv_rkey(struct ib_send_wr *inv_wr, struct ib_mr *mr)
2506 {
2507         u32 rkey;
2508
2509         memset(inv_wr, 0, sizeof(*inv_wr));
2510         inv_wr->wr_id = ISER_FASTREG_LI_WRID;
2511         inv_wr->opcode = IB_WR_LOCAL_INV;
2512         inv_wr->ex.invalidate_rkey = mr->rkey;
2513
2514         /* Bump the key */
2515         rkey = ib_inc_rkey(mr->rkey);
2516         ib_update_fast_reg_key(mr, rkey);
2517 }
2518
2519 static int
2520 isert_fast_reg_mr(struct isert_conn *isert_conn,
2521                   struct fast_reg_descriptor *fr_desc,
2522                   struct isert_data_buf *mem,
2523                   enum isert_indicator ind,
2524                   struct ib_sge *sge)
2525 {
2526         struct isert_device *device = isert_conn->device;
2527         struct ib_device *ib_dev = device->ib_device;
2528         struct ib_mr *mr;
2529         struct ib_reg_wr reg_wr;
2530         struct ib_send_wr inv_wr, *bad_wr, *wr = NULL;
2531         int ret, n;
2532
2533         if (mem->dma_nents == 1) {
2534                 sge->lkey = device->pd->local_dma_lkey;
2535                 sge->addr = ib_sg_dma_address(ib_dev, &mem->sg[0]);
2536                 sge->length = ib_sg_dma_len(ib_dev, &mem->sg[0]);
2537                 isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2538                          sge->addr, sge->length, sge->lkey);
2539                 return 0;
2540         }
2541
2542         if (ind == ISERT_DATA_KEY_VALID)
2543                 /* Registering data buffer */
2544                 mr = fr_desc->data_mr;
2545         else
2546                 /* Registering protection buffer */
2547                 mr = fr_desc->pi_ctx->prot_mr;
2548
2549         if (!(fr_desc->ind & ind)) {
2550                 isert_inv_rkey(&inv_wr, mr);
2551                 wr = &inv_wr;
2552         }
2553
2554         n = ib_map_mr_sg(mr, mem->sg, mem->nents, PAGE_SIZE);
2555         if (unlikely(n != mem->nents)) {
2556                 isert_err("failed to map mr sg (%d/%d)\n",
2557                          n, mem->nents);
2558                 return n < 0 ? n : -EINVAL;
2559         }
2560
2561         isert_dbg("Use fr_desc %p sg_nents %d offset %u\n",
2562                   fr_desc, mem->nents, mem->offset);
2563
2564         reg_wr.wr.next = NULL;
2565         reg_wr.wr.opcode = IB_WR_REG_MR;
2566         reg_wr.wr.wr_id = ISER_FASTREG_LI_WRID;
2567         reg_wr.wr.send_flags = 0;
2568         reg_wr.wr.num_sge = 0;
2569         reg_wr.mr = mr;
2570         reg_wr.key = mr->lkey;
2571         reg_wr.access = IB_ACCESS_LOCAL_WRITE;
2572
2573         if (!wr)
2574                 wr = &reg_wr.wr;
2575         else
2576                 wr->next = &reg_wr.wr;
2577
2578         ret = ib_post_send(isert_conn->qp, wr, &bad_wr);
2579         if (ret) {
2580                 isert_err("fast registration failed, ret:%d\n", ret);
2581                 return ret;
2582         }
2583         fr_desc->ind &= ~ind;
2584
2585         sge->lkey = mr->lkey;
2586         sge->addr = mr->iova;
2587         sge->length = mr->length;
2588
2589         isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2590                   sge->addr, sge->length, sge->lkey);
2591
2592         return ret;
2593 }
2594
2595 static inline void
2596 isert_set_dif_domain(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs,
2597                      struct ib_sig_domain *domain)
2598 {
2599         domain->sig_type = IB_SIG_TYPE_T10_DIF;
2600         domain->sig.dif.bg_type = IB_T10DIF_CRC;
2601         domain->sig.dif.pi_interval = se_cmd->se_dev->dev_attrib.block_size;
2602         domain->sig.dif.ref_tag = se_cmd->reftag_seed;
2603         /*
2604          * At the moment we hard code those, but if in the future
2605          * the target core would like to use it, we will take it
2606          * from se_cmd.
2607          */
2608         domain->sig.dif.apptag_check_mask = 0xffff;
2609         domain->sig.dif.app_escape = true;
2610         domain->sig.dif.ref_escape = true;
2611         if (se_cmd->prot_type == TARGET_DIF_TYPE1_PROT ||
2612             se_cmd->prot_type == TARGET_DIF_TYPE2_PROT)
2613                 domain->sig.dif.ref_remap = true;
2614 };
2615
2616 static int
2617 isert_set_sig_attrs(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs)
2618 {
2619         switch (se_cmd->prot_op) {
2620         case TARGET_PROT_DIN_INSERT:
2621         case TARGET_PROT_DOUT_STRIP:
2622                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
2623                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2624                 break;
2625         case TARGET_PROT_DOUT_INSERT:
2626         case TARGET_PROT_DIN_STRIP:
2627                 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
2628                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2629                 break;
2630         case TARGET_PROT_DIN_PASS:
2631         case TARGET_PROT_DOUT_PASS:
2632                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2633                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2634                 break;
2635         default:
2636                 isert_err("Unsupported PI operation %d\n", se_cmd->prot_op);
2637                 return -EINVAL;
2638         }
2639
2640         return 0;
2641 }
2642
2643 static inline u8
2644 isert_set_prot_checks(u8 prot_checks)
2645 {
2646         return (prot_checks & TARGET_DIF_CHECK_GUARD  ? 0xc0 : 0) |
2647                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x30 : 0) |
2648                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x0f : 0);
2649 }
2650
2651 static int
2652 isert_reg_sig_mr(struct isert_conn *isert_conn,
2653                  struct se_cmd *se_cmd,
2654                  struct isert_rdma_wr *rdma_wr,
2655                  struct fast_reg_descriptor *fr_desc)
2656 {
2657         struct ib_sig_handover_wr sig_wr;
2658         struct ib_send_wr inv_wr, *bad_wr, *wr = NULL;
2659         struct pi_context *pi_ctx = fr_desc->pi_ctx;
2660         struct ib_sig_attrs sig_attrs;
2661         int ret;
2662
2663         memset(&sig_attrs, 0, sizeof(sig_attrs));
2664         ret = isert_set_sig_attrs(se_cmd, &sig_attrs);
2665         if (ret)
2666                 goto err;
2667
2668         sig_attrs.check_mask = isert_set_prot_checks(se_cmd->prot_checks);
2669
2670         if (!(fr_desc->ind & ISERT_SIG_KEY_VALID)) {
2671                 isert_inv_rkey(&inv_wr, pi_ctx->sig_mr);
2672                 wr = &inv_wr;
2673         }
2674
2675         memset(&sig_wr, 0, sizeof(sig_wr));
2676         sig_wr.wr.opcode = IB_WR_REG_SIG_MR;
2677         sig_wr.wr.wr_id = ISER_FASTREG_LI_WRID;
2678         sig_wr.wr.sg_list = &rdma_wr->ib_sg[DATA];
2679         sig_wr.wr.num_sge = 1;
2680         sig_wr.access_flags = IB_ACCESS_LOCAL_WRITE;
2681         sig_wr.sig_attrs = &sig_attrs;
2682         sig_wr.sig_mr = pi_ctx->sig_mr;
2683         if (se_cmd->t_prot_sg)
2684                 sig_wr.prot = &rdma_wr->ib_sg[PROT];
2685
2686         if (!wr)
2687                 wr = &sig_wr.wr;
2688         else
2689                 wr->next = &sig_wr.wr;
2690
2691         ret = ib_post_send(isert_conn->qp, wr, &bad_wr);
2692         if (ret) {
2693                 isert_err("fast registration failed, ret:%d\n", ret);
2694                 goto err;
2695         }
2696         fr_desc->ind &= ~ISERT_SIG_KEY_VALID;
2697
2698         rdma_wr->ib_sg[SIG].lkey = pi_ctx->sig_mr->lkey;
2699         rdma_wr->ib_sg[SIG].addr = 0;
2700         rdma_wr->ib_sg[SIG].length = se_cmd->data_length;
2701         if (se_cmd->prot_op != TARGET_PROT_DIN_STRIP &&
2702             se_cmd->prot_op != TARGET_PROT_DOUT_INSERT)
2703                 /*
2704                  * We have protection guards on the wire
2705                  * so we need to set a larget transfer
2706                  */
2707                 rdma_wr->ib_sg[SIG].length += se_cmd->prot_length;
2708
2709         isert_dbg("sig_sge: addr: 0x%llx  length: %u lkey: %x\n",
2710                   rdma_wr->ib_sg[SIG].addr, rdma_wr->ib_sg[SIG].length,
2711                   rdma_wr->ib_sg[SIG].lkey);
2712 err:
2713         return ret;
2714 }
2715
2716 static int
2717 isert_handle_prot_cmd(struct isert_conn *isert_conn,
2718                       struct isert_cmd *isert_cmd,
2719                       struct isert_rdma_wr *wr)
2720 {
2721         struct isert_device *device = isert_conn->device;
2722         struct se_cmd *se_cmd = &isert_cmd->iscsi_cmd->se_cmd;
2723         int ret;
2724
2725         if (!wr->fr_desc->pi_ctx) {
2726                 ret = isert_create_pi_ctx(wr->fr_desc,
2727                                           device->ib_device,
2728                                           device->pd);
2729                 if (ret) {
2730                         isert_err("conn %p failed to allocate pi_ctx\n",
2731                                   isert_conn);
2732                         return ret;
2733                 }
2734         }
2735
2736         if (se_cmd->t_prot_sg) {
2737                 ret = isert_map_data_buf(isert_conn, isert_cmd,
2738                                          se_cmd->t_prot_sg,
2739                                          se_cmd->t_prot_nents,
2740                                          se_cmd->prot_length,
2741                                          0, wr->iser_ib_op, &wr->prot);
2742                 if (ret) {
2743                         isert_err("conn %p failed to map protection buffer\n",
2744                                   isert_conn);
2745                         return ret;
2746                 }
2747
2748                 memset(&wr->ib_sg[PROT], 0, sizeof(wr->ib_sg[PROT]));
2749                 ret = isert_fast_reg_mr(isert_conn, wr->fr_desc, &wr->prot,
2750                                         ISERT_PROT_KEY_VALID, &wr->ib_sg[PROT]);
2751                 if (ret) {
2752                         isert_err("conn %p failed to fast reg mr\n",
2753                                   isert_conn);
2754                         goto unmap_prot_cmd;
2755                 }
2756         }
2757
2758         ret = isert_reg_sig_mr(isert_conn, se_cmd, wr, wr->fr_desc);
2759         if (ret) {
2760                 isert_err("conn %p failed to fast reg mr\n",
2761                           isert_conn);
2762                 goto unmap_prot_cmd;
2763         }
2764         wr->fr_desc->ind |= ISERT_PROTECTED;
2765
2766         return 0;
2767
2768 unmap_prot_cmd:
2769         if (se_cmd->t_prot_sg)
2770                 isert_unmap_data_buf(isert_conn, &wr->prot);
2771
2772         return ret;
2773 }
2774
2775 static int
2776 isert_reg_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2777                struct isert_rdma_wr *wr)
2778 {
2779         struct se_cmd *se_cmd = &cmd->se_cmd;
2780         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2781         struct isert_conn *isert_conn = conn->context;
2782         struct fast_reg_descriptor *fr_desc = NULL;
2783         struct ib_rdma_wr *rdma_wr;
2784         struct ib_sge *ib_sg;
2785         u32 offset;
2786         int ret = 0;
2787         unsigned long flags;
2788
2789         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2790
2791         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2792         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2793                                  se_cmd->t_data_nents, se_cmd->data_length,
2794                                  offset, wr->iser_ib_op, &wr->data);
2795         if (ret)
2796                 return ret;
2797
2798         if (wr->data.dma_nents != 1 || isert_prot_cmd(isert_conn, se_cmd)) {
2799                 spin_lock_irqsave(&isert_conn->pool_lock, flags);
2800                 fr_desc = list_first_entry(&isert_conn->fr_pool,
2801                                            struct fast_reg_descriptor, list);
2802                 list_del(&fr_desc->list);
2803                 spin_unlock_irqrestore(&isert_conn->pool_lock, flags);
2804                 wr->fr_desc = fr_desc;
2805         }
2806
2807         ret = isert_fast_reg_mr(isert_conn, fr_desc, &wr->data,
2808                                 ISERT_DATA_KEY_VALID, &wr->ib_sg[DATA]);
2809         if (ret)
2810                 goto unmap_cmd;
2811
2812         if (isert_prot_cmd(isert_conn, se_cmd)) {
2813                 ret = isert_handle_prot_cmd(isert_conn, isert_cmd, wr);
2814                 if (ret)
2815                         goto unmap_cmd;
2816
2817                 ib_sg = &wr->ib_sg[SIG];
2818         } else {
2819                 ib_sg = &wr->ib_sg[DATA];
2820         }
2821
2822         memcpy(&wr->s_ib_sge, ib_sg, sizeof(*ib_sg));
2823         wr->ib_sge = &wr->s_ib_sge;
2824         wr->rdma_wr_num = 1;
2825         memset(&wr->s_rdma_wr, 0, sizeof(wr->s_rdma_wr));
2826         wr->rdma_wr = &wr->s_rdma_wr;
2827         wr->isert_cmd = isert_cmd;
2828
2829         rdma_wr = &isert_cmd->rdma_wr.s_rdma_wr;
2830         rdma_wr->wr.sg_list = &wr->s_ib_sge;
2831         rdma_wr->wr.num_sge = 1;
2832         rdma_wr->wr.wr_id = (uintptr_t)&isert_cmd->tx_desc;
2833         if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2834                 rdma_wr->wr.opcode = IB_WR_RDMA_WRITE;
2835                 rdma_wr->remote_addr = isert_cmd->read_va;
2836                 rdma_wr->rkey = isert_cmd->read_stag;
2837                 rdma_wr->wr.send_flags = !isert_prot_cmd(isert_conn, se_cmd) ?
2838                                       0 : IB_SEND_SIGNALED;
2839         } else {
2840                 rdma_wr->wr.opcode = IB_WR_RDMA_READ;
2841                 rdma_wr->remote_addr = isert_cmd->write_va;
2842                 rdma_wr->rkey = isert_cmd->write_stag;
2843                 rdma_wr->wr.send_flags = IB_SEND_SIGNALED;
2844         }
2845
2846         return 0;
2847
2848 unmap_cmd:
2849         if (fr_desc) {
2850                 spin_lock_irqsave(&isert_conn->pool_lock, flags);
2851                 list_add_tail(&fr_desc->list, &isert_conn->fr_pool);
2852                 spin_unlock_irqrestore(&isert_conn->pool_lock, flags);
2853         }
2854         isert_unmap_data_buf(isert_conn, &wr->data);
2855
2856         return ret;
2857 }
2858
2859 static int
2860 isert_put_datain(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2861 {
2862         struct se_cmd *se_cmd = &cmd->se_cmd;
2863         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2864         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2865         struct isert_conn *isert_conn = conn->context;
2866         struct isert_device *device = isert_conn->device;
2867         struct ib_send_wr *wr_failed;
2868         int rc;
2869
2870         isert_dbg("Cmd: %p RDMA_WRITE data_length: %u\n",
2871                  isert_cmd, se_cmd->data_length);
2872
2873         wr->iser_ib_op = ISER_IB_RDMA_WRITE;
2874         rc = device->reg_rdma_mem(conn, cmd, wr);
2875         if (rc) {
2876                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2877                 return rc;
2878         }
2879
2880         if (!isert_prot_cmd(isert_conn, se_cmd)) {
2881                 /*
2882                  * Build isert_conn->tx_desc for iSCSI response PDU and attach
2883                  */
2884                 isert_create_send_desc(isert_conn, isert_cmd,
2885                                        &isert_cmd->tx_desc);
2886                 iscsit_build_rsp_pdu(cmd, conn, true, (struct iscsi_scsi_rsp *)
2887                                      &isert_cmd->tx_desc.iscsi_header);
2888                 isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2889                 isert_init_send_wr(isert_conn, isert_cmd,
2890                                    &isert_cmd->tx_desc.send_wr);
2891                 isert_cmd->rdma_wr.s_rdma_wr.wr.next = &isert_cmd->tx_desc.send_wr;
2892                 wr->rdma_wr_num += 1;
2893
2894                 rc = isert_post_recv(isert_conn, isert_cmd->rx_desc);
2895                 if (rc) {
2896                         isert_err("ib_post_recv failed with %d\n", rc);
2897                         return rc;
2898                 }
2899         }
2900
2901         rc = ib_post_send(isert_conn->qp, &wr->rdma_wr->wr, &wr_failed);
2902         if (rc)
2903                 isert_warn("ib_post_send() failed for IB_WR_RDMA_WRITE\n");
2904
2905         if (!isert_prot_cmd(isert_conn, se_cmd))
2906                 isert_dbg("Cmd: %p posted RDMA_WRITE + Response for iSER Data "
2907                          "READ\n", isert_cmd);
2908         else
2909                 isert_dbg("Cmd: %p posted RDMA_WRITE for iSER Data READ\n",
2910                          isert_cmd);
2911
2912         return 1;
2913 }
2914
2915 static int
2916 isert_get_dataout(struct iscsi_conn *conn, struct iscsi_cmd *cmd, bool recovery)
2917 {
2918         struct se_cmd *se_cmd = &cmd->se_cmd;
2919         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2920         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2921         struct isert_conn *isert_conn = conn->context;
2922         struct isert_device *device = isert_conn->device;
2923         struct ib_send_wr *wr_failed;
2924         int rc;
2925
2926         isert_dbg("Cmd: %p RDMA_READ data_length: %u write_data_done: %u\n",
2927                  isert_cmd, se_cmd->data_length, cmd->write_data_done);
2928         wr->iser_ib_op = ISER_IB_RDMA_READ;
2929         rc = device->reg_rdma_mem(conn, cmd, wr);
2930         if (rc) {
2931                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2932                 return rc;
2933         }
2934
2935         rc = ib_post_send(isert_conn->qp, &wr->rdma_wr->wr, &wr_failed);
2936         if (rc)
2937                 isert_warn("ib_post_send() failed for IB_WR_RDMA_READ\n");
2938
2939         isert_dbg("Cmd: %p posted RDMA_READ memory for ISER Data WRITE\n",
2940                  isert_cmd);
2941
2942         return 0;
2943 }
2944
2945 static int
2946 isert_immediate_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2947 {
2948         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2949         int ret = 0;
2950
2951         switch (state) {
2952         case ISTATE_REMOVE:
2953                 spin_lock_bh(&conn->cmd_lock);
2954                 list_del_init(&cmd->i_conn_node);
2955                 spin_unlock_bh(&conn->cmd_lock);
2956                 isert_put_cmd(isert_cmd, true);
2957                 break;
2958         case ISTATE_SEND_NOPIN_WANT_RESPONSE:
2959                 ret = isert_put_nopin(cmd, conn, false);
2960                 break;
2961         default:
2962                 isert_err("Unknown immediate state: 0x%02x\n", state);
2963                 ret = -EINVAL;
2964                 break;
2965         }
2966
2967         return ret;
2968 }
2969
2970 static int
2971 isert_response_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2972 {
2973         struct isert_conn *isert_conn = conn->context;
2974         int ret;
2975
2976         switch (state) {
2977         case ISTATE_SEND_LOGOUTRSP:
2978                 ret = isert_put_logout_rsp(cmd, conn);
2979                 if (!ret)
2980                         isert_conn->logout_posted = true;
2981                 break;
2982         case ISTATE_SEND_NOPIN:
2983                 ret = isert_put_nopin(cmd, conn, true);
2984                 break;
2985         case ISTATE_SEND_TASKMGTRSP:
2986                 ret = isert_put_tm_rsp(cmd, conn);
2987                 break;
2988         case ISTATE_SEND_REJECT:
2989                 ret = isert_put_reject(cmd, conn);
2990                 break;
2991         case ISTATE_SEND_TEXTRSP:
2992                 ret = isert_put_text_rsp(cmd, conn);
2993                 break;
2994         case ISTATE_SEND_STATUS:
2995                 /*
2996                  * Special case for sending non GOOD SCSI status from TX thread
2997                  * context during pre se_cmd excecution failure.
2998                  */
2999                 ret = isert_put_response(conn, cmd);
3000                 break;
3001         default:
3002                 isert_err("Unknown response state: 0x%02x\n", state);
3003                 ret = -EINVAL;
3004                 break;
3005         }
3006
3007         return ret;
3008 }
3009
3010 struct rdma_cm_id *
3011 isert_setup_id(struct isert_np *isert_np)
3012 {
3013         struct iscsi_np *np = isert_np->np;
3014         struct rdma_cm_id *id;
3015         struct sockaddr *sa;
3016         int ret;
3017
3018         sa = (struct sockaddr *)&np->np_sockaddr;
3019         isert_dbg("ksockaddr: %p, sa: %p\n", &np->np_sockaddr, sa);
3020
3021         id = rdma_create_id(&init_net, isert_cma_handler, isert_np,
3022                             RDMA_PS_TCP, IB_QPT_RC);
3023         if (IS_ERR(id)) {
3024                 isert_err("rdma_create_id() failed: %ld\n", PTR_ERR(id));
3025                 ret = PTR_ERR(id);
3026                 goto out;
3027         }
3028         isert_dbg("id %p context %p\n", id, id->context);
3029
3030         ret = rdma_bind_addr(id, sa);
3031         if (ret) {
3032                 isert_err("rdma_bind_addr() failed: %d\n", ret);
3033                 goto out_id;
3034         }
3035
3036         ret = rdma_listen(id, 0);
3037         if (ret) {
3038                 isert_err("rdma_listen() failed: %d\n", ret);
3039                 goto out_id;
3040         }
3041
3042         return id;
3043 out_id:
3044         rdma_destroy_id(id);
3045 out:
3046         return ERR_PTR(ret);
3047 }
3048
3049 static int
3050 isert_setup_np(struct iscsi_np *np,
3051                struct sockaddr_storage *ksockaddr)
3052 {
3053         struct isert_np *isert_np;
3054         struct rdma_cm_id *isert_lid;
3055         int ret;
3056
3057         isert_np = kzalloc(sizeof(struct isert_np), GFP_KERNEL);
3058         if (!isert_np) {
3059                 isert_err("Unable to allocate struct isert_np\n");
3060                 return -ENOMEM;
3061         }
3062         sema_init(&isert_np->sem, 0);
3063         mutex_init(&isert_np->mutex);
3064         INIT_LIST_HEAD(&isert_np->accepted);
3065         INIT_LIST_HEAD(&isert_np->pending);
3066         isert_np->np = np;
3067
3068         /*
3069          * Setup the np->np_sockaddr from the passed sockaddr setup
3070          * in iscsi_target_configfs.c code..
3071          */
3072         memcpy(&np->np_sockaddr, ksockaddr,
3073                sizeof(struct sockaddr_storage));
3074
3075         isert_lid = isert_setup_id(isert_np);
3076         if (IS_ERR(isert_lid)) {
3077                 ret = PTR_ERR(isert_lid);
3078                 goto out;
3079         }
3080
3081         isert_np->cm_id = isert_lid;
3082         np->np_context = isert_np;
3083
3084         return 0;
3085
3086 out:
3087         kfree(isert_np);
3088
3089         return ret;
3090 }
3091
3092 static int
3093 isert_rdma_accept(struct isert_conn *isert_conn)
3094 {
3095         struct rdma_cm_id *cm_id = isert_conn->cm_id;
3096         struct rdma_conn_param cp;
3097         int ret;
3098
3099         memset(&cp, 0, sizeof(struct rdma_conn_param));
3100         cp.initiator_depth = isert_conn->initiator_depth;
3101         cp.retry_count = 7;
3102         cp.rnr_retry_count = 7;
3103
3104         ret = rdma_accept(cm_id, &cp);
3105         if (ret) {
3106                 isert_err("rdma_accept() failed with: %d\n", ret);
3107                 return ret;
3108         }
3109
3110         return 0;
3111 }
3112
3113 static int
3114 isert_get_login_rx(struct iscsi_conn *conn, struct iscsi_login *login)
3115 {
3116         struct isert_conn *isert_conn = conn->context;
3117         int ret;
3118
3119         isert_info("before login_req comp conn: %p\n", isert_conn);
3120         ret = wait_for_completion_interruptible(&isert_conn->login_req_comp);
3121         if (ret) {
3122                 isert_err("isert_conn %p interrupted before got login req\n",
3123                           isert_conn);
3124                 return ret;
3125         }
3126         reinit_completion(&isert_conn->login_req_comp);
3127
3128         /*
3129          * For login requests after the first PDU, isert_rx_login_req() will
3130          * kick schedule_delayed_work(&conn->login_work) as the packet is
3131          * received, which turns this callback from iscsi_target_do_login_rx()
3132          * into a NOP.
3133          */
3134         if (!login->first_request)
3135                 return 0;
3136
3137         isert_rx_login_req(isert_conn);
3138
3139         isert_info("before login_comp conn: %p\n", conn);
3140         ret = wait_for_completion_interruptible(&isert_conn->login_comp);
3141         if (ret)
3142                 return ret;
3143
3144         isert_info("processing login->req: %p\n", login->req);
3145
3146         return 0;
3147 }
3148
3149 static void
3150 isert_set_conn_info(struct iscsi_np *np, struct iscsi_conn *conn,
3151                     struct isert_conn *isert_conn)
3152 {
3153         struct rdma_cm_id *cm_id = isert_conn->cm_id;
3154         struct rdma_route *cm_route = &cm_id->route;
3155
3156         conn->login_family = np->np_sockaddr.ss_family;
3157
3158         conn->login_sockaddr = cm_route->addr.dst_addr;
3159         conn->local_sockaddr = cm_route->addr.src_addr;
3160 }
3161
3162 static int
3163 isert_accept_np(struct iscsi_np *np, struct iscsi_conn *conn)
3164 {
3165         struct isert_np *isert_np = np->np_context;
3166         struct isert_conn *isert_conn;
3167         int ret;
3168
3169 accept_wait:
3170         ret = down_interruptible(&isert_np->sem);
3171         if (ret)
3172                 return -ENODEV;
3173
3174         spin_lock_bh(&np->np_thread_lock);
3175         if (np->np_thread_state >= ISCSI_NP_THREAD_RESET) {
3176                 spin_unlock_bh(&np->np_thread_lock);
3177                 isert_dbg("np_thread_state %d\n",
3178                          np->np_thread_state);
3179                 /**
3180                  * No point in stalling here when np_thread
3181                  * is in state RESET/SHUTDOWN/EXIT - bail
3182                  **/
3183                 return -ENODEV;
3184         }
3185         spin_unlock_bh(&np->np_thread_lock);
3186
3187         mutex_lock(&isert_np->mutex);
3188         if (list_empty(&isert_np->pending)) {
3189                 mutex_unlock(&isert_np->mutex);
3190                 goto accept_wait;
3191         }
3192         isert_conn = list_first_entry(&isert_np->pending,
3193                         struct isert_conn, node);
3194         list_del_init(&isert_conn->node);
3195         mutex_unlock(&isert_np->mutex);
3196
3197         conn->context = isert_conn;
3198         isert_conn->conn = conn;
3199         isert_conn->state = ISER_CONN_BOUND;
3200
3201         isert_set_conn_info(np, conn, isert_conn);
3202
3203         isert_dbg("Processing isert_conn: %p\n", isert_conn);
3204
3205         return 0;
3206 }
3207
3208 static void
3209 isert_free_np(struct iscsi_np *np)
3210 {
3211         struct isert_np *isert_np = np->np_context;
3212         struct isert_conn *isert_conn, *n;
3213
3214         if (isert_np->cm_id)
3215                 rdma_destroy_id(isert_np->cm_id);
3216
3217         /*
3218          * FIXME: At this point we don't have a good way to insure
3219          * that at this point we don't have hanging connections that
3220          * completed RDMA establishment but didn't start iscsi login
3221          * process. So work-around this by cleaning up what ever piled
3222          * up in accepted and pending lists.
3223          */
3224         mutex_lock(&isert_np->mutex);
3225         if (!list_empty(&isert_np->pending)) {
3226                 isert_info("Still have isert pending connections\n");
3227                 list_for_each_entry_safe(isert_conn, n,
3228                                          &isert_np->pending,
3229                                          node) {
3230                         isert_info("cleaning isert_conn %p state (%d)\n",
3231                                    isert_conn, isert_conn->state);
3232                         isert_connect_release(isert_conn);
3233                 }
3234         }
3235
3236         if (!list_empty(&isert_np->accepted)) {
3237                 isert_info("Still have isert accepted connections\n");
3238                 list_for_each_entry_safe(isert_conn, n,
3239                                          &isert_np->accepted,
3240                                          node) {
3241                         isert_info("cleaning isert_conn %p state (%d)\n",
3242                                    isert_conn, isert_conn->state);
3243                         isert_connect_release(isert_conn);
3244                 }
3245         }
3246         mutex_unlock(&isert_np->mutex);
3247
3248         np->np_context = NULL;
3249         kfree(isert_np);
3250 }
3251
3252 static void isert_release_work(struct work_struct *work)
3253 {
3254         struct isert_conn *isert_conn = container_of(work,
3255                                                      struct isert_conn,
3256                                                      release_work);
3257
3258         isert_info("Starting release conn %p\n", isert_conn);
3259
3260         wait_for_completion(&isert_conn->wait);
3261
3262         mutex_lock(&isert_conn->mutex);
3263         isert_conn->state = ISER_CONN_DOWN;
3264         mutex_unlock(&isert_conn->mutex);
3265
3266         isert_info("Destroying conn %p\n", isert_conn);
3267         isert_put_conn(isert_conn);
3268 }
3269
3270 static void
3271 isert_wait4logout(struct isert_conn *isert_conn)
3272 {
3273         struct iscsi_conn *conn = isert_conn->conn;
3274
3275         isert_info("conn %p\n", isert_conn);
3276
3277         if (isert_conn->logout_posted) {
3278                 isert_info("conn %p wait for conn_logout_comp\n", isert_conn);
3279                 wait_for_completion_timeout(&conn->conn_logout_comp,
3280                                             SECONDS_FOR_LOGOUT_COMP * HZ);
3281         }
3282 }
3283
3284 static void
3285 isert_wait4cmds(struct iscsi_conn *conn)
3286 {
3287         isert_info("iscsi_conn %p\n", conn);
3288
3289         if (conn->sess) {
3290                 target_sess_cmd_list_set_waiting(conn->sess->se_sess);
3291                 target_wait_for_sess_cmds(conn->sess->se_sess);
3292         }
3293 }
3294
3295 static void
3296 isert_wait4flush(struct isert_conn *isert_conn)
3297 {
3298         struct ib_recv_wr *bad_wr;
3299
3300         isert_info("conn %p\n", isert_conn);
3301
3302         init_completion(&isert_conn->wait_comp_err);
3303         isert_conn->beacon.wr_id = ISER_BEACON_WRID;
3304         /* post an indication that all flush errors were consumed */
3305         if (ib_post_recv(isert_conn->qp, &isert_conn->beacon, &bad_wr)) {
3306                 isert_err("conn %p failed to post beacon", isert_conn);
3307                 return;
3308         }
3309
3310         wait_for_completion(&isert_conn->wait_comp_err);
3311 }
3312
3313 /**
3314  * isert_put_unsol_pending_cmds() - Drop commands waiting for
3315  *     unsolicitate dataout
3316  * @conn:    iscsi connection
3317  *
3318  * We might still have commands that are waiting for unsolicited
3319  * dataouts messages. We must put the extra reference on those
3320  * before blocking on the target_wait_for_session_cmds
3321  */
3322 static void
3323 isert_put_unsol_pending_cmds(struct iscsi_conn *conn)
3324 {
3325         struct iscsi_cmd *cmd, *tmp;
3326         static LIST_HEAD(drop_cmd_list);
3327
3328         spin_lock_bh(&conn->cmd_lock);
3329         list_for_each_entry_safe(cmd, tmp, &conn->conn_cmd_list, i_conn_node) {
3330                 if ((cmd->cmd_flags & ICF_NON_IMMEDIATE_UNSOLICITED_DATA) &&
3331                     (cmd->write_data_done < conn->sess->sess_ops->FirstBurstLength) &&
3332                     (cmd->write_data_done < cmd->se_cmd.data_length))
3333                         list_move_tail(&cmd->i_conn_node, &drop_cmd_list);
3334         }
3335         spin_unlock_bh(&conn->cmd_lock);
3336
3337         list_for_each_entry_safe(cmd, tmp, &drop_cmd_list, i_conn_node) {
3338                 list_del_init(&cmd->i_conn_node);
3339                 if (cmd->i_state != ISTATE_REMOVE) {
3340                         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
3341
3342                         isert_info("conn %p dropping cmd %p\n", conn, cmd);
3343                         isert_put_cmd(isert_cmd, true);
3344                 }
3345         }
3346 }
3347
3348 static void isert_wait_conn(struct iscsi_conn *conn)
3349 {
3350         struct isert_conn *isert_conn = conn->context;
3351
3352         isert_info("Starting conn %p\n", isert_conn);
3353
3354         mutex_lock(&isert_conn->mutex);
3355         /*
3356          * Only wait for wait_comp_err if the isert_conn made it
3357          * into full feature phase..
3358          */
3359         if (isert_conn->state == ISER_CONN_INIT) {
3360                 mutex_unlock(&isert_conn->mutex);
3361                 return;
3362         }
3363         isert_conn_terminate(isert_conn);
3364         mutex_unlock(&isert_conn->mutex);
3365
3366         isert_wait4flush(isert_conn);
3367         isert_put_unsol_pending_cmds(conn);
3368         isert_wait4cmds(conn);
3369         isert_wait4logout(isert_conn);
3370
3371         queue_work(isert_release_wq, &isert_conn->release_work);
3372 }
3373
3374 static void isert_free_conn(struct iscsi_conn *conn)
3375 {
3376         struct isert_conn *isert_conn = conn->context;
3377
3378         isert_wait4flush(isert_conn);
3379         isert_put_conn(isert_conn);
3380 }
3381
3382 static struct iscsit_transport iser_target_transport = {
3383         .name                   = "IB/iSER",
3384         .transport_type         = ISCSI_INFINIBAND,
3385         .priv_size              = sizeof(struct isert_cmd),
3386         .owner                  = THIS_MODULE,
3387         .iscsit_setup_np        = isert_setup_np,
3388         .iscsit_accept_np       = isert_accept_np,
3389         .iscsit_free_np         = isert_free_np,
3390         .iscsit_wait_conn       = isert_wait_conn,
3391         .iscsit_free_conn       = isert_free_conn,
3392         .iscsit_get_login_rx    = isert_get_login_rx,
3393         .iscsit_put_login_tx    = isert_put_login_tx,
3394         .iscsit_immediate_queue = isert_immediate_queue,
3395         .iscsit_response_queue  = isert_response_queue,
3396         .iscsit_get_dataout     = isert_get_dataout,
3397         .iscsit_queue_data_in   = isert_put_datain,
3398         .iscsit_queue_status    = isert_put_response,
3399         .iscsit_aborted_task    = isert_aborted_task,
3400         .iscsit_get_sup_prot_ops = isert_get_sup_prot_ops,
3401 };
3402
3403 static int __init isert_init(void)
3404 {
3405         int ret;
3406
3407         isert_comp_wq = alloc_workqueue("isert_comp_wq",
3408                                         WQ_UNBOUND | WQ_HIGHPRI, 0);
3409         if (!isert_comp_wq) {
3410                 isert_err("Unable to allocate isert_comp_wq\n");
3411                 ret = -ENOMEM;
3412                 return -ENOMEM;
3413         }
3414
3415         isert_release_wq = alloc_workqueue("isert_release_wq", WQ_UNBOUND,
3416                                         WQ_UNBOUND_MAX_ACTIVE);
3417         if (!isert_release_wq) {
3418                 isert_err("Unable to allocate isert_release_wq\n");
3419                 ret = -ENOMEM;
3420                 goto destroy_comp_wq;
3421         }
3422
3423         iscsit_register_transport(&iser_target_transport);
3424         isert_info("iSER_TARGET[0] - Loaded iser_target_transport\n");
3425
3426         return 0;
3427
3428 destroy_comp_wq:
3429         destroy_workqueue(isert_comp_wq);
3430
3431         return ret;
3432 }
3433
3434 static void __exit isert_exit(void)
3435 {
3436         flush_scheduled_work();
3437         destroy_workqueue(isert_release_wq);
3438         destroy_workqueue(isert_comp_wq);
3439         iscsit_unregister_transport(&iser_target_transport);
3440         isert_info("iSER_TARGET[0] - Released iser_target_transport\n");
3441 }
3442
3443 MODULE_DESCRIPTION("iSER-Target for mainline target infrastructure");
3444 MODULE_VERSION("1.0");
3445 MODULE_AUTHOR("nab@Linux-iSCSI.org");
3446 MODULE_LICENSE("GPL");
3447
3448 module_init(isert_init);
3449 module_exit(isert_exit);