kernel/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 FULL_FEATURE, 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_FULL_FEATURE: /* FALLTHRU */
 840                 isert_info("Terminating conn %p state %d\n",
 841                            isert_conn, isert_conn->state);
 842                 isert_conn->state = ISER_CONN_TERMINATING;
 843                 err = rdma_disconnect(isert_conn->cm_id);
 844                 if (err)
 845                         isert_warn("Failed rdma_disconnect isert_conn %p\n",
 846                                    isert_conn);
 847                 break;
 848         default:
 849                 isert_warn("conn %p teminating in state %d\n",
 850                            isert_conn, isert_conn->state);
 851         }
 852 }
 853
 854 static int
 855 isert_np_cma_handler(struct isert_np *isert_np,
 856                      enum rdma_cm_event_type event)
 857 {
 858         isert_dbg("%s (%d): isert np %p\n",
 859                   rdma_event_msg(event), event, isert_np);
 860
 861         switch (event) {
 862         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 863                 isert_np->cm_id = NULL;
 864                 break;
 865         case RDMA_CM_EVENT_ADDR_CHANGE:
 866                 isert_np->cm_id = isert_setup_id(isert_np);
 867                 if (IS_ERR(isert_np->cm_id)) {
 868                         isert_err("isert np %p setup id failed: %ld\n",
 869                                   isert_np, PTR_ERR(isert_np->cm_id));
 870                         isert_np->cm_id = NULL;
 871                 }
 872                 break;
 873         default:
 874                 isert_err("isert np %p Unexpected event %d\n",
 875                           isert_np, event);
 876         }
 877
 878         return -1;
 879 }
 880
 881 static int
 882 isert_disconnected_handler(struct rdma_cm_id *cma_id,
 883                            enum rdma_cm_event_type event)
 884 {
 885         struct isert_np *isert_np = cma_id->context;
 886         struct isert_conn *isert_conn;
 887         bool terminating = false;
 888
 889         if (isert_np->cm_id == cma_id)
 890                 return isert_np_cma_handler(cma_id->context, event);
 891
 892         isert_conn = cma_id->qp->qp_context;
 893
 894         mutex_lock(&isert_conn->mutex);
 895         terminating = (isert_conn->state == ISER_CONN_TERMINATING);
 896         isert_conn_terminate(isert_conn);
 897         mutex_unlock(&isert_conn->mutex);
 898
 899         isert_info("conn %p completing wait\n", isert_conn);
 900         complete(&isert_conn->wait);
 901
 902         if (terminating)
 903                 goto out;
 904
 905         mutex_lock(&isert_np->mutex);
 906         if (!list_empty(&isert_conn->node)) {
 907                 list_del_init(&isert_conn->node);
 908                 isert_put_conn(isert_conn);
 909                 queue_work(isert_release_wq, &isert_conn->release_work);
 910         }
 911         mutex_unlock(&isert_np->mutex);
 912
 913 out:
 914         return 0;
 915 }
 916
 917 static int
 918 isert_connect_error(struct rdma_cm_id *cma_id)
 919 {
 920         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 921
 922         list_del_init(&isert_conn->node);
 923         isert_conn->cm_id = NULL;
 924         isert_put_conn(isert_conn);
 925
 926         return -1;
 927 }
 928
 929 static int
 930 isert_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 931 {
 932         int ret = 0;
 933
 934         isert_info("%s (%d): status %d id %p np %p\n",
 935                    rdma_event_msg(event->event), event->event,
 936                    event->status, cma_id, cma_id->context);
 937
 938         switch (event->event) {
 939         case RDMA_CM_EVENT_CONNECT_REQUEST:
 940                 ret = isert_connect_request(cma_id, event);
 941                 if (ret)
 942                         isert_err("failed handle connect request %d\n", ret);
 943                 break;
 944         case RDMA_CM_EVENT_ESTABLISHED:
 945                 isert_connected_handler(cma_id);
 946                 break;
 947         case RDMA_CM_EVENT_ADDR_CHANGE:    /* FALLTHRU */
 948         case RDMA_CM_EVENT_DISCONNECTED:   /* FALLTHRU */
 949         case RDMA_CM_EVENT_DEVICE_REMOVAL: /* FALLTHRU */
 950         case RDMA_CM_EVENT_TIMEWAIT_EXIT:  /* FALLTHRU */
 951                 ret = isert_disconnected_handler(cma_id, event->event);
 952                 break;
 953         case RDMA_CM_EVENT_REJECTED:       /* FALLTHRU */
 954         case RDMA_CM_EVENT_UNREACHABLE:    /* FALLTHRU */
 955         case RDMA_CM_EVENT_CONNECT_ERROR:
 956                 ret = isert_connect_error(cma_id);
 957                 break;
 958         default:
 959                 isert_err("Unhandled RDMA CMA event: %d\n", event->event);
 960                 break;
 961         }
 962
 963         return ret;
 964 }
 965
 966 static int
 967 isert_post_recvm(struct isert_conn *isert_conn, u32 count)
 968 {
 969         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 970         int i, ret;
 971         struct iser_rx_desc *rx_desc;
 972
 973         for (rx_wr = isert_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
 974                 rx_desc = &isert_conn->rx_descs[i];
 975                 rx_wr->wr_id = (uintptr_t)rx_desc;
 976                 rx_wr->sg_list = &rx_desc->rx_sg;
 977                 rx_wr->num_sge = 1;
 978                 rx_wr->next = rx_wr + 1;
 979         }
 980         rx_wr--;
 981         rx_wr->next = NULL; /* mark end of work requests list */
 982
 983         isert_conn->post_recv_buf_count += count;
 984         ret = ib_post_recv(isert_conn->qp, isert_conn->rx_wr,
 985                            &rx_wr_failed);
 986         if (ret) {
 987                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
 988                 isert_conn->post_recv_buf_count -= count;
 989         }
 990
 991         return ret;
 992 }
 993
 994 static int
 995 isert_post_recv(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc)
 996 {
 997         struct ib_recv_wr *rx_wr_failed, rx_wr;
 998         int ret;
 999
1000         rx_wr.wr_id = (uintptr_t)rx_desc;
1001         rx_wr.sg_list = &rx_desc->rx_sg;
1002         rx_wr.num_sge = 1;
1003         rx_wr.next = NULL;
1004
1005         isert_conn->post_recv_buf_count++;
1006         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_failed);
1007         if (ret) {
1008                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
1009                 isert_conn->post_recv_buf_count--;
1010         }
1011
1012         return ret;
1013 }
1014
1015 static int
1016 isert_post_send(struct isert_conn *isert_conn, struct iser_tx_desc *tx_desc)
1017 {
1018         struct ib_device *ib_dev = isert_conn->cm_id->device;
1019         struct ib_send_wr send_wr, *send_wr_failed;
1020         int ret;
1021
1022         ib_dma_sync_single_for_device(ib_dev, tx_desc->dma_addr,
1023                                       ISER_HEADERS_LEN, DMA_TO_DEVICE);
1024
1025         send_wr.next    = NULL;
1026         send_wr.wr_id   = (uintptr_t)tx_desc;
1027         send_wr.sg_list = tx_desc->tx_sg;
1028         send_wr.num_sge = tx_desc->num_sge;
1029         send_wr.opcode  = IB_WR_SEND;
1030         send_wr.send_flags = IB_SEND_SIGNALED;
1031
1032         ret = ib_post_send(isert_conn->qp, &send_wr, &send_wr_failed);
1033         if (ret)
1034                 isert_err("ib_post_send() failed, ret: %d\n", ret);
1035
1036         return ret;
1037 }
1038
1039 static void
1040 isert_create_send_desc(struct isert_conn *isert_conn,
1041                        struct isert_cmd *isert_cmd,
1042                        struct iser_tx_desc *tx_desc)
1043 {
1044         struct isert_device *device = isert_conn->device;
1045         struct ib_device *ib_dev = device->ib_device;
1046
1047         ib_dma_sync_single_for_cpu(ib_dev, tx_desc->dma_addr,
1048                                    ISER_HEADERS_LEN, DMA_TO_DEVICE);
1049
1050         memset(&tx_desc->iser_header, 0, sizeof(struct iser_hdr));
1051         tx_desc->iser_header.flags = ISER_VER;
1052
1053         tx_desc->num_sge = 1;
1054         tx_desc->isert_cmd = isert_cmd;
1055
1056         if (tx_desc->tx_sg[0].lkey != device->pd->local_dma_lkey) {
1057                 tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
1058                 isert_dbg("tx_desc %p lkey mismatch, fixing\n", tx_desc);
1059         }
1060 }
1061
1062 static int
1063 isert_init_tx_hdrs(struct isert_conn *isert_conn,
1064                    struct iser_tx_desc *tx_desc)
1065 {
1066         struct isert_device *device = isert_conn->device;
1067         struct ib_device *ib_dev = device->ib_device;
1068         u64 dma_addr;
1069
1070         dma_addr = ib_dma_map_single(ib_dev, (void *)tx_desc,
1071                         ISER_HEADERS_LEN, DMA_TO_DEVICE);
1072         if (ib_dma_mapping_error(ib_dev, dma_addr)) {
1073                 isert_err("ib_dma_mapping_error() failed\n");
1074                 return -ENOMEM;
1075         }
1076
1077         tx_desc->dma_addr = dma_addr;
1078         tx_desc->tx_sg[0].addr  = tx_desc->dma_addr;
1079         tx_desc->tx_sg[0].length = ISER_HEADERS_LEN;
1080         tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
1081
1082         isert_dbg("Setup tx_sg[0].addr: 0x%llx length: %u lkey: 0x%x\n",
1083                   tx_desc->tx_sg[0].addr, tx_desc->tx_sg[0].length,
1084                   tx_desc->tx_sg[0].lkey);
1085
1086         return 0;
1087 }
1088
1089 static void
1090 isert_init_send_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1091                    struct ib_send_wr *send_wr)
1092 {
1093         struct iser_tx_desc *tx_desc = &isert_cmd->tx_desc;
1094
1095         isert_cmd->rdma_wr.iser_ib_op = ISER_IB_SEND;
1096         send_wr->wr_id = (uintptr_t)&isert_cmd->tx_desc;
1097         send_wr->opcode = IB_WR_SEND;
1098         send_wr->sg_list = &tx_desc->tx_sg[0];
1099         send_wr->num_sge = isert_cmd->tx_desc.num_sge;
1100         send_wr->send_flags = IB_SEND_SIGNALED;
1101 }
1102
1103 static int
1104 isert_rdma_post_recvl(struct isert_conn *isert_conn)
1105 {
1106         struct ib_recv_wr rx_wr, *rx_wr_fail;
1107         struct ib_sge sge;
1108         int ret;
1109
1110         memset(&sge, 0, sizeof(struct ib_sge));
1111         sge.addr = isert_conn->login_req_dma;
1112         sge.length = ISER_RX_LOGIN_SIZE;
1113         sge.lkey = isert_conn->device->pd->local_dma_lkey;
1114
1115         isert_dbg("Setup sge: addr: %llx length: %d 0x%08x\n",
1116                 sge.addr, sge.length, sge.lkey);
1117
1118         memset(&rx_wr, 0, sizeof(struct ib_recv_wr));
1119         rx_wr.wr_id = (uintptr_t)isert_conn->login_req_buf;
1120         rx_wr.sg_list = &sge;
1121         rx_wr.num_sge = 1;
1122
1123         isert_conn->post_recv_buf_count++;
1124         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_fail);
1125         if (ret) {
1126                 isert_err("ib_post_recv() failed: %d\n", ret);
1127                 isert_conn->post_recv_buf_count--;
1128         }
1129
1130         return ret;
1131 }
1132
1133 static int
1134 isert_put_login_tx(struct iscsi_conn *conn, struct iscsi_login *login,
1135                    u32 length)
1136 {
1137         struct isert_conn *isert_conn = conn->context;
1138         struct isert_device *device = isert_conn->device;
1139         struct ib_device *ib_dev = device->ib_device;
1140         struct iser_tx_desc *tx_desc = &isert_conn->login_tx_desc;
1141         int ret;
1142
1143         isert_create_send_desc(isert_conn, NULL, tx_desc);
1144
1145         memcpy(&tx_desc->iscsi_header, &login->rsp[0],
1146                sizeof(struct iscsi_hdr));
1147
1148         isert_init_tx_hdrs(isert_conn, tx_desc);
1149
1150         if (length > 0) {
1151                 struct ib_sge *tx_dsg = &tx_desc->tx_sg[1];
1152
1153                 ib_dma_sync_single_for_cpu(ib_dev, isert_conn->login_rsp_dma,
1154                                            length, DMA_TO_DEVICE);
1155
1156                 memcpy(isert_conn->login_rsp_buf, login->rsp_buf, length);
1157
1158                 ib_dma_sync_single_for_device(ib_dev, isert_conn->login_rsp_dma,
1159                                               length, DMA_TO_DEVICE);
1160
1161                 tx_dsg->addr    = isert_conn->login_rsp_dma;
1162                 tx_dsg->length  = length;
1163                 tx_dsg->lkey    = isert_conn->device->pd->local_dma_lkey;
1164                 tx_desc->num_sge = 2;
1165         }
1166         if (!login->login_failed) {
1167                 if (login->login_complete) {
1168                         if (!conn->sess->sess_ops->SessionType &&
1169                             isert_conn->device->use_fastreg) {
1170                                 ret = isert_conn_create_fastreg_pool(isert_conn);
1171                                 if (ret) {
1172                                         isert_err("Conn: %p failed to create"
1173                                                " fastreg pool\n", isert_conn);
1174                                         return ret;
1175                                 }
1176                         }
1177
1178                         ret = isert_alloc_rx_descriptors(isert_conn);
1179                         if (ret)
1180                                 return ret;
1181
1182                         ret = isert_post_recvm(isert_conn,
1183                                                ISERT_QP_MAX_RECV_DTOS);
1184                         if (ret)
1185                                 return ret;
1186
1187                         /* Now we are in FULL_FEATURE phase */
1188                         mutex_lock(&isert_conn->mutex);
1189                         isert_conn->state = ISER_CONN_FULL_FEATURE;
1190                         mutex_unlock(&isert_conn->mutex);
1191                         goto post_send;
1192                 }
1193
1194                 ret = isert_rdma_post_recvl(isert_conn);
1195                 if (ret)
1196                         return ret;
1197         }
1198 post_send:
1199         ret = isert_post_send(isert_conn, tx_desc);
1200         if (ret)
1201                 return ret;
1202
1203         return 0;
1204 }
1205
1206 static void
1207 isert_rx_login_req(struct isert_conn *isert_conn)
1208 {
1209         struct iser_rx_desc *rx_desc = (void *)isert_conn->login_req_buf;
1210         int rx_buflen = isert_conn->login_req_len;
1211         struct iscsi_conn *conn = isert_conn->conn;
1212         struct iscsi_login *login = conn->conn_login;
1213         int size;
1214
1215         isert_info("conn %p\n", isert_conn);
1216
1217         WARN_ON_ONCE(!login);
1218
1219         if (login->first_request) {
1220                 struct iscsi_login_req *login_req =
1221                         (struct iscsi_login_req *)&rx_desc->iscsi_header;
1222                 /*
1223                  * Setup the initial iscsi_login values from the leading
1224                  * login request PDU.
1225                  */
1226                 login->leading_connection = (!login_req->tsih) ? 1 : 0;
1227                 login->current_stage =
1228                         (login_req->flags & ISCSI_FLAG_LOGIN_CURRENT_STAGE_MASK)
1229                          >> 2;
1230                 login->version_min      = login_req->min_version;
1231                 login->version_max      = login_req->max_version;
1232                 memcpy(login->isid, login_req->isid, 6);
1233                 login->cmd_sn           = be32_to_cpu(login_req->cmdsn);
1234                 login->init_task_tag    = login_req->itt;
1235                 login->initial_exp_statsn = be32_to_cpu(login_req->exp_statsn);
1236                 login->cid              = be16_to_cpu(login_req->cid);
1237                 login->tsih             = be16_to_cpu(login_req->tsih);
1238         }
1239
1240         memcpy(&login->req[0], (void *)&rx_desc->iscsi_header, ISCSI_HDR_LEN);
1241
1242         size = min(rx_buflen, MAX_KEY_VALUE_PAIRS);
1243         isert_dbg("Using login payload size: %d, rx_buflen: %d "
1244                   "MAX_KEY_VALUE_PAIRS: %d\n", size, rx_buflen,
1245                   MAX_KEY_VALUE_PAIRS);
1246         memcpy(login->req_buf, &rx_desc->data[0], size);
1247
1248         if (login->first_request) {
1249                 complete(&isert_conn->login_comp);
1250                 return;
1251         }
1252         schedule_delayed_work(&conn->login_work, 0);
1253 }
1254
1255 static struct iscsi_cmd
1256 *isert_allocate_cmd(struct iscsi_conn *conn, struct iser_rx_desc *rx_desc)
1257 {
1258         struct isert_conn *isert_conn = conn->context;
1259         struct isert_cmd *isert_cmd;
1260         struct iscsi_cmd *cmd;
1261
1262         cmd = iscsit_allocate_cmd(conn, TASK_INTERRUPTIBLE);
1263         if (!cmd) {
1264                 isert_err("Unable to allocate iscsi_cmd + isert_cmd\n");
1265                 return NULL;
1266         }
1267         isert_cmd = iscsit_priv_cmd(cmd);
1268         isert_cmd->conn = isert_conn;
1269         isert_cmd->iscsi_cmd = cmd;
1270         isert_cmd->rx_desc = rx_desc;
1271
1272         return cmd;
1273 }
1274
1275 static int
1276 isert_handle_scsi_cmd(struct isert_conn *isert_conn,
1277                       struct isert_cmd *isert_cmd, struct iscsi_cmd *cmd,
1278                       struct iser_rx_desc *rx_desc, unsigned char *buf)
1279 {
1280         struct iscsi_conn *conn = isert_conn->conn;
1281         struct iscsi_scsi_req *hdr = (struct iscsi_scsi_req *)buf;
1282         int imm_data, imm_data_len, unsol_data, sg_nents, rc;
1283         bool dump_payload = false;
1284         unsigned int data_len;
1285
1286         rc = iscsit_setup_scsi_cmd(conn, cmd, buf);
1287         if (rc < 0)
1288                 return rc;
1289
1290         imm_data = cmd->immediate_data;
1291         imm_data_len = cmd->first_burst_len;
1292         unsol_data = cmd->unsolicited_data;
1293         data_len = cmd->se_cmd.data_length;
1294
1295         if (imm_data && imm_data_len == data_len)
1296                 cmd->se_cmd.se_cmd_flags |= SCF_PASSTHROUGH_SG_TO_MEM_NOALLOC;
1297         rc = iscsit_process_scsi_cmd(conn, cmd, hdr);
1298         if (rc < 0) {
1299                 return 0;
1300         } else if (rc > 0) {
1301                 dump_payload = true;
1302                 goto sequence_cmd;
1303         }
1304
1305         if (!imm_data)
1306                 return 0;
1307
1308         if (imm_data_len != data_len) {
1309                 sg_nents = max(1UL, DIV_ROUND_UP(imm_data_len, PAGE_SIZE));
1310                 sg_copy_from_buffer(cmd->se_cmd.t_data_sg, sg_nents,
1311                                     &rx_desc->data[0], imm_data_len);
1312                 isert_dbg("Copy Immediate sg_nents: %u imm_data_len: %d\n",
1313                           sg_nents, imm_data_len);
1314         } else {
1315                 sg_init_table(&isert_cmd->sg, 1);
1316                 cmd->se_cmd.t_data_sg = &isert_cmd->sg;
1317                 cmd->se_cmd.t_data_nents = 1;
1318                 sg_set_buf(&isert_cmd->sg, &rx_desc->data[0], imm_data_len);
1319                 isert_dbg("Transfer Immediate imm_data_len: %d\n",
1320                           imm_data_len);
1321         }
1322
1323         cmd->write_data_done += imm_data_len;
1324
1325         if (cmd->write_data_done == cmd->se_cmd.data_length) {
1326                 spin_lock_bh(&cmd->istate_lock);
1327                 cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1328                 cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1329                 spin_unlock_bh(&cmd->istate_lock);
1330         }
1331
1332 sequence_cmd:
1333         rc = iscsit_sequence_cmd(conn, cmd, buf, hdr->cmdsn);
1334
1335         if (!rc && dump_payload == false && unsol_data)
1336                 iscsit_set_unsoliticed_dataout(cmd);
1337         else if (dump_payload && imm_data)
1338                 target_put_sess_cmd(&cmd->se_cmd);
1339
1340         return 0;
1341 }
1342
1343 static int
1344 isert_handle_iscsi_dataout(struct isert_conn *isert_conn,
1345                            struct iser_rx_desc *rx_desc, unsigned char *buf)
1346 {
1347         struct scatterlist *sg_start;
1348         struct iscsi_conn *conn = isert_conn->conn;
1349         struct iscsi_cmd *cmd = NULL;
1350         struct iscsi_data *hdr = (struct iscsi_data *)buf;
1351         u32 unsol_data_len = ntoh24(hdr->dlength);
1352         int rc, sg_nents, sg_off, page_off;
1353
1354         rc = iscsit_check_dataout_hdr(conn, buf, &cmd);
1355         if (rc < 0)
1356                 return rc;
1357         else if (!cmd)
1358                 return 0;
1359         /*
1360          * FIXME: Unexpected unsolicited_data out
1361          */
1362         if (!cmd->unsolicited_data) {
1363                 isert_err("Received unexpected solicited data payload\n");
1364                 dump_stack();
1365                 return -1;
1366         }
1367
1368         isert_dbg("Unsolicited DataOut unsol_data_len: %u, "
1369                   "write_data_done: %u, data_length: %u\n",
1370                   unsol_data_len,  cmd->write_data_done,
1371                   cmd->se_cmd.data_length);
1372
1373         sg_off = cmd->write_data_done / PAGE_SIZE;
1374         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
1375         sg_nents = max(1UL, DIV_ROUND_UP(unsol_data_len, PAGE_SIZE));
1376         page_off = cmd->write_data_done % PAGE_SIZE;
1377         /*
1378          * FIXME: Non page-aligned unsolicited_data out
1379          */
1380         if (page_off) {
1381                 isert_err("unexpected non-page aligned data payload\n");
1382                 dump_stack();
1383                 return -1;
1384         }
1385         isert_dbg("Copying DataOut: sg_start: %p, sg_off: %u "
1386                   "sg_nents: %u from %p %u\n", sg_start, sg_off,
1387                   sg_nents, &rx_desc->data[0], unsol_data_len);
1388
1389         sg_copy_from_buffer(sg_start, sg_nents, &rx_desc->data[0],
1390                             unsol_data_len);
1391
1392         rc = iscsit_check_dataout_payload(cmd, hdr, false);
1393         if (rc < 0)
1394                 return rc;
1395
1396         /*
1397          * multiple data-outs on the same command can arrive -
1398          * so post the buffer before hand
1399          */
1400         rc = isert_post_recv(isert_conn, rx_desc);
1401         if (rc) {
1402                 isert_err("ib_post_recv failed with %d\n", rc);
1403                 return rc;
1404         }
1405         return 0;
1406 }
1407
1408 static int
1409 isert_handle_nop_out(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1410                      struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1411                      unsigned char *buf)
1412 {
1413         struct iscsi_conn *conn = isert_conn->conn;
1414         struct iscsi_nopout *hdr = (struct iscsi_nopout *)buf;
1415         int rc;
1416
1417         rc = iscsit_setup_nop_out(conn, cmd, hdr);
1418         if (rc < 0)
1419                 return rc;
1420         /*
1421          * FIXME: Add support for NOPOUT payload using unsolicited RDMA payload
1422          */
1423
1424         return iscsit_process_nop_out(conn, cmd, hdr);
1425 }
1426
1427 static int
1428 isert_handle_text_cmd(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1429                       struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1430                       struct iscsi_text *hdr)
1431 {
1432         struct iscsi_conn *conn = isert_conn->conn;
1433         u32 payload_length = ntoh24(hdr->dlength);
1434         int rc;
1435         unsigned char *text_in = NULL;
1436
1437         rc = iscsit_setup_text_cmd(conn, cmd, hdr);
1438         if (rc < 0)
1439                 return rc;
1440
1441         if (payload_length) {
1442                 text_in = kzalloc(payload_length, GFP_KERNEL);
1443                 if (!text_in) {
1444                         isert_err("Unable to allocate text_in of payload_length: %u\n",
1445                                   payload_length);
1446                         return -ENOMEM;
1447                 }
1448         }
1449         cmd->text_in_ptr = text_in;
1450
1451         memcpy(cmd->text_in_ptr, &rx_desc->data[0], payload_length);
1452
1453         return iscsit_process_text_cmd(conn, cmd, hdr);
1454 }
1455
1456 static int
1457 isert_rx_opcode(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc,
1458                 uint32_t read_stag, uint64_t read_va,
1459                 uint32_t write_stag, uint64_t write_va)
1460 {
1461         struct iscsi_hdr *hdr = &rx_desc->iscsi_header;
1462         struct iscsi_conn *conn = isert_conn->conn;
1463         struct iscsi_cmd *cmd;
1464         struct isert_cmd *isert_cmd;
1465         int ret = -EINVAL;
1466         u8 opcode = (hdr->opcode & ISCSI_OPCODE_MASK);
1467
1468         if (conn->sess->sess_ops->SessionType &&
1469            (!(opcode & ISCSI_OP_TEXT) || !(opcode & ISCSI_OP_LOGOUT))) {
1470                 isert_err("Got illegal opcode: 0x%02x in SessionType=Discovery,"
1471                           " ignoring\n", opcode);
1472                 return 0;
1473         }
1474
1475         switch (opcode) {
1476         case ISCSI_OP_SCSI_CMD:
1477                 cmd = isert_allocate_cmd(conn, rx_desc);
1478                 if (!cmd)
1479                         break;
1480
1481                 isert_cmd = iscsit_priv_cmd(cmd);
1482                 isert_cmd->read_stag = read_stag;
1483                 isert_cmd->read_va = read_va;
1484                 isert_cmd->write_stag = write_stag;
1485                 isert_cmd->write_va = write_va;
1486
1487                 ret = isert_handle_scsi_cmd(isert_conn, isert_cmd, cmd,
1488                                         rx_desc, (unsigned char *)hdr);
1489                 break;
1490         case ISCSI_OP_NOOP_OUT:
1491                 cmd = isert_allocate_cmd(conn, rx_desc);
1492                 if (!cmd)
1493                         break;
1494
1495                 isert_cmd = iscsit_priv_cmd(cmd);
1496                 ret = isert_handle_nop_out(isert_conn, isert_cmd, cmd,
1497                                            rx_desc, (unsigned char *)hdr);
1498                 break;
1499         case ISCSI_OP_SCSI_DATA_OUT:
1500                 ret = isert_handle_iscsi_dataout(isert_conn, rx_desc,
1501                                                 (unsigned char *)hdr);
1502                 break;
1503         case ISCSI_OP_SCSI_TMFUNC:
1504                 cmd = isert_allocate_cmd(conn, rx_desc);
1505                 if (!cmd)
1506                         break;
1507
1508                 ret = iscsit_handle_task_mgt_cmd(conn, cmd,
1509                                                 (unsigned char *)hdr);
1510                 break;
1511         case ISCSI_OP_LOGOUT:
1512                 cmd = isert_allocate_cmd(conn, rx_desc);
1513                 if (!cmd)
1514                         break;
1515
1516                 ret = iscsit_handle_logout_cmd(conn, cmd, (unsigned char *)hdr);
1517                 break;
1518         case ISCSI_OP_TEXT:
1519                 if (be32_to_cpu(hdr->ttt) != 0xFFFFFFFF)
1520                         cmd = iscsit_find_cmd_from_itt(conn, hdr->itt);
1521                 else
1522                         cmd = isert_allocate_cmd(conn, rx_desc);
1523
1524                 if (!cmd)
1525                         break;
1526
1527                 isert_cmd = iscsit_priv_cmd(cmd);
1528                 ret = isert_handle_text_cmd(isert_conn, isert_cmd, cmd,
1529                                             rx_desc, (struct iscsi_text *)hdr);
1530                 break;
1531         default:
1532                 isert_err("Got unknown iSCSI OpCode: 0x%02x\n", opcode);
1533                 dump_stack();
1534                 break;
1535         }
1536
1537         return ret;
1538 }
1539
1540 static void
1541 isert_rx_do_work(struct iser_rx_desc *rx_desc, struct isert_conn *isert_conn)
1542 {
1543         struct iser_hdr *iser_hdr = &rx_desc->iser_header;
1544         uint64_t read_va = 0, write_va = 0;
1545         uint32_t read_stag = 0, write_stag = 0;
1546
1547         switch (iser_hdr->flags & 0xF0) {
1548         case ISCSI_CTRL:
1549                 if (iser_hdr->flags & ISER_RSV) {
1550                         read_stag = be32_to_cpu(iser_hdr->read_stag);
1551                         read_va = be64_to_cpu(iser_hdr->read_va);
1552                         isert_dbg("ISER_RSV: read_stag: 0x%x read_va: 0x%llx\n",
1553                                   read_stag, (unsigned long long)read_va);
1554                 }
1555                 if (iser_hdr->flags & ISER_WSV) {
1556                         write_stag = be32_to_cpu(iser_hdr->write_stag);
1557                         write_va = be64_to_cpu(iser_hdr->write_va);
1558                         isert_dbg("ISER_WSV: write_stag: 0x%x write_va: 0x%llx\n",
1559                                   write_stag, (unsigned long long)write_va);
1560                 }
1561
1562                 isert_dbg("ISER ISCSI_CTRL PDU\n");
1563                 break;
1564         case ISER_HELLO:
1565                 isert_err("iSER Hello message\n");
1566                 break;
1567         default:
1568                 isert_warn("Unknown iSER hdr flags: 0x%02x\n", iser_hdr->flags);
1569                 break;
1570         }
1571
1572         isert_rx_opcode(isert_conn, rx_desc,
1573                         read_stag, read_va, write_stag, write_va);
1574 }
1575
1576 static void
1577 isert_rcv_completion(struct iser_rx_desc *desc,
1578                      struct isert_conn *isert_conn,
1579                      u32 xfer_len)
1580 {
1581         struct ib_device *ib_dev = isert_conn->cm_id->device;
1582         struct iscsi_hdr *hdr;
1583         u64 rx_dma;
1584         int rx_buflen;
1585
1586         if ((char *)desc == isert_conn->login_req_buf) {
1587                 rx_dma = isert_conn->login_req_dma;
1588                 rx_buflen = ISER_RX_LOGIN_SIZE;
1589                 isert_dbg("login_buf: Using rx_dma: 0x%llx, rx_buflen: %d\n",
1590                          rx_dma, rx_buflen);
1591         } else {
1592                 rx_dma = desc->dma_addr;
1593                 rx_buflen = ISER_RX_PAYLOAD_SIZE;
1594                 isert_dbg("req_buf: Using rx_dma: 0x%llx, rx_buflen: %d\n",
1595                          rx_dma, rx_buflen);
1596         }
1597
1598         ib_dma_sync_single_for_cpu(ib_dev, rx_dma, rx_buflen, DMA_FROM_DEVICE);
1599
1600         hdr = &desc->iscsi_header;
1601         isert_dbg("iSCSI opcode: 0x%02x, ITT: 0x%08x, flags: 0x%02x dlen: %d\n",
1602                  hdr->opcode, hdr->itt, hdr->flags,
1603                  (int)(xfer_len - ISER_HEADERS_LEN));
1604
1605         if ((char *)desc == isert_conn->login_req_buf) {
1606                 isert_conn->login_req_len = xfer_len - ISER_HEADERS_LEN;
1607                 if (isert_conn->conn) {
1608                         struct iscsi_login *login = isert_conn->conn->conn_login;
1609
1610                         if (login && !login->first_request)
1611                                 isert_rx_login_req(isert_conn);
1612                 }
1613                 mutex_lock(&isert_conn->mutex);
1614                 complete(&isert_conn->login_req_comp);
1615                 mutex_unlock(&isert_conn->mutex);
1616         } else {
1617                 isert_rx_do_work(desc, isert_conn);
1618         }
1619
1620         ib_dma_sync_single_for_device(ib_dev, rx_dma, rx_buflen,
1621                                       DMA_FROM_DEVICE);
1622
1623         isert_conn->post_recv_buf_count--;
1624 }
1625
1626 static int
1627 isert_map_data_buf(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1628                    struct scatterlist *sg, u32 nents, u32 length, u32 offset,
1629                    enum iser_ib_op_code op, struct isert_data_buf *data)
1630 {
1631         struct ib_device *ib_dev = isert_conn->cm_id->device;
1632
1633         data->dma_dir = op == ISER_IB_RDMA_WRITE ?
1634                               DMA_TO_DEVICE : DMA_FROM_DEVICE;
1635
1636         data->len = length - offset;
1637         data->offset = offset;
1638         data->sg_off = data->offset / PAGE_SIZE;
1639
1640         data->sg = &sg[data->sg_off];
1641         data->nents = min_t(unsigned int, nents - data->sg_off,
1642                                           ISCSI_ISER_SG_TABLESIZE);
1643         data->len = min_t(unsigned int, data->len, ISCSI_ISER_SG_TABLESIZE *
1644                                         PAGE_SIZE);
1645
1646         data->dma_nents = ib_dma_map_sg(ib_dev, data->sg, data->nents,
1647                                         data->dma_dir);
1648         if (unlikely(!data->dma_nents)) {
1649                 isert_err("Cmd: unable to dma map SGs %p\n", sg);
1650                 return -EINVAL;
1651         }
1652
1653         isert_dbg("Mapped cmd: %p count: %u sg: %p sg_nents: %u rdma_len %d\n",
1654                   isert_cmd, data->dma_nents, data->sg, data->nents, data->len);
1655
1656         return 0;
1657 }
1658
1659 static void
1660 isert_unmap_data_buf(struct isert_conn *isert_conn, struct isert_data_buf *data)
1661 {
1662         struct ib_device *ib_dev = isert_conn->cm_id->device;
1663
1664         ib_dma_unmap_sg(ib_dev, data->sg, data->nents, data->dma_dir);
1665         memset(data, 0, sizeof(*data));
1666 }
1667
1668
1669
1670 static void
1671 isert_unmap_cmd(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn)
1672 {
1673         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1674
1675         isert_dbg("Cmd %p\n", isert_cmd);
1676
1677         if (wr->data.sg) {
1678                 isert_dbg("Cmd %p unmap_sg op\n", isert_cmd);
1679                 isert_unmap_data_buf(isert_conn, &wr->data);
1680         }
1681
1682         if (wr->rdma_wr) {
1683                 isert_dbg("Cmd %p free send_wr\n", isert_cmd);
1684                 kfree(wr->rdma_wr);
1685                 wr->rdma_wr = NULL;
1686         }
1687
1688         if (wr->ib_sge) {
1689                 isert_dbg("Cmd %p free ib_sge\n", isert_cmd);
1690                 kfree(wr->ib_sge);
1691                 wr->ib_sge = NULL;
1692         }
1693 }
1694
1695 static void
1696 isert_unreg_rdma(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn)
1697 {
1698         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1699
1700         isert_dbg("Cmd %p\n", isert_cmd);
1701
1702         if (wr->fr_desc) {
1703                 isert_dbg("Cmd %p free fr_desc %p\n", isert_cmd, wr->fr_desc);
1704                 if (wr->fr_desc->ind & ISERT_PROTECTED) {
1705                         isert_unmap_data_buf(isert_conn, &wr->prot);
1706                         wr->fr_desc->ind &= ~ISERT_PROTECTED;
1707                 }
1708                 spin_lock_bh(&isert_conn->pool_lock);
1709                 list_add_tail(&wr->fr_desc->list, &isert_conn->fr_pool);
1710                 spin_unlock_bh(&isert_conn->pool_lock);
1711                 wr->fr_desc = NULL;
1712         }
1713
1714         if (wr->data.sg) {
1715                 isert_dbg("Cmd %p unmap_sg op\n", isert_cmd);
1716                 isert_unmap_data_buf(isert_conn, &wr->data);
1717         }
1718
1719         wr->ib_sge = NULL;
1720         wr->rdma_wr = NULL;
1721 }
1722
1723 static void
1724 isert_put_cmd(struct isert_cmd *isert_cmd, bool comp_err)
1725 {
1726         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1727         struct isert_conn *isert_conn = isert_cmd->conn;
1728         struct iscsi_conn *conn = isert_conn->conn;
1729         struct isert_device *device = isert_conn->device;
1730         struct iscsi_text_rsp *hdr;
1731
1732         isert_dbg("Cmd %p\n", isert_cmd);
1733
1734         switch (cmd->iscsi_opcode) {
1735         case ISCSI_OP_SCSI_CMD:
1736                 spin_lock_bh(&conn->cmd_lock);
1737                 if (!list_empty(&cmd->i_conn_node))
1738                         list_del_init(&cmd->i_conn_node);
1739                 spin_unlock_bh(&conn->cmd_lock);
1740
1741                 if (cmd->data_direction == DMA_TO_DEVICE) {
1742                         iscsit_stop_dataout_timer(cmd);
1743                         /*
1744                          * Check for special case during comp_err where
1745                          * WRITE_PENDING has been handed off from core,
1746                          * but requires an extra target_put_sess_cmd()
1747                          * before transport_generic_free_cmd() below.
1748                          */
1749                         if (comp_err &&
1750                             cmd->se_cmd.t_state == TRANSPORT_WRITE_PENDING) {
1751                                 struct se_cmd *se_cmd = &cmd->se_cmd;
1752
1753                                 target_put_sess_cmd(se_cmd);
1754                         }
1755                 }
1756
1757                 device->unreg_rdma_mem(isert_cmd, isert_conn);
1758                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1759                 break;
1760         case ISCSI_OP_SCSI_TMFUNC:
1761                 spin_lock_bh(&conn->cmd_lock);
1762                 if (!list_empty(&cmd->i_conn_node))
1763                         list_del_init(&cmd->i_conn_node);
1764                 spin_unlock_bh(&conn->cmd_lock);
1765
1766                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1767                 break;
1768         case ISCSI_OP_REJECT:
1769         case ISCSI_OP_NOOP_OUT:
1770         case ISCSI_OP_TEXT:
1771                 hdr = (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
1772                 /* If the continue bit is on, keep the command alive */
1773                 if (hdr->flags & ISCSI_FLAG_TEXT_CONTINUE)
1774                         break;
1775
1776                 spin_lock_bh(&conn->cmd_lock);
1777                 if (!list_empty(&cmd->i_conn_node))
1778                         list_del_init(&cmd->i_conn_node);
1779                 spin_unlock_bh(&conn->cmd_lock);
1780
1781                 /*
1782                  * Handle special case for REJECT when iscsi_add_reject*() has
1783                  * overwritten the original iscsi_opcode assignment, and the
1784                  * associated cmd->se_cmd needs to be released.
1785                  */
1786                 if (cmd->se_cmd.se_tfo != NULL) {
1787                         isert_dbg("Calling transport_generic_free_cmd for 0x%02x\n",
1788                                  cmd->iscsi_opcode);
1789                         transport_generic_free_cmd(&cmd->se_cmd, 0);
1790                         break;
1791                 }
1792                 /*
1793                  * Fall-through
1794                  */
1795         default:
1796                 iscsit_release_cmd(cmd);
1797                 break;
1798         }
1799 }
1800
1801 static void
1802 isert_unmap_tx_desc(struct iser_tx_desc *tx_desc, struct ib_device *ib_dev)
1803 {
1804         if (tx_desc->dma_addr != 0) {
1805                 isert_dbg("unmap single for tx_desc->dma_addr\n");
1806                 ib_dma_unmap_single(ib_dev, tx_desc->dma_addr,
1807                                     ISER_HEADERS_LEN, DMA_TO_DEVICE);
1808                 tx_desc->dma_addr = 0;
1809         }
1810 }
1811
1812 static void
1813 isert_completion_put(struct iser_tx_desc *tx_desc, struct isert_cmd *isert_cmd,
1814                      struct ib_device *ib_dev, bool comp_err)
1815 {
1816         if (isert_cmd->pdu_buf_dma != 0) {
1817                 isert_dbg("unmap single for isert_cmd->pdu_buf_dma\n");
1818                 ib_dma_unmap_single(ib_dev, isert_cmd->pdu_buf_dma,
1819                                     isert_cmd->pdu_buf_len, DMA_TO_DEVICE);
1820                 isert_cmd->pdu_buf_dma = 0;
1821         }
1822
1823         isert_unmap_tx_desc(tx_desc, ib_dev);
1824         isert_put_cmd(isert_cmd, comp_err);
1825 }
1826
1827 static int
1828 isert_check_pi_status(struct se_cmd *se_cmd, struct ib_mr *sig_mr)
1829 {
1830         struct ib_mr_status mr_status;
1831         int ret;
1832
1833         ret = ib_check_mr_status(sig_mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
1834         if (ret) {
1835                 isert_err("ib_check_mr_status failed, ret %d\n", ret);
1836                 goto fail_mr_status;
1837         }
1838
1839         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1840                 u64 sec_offset_err;
1841                 u32 block_size = se_cmd->se_dev->dev_attrib.block_size + 8;
1842
1843                 switch (mr_status.sig_err.err_type) {
1844                 case IB_SIG_BAD_GUARD:
1845                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_GUARD_CHECK_FAILED;
1846                         break;
1847                 case IB_SIG_BAD_REFTAG:
1848                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_REF_TAG_CHECK_FAILED;
1849                         break;
1850                 case IB_SIG_BAD_APPTAG:
1851                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_APP_TAG_CHECK_FAILED;
1852                         break;
1853                 }
1854                 sec_offset_err = mr_status.sig_err.sig_err_offset;
1855                 do_div(sec_offset_err, block_size);
1856                 se_cmd->bad_sector = sec_offset_err + se_cmd->t_task_lba;
1857
1858                 isert_err("PI error found type %d at sector 0x%llx "
1859                           "expected 0x%x vs actual 0x%x\n",
1860                           mr_status.sig_err.err_type,
1861                           (unsigned long long)se_cmd->bad_sector,
1862                           mr_status.sig_err.expected,
1863                           mr_status.sig_err.actual);
1864                 ret = 1;
1865         }
1866
1867 fail_mr_status:
1868         return ret;
1869 }
1870
1871 static void
1872 isert_completion_rdma_write(struct iser_tx_desc *tx_desc,
1873                             struct isert_cmd *isert_cmd)
1874 {
1875         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1876         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1877         struct se_cmd *se_cmd = &cmd->se_cmd;
1878         struct isert_conn *isert_conn = isert_cmd->conn;
1879         struct isert_device *device = isert_conn->device;
1880         int ret = 0;
1881
1882         if (wr->fr_desc && wr->fr_desc->ind & ISERT_PROTECTED) {
1883                 ret = isert_check_pi_status(se_cmd,
1884                                             wr->fr_desc->pi_ctx->sig_mr);
1885                 wr->fr_desc->ind &= ~ISERT_PROTECTED;
1886         }
1887
1888         device->unreg_rdma_mem(isert_cmd, isert_conn);
1889         wr->rdma_wr_num = 0;
1890         if (ret)
1891                 transport_send_check_condition_and_sense(se_cmd,
1892                                                          se_cmd->pi_err, 0);
1893         else
1894                 isert_put_response(isert_conn->conn, cmd);
1895 }
1896
1897 static void
1898 isert_completion_rdma_read(struct iser_tx_desc *tx_desc,
1899                            struct isert_cmd *isert_cmd)
1900 {
1901         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1902         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1903         struct se_cmd *se_cmd = &cmd->se_cmd;
1904         struct isert_conn *isert_conn = isert_cmd->conn;
1905         struct isert_device *device = isert_conn->device;
1906         int ret = 0;
1907
1908         if (wr->fr_desc && wr->fr_desc->ind & ISERT_PROTECTED) {
1909                 ret = isert_check_pi_status(se_cmd,
1910                                             wr->fr_desc->pi_ctx->sig_mr);
1911                 wr->fr_desc->ind &= ~ISERT_PROTECTED;
1912         }
1913
1914         iscsit_stop_dataout_timer(cmd);
1915         device->unreg_rdma_mem(isert_cmd, isert_conn);
1916         cmd->write_data_done = wr->data.len;
1917         wr->rdma_wr_num = 0;
1918
1919         isert_dbg("Cmd: %p RDMA_READ comp calling execute_cmd\n", isert_cmd);
1920         spin_lock_bh(&cmd->istate_lock);
1921         cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1922         cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1923         spin_unlock_bh(&cmd->istate_lock);
1924
1925         if (ret) {
1926                 target_put_sess_cmd(se_cmd);
1927                 transport_send_check_condition_and_sense(se_cmd,
1928                                                          se_cmd->pi_err, 0);
1929         } else {
1930                 target_execute_cmd(se_cmd);
1931         }
1932 }
1933
1934 static void
1935 isert_do_control_comp(struct work_struct *work)
1936 {
1937         struct isert_cmd *isert_cmd = container_of(work,
1938                         struct isert_cmd, comp_work);
1939         struct isert_conn *isert_conn = isert_cmd->conn;
1940         struct ib_device *ib_dev = isert_conn->cm_id->device;
1941         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1942
1943         isert_dbg("Cmd %p i_state %d\n", isert_cmd, cmd->i_state);
1944
1945         switch (cmd->i_state) {
1946         case ISTATE_SEND_TASKMGTRSP:
1947                 iscsit_tmr_post_handler(cmd, cmd->conn);
1948         case ISTATE_SEND_REJECT:   /* FALLTHRU */
1949         case ISTATE_SEND_TEXTRSP:  /* FALLTHRU */
1950                 cmd->i_state = ISTATE_SENT_STATUS;
1951                 isert_completion_put(&isert_cmd->tx_desc, isert_cmd,
1952                                      ib_dev, false);
1953                 break;
1954         case ISTATE_SEND_LOGOUTRSP:
1955                 iscsit_logout_post_handler(cmd, cmd->conn);
1956                 break;
1957         default:
1958                 isert_err("Unknown i_state %d\n", cmd->i_state);
1959                 dump_stack();
1960                 break;
1961         }
1962 }
1963
1964 static void
1965 isert_response_completion(struct iser_tx_desc *tx_desc,
1966                           struct isert_cmd *isert_cmd,
1967                           struct isert_conn *isert_conn,
1968                           struct ib_device *ib_dev)
1969 {
1970         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1971
1972         if (cmd->i_state == ISTATE_SEND_TASKMGTRSP ||
1973             cmd->i_state == ISTATE_SEND_LOGOUTRSP ||
1974             cmd->i_state == ISTATE_SEND_REJECT ||
1975             cmd->i_state == ISTATE_SEND_TEXTRSP) {
1976                 isert_unmap_tx_desc(tx_desc, ib_dev);
1977
1978                 INIT_WORK(&isert_cmd->comp_work, isert_do_control_comp);
1979                 queue_work(isert_comp_wq, &isert_cmd->comp_work);
1980                 return;
1981         }
1982
1983         cmd->i_state = ISTATE_SENT_STATUS;
1984         isert_completion_put(tx_desc, isert_cmd, ib_dev, false);
1985 }
1986
1987 static void
1988 isert_snd_completion(struct iser_tx_desc *tx_desc,
1989                       struct isert_conn *isert_conn)
1990 {
1991         struct ib_device *ib_dev = isert_conn->cm_id->device;
1992         struct isert_cmd *isert_cmd = tx_desc->isert_cmd;
1993         struct isert_rdma_wr *wr;
1994
1995         if (!isert_cmd) {
1996                 isert_unmap_tx_desc(tx_desc, ib_dev);
1997                 return;
1998         }
1999         wr = &isert_cmd->rdma_wr;
2000
2001         isert_dbg("Cmd %p iser_ib_op %d\n", isert_cmd, wr->iser_ib_op);
2002
2003         switch (wr->iser_ib_op) {
2004         case ISER_IB_SEND:
2005                 isert_response_completion(tx_desc, isert_cmd,
2006                                           isert_conn, ib_dev);
2007                 break;
2008         case ISER_IB_RDMA_WRITE:
2009                 isert_completion_rdma_write(tx_desc, isert_cmd);
2010                 break;
2011         case ISER_IB_RDMA_READ:
2012                 isert_completion_rdma_read(tx_desc, isert_cmd);
2013                 break;
2014         default:
2015                 isert_err("Unknown wr->iser_ib_op: 0x%x\n", wr->iser_ib_op);
2016                 dump_stack();
2017                 break;
2018         }
2019 }
2020
2021 /**
2022  * is_isert_tx_desc() - Indicate if the completion wr_id
2023  *     is a TX descriptor or not.
2024  * @isert_conn: iser connection
2025  * @wr_id: completion WR identifier
2026  *
2027  * Since we cannot rely on wc opcode in FLUSH errors
2028  * we must work around it by checking if the wr_id address
2029  * falls in the iser connection rx_descs buffer. If so
2030  * it is an RX descriptor, otherwize it is a TX.
2031  */
2032 static inline bool
2033 is_isert_tx_desc(struct isert_conn *isert_conn, void *wr_id)
2034 {
2035         void *start = isert_conn->rx_descs;
2036         int len = ISERT_QP_MAX_RECV_DTOS * sizeof(*isert_conn->rx_descs);
2037
2038         if (wr_id >= start && wr_id < start + len)
2039                 return false;
2040
2041         return true;
2042 }
2043
2044 static void
2045 isert_cq_comp_err(struct isert_conn *isert_conn, struct ib_wc *wc)
2046 {
2047         if (wc->wr_id == ISER_BEACON_WRID) {
2048                 isert_info("conn %p completing wait_comp_err\n",
2049                            isert_conn);
2050                 complete(&isert_conn->wait_comp_err);
2051         } else if (is_isert_tx_desc(isert_conn, (void *)(uintptr_t)wc->wr_id)) {
2052                 struct ib_device *ib_dev = isert_conn->cm_id->device;
2053                 struct isert_cmd *isert_cmd;
2054                 struct iser_tx_desc *desc;
2055
2056                 desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
2057                 isert_cmd = desc->isert_cmd;
2058                 if (!isert_cmd)
2059                         isert_unmap_tx_desc(desc, ib_dev);
2060                 else
2061                         isert_completion_put(desc, isert_cmd, ib_dev, true);
2062         } else {
2063                 isert_conn->post_recv_buf_count--;
2064                 if (!isert_conn->post_recv_buf_count)
2065                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
2066         }
2067 }
2068
2069 static void
2070 isert_handle_wc(struct ib_wc *wc)
2071 {
2072         struct isert_conn *isert_conn;
2073         struct iser_tx_desc *tx_desc;
2074         struct iser_rx_desc *rx_desc;
2075
2076         isert_conn = wc->qp->qp_context;
2077         if (likely(wc->status == IB_WC_SUCCESS)) {
2078                 if (wc->opcode == IB_WC_RECV) {
2079                         rx_desc = (struct iser_rx_desc *)(uintptr_t)wc->wr_id;
2080                         isert_rcv_completion(rx_desc, isert_conn, wc->byte_len);
2081                 } else {
2082                         tx_desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
2083                         isert_snd_completion(tx_desc, isert_conn);
2084                 }
2085         } else {
2086                 if (wc->status != IB_WC_WR_FLUSH_ERR)
2087                         isert_err("%s (%d): wr id %llx vend_err %x\n",
2088                                   ib_wc_status_msg(wc->status), wc->status,
2089                                   wc->wr_id, wc->vendor_err);
2090                 else
2091                         isert_dbg("%s (%d): wr id %llx\n",
2092                                   ib_wc_status_msg(wc->status), wc->status,
2093                                   wc->wr_id);
2094
2095                 if (wc->wr_id != ISER_FASTREG_LI_WRID)
2096                         isert_cq_comp_err(isert_conn, wc);
2097         }
2098 }
2099
2100 static void
2101 isert_cq_work(struct work_struct *work)
2102 {
2103         enum { isert_poll_budget = 65536 };
2104         struct isert_comp *comp = container_of(work, struct isert_comp,
2105                                                work);
2106         struct ib_wc *const wcs = comp->wcs;
2107         int i, n, completed = 0;
2108
2109         while ((n = ib_poll_cq(comp->cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
2110                 for (i = 0; i < n; i++)
2111                         isert_handle_wc(&wcs[i]);
2112
2113                 completed += n;
2114                 if (completed >= isert_poll_budget)
2115                         break;
2116         }
2117
2118         ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP);
2119 }
2120
2121 static void
2122 isert_cq_callback(struct ib_cq *cq, void *context)
2123 {
2124         struct isert_comp *comp = context;
2125
2126         queue_work(isert_comp_wq, &comp->work);
2127 }
2128
2129 static int
2130 isert_post_response(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd)
2131 {
2132         struct ib_send_wr *wr_failed;
2133         int ret;
2134
2135         ret = isert_post_recv(isert_conn, isert_cmd->rx_desc);
2136         if (ret) {
2137                 isert_err("ib_post_recv failed with %d\n", ret);
2138                 return ret;
2139         }
2140
2141         ret = ib_post_send(isert_conn->qp, &isert_cmd->tx_desc.send_wr,
2142                            &wr_failed);
2143         if (ret) {
2144                 isert_err("ib_post_send failed with %d\n", ret);
2145                 return ret;
2146         }
2147         return ret;
2148 }
2149
2150 static int
2151 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2152 {
2153         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2154         struct isert_conn *isert_conn = conn->context;
2155         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2156         struct iscsi_scsi_rsp *hdr = (struct iscsi_scsi_rsp *)
2157                                 &isert_cmd->tx_desc.iscsi_header;
2158
2159         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2160         iscsit_build_rsp_pdu(cmd, conn, true, hdr);
2161         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2162         /*
2163          * Attach SENSE DATA payload to iSCSI Response PDU
2164          */
2165         if (cmd->se_cmd.sense_buffer &&
2166             ((cmd->se_cmd.se_cmd_flags & SCF_TRANSPORT_TASK_SENSE) ||
2167             (cmd->se_cmd.se_cmd_flags & SCF_EMULATED_TASK_SENSE))) {
2168                 struct isert_device *device = isert_conn->device;
2169                 struct ib_device *ib_dev = device->ib_device;
2170                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2171                 u32 padding, pdu_len;
2172
2173                 put_unaligned_be16(cmd->se_cmd.scsi_sense_length,
2174                                    cmd->sense_buffer);
2175                 cmd->se_cmd.scsi_sense_length += sizeof(__be16);
2176
2177                 padding = -(cmd->se_cmd.scsi_sense_length) & 3;
2178                 hton24(hdr->dlength, (u32)cmd->se_cmd.scsi_sense_length);
2179                 pdu_len = cmd->se_cmd.scsi_sense_length + padding;
2180
2181                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2182                                 (void *)cmd->sense_buffer, pdu_len,
2183                                 DMA_TO_DEVICE);
2184
2185                 isert_cmd->pdu_buf_len = pdu_len;
2186                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2187                 tx_dsg->length  = pdu_len;
2188                 tx_dsg->lkey    = device->pd->local_dma_lkey;
2189                 isert_cmd->tx_desc.num_sge = 2;
2190         }
2191
2192         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2193
2194         isert_dbg("Posting SCSI Response\n");
2195
2196         return isert_post_response(isert_conn, isert_cmd);
2197 }
2198
2199 static void
2200 isert_aborted_task(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2201 {
2202         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2203         struct isert_conn *isert_conn = conn->context;
2204         struct isert_device *device = isert_conn->device;
2205
2206         spin_lock_bh(&conn->cmd_lock);
2207         if (!list_empty(&cmd->i_conn_node))
2208                 list_del_init(&cmd->i_conn_node);
2209         spin_unlock_bh(&conn->cmd_lock);
2210
2211         if (cmd->data_direction == DMA_TO_DEVICE)
2212                 iscsit_stop_dataout_timer(cmd);
2213
2214         device->unreg_rdma_mem(isert_cmd, isert_conn);
2215 }
2216
2217 static enum target_prot_op
2218 isert_get_sup_prot_ops(struct iscsi_conn *conn)
2219 {
2220         struct isert_conn *isert_conn = conn->context;
2221         struct isert_device *device = isert_conn->device;
2222
2223         if (conn->tpg->tpg_attrib.t10_pi) {
2224                 if (device->pi_capable) {
2225                         isert_info("conn %p PI offload enabled\n", isert_conn);
2226                         isert_conn->pi_support = true;
2227                         return TARGET_PROT_ALL;
2228                 }
2229         }
2230
2231         isert_info("conn %p PI offload disabled\n", isert_conn);
2232         isert_conn->pi_support = false;
2233
2234         return TARGET_PROT_NORMAL;
2235 }
2236
2237 static int
2238 isert_put_nopin(struct iscsi_cmd *cmd, struct iscsi_conn *conn,
2239                 bool nopout_response)
2240 {
2241         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2242         struct isert_conn *isert_conn = conn->context;
2243         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2244
2245         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2246         iscsit_build_nopin_rsp(cmd, conn, (struct iscsi_nopin *)
2247                                &isert_cmd->tx_desc.iscsi_header,
2248                                nopout_response);
2249         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2250         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2251
2252         isert_dbg("conn %p Posting NOPIN Response\n", isert_conn);
2253
2254         return isert_post_response(isert_conn, isert_cmd);
2255 }
2256
2257 static int
2258 isert_put_logout_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2259 {
2260         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2261         struct isert_conn *isert_conn = conn->context;
2262         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2263
2264         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2265         iscsit_build_logout_rsp(cmd, conn, (struct iscsi_logout_rsp *)
2266                                 &isert_cmd->tx_desc.iscsi_header);
2267         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2268         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2269
2270         isert_dbg("conn %p Posting Logout Response\n", isert_conn);
2271
2272         return isert_post_response(isert_conn, isert_cmd);
2273 }
2274
2275 static int
2276 isert_put_tm_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2277 {
2278         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2279         struct isert_conn *isert_conn = conn->context;
2280         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2281
2282         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2283         iscsit_build_task_mgt_rsp(cmd, conn, (struct iscsi_tm_rsp *)
2284                                   &isert_cmd->tx_desc.iscsi_header);
2285         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2286         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2287
2288         isert_dbg("conn %p Posting Task Management Response\n", isert_conn);
2289
2290         return isert_post_response(isert_conn, isert_cmd);
2291 }
2292
2293 static int
2294 isert_put_reject(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2295 {
2296         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2297         struct isert_conn *isert_conn = conn->context;
2298         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2299         struct isert_device *device = isert_conn->device;
2300         struct ib_device *ib_dev = device->ib_device;
2301         struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2302         struct iscsi_reject *hdr =
2303                 (struct iscsi_reject *)&isert_cmd->tx_desc.iscsi_header;
2304
2305         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2306         iscsit_build_reject(cmd, conn, hdr);
2307         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2308
2309         hton24(hdr->dlength, ISCSI_HDR_LEN);
2310         isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2311                         (void *)cmd->buf_ptr, ISCSI_HDR_LEN,
2312                         DMA_TO_DEVICE);
2313         isert_cmd->pdu_buf_len = ISCSI_HDR_LEN;
2314         tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2315         tx_dsg->length  = ISCSI_HDR_LEN;
2316         tx_dsg->lkey    = device->pd->local_dma_lkey;
2317         isert_cmd->tx_desc.num_sge = 2;
2318
2319         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2320
2321         isert_dbg("conn %p Posting Reject\n", isert_conn);
2322
2323         return isert_post_response(isert_conn, isert_cmd);
2324 }
2325
2326 static int
2327 isert_put_text_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2328 {
2329         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2330         struct isert_conn *isert_conn = conn->context;
2331         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2332         struct iscsi_text_rsp *hdr =
2333                 (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
2334         u32 txt_rsp_len;
2335         int rc;
2336
2337         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2338         rc = iscsit_build_text_rsp(cmd, conn, hdr, ISCSI_INFINIBAND);
2339         if (rc < 0)
2340                 return rc;
2341
2342         txt_rsp_len = rc;
2343         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2344
2345         if (txt_rsp_len) {
2346                 struct isert_device *device = isert_conn->device;
2347                 struct ib_device *ib_dev = device->ib_device;
2348                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2349                 void *txt_rsp_buf = cmd->buf_ptr;
2350
2351                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2352                                 txt_rsp_buf, txt_rsp_len, DMA_TO_DEVICE);
2353
2354                 isert_cmd->pdu_buf_len = txt_rsp_len;
2355                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2356                 tx_dsg->length  = txt_rsp_len;
2357                 tx_dsg->lkey    = device->pd->local_dma_lkey;
2358                 isert_cmd->tx_desc.num_sge = 2;
2359         }
2360         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2361
2362         isert_dbg("conn %p Text Response\n", isert_conn);
2363
2364         return isert_post_response(isert_conn, isert_cmd);
2365 }
2366
2367 static int
2368 isert_build_rdma_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
2369                     struct ib_sge *ib_sge, struct ib_rdma_wr *rdma_wr,
2370                     u32 data_left, u32 offset)
2371 {
2372         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
2373         struct scatterlist *sg_start, *tmp_sg;
2374         struct isert_device *device = isert_conn->device;
2375         struct ib_device *ib_dev = device->ib_device;
2376         u32 sg_off, page_off;
2377         int i = 0, sg_nents;
2378
2379         sg_off = offset / PAGE_SIZE;
2380         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
2381         sg_nents = min(cmd->se_cmd.t_data_nents - sg_off, isert_conn->max_sge);
2382         page_off = offset % PAGE_SIZE;
2383
2384         rdma_wr->wr.sg_list = ib_sge;
2385         rdma_wr->wr.wr_id = (uintptr_t)&isert_cmd->tx_desc;
2386         /*
2387          * Perform mapping of TCM scatterlist memory ib_sge dma_addr.
2388          */
2389         for_each_sg(sg_start, tmp_sg, sg_nents, i) {
2390                 isert_dbg("RDMA from SGL dma_addr: 0x%llx dma_len: %u, "
2391                           "page_off: %u\n",
2392                           (unsigned long long)tmp_sg->dma_address,
2393                           tmp_sg->length, page_off);
2394
2395                 ib_sge->addr = ib_sg_dma_address(ib_dev, tmp_sg) + page_off;
2396                 ib_sge->length = min_t(u32, data_left,
2397                                 ib_sg_dma_len(ib_dev, tmp_sg) - page_off);
2398                 ib_sge->lkey = device->pd->local_dma_lkey;
2399
2400                 isert_dbg("RDMA ib_sge: addr: 0x%llx  length: %u lkey: %x\n",
2401                           ib_sge->addr, ib_sge->length, ib_sge->lkey);
2402                 page_off = 0;
2403                 data_left -= ib_sge->length;
2404                 if (!data_left)
2405                         break;
2406                 ib_sge++;
2407                 isert_dbg("Incrementing ib_sge pointer to %p\n", ib_sge);
2408         }
2409
2410         rdma_wr->wr.num_sge = ++i;
2411         isert_dbg("Set outgoing sg_list: %p num_sg: %u from TCM SGLs\n",
2412                   rdma_wr->wr.sg_list, rdma_wr->wr.num_sge);
2413
2414         return rdma_wr->wr.num_sge;
2415 }
2416
2417 static int
2418 isert_map_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2419                struct isert_rdma_wr *wr)
2420 {
2421         struct se_cmd *se_cmd = &cmd->se_cmd;
2422         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2423         struct isert_conn *isert_conn = conn->context;
2424         struct isert_data_buf *data = &wr->data;
2425         struct ib_rdma_wr *rdma_wr;
2426         struct ib_sge *ib_sge;
2427         u32 offset, data_len, data_left, rdma_write_max, va_offset = 0;
2428         int ret = 0, i, ib_sge_cnt;
2429
2430         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2431
2432         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2433         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2434                                  se_cmd->t_data_nents, se_cmd->data_length,
2435                                  offset, wr->iser_ib_op, &wr->data);
2436         if (ret)
2437                 return ret;
2438
2439         data_left = data->len;
2440         offset = data->offset;
2441
2442         ib_sge = kzalloc(sizeof(struct ib_sge) * data->nents, GFP_KERNEL);
2443         if (!ib_sge) {
2444                 isert_warn("Unable to allocate ib_sge\n");
2445                 ret = -ENOMEM;
2446                 goto unmap_cmd;
2447         }
2448         wr->ib_sge = ib_sge;
2449
2450         wr->rdma_wr_num = DIV_ROUND_UP(data->nents, isert_conn->max_sge);
2451         wr->rdma_wr = kzalloc(sizeof(struct ib_rdma_wr) * wr->rdma_wr_num,
2452                                 GFP_KERNEL);
2453         if (!wr->rdma_wr) {
2454                 isert_dbg("Unable to allocate wr->rdma_wr\n");
2455                 ret = -ENOMEM;
2456                 goto unmap_cmd;
2457         }
2458
2459         wr->isert_cmd = isert_cmd;
2460         rdma_write_max = isert_conn->max_sge * PAGE_SIZE;
2461
2462         for (i = 0; i < wr->rdma_wr_num; i++) {
2463                 rdma_wr = &isert_cmd->rdma_wr.rdma_wr[i];
2464                 data_len = min(data_left, rdma_write_max);
2465
2466                 rdma_wr->wr.send_flags = 0;
2467                 if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2468                         rdma_wr->wr.opcode = IB_WR_RDMA_WRITE;
2469                         rdma_wr->remote_addr = isert_cmd->read_va + offset;
2470                         rdma_wr->rkey = isert_cmd->read_stag;
2471                         if (i + 1 == wr->rdma_wr_num)
2472                                 rdma_wr->wr.next = &isert_cmd->tx_desc.send_wr;
2473                         else
2474                                 rdma_wr->wr.next = &wr->rdma_wr[i + 1].wr;
2475                 } else {
2476                         rdma_wr->wr.opcode = IB_WR_RDMA_READ;
2477                         rdma_wr->remote_addr = isert_cmd->write_va + va_offset;
2478                         rdma_wr->rkey = isert_cmd->write_stag;
2479                         if (i + 1 == wr->rdma_wr_num)
2480                                 rdma_wr->wr.send_flags = IB_SEND_SIGNALED;
2481                         else
2482                                 rdma_wr->wr.next = &wr->rdma_wr[i + 1].wr;
2483                 }
2484
2485                 ib_sge_cnt = isert_build_rdma_wr(isert_conn, isert_cmd, ib_sge,
2486                                         rdma_wr, data_len, offset);
2487                 ib_sge += ib_sge_cnt;
2488
2489                 offset += data_len;
2490                 va_offset += data_len;
2491                 data_left -= data_len;
2492         }
2493
2494         return 0;
2495 unmap_cmd:
2496         isert_unmap_data_buf(isert_conn, data);
2497
2498         return ret;
2499 }
2500
2501 static inline void
2502 isert_inv_rkey(struct ib_send_wr *inv_wr, struct ib_mr *mr)
2503 {
2504         u32 rkey;
2505
2506         memset(inv_wr, 0, sizeof(*inv_wr));
2507         inv_wr->wr_id = ISER_FASTREG_LI_WRID;
2508         inv_wr->opcode = IB_WR_LOCAL_INV;
2509         inv_wr->ex.invalidate_rkey = mr->rkey;
2510
2511         /* Bump the key */
2512         rkey = ib_inc_rkey(mr->rkey);
2513         ib_update_fast_reg_key(mr, rkey);
2514 }
2515
2516 static int
2517 isert_fast_reg_mr(struct isert_conn *isert_conn,
2518                   struct fast_reg_descriptor *fr_desc,
2519                   struct isert_data_buf *mem,
2520                   enum isert_indicator ind,
2521                   struct ib_sge *sge)
2522 {
2523         struct isert_device *device = isert_conn->device;
2524         struct ib_device *ib_dev = device->ib_device;
2525         struct ib_mr *mr;
2526         struct ib_reg_wr reg_wr;
2527         struct ib_send_wr inv_wr, *bad_wr, *wr = NULL;
2528         int ret, n;
2529
2530         if (mem->dma_nents == 1) {
2531                 sge->lkey = device->pd->local_dma_lkey;
2532                 sge->addr = ib_sg_dma_address(ib_dev, &mem->sg[0]);
2533                 sge->length = ib_sg_dma_len(ib_dev, &mem->sg[0]);
2534                 isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2535                          sge->addr, sge->length, sge->lkey);
2536                 return 0;
2537         }
2538
2539         if (ind == ISERT_DATA_KEY_VALID)
2540                 /* Registering data buffer */
2541                 mr = fr_desc->data_mr;
2542         else
2543                 /* Registering protection buffer */
2544                 mr = fr_desc->pi_ctx->prot_mr;
2545
2546         if (!(fr_desc->ind & ind)) {
2547                 isert_inv_rkey(&inv_wr, mr);
2548                 wr = &inv_wr;
2549         }
2550
2551         n = ib_map_mr_sg(mr, mem->sg, mem->nents, PAGE_SIZE);
2552         if (unlikely(n != mem->nents)) {
2553                 isert_err("failed to map mr sg (%d/%d)\n",
2554                          n, mem->nents);
2555                 return n < 0 ? n : -EINVAL;
2556         }
2557
2558         isert_dbg("Use fr_desc %p sg_nents %d offset %u\n",
2559                   fr_desc, mem->nents, mem->offset);
2560
2561         reg_wr.wr.next = NULL;
2562         reg_wr.wr.opcode = IB_WR_REG_MR;
2563         reg_wr.wr.wr_id = ISER_FASTREG_LI_WRID;
2564         reg_wr.wr.send_flags = 0;
2565         reg_wr.wr.num_sge = 0;
2566         reg_wr.mr = mr;
2567         reg_wr.key = mr->lkey;
2568         reg_wr.access = IB_ACCESS_LOCAL_WRITE;
2569
2570         if (!wr)
2571                 wr = &reg_wr.wr;
2572         else
2573                 wr->next = &reg_wr.wr;
2574
2575         ret = ib_post_send(isert_conn->qp, wr, &bad_wr);
2576         if (ret) {
2577                 isert_err("fast registration failed, ret:%d\n", ret);
2578                 return ret;
2579         }
2580         fr_desc->ind &= ~ind;
2581
2582         sge->lkey = mr->lkey;
2583         sge->addr = mr->iova;
2584         sge->length = mr->length;
2585
2586         isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2587                   sge->addr, sge->length, sge->lkey);
2588
2589         return ret;
2590 }
2591
2592 static inline void
2593 isert_set_dif_domain(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs,
2594                      struct ib_sig_domain *domain)
2595 {
2596         domain->sig_type = IB_SIG_TYPE_T10_DIF;
2597         domain->sig.dif.bg_type = IB_T10DIF_CRC;
2598         domain->sig.dif.pi_interval = se_cmd->se_dev->dev_attrib.block_size;
2599         domain->sig.dif.ref_tag = se_cmd->reftag_seed;
2600         /*
2601          * At the moment we hard code those, but if in the future
2602          * the target core would like to use it, we will take it
2603          * from se_cmd.
2604          */
2605         domain->sig.dif.apptag_check_mask = 0xffff;
2606         domain->sig.dif.app_escape = true;
2607         domain->sig.dif.ref_escape = true;
2608         if (se_cmd->prot_type == TARGET_DIF_TYPE1_PROT ||
2609             se_cmd->prot_type == TARGET_DIF_TYPE2_PROT)
2610                 domain->sig.dif.ref_remap = true;
2611 };
2612
2613 static int
2614 isert_set_sig_attrs(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs)
2615 {
2616         switch (se_cmd->prot_op) {
2617         case TARGET_PROT_DIN_INSERT:
2618         case TARGET_PROT_DOUT_STRIP:
2619                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
2620                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2621                 break;
2622         case TARGET_PROT_DOUT_INSERT:
2623         case TARGET_PROT_DIN_STRIP:
2624                 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
2625                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2626                 break;
2627         case TARGET_PROT_DIN_PASS:
2628         case TARGET_PROT_DOUT_PASS:
2629                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2630                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2631                 break;
2632         default:
2633                 isert_err("Unsupported PI operation %d\n", se_cmd->prot_op);
2634                 return -EINVAL;
2635         }
2636
2637         return 0;
2638 }
2639
2640 static inline u8
2641 isert_set_prot_checks(u8 prot_checks)
2642 {
2643         return (prot_checks & TARGET_DIF_CHECK_GUARD  ? 0xc0 : 0) |
2644                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x30 : 0) |
2645                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x0f : 0);
2646 }
2647
2648 static int
2649 isert_reg_sig_mr(struct isert_conn *isert_conn,
2650                  struct se_cmd *se_cmd,
2651                  struct isert_rdma_wr *rdma_wr,
2652                  struct fast_reg_descriptor *fr_desc)
2653 {
2654         struct ib_sig_handover_wr sig_wr;
2655         struct ib_send_wr inv_wr, *bad_wr, *wr = NULL;
2656         struct pi_context *pi_ctx = fr_desc->pi_ctx;
2657         struct ib_sig_attrs sig_attrs;
2658         int ret;
2659
2660         memset(&sig_attrs, 0, sizeof(sig_attrs));
2661         ret = isert_set_sig_attrs(se_cmd, &sig_attrs);
2662         if (ret)
2663                 goto err;
2664
2665         sig_attrs.check_mask = isert_set_prot_checks(se_cmd->prot_checks);
2666
2667         if (!(fr_desc->ind & ISERT_SIG_KEY_VALID)) {
2668                 isert_inv_rkey(&inv_wr, pi_ctx->sig_mr);
2669                 wr = &inv_wr;
2670         }
2671
2672         memset(&sig_wr, 0, sizeof(sig_wr));
2673         sig_wr.wr.opcode = IB_WR_REG_SIG_MR;
2674         sig_wr.wr.wr_id = ISER_FASTREG_LI_WRID;
2675         sig_wr.wr.sg_list = &rdma_wr->ib_sg[DATA];
2676         sig_wr.wr.num_sge = 1;
2677         sig_wr.access_flags = IB_ACCESS_LOCAL_WRITE;
2678         sig_wr.sig_attrs = &sig_attrs;
2679         sig_wr.sig_mr = pi_ctx->sig_mr;
2680         if (se_cmd->t_prot_sg)
2681                 sig_wr.prot = &rdma_wr->ib_sg[PROT];
2682
2683         if (!wr)
2684                 wr = &sig_wr.wr;
2685         else
2686                 wr->next = &sig_wr.wr;
2687
2688         ret = ib_post_send(isert_conn->qp, wr, &bad_wr);
2689         if (ret) {
2690                 isert_err("fast registration failed, ret:%d\n", ret);
2691                 goto err;
2692         }
2693         fr_desc->ind &= ~ISERT_SIG_KEY_VALID;
2694
2695         rdma_wr->ib_sg[SIG].lkey = pi_ctx->sig_mr->lkey;
2696         rdma_wr->ib_sg[SIG].addr = 0;
2697         rdma_wr->ib_sg[SIG].length = se_cmd->data_length;
2698         if (se_cmd->prot_op != TARGET_PROT_DIN_STRIP &&
2699             se_cmd->prot_op != TARGET_PROT_DOUT_INSERT)
2700                 /*
2701                  * We have protection guards on the wire
2702                  * so we need to set a larget transfer
2703                  */
2704                 rdma_wr->ib_sg[SIG].length += se_cmd->prot_length;
2705
2706         isert_dbg("sig_sge: addr: 0x%llx  length: %u lkey: %x\n",
2707                   rdma_wr->ib_sg[SIG].addr, rdma_wr->ib_sg[SIG].length,
2708                   rdma_wr->ib_sg[SIG].lkey);
2709 err:
2710         return ret;
2711 }
2712
2713 static int
2714 isert_handle_prot_cmd(struct isert_conn *isert_conn,
2715                       struct isert_cmd *isert_cmd,
2716                       struct isert_rdma_wr *wr)
2717 {
2718         struct isert_device *device = isert_conn->device;
2719         struct se_cmd *se_cmd = &isert_cmd->iscsi_cmd->se_cmd;
2720         int ret;
2721
2722         if (!wr->fr_desc->pi_ctx) {
2723                 ret = isert_create_pi_ctx(wr->fr_desc,
2724                                           device->ib_device,
2725                                           device->pd);
2726                 if (ret) {
2727                         isert_err("conn %p failed to allocate pi_ctx\n",
2728                                   isert_conn);
2729                         return ret;
2730                 }
2731         }
2732
2733         if (se_cmd->t_prot_sg) {
2734                 ret = isert_map_data_buf(isert_conn, isert_cmd,
2735                                          se_cmd->t_prot_sg,
2736                                          se_cmd->t_prot_nents,
2737                                          se_cmd->prot_length,
2738                                          0, wr->iser_ib_op, &wr->prot);
2739                 if (ret) {
2740                         isert_err("conn %p failed to map protection buffer\n",
2741                                   isert_conn);
2742                         return ret;
2743                 }
2744
2745                 memset(&wr->ib_sg[PROT], 0, sizeof(wr->ib_sg[PROT]));
2746                 ret = isert_fast_reg_mr(isert_conn, wr->fr_desc, &wr->prot,
2747                                         ISERT_PROT_KEY_VALID, &wr->ib_sg[PROT]);
2748                 if (ret) {
2749                         isert_err("conn %p failed to fast reg mr\n",
2750                                   isert_conn);
2751                         goto unmap_prot_cmd;
2752                 }
2753         }
2754
2755         ret = isert_reg_sig_mr(isert_conn, se_cmd, wr, wr->fr_desc);
2756         if (ret) {
2757                 isert_err("conn %p failed to fast reg mr\n",
2758                           isert_conn);
2759                 goto unmap_prot_cmd;
2760         }
2761         wr->fr_desc->ind |= ISERT_PROTECTED;
2762
2763         return 0;
2764
2765 unmap_prot_cmd:
2766         if (se_cmd->t_prot_sg)
2767                 isert_unmap_data_buf(isert_conn, &wr->prot);
2768
2769         return ret;
2770 }
2771
2772 static int
2773 isert_reg_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2774                struct isert_rdma_wr *wr)
2775 {
2776         struct se_cmd *se_cmd = &cmd->se_cmd;
2777         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2778         struct isert_conn *isert_conn = conn->context;
2779         struct fast_reg_descriptor *fr_desc = NULL;
2780         struct ib_rdma_wr *rdma_wr;
2781         struct ib_sge *ib_sg;
2782         u32 offset;
2783         int ret = 0;
2784         unsigned long flags;
2785
2786         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2787
2788         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2789         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2790                                  se_cmd->t_data_nents, se_cmd->data_length,
2791                                  offset, wr->iser_ib_op, &wr->data);
2792         if (ret)
2793                 return ret;
2794
2795         if (wr->data.dma_nents != 1 || isert_prot_cmd(isert_conn, se_cmd)) {
2796                 spin_lock_irqsave(&isert_conn->pool_lock, flags);
2797                 fr_desc = list_first_entry(&isert_conn->fr_pool,
2798                                            struct fast_reg_descriptor, list);
2799                 list_del(&fr_desc->list);
2800                 spin_unlock_irqrestore(&isert_conn->pool_lock, flags);
2801                 wr->fr_desc = fr_desc;
2802         }
2803
2804         ret = isert_fast_reg_mr(isert_conn, fr_desc, &wr->data,
2805                                 ISERT_DATA_KEY_VALID, &wr->ib_sg[DATA]);
2806         if (ret)
2807                 goto unmap_cmd;
2808
2809         if (isert_prot_cmd(isert_conn, se_cmd)) {
2810                 ret = isert_handle_prot_cmd(isert_conn, isert_cmd, wr);
2811                 if (ret)
2812                         goto unmap_cmd;
2813
2814                 ib_sg = &wr->ib_sg[SIG];
2815         } else {
2816                 ib_sg = &wr->ib_sg[DATA];
2817         }
2818
2819         memcpy(&wr->s_ib_sge, ib_sg, sizeof(*ib_sg));
2820         wr->ib_sge = &wr->s_ib_sge;
2821         wr->rdma_wr_num = 1;
2822         memset(&wr->s_rdma_wr, 0, sizeof(wr->s_rdma_wr));
2823         wr->rdma_wr = &wr->s_rdma_wr;
2824         wr->isert_cmd = isert_cmd;
2825
2826         rdma_wr = &isert_cmd->rdma_wr.s_rdma_wr;
2827         rdma_wr->wr.sg_list = &wr->s_ib_sge;
2828         rdma_wr->wr.num_sge = 1;
2829         rdma_wr->wr.wr_id = (uintptr_t)&isert_cmd->tx_desc;
2830         if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2831                 rdma_wr->wr.opcode = IB_WR_RDMA_WRITE;
2832                 rdma_wr->remote_addr = isert_cmd->read_va;
2833                 rdma_wr->rkey = isert_cmd->read_stag;
2834                 rdma_wr->wr.send_flags = !isert_prot_cmd(isert_conn, se_cmd) ?
2835                                       0 : IB_SEND_SIGNALED;
2836         } else {
2837                 rdma_wr->wr.opcode = IB_WR_RDMA_READ;
2838                 rdma_wr->remote_addr = isert_cmd->write_va;
2839                 rdma_wr->rkey = isert_cmd->write_stag;
2840                 rdma_wr->wr.send_flags = IB_SEND_SIGNALED;
2841         }
2842
2843         return 0;
2844
2845 unmap_cmd:
2846         if (fr_desc) {
2847                 spin_lock_irqsave(&isert_conn->pool_lock, flags);
2848                 list_add_tail(&fr_desc->list, &isert_conn->fr_pool);
2849                 spin_unlock_irqrestore(&isert_conn->pool_lock, flags);
2850         }
2851         isert_unmap_data_buf(isert_conn, &wr->data);
2852
2853         return ret;
2854 }
2855
2856 static int
2857 isert_put_datain(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2858 {
2859         struct se_cmd *se_cmd = &cmd->se_cmd;
2860         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2861         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2862         struct isert_conn *isert_conn = conn->context;
2863         struct isert_device *device = isert_conn->device;
2864         struct ib_send_wr *wr_failed;
2865         int rc;
2866
2867         isert_dbg("Cmd: %p RDMA_WRITE data_length: %u\n",
2868                  isert_cmd, se_cmd->data_length);
2869
2870         wr->iser_ib_op = ISER_IB_RDMA_WRITE;
2871         rc = device->reg_rdma_mem(conn, cmd, wr);
2872         if (rc) {
2873                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2874                 return rc;
2875         }
2876
2877         if (!isert_prot_cmd(isert_conn, se_cmd)) {
2878                 /*
2879                  * Build isert_conn->tx_desc for iSCSI response PDU and attach
2880                  */
2881                 isert_create_send_desc(isert_conn, isert_cmd,
2882                                        &isert_cmd->tx_desc);
2883                 iscsit_build_rsp_pdu(cmd, conn, true, (struct iscsi_scsi_rsp *)
2884                                      &isert_cmd->tx_desc.iscsi_header);
2885                 isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2886                 isert_init_send_wr(isert_conn, isert_cmd,
2887                                    &isert_cmd->tx_desc.send_wr);
2888                 isert_cmd->rdma_wr.s_rdma_wr.wr.next = &isert_cmd->tx_desc.send_wr;
2889                 wr->rdma_wr_num += 1;
2890
2891                 rc = isert_post_recv(isert_conn, isert_cmd->rx_desc);
2892                 if (rc) {
2893                         isert_err("ib_post_recv failed with %d\n", rc);
2894                         return rc;
2895                 }
2896         }
2897
2898         rc = ib_post_send(isert_conn->qp, &wr->rdma_wr->wr, &wr_failed);
2899         if (rc)
2900                 isert_warn("ib_post_send() failed for IB_WR_RDMA_WRITE\n");
2901
2902         if (!isert_prot_cmd(isert_conn, se_cmd))
2903                 isert_dbg("Cmd: %p posted RDMA_WRITE + Response for iSER Data "
2904                          "READ\n", isert_cmd);
2905         else
2906                 isert_dbg("Cmd: %p posted RDMA_WRITE for iSER Data READ\n",
2907                          isert_cmd);
2908
2909         return 1;
2910 }
2911
2912 static int
2913 isert_get_dataout(struct iscsi_conn *conn, struct iscsi_cmd *cmd, bool recovery)
2914 {
2915         struct se_cmd *se_cmd = &cmd->se_cmd;
2916         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2917         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2918         struct isert_conn *isert_conn = conn->context;
2919         struct isert_device *device = isert_conn->device;
2920         struct ib_send_wr *wr_failed;
2921         int rc;
2922
2923         isert_dbg("Cmd: %p RDMA_READ data_length: %u write_data_done: %u\n",
2924                  isert_cmd, se_cmd->data_length, cmd->write_data_done);
2925         wr->iser_ib_op = ISER_IB_RDMA_READ;
2926         rc = device->reg_rdma_mem(conn, cmd, wr);
2927         if (rc) {
2928                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2929                 return rc;
2930         }
2931
2932         rc = ib_post_send(isert_conn->qp, &wr->rdma_wr->wr, &wr_failed);
2933         if (rc)
2934                 isert_warn("ib_post_send() failed for IB_WR_RDMA_READ\n");
2935
2936         isert_dbg("Cmd: %p posted RDMA_READ memory for ISER Data WRITE\n",
2937                  isert_cmd);
2938
2939         return 0;
2940 }
2941
2942 static int
2943 isert_immediate_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2944 {
2945         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2946         int ret = 0;
2947
2948         switch (state) {
2949         case ISTATE_REMOVE:
2950                 spin_lock_bh(&conn->cmd_lock);
2951                 list_del_init(&cmd->i_conn_node);
2952                 spin_unlock_bh(&conn->cmd_lock);
2953                 isert_put_cmd(isert_cmd, true);
2954                 break;
2955         case ISTATE_SEND_NOPIN_WANT_RESPONSE:
2956                 ret = isert_put_nopin(cmd, conn, false);
2957                 break;
2958         default:
2959                 isert_err("Unknown immediate state: 0x%02x\n", state);
2960                 ret = -EINVAL;
2961                 break;
2962         }
2963
2964         return ret;
2965 }
2966
2967 static int
2968 isert_response_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2969 {
2970         struct isert_conn *isert_conn = conn->context;
2971         int ret;
2972
2973         switch (state) {
2974         case ISTATE_SEND_LOGOUTRSP:
2975                 ret = isert_put_logout_rsp(cmd, conn);
2976                 if (!ret)
2977                         isert_conn->logout_posted = true;
2978                 break;
2979         case ISTATE_SEND_NOPIN:
2980                 ret = isert_put_nopin(cmd, conn, true);
2981                 break;
2982         case ISTATE_SEND_TASKMGTRSP:
2983                 ret = isert_put_tm_rsp(cmd, conn);
2984                 break;
2985         case ISTATE_SEND_REJECT:
2986                 ret = isert_put_reject(cmd, conn);
2987                 break;
2988         case ISTATE_SEND_TEXTRSP:
2989                 ret = isert_put_text_rsp(cmd, conn);
2990                 break;
2991         case ISTATE_SEND_STATUS:
2992                 /*
2993                  * Special case for sending non GOOD SCSI status from TX thread
2994                  * context during pre se_cmd excecution failure.
2995                  */
2996                 ret = isert_put_response(conn, cmd);
2997                 break;
2998         default:
2999                 isert_err("Unknown response state: 0x%02x\n", state);
3000                 ret = -EINVAL;
3001                 break;
3002         }
3003
3004         return ret;
3005 }
3006
3007 struct rdma_cm_id *
3008 isert_setup_id(struct isert_np *isert_np)
3009 {
3010         struct iscsi_np *np = isert_np->np;
3011         struct rdma_cm_id *id;
3012         struct sockaddr *sa;
3013         int ret;
3014
3015         sa = (struct sockaddr *)&np->np_sockaddr;
3016         isert_dbg("ksockaddr: %p, sa: %p\n", &np->np_sockaddr, sa);
3017
3018         id = rdma_create_id(&init_net, isert_cma_handler, isert_np,
3019                             RDMA_PS_TCP, IB_QPT_RC);
3020         if (IS_ERR(id)) {
3021                 isert_err("rdma_create_id() failed: %ld\n", PTR_ERR(id));
3022                 ret = PTR_ERR(id);
3023                 goto out;
3024         }
3025         isert_dbg("id %p context %p\n", id, id->context);
3026
3027         ret = rdma_bind_addr(id, sa);
3028         if (ret) {
3029                 isert_err("rdma_bind_addr() failed: %d\n", ret);
3030                 goto out_id;
3031         }
3032
3033         ret = rdma_listen(id, 0);
3034         if (ret) {
3035                 isert_err("rdma_listen() failed: %d\n", ret);
3036                 goto out_id;
3037         }
3038
3039         return id;
3040 out_id:
3041         rdma_destroy_id(id);
3042 out:
3043         return ERR_PTR(ret);
3044 }
3045
3046 static int
3047 isert_setup_np(struct iscsi_np *np,
3048                struct sockaddr_storage *ksockaddr)
3049 {
3050         struct isert_np *isert_np;
3051         struct rdma_cm_id *isert_lid;
3052         int ret;
3053
3054         isert_np = kzalloc(sizeof(struct isert_np), GFP_KERNEL);
3055         if (!isert_np) {
3056                 isert_err("Unable to allocate struct isert_np\n");
3057                 return -ENOMEM;
3058         }
3059         sema_init(&isert_np->sem, 0);
3060         mutex_init(&isert_np->mutex);
3061         INIT_LIST_HEAD(&isert_np->accepted);
3062         INIT_LIST_HEAD(&isert_np->pending);
3063         isert_np->np = np;
3064
3065         /*
3066          * Setup the np->np_sockaddr from the passed sockaddr setup
3067          * in iscsi_target_configfs.c code..
3068          */
3069         memcpy(&np->np_sockaddr, ksockaddr,
3070                sizeof(struct sockaddr_storage));
3071
3072         isert_lid = isert_setup_id(isert_np);
3073         if (IS_ERR(isert_lid)) {
3074                 ret = PTR_ERR(isert_lid);
3075                 goto out;
3076         }
3077
3078         isert_np->cm_id = isert_lid;
3079         np->np_context = isert_np;
3080
3081         return 0;
3082
3083 out:
3084         kfree(isert_np);
3085
3086         return ret;
3087 }
3088
3089 static int
3090 isert_rdma_accept(struct isert_conn *isert_conn)
3091 {
3092         struct rdma_cm_id *cm_id = isert_conn->cm_id;
3093         struct rdma_conn_param cp;
3094         int ret;
3095
3096         memset(&cp, 0, sizeof(struct rdma_conn_param));
3097         cp.initiator_depth = isert_conn->initiator_depth;
3098         cp.retry_count = 7;
3099         cp.rnr_retry_count = 7;
3100
3101         ret = rdma_accept(cm_id, &cp);
3102         if (ret) {
3103                 isert_err("rdma_accept() failed with: %d\n", ret);
3104                 return ret;
3105         }
3106
3107         return 0;
3108 }
3109
3110 static int
3111 isert_get_login_rx(struct iscsi_conn *conn, struct iscsi_login *login)
3112 {
3113         struct isert_conn *isert_conn = conn->context;
3114         int ret;
3115
3116         isert_info("before login_req comp conn: %p\n", isert_conn);
3117         ret = wait_for_completion_interruptible(&isert_conn->login_req_comp);
3118         if (ret) {
3119                 isert_err("isert_conn %p interrupted before got login req\n",
3120                           isert_conn);
3121                 return ret;
3122         }
3123         reinit_completion(&isert_conn->login_req_comp);
3124
3125         /*
3126          * For login requests after the first PDU, isert_rx_login_req() will
3127          * kick schedule_delayed_work(&conn->login_work) as the packet is
3128          * received, which turns this callback from iscsi_target_do_login_rx()
3129          * into a NOP.
3130          */
3131         if (!login->first_request)
3132                 return 0;
3133
3134         isert_rx_login_req(isert_conn);
3135
3136         isert_info("before login_comp conn: %p\n", conn);
3137         ret = wait_for_completion_interruptible(&isert_conn->login_comp);
3138         if (ret)
3139                 return ret;
3140
3141         isert_info("processing login->req: %p\n", login->req);
3142
3143         return 0;
3144 }
3145
3146 static void
3147 isert_set_conn_info(struct iscsi_np *np, struct iscsi_conn *conn,
3148                     struct isert_conn *isert_conn)
3149 {
3150         struct rdma_cm_id *cm_id = isert_conn->cm_id;
3151         struct rdma_route *cm_route = &cm_id->route;
3152
3153         conn->login_family = np->np_sockaddr.ss_family;
3154
3155         conn->login_sockaddr = cm_route->addr.dst_addr;
3156         conn->local_sockaddr = cm_route->addr.src_addr;
3157 }
3158
3159 static int
3160 isert_accept_np(struct iscsi_np *np, struct iscsi_conn *conn)
3161 {
3162         struct isert_np *isert_np = np->np_context;
3163         struct isert_conn *isert_conn;
3164         int ret;
3165
3166 accept_wait:
3167         ret = down_interruptible(&isert_np->sem);
3168         if (ret)
3169                 return -ENODEV;
3170
3171         spin_lock_bh(&np->np_thread_lock);
3172         if (np->np_thread_state >= ISCSI_NP_THREAD_RESET) {
3173                 spin_unlock_bh(&np->np_thread_lock);
3174                 isert_dbg("np_thread_state %d\n",
3175                          np->np_thread_state);
3176                 /**
3177                  * No point in stalling here when np_thread
3178                  * is in state RESET/SHUTDOWN/EXIT - bail
3179                  **/
3180                 return -ENODEV;
3181         }
3182         spin_unlock_bh(&np->np_thread_lock);
3183
3184         mutex_lock(&isert_np->mutex);
3185         if (list_empty(&isert_np->pending)) {
3186                 mutex_unlock(&isert_np->mutex);
3187                 goto accept_wait;
3188         }
3189         isert_conn = list_first_entry(&isert_np->pending,
3190                         struct isert_conn, node);
3191         list_del_init(&isert_conn->node);
3192         mutex_unlock(&isert_np->mutex);
3193
3194         conn->context = isert_conn;
3195         isert_conn->conn = conn;
3196
3197         isert_set_conn_info(np, conn, isert_conn);
3198
3199         isert_dbg("Processing isert_conn: %p\n", isert_conn);
3200
3201         return 0;
3202 }
3203
3204 static void
3205 isert_free_np(struct iscsi_np *np)
3206 {
3207         struct isert_np *isert_np = np->np_context;
3208         struct isert_conn *isert_conn, *n;
3209
3210         if (isert_np->cm_id)
3211                 rdma_destroy_id(isert_np->cm_id);
3212
3213         /*
3214          * FIXME: At this point we don't have a good way to insure
3215          * that at this point we don't have hanging connections that
3216          * completed RDMA establishment but didn't start iscsi login
3217          * process. So work-around this by cleaning up what ever piled
3218          * up in accepted and pending lists.
3219          */
3220         mutex_lock(&isert_np->mutex);
3221         if (!list_empty(&isert_np->pending)) {
3222                 isert_info("Still have isert pending connections\n");
3223                 list_for_each_entry_safe(isert_conn, n,
3224                                          &isert_np->pending,
3225                                          node) {
3226                         isert_info("cleaning isert_conn %p state (%d)\n",
3227                                    isert_conn, isert_conn->state);
3228                         isert_connect_release(isert_conn);
3229                 }
3230         }
3231
3232         if (!list_empty(&isert_np->accepted)) {
3233                 isert_info("Still have isert accepted connections\n");
3234                 list_for_each_entry_safe(isert_conn, n,
3235                                          &isert_np->accepted,
3236                                          node) {
3237                         isert_info("cleaning isert_conn %p state (%d)\n",
3238                                    isert_conn, isert_conn->state);
3239                         isert_connect_release(isert_conn);
3240                 }
3241         }
3242         mutex_unlock(&isert_np->mutex);
3243
3244         np->np_context = NULL;
3245         kfree(isert_np);
3246 }
3247
3248 static void isert_release_work(struct work_struct *work)
3249 {
3250         struct isert_conn *isert_conn = container_of(work,
3251                                                      struct isert_conn,
3252                                                      release_work);
3253
3254         isert_info("Starting release conn %p\n", isert_conn);
3255
3256         wait_for_completion(&isert_conn->wait);
3257
3258         mutex_lock(&isert_conn->mutex);
3259         isert_conn->state = ISER_CONN_DOWN;
3260         mutex_unlock(&isert_conn->mutex);
3261
3262         isert_info("Destroying conn %p\n", isert_conn);
3263         isert_put_conn(isert_conn);
3264 }
3265
3266 static void
3267 isert_wait4logout(struct isert_conn *isert_conn)
3268 {
3269         struct iscsi_conn *conn = isert_conn->conn;
3270
3271         isert_info("conn %p\n", isert_conn);
3272
3273         if (isert_conn->logout_posted) {
3274                 isert_info("conn %p wait for conn_logout_comp\n", isert_conn);
3275                 wait_for_completion_timeout(&conn->conn_logout_comp,
3276                                             SECONDS_FOR_LOGOUT_COMP * HZ);
3277         }
3278 }
3279
3280 static void
3281 isert_wait4cmds(struct iscsi_conn *conn)
3282 {
3283         isert_info("iscsi_conn %p\n", conn);
3284
3285         if (conn->sess) {
3286                 target_sess_cmd_list_set_waiting(conn->sess->se_sess);
3287                 target_wait_for_sess_cmds(conn->sess->se_sess);
3288         }
3289 }
3290
3291 static void
3292 isert_wait4flush(struct isert_conn *isert_conn)
3293 {
3294         struct ib_recv_wr *bad_wr;
3295
3296         isert_info("conn %p\n", isert_conn);
3297
3298         init_completion(&isert_conn->wait_comp_err);
3299         isert_conn->beacon.wr_id = ISER_BEACON_WRID;
3300         /* post an indication that all flush errors were consumed */
3301         if (ib_post_recv(isert_conn->qp, &isert_conn->beacon, &bad_wr)) {
3302                 isert_err("conn %p failed to post beacon", isert_conn);
3303                 return;
3304         }
3305
3306         wait_for_completion(&isert_conn->wait_comp_err);
3307 }
3308
3309 /**
3310  * isert_put_unsol_pending_cmds() - Drop commands waiting for
3311  *     unsolicitate dataout
3312  * @conn:    iscsi connection
3313  *
3314  * We might still have commands that are waiting for unsolicited
3315  * dataouts messages. We must put the extra reference on those
3316  * before blocking on the target_wait_for_session_cmds
3317  */
3318 static void
3319 isert_put_unsol_pending_cmds(struct iscsi_conn *conn)
3320 {
3321         struct iscsi_cmd *cmd, *tmp;
3322         static LIST_HEAD(drop_cmd_list);
3323
3324         spin_lock_bh(&conn->cmd_lock);
3325         list_for_each_entry_safe(cmd, tmp, &conn->conn_cmd_list, i_conn_node) {
3326                 if ((cmd->cmd_flags & ICF_NON_IMMEDIATE_UNSOLICITED_DATA) &&
3327                     (cmd->write_data_done < conn->sess->sess_ops->FirstBurstLength) &&
3328                     (cmd->write_data_done < cmd->se_cmd.data_length))
3329                         list_move_tail(&cmd->i_conn_node, &drop_cmd_list);
3330         }
3331         spin_unlock_bh(&conn->cmd_lock);
3332
3333         list_for_each_entry_safe(cmd, tmp, &drop_cmd_list, i_conn_node) {
3334                 list_del_init(&cmd->i_conn_node);
3335                 if (cmd->i_state != ISTATE_REMOVE) {
3336                         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
3337
3338                         isert_info("conn %p dropping cmd %p\n", conn, cmd);
3339                         isert_put_cmd(isert_cmd, true);
3340                 }
3341         }
3342 }
3343
3344 static void isert_wait_conn(struct iscsi_conn *conn)
3345 {
3346         struct isert_conn *isert_conn = conn->context;
3347
3348         isert_info("Starting conn %p\n", isert_conn);
3349
3350         mutex_lock(&isert_conn->mutex);
3351         /*
3352          * Only wait for wait_comp_err if the isert_conn made it
3353          * into full feature phase..
3354          */
3355         if (isert_conn->state == ISER_CONN_INIT) {
3356                 mutex_unlock(&isert_conn->mutex);
3357                 return;
3358         }
3359         isert_conn_terminate(isert_conn);
3360         mutex_unlock(&isert_conn->mutex);
3361
3362         isert_wait4flush(isert_conn);
3363         isert_put_unsol_pending_cmds(conn);
3364         isert_wait4cmds(conn);
3365         isert_wait4logout(isert_conn);
3366
3367         queue_work(isert_release_wq, &isert_conn->release_work);
3368 }
3369
3370 static void isert_free_conn(struct iscsi_conn *conn)
3371 {
3372         struct isert_conn *isert_conn = conn->context;
3373
3374         isert_wait4flush(isert_conn);
3375         isert_put_conn(isert_conn);
3376 }
3377
3378 static struct iscsit_transport iser_target_transport = {
3379         .name                   = "IB/iSER",
3380         .transport_type         = ISCSI_INFINIBAND,
3381         .priv_size              = sizeof(struct isert_cmd),
3382         .owner                  = THIS_MODULE,
3383         .iscsit_setup_np        = isert_setup_np,
3384         .iscsit_accept_np       = isert_accept_np,
3385         .iscsit_free_np         = isert_free_np,
3386         .iscsit_wait_conn       = isert_wait_conn,
3387         .iscsit_free_conn       = isert_free_conn,
3388         .iscsit_get_login_rx    = isert_get_login_rx,
3389         .iscsit_put_login_tx    = isert_put_login_tx,
3390         .iscsit_immediate_queue = isert_immediate_queue,
3391         .iscsit_response_queue  = isert_response_queue,
3392         .iscsit_get_dataout     = isert_get_dataout,
3393         .iscsit_queue_data_in   = isert_put_datain,
3394         .iscsit_queue_status    = isert_put_response,
3395         .iscsit_aborted_task    = isert_aborted_task,
3396         .iscsit_get_sup_prot_ops = isert_get_sup_prot_ops,
3397 };
3398
3399 static int __init isert_init(void)
3400 {
3401         int ret;
3402
3403         isert_comp_wq = alloc_workqueue("isert_comp_wq",
3404                                         WQ_UNBOUND | WQ_HIGHPRI, 0);
3405         if (!isert_comp_wq) {
3406                 isert_err("Unable to allocate isert_comp_wq\n");
3407                 ret = -ENOMEM;
3408                 return -ENOMEM;
3409         }
3410
3411         isert_release_wq = alloc_workqueue("isert_release_wq", WQ_UNBOUND,
3412                                         WQ_UNBOUND_MAX_ACTIVE);
3413         if (!isert_release_wq) {
3414                 isert_err("Unable to allocate isert_release_wq\n");
3415                 ret = -ENOMEM;
3416                 goto destroy_comp_wq;
3417         }
3418
3419         iscsit_register_transport(&iser_target_transport);
3420         isert_info("iSER_TARGET[0] - Loaded iser_target_transport\n");
3421
3422         return 0;
3423
3424 destroy_comp_wq:
3425         destroy_workqueue(isert_comp_wq);
3426
3427         return ret;
3428 }
3429
3430 static void __exit isert_exit(void)
3431 {
3432         flush_scheduled_work();
3433         destroy_workqueue(isert_release_wq);
3434         destroy_workqueue(isert_comp_wq);
3435         iscsit_unregister_transport(&iser_target_transport);
3436         isert_info("iSER_TARGET[0] - Released iser_target_transport\n");
3437 }
3438
3439 MODULE_DESCRIPTION("iSER-Target for mainline target infrastructure");
3440 MODULE_VERSION("1.0");
3441 MODULE_AUTHOR("nab@Linux-iSCSI.org");
3442 MODULE_LICENSE("GPL");
3443
3444 module_init(isert_init);
3445 module_exit(isert_exit);