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 static void isert_wait4flush(struct isert_conn *isert_conn);
  70
  71 static inline bool
  72 isert_prot_cmd(struct isert_conn *conn, struct se_cmd *cmd)
  73 {
  74         return (conn->pi_support &&
  75                 cmd->prot_op != TARGET_PROT_NORMAL);
  76 }
  77
  78
  79 static void
  80 isert_qp_event_callback(struct ib_event *e, void *context)
  81 {
  82         struct isert_conn *isert_conn = context;
  83
  84         isert_err("%s (%d): conn %p\n",
  85                   ib_event_msg(e->event), e->event, isert_conn);
  86
  87         switch (e->event) {
  88         case IB_EVENT_COMM_EST:
  89                 rdma_notify(isert_conn->cm_id, IB_EVENT_COMM_EST);
  90                 break;
  91         case IB_EVENT_QP_LAST_WQE_REACHED:
  92                 isert_warn("Reached TX IB_EVENT_QP_LAST_WQE_REACHED\n");
  93                 break;
  94         default:
  95                 break;
  96         }
  97 }
  98
  99 static int
 100 isert_query_device(struct ib_device *ib_dev, struct ib_device_attr *devattr)
 101 {
 102         int ret;
 103
 104         ret = ib_query_device(ib_dev, devattr);
 105         if (ret) {
 106                 isert_err("ib_query_device() failed: %d\n", ret);
 107                 return ret;
 108         }
 109         isert_dbg("devattr->max_sge: %d\n", devattr->max_sge);
 110         isert_dbg("devattr->max_sge_rd: %d\n", devattr->max_sge_rd);
 111
 112         return 0;
 113 }
 114
 115 static struct isert_comp *
 116 isert_comp_get(struct isert_conn *isert_conn)
 117 {
 118         struct isert_device *device = isert_conn->device;
 119         struct isert_comp *comp;
 120         int i, min = 0;
 121
 122         mutex_lock(&device_list_mutex);
 123         for (i = 0; i < device->comps_used; i++)
 124                 if (device->comps[i].active_qps <
 125                     device->comps[min].active_qps)
 126                         min = i;
 127         comp = &device->comps[min];
 128         comp->active_qps++;
 129         mutex_unlock(&device_list_mutex);
 130
 131         isert_info("conn %p, using comp %p min_index: %d\n",
 132                    isert_conn, comp, min);
 133
 134         return comp;
 135 }
 136
 137 static void
 138 isert_comp_put(struct isert_comp *comp)
 139 {
 140         mutex_lock(&device_list_mutex);
 141         comp->active_qps--;
 142         mutex_unlock(&device_list_mutex);
 143 }
 144
 145 static struct ib_qp *
 146 isert_create_qp(struct isert_conn *isert_conn,
 147                 struct isert_comp *comp,
 148                 struct rdma_cm_id *cma_id)
 149 {
 150         struct isert_device *device = isert_conn->device;
 151         struct ib_qp_init_attr attr;
 152         int ret;
 153
 154         memset(&attr, 0, sizeof(struct ib_qp_init_attr));
 155         attr.event_handler = isert_qp_event_callback;
 156         attr.qp_context = isert_conn;
 157         attr.send_cq = comp->cq;
 158         attr.recv_cq = comp->cq;
 159         attr.cap.max_send_wr = ISERT_QP_MAX_REQ_DTOS;
 160         attr.cap.max_recv_wr = ISERT_QP_MAX_RECV_DTOS + 1;
 161         attr.cap.max_send_sge = device->dev_attr.max_sge;
 162         isert_conn->max_sge = min(device->dev_attr.max_sge,
 163                                   device->dev_attr.max_sge_rd);
 164         attr.cap.max_recv_sge = 1;
 165         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 166         attr.qp_type = IB_QPT_RC;
 167         if (device->pi_capable)
 168                 attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
 169
 170         ret = rdma_create_qp(cma_id, device->pd, &attr);
 171         if (ret) {
 172                 isert_err("rdma_create_qp failed for cma_id %d\n", ret);
 173                 return ERR_PTR(ret);
 174         }
 175
 176         return cma_id->qp;
 177 }
 178
 179 static int
 180 isert_conn_setup_qp(struct isert_conn *isert_conn, struct rdma_cm_id *cma_id)
 181 {
 182         struct isert_comp *comp;
 183         int ret;
 184
 185         comp = isert_comp_get(isert_conn);
 186         isert_conn->qp = isert_create_qp(isert_conn, comp, cma_id);
 187         if (IS_ERR(isert_conn->qp)) {
 188                 ret = PTR_ERR(isert_conn->qp);
 189                 goto err;
 190         }
 191
 192         return 0;
 193 err:
 194         isert_comp_put(comp);
 195         return ret;
 196 }
 197
 198 static void
 199 isert_cq_event_callback(struct ib_event *e, void *context)
 200 {
 201         isert_dbg("event: %d\n", e->event);
 202 }
 203
 204 static int
 205 isert_alloc_rx_descriptors(struct isert_conn *isert_conn)
 206 {
 207         struct isert_device *device = isert_conn->device;
 208         struct ib_device *ib_dev = device->ib_device;
 209         struct iser_rx_desc *rx_desc;
 210         struct ib_sge *rx_sg;
 211         u64 dma_addr;
 212         int i, j;
 213
 214         isert_conn->rx_descs = kzalloc(ISERT_QP_MAX_RECV_DTOS *
 215                                 sizeof(struct iser_rx_desc), GFP_KERNEL);
 216         if (!isert_conn->rx_descs)
 217                 goto fail;
 218
 219         rx_desc = isert_conn->rx_descs;
 220
 221         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 222                 dma_addr = ib_dma_map_single(ib_dev, (void *)rx_desc,
 223                                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 224                 if (ib_dma_mapping_error(ib_dev, dma_addr))
 225                         goto dma_map_fail;
 226
 227                 rx_desc->dma_addr = dma_addr;
 228
 229                 rx_sg = &rx_desc->rx_sg;
 230                 rx_sg->addr = rx_desc->dma_addr;
 231                 rx_sg->length = ISER_RX_PAYLOAD_SIZE;
 232                 rx_sg->lkey = device->pd->local_dma_lkey;
 233         }
 234
 235         return 0;
 236
 237 dma_map_fail:
 238         rx_desc = isert_conn->rx_descs;
 239         for (j = 0; j < i; j++, rx_desc++) {
 240                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 241                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 242         }
 243         kfree(isert_conn->rx_descs);
 244         isert_conn->rx_descs = NULL;
 245 fail:
 246         isert_err("conn %p failed to allocate rx descriptors\n", isert_conn);
 247
 248         return -ENOMEM;
 249 }
 250
 251 static void
 252 isert_free_rx_descriptors(struct isert_conn *isert_conn)
 253 {
 254         struct ib_device *ib_dev = isert_conn->device->ib_device;
 255         struct iser_rx_desc *rx_desc;
 256         int i;
 257
 258         if (!isert_conn->rx_descs)
 259                 return;
 260
 261         rx_desc = isert_conn->rx_descs;
 262         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 263                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 264                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 265         }
 266
 267         kfree(isert_conn->rx_descs);
 268         isert_conn->rx_descs = NULL;
 269 }
 270
 271 static void isert_cq_work(struct work_struct *);
 272 static void isert_cq_callback(struct ib_cq *, void *);
 273
 274 static void
 275 isert_free_comps(struct isert_device *device)
 276 {
 277         int i;
 278
 279         for (i = 0; i < device->comps_used; i++) {
 280                 struct isert_comp *comp = &device->comps[i];
 281
 282                 if (comp->cq) {
 283                         cancel_work_sync(&comp->work);
 284                         ib_destroy_cq(comp->cq);
 285                 }
 286         }
 287         kfree(device->comps);
 288 }
 289
 290 static int
 291 isert_alloc_comps(struct isert_device *device,
 292                   struct ib_device_attr *attr)
 293 {
 294         int i, max_cqe, ret = 0;
 295
 296         device->comps_used = min(ISERT_MAX_CQ, min_t(int, num_online_cpus(),
 297                                  device->ib_device->num_comp_vectors));
 298
 299         isert_info("Using %d CQs, %s supports %d vectors support "
 300                    "Fast registration %d pi_capable %d\n",
 301                    device->comps_used, device->ib_device->name,
 302                    device->ib_device->num_comp_vectors, device->use_fastreg,
 303                    device->pi_capable);
 304
 305         device->comps = kcalloc(device->comps_used, sizeof(struct isert_comp),
 306                                 GFP_KERNEL);
 307         if (!device->comps) {
 308                 isert_err("Unable to allocate completion contexts\n");
 309                 return -ENOMEM;
 310         }
 311
 312         max_cqe = min(ISER_MAX_CQ_LEN, attr->max_cqe);
 313
 314         for (i = 0; i < device->comps_used; i++) {
 315                 struct ib_cq_init_attr cq_attr = {};
 316                 struct isert_comp *comp = &device->comps[i];
 317
 318                 comp->device = device;
 319                 INIT_WORK(&comp->work, isert_cq_work);
 320                 cq_attr.cqe = max_cqe;
 321                 cq_attr.comp_vector = i;
 322                 comp->cq = ib_create_cq(device->ib_device,
 323                                         isert_cq_callback,
 324                                         isert_cq_event_callback,
 325                                         (void *)comp,
 326                                         &cq_attr);
 327                 if (IS_ERR(comp->cq)) {
 328                         isert_err("Unable to allocate cq\n");
 329                         ret = PTR_ERR(comp->cq);
 330                         comp->cq = NULL;
 331                         goto out_cq;
 332                 }
 333
 334                 ret = ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP);
 335                 if (ret)
 336                         goto out_cq;
 337         }
 338
 339         return 0;
 340 out_cq:
 341         isert_free_comps(device);
 342         return ret;
 343 }
 344
 345 static int
 346 isert_create_device_ib_res(struct isert_device *device)
 347 {
 348         struct ib_device_attr *dev_attr;
 349         int ret;
 350
 351         dev_attr = &device->dev_attr;
 352         ret = isert_query_device(device->ib_device, dev_attr);
 353         if (ret)
 354                 return ret;
 355
 356         /* asign function handlers */
 357         if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS &&
 358             dev_attr->device_cap_flags & IB_DEVICE_SIGNATURE_HANDOVER) {
 359                 device->use_fastreg = 1;
 360                 device->reg_rdma_mem = isert_reg_rdma;
 361                 device->unreg_rdma_mem = isert_unreg_rdma;
 362         } else {
 363                 device->use_fastreg = 0;
 364                 device->reg_rdma_mem = isert_map_rdma;
 365                 device->unreg_rdma_mem = isert_unmap_cmd;
 366         }
 367
 368         ret = isert_alloc_comps(device, dev_attr);
 369         if (ret)
 370                 return ret;
 371
 372         device->pd = ib_alloc_pd(device->ib_device);
 373         if (IS_ERR(device->pd)) {
 374                 ret = PTR_ERR(device->pd);
 375                 isert_err("failed to allocate pd, device %p, ret=%d\n",
 376                           device, ret);
 377                 goto out_cq;
 378         }
 379
 380         /* Check signature cap */
 381         device->pi_capable = dev_attr->device_cap_flags &
 382                              IB_DEVICE_SIGNATURE_HANDOVER ? true : false;
 383
 384         return 0;
 385
 386 out_cq:
 387         isert_free_comps(device);
 388         return ret;
 389 }
 390
 391 static void
 392 isert_free_device_ib_res(struct isert_device *device)
 393 {
 394         isert_info("device %p\n", device);
 395
 396         ib_dealloc_pd(device->pd);
 397         isert_free_comps(device);
 398 }
 399
 400 static void
 401 isert_device_put(struct isert_device *device)
 402 {
 403         mutex_lock(&device_list_mutex);
 404         device->refcount--;
 405         isert_info("device %p refcount %d\n", device, device->refcount);
 406         if (!device->refcount) {
 407                 isert_free_device_ib_res(device);
 408                 list_del(&device->dev_node);
 409                 kfree(device);
 410         }
 411         mutex_unlock(&device_list_mutex);
 412 }
 413
 414 static struct isert_device *
 415 isert_device_get(struct rdma_cm_id *cma_id)
 416 {
 417         struct isert_device *device;
 418         int ret;
 419
 420         mutex_lock(&device_list_mutex);
 421         list_for_each_entry(device, &device_list, dev_node) {
 422                 if (device->ib_device->node_guid == cma_id->device->node_guid) {
 423                         device->refcount++;
 424                         isert_info("Found iser device %p refcount %d\n",
 425                                    device, device->refcount);
 426                         mutex_unlock(&device_list_mutex);
 427                         return device;
 428                 }
 429         }
 430
 431         device = kzalloc(sizeof(struct isert_device), GFP_KERNEL);
 432         if (!device) {
 433                 mutex_unlock(&device_list_mutex);
 434                 return ERR_PTR(-ENOMEM);
 435         }
 436
 437         INIT_LIST_HEAD(&device->dev_node);
 438
 439         device->ib_device = cma_id->device;
 440         ret = isert_create_device_ib_res(device);
 441         if (ret) {
 442                 kfree(device);
 443                 mutex_unlock(&device_list_mutex);
 444                 return ERR_PTR(ret);
 445         }
 446
 447         device->refcount++;
 448         list_add_tail(&device->dev_node, &device_list);
 449         isert_info("Created a new iser device %p refcount %d\n",
 450                    device, device->refcount);
 451         mutex_unlock(&device_list_mutex);
 452
 453         return device;
 454 }
 455
 456 static void
 457 isert_conn_free_fastreg_pool(struct isert_conn *isert_conn)
 458 {
 459         struct fast_reg_descriptor *fr_desc, *tmp;
 460         int i = 0;
 461
 462         if (list_empty(&isert_conn->fr_pool))
 463                 return;
 464
 465         isert_info("Freeing conn %p fastreg pool", isert_conn);
 466
 467         list_for_each_entry_safe(fr_desc, tmp,
 468                                  &isert_conn->fr_pool, list) {
 469                 list_del(&fr_desc->list);
 470                 ib_dereg_mr(fr_desc->data_mr);
 471                 if (fr_desc->pi_ctx) {
 472                         ib_dereg_mr(fr_desc->pi_ctx->prot_mr);
 473                         ib_dereg_mr(fr_desc->pi_ctx->sig_mr);
 474                         kfree(fr_desc->pi_ctx);
 475                 }
 476                 kfree(fr_desc);
 477                 ++i;
 478         }
 479
 480         if (i < isert_conn->fr_pool_size)
 481                 isert_warn("Pool still has %d regions registered\n",
 482                         isert_conn->fr_pool_size - i);
 483 }
 484
 485 static int
 486 isert_create_pi_ctx(struct fast_reg_descriptor *desc,
 487                     struct ib_device *device,
 488                     struct ib_pd *pd)
 489 {
 490         struct pi_context *pi_ctx;
 491         int ret;
 492
 493         pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
 494         if (!pi_ctx) {
 495                 isert_err("Failed to allocate pi context\n");
 496                 return -ENOMEM;
 497         }
 498
 499         pi_ctx->prot_mr = ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG,
 500                                       ISCSI_ISER_SG_TABLESIZE);
 501         if (IS_ERR(pi_ctx->prot_mr)) {
 502                 isert_err("Failed to allocate prot frmr err=%ld\n",
 503                           PTR_ERR(pi_ctx->prot_mr));
 504                 ret = PTR_ERR(pi_ctx->prot_mr);
 505                 goto err_pi_ctx;
 506         }
 507         desc->ind |= ISERT_PROT_KEY_VALID;
 508
 509         pi_ctx->sig_mr = ib_alloc_mr(pd, IB_MR_TYPE_SIGNATURE, 2);
 510         if (IS_ERR(pi_ctx->sig_mr)) {
 511                 isert_err("Failed to allocate signature enabled mr err=%ld\n",
 512                           PTR_ERR(pi_ctx->sig_mr));
 513                 ret = PTR_ERR(pi_ctx->sig_mr);
 514                 goto err_prot_mr;
 515         }
 516
 517         desc->pi_ctx = pi_ctx;
 518         desc->ind |= ISERT_SIG_KEY_VALID;
 519         desc->ind &= ~ISERT_PROTECTED;
 520
 521         return 0;
 522
 523 err_prot_mr:
 524         ib_dereg_mr(pi_ctx->prot_mr);
 525 err_pi_ctx:
 526         kfree(pi_ctx);
 527
 528         return ret;
 529 }
 530
 531 static int
 532 isert_create_fr_desc(struct ib_device *ib_device, struct ib_pd *pd,
 533                      struct fast_reg_descriptor *fr_desc)
 534 {
 535         fr_desc->data_mr = ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG,
 536                                        ISCSI_ISER_SG_TABLESIZE);
 537         if (IS_ERR(fr_desc->data_mr)) {
 538                 isert_err("Failed to allocate data frmr err=%ld\n",
 539                           PTR_ERR(fr_desc->data_mr));
 540                 return PTR_ERR(fr_desc->data_mr);
 541         }
 542         fr_desc->ind |= ISERT_DATA_KEY_VALID;
 543
 544         isert_dbg("Created fr_desc %p\n", fr_desc);
 545
 546         return 0;
 547 }
 548
 549 static int
 550 isert_conn_create_fastreg_pool(struct isert_conn *isert_conn)
 551 {
 552         struct fast_reg_descriptor *fr_desc;
 553         struct isert_device *device = isert_conn->device;
 554         struct se_session *se_sess = isert_conn->conn->sess->se_sess;
 555         struct se_node_acl *se_nacl = se_sess->se_node_acl;
 556         int i, ret, tag_num;
 557         /*
 558          * Setup the number of FRMRs based upon the number of tags
 559          * available to session in iscsi_target_locate_portal().
 560          */
 561         tag_num = max_t(u32, ISCSIT_MIN_TAGS, se_nacl->queue_depth);
 562         tag_num = (tag_num * 2) + ISCSIT_EXTRA_TAGS;
 563
 564         isert_conn->fr_pool_size = 0;
 565         for (i = 0; i < tag_num; i++) {
 566                 fr_desc = kzalloc(sizeof(*fr_desc), GFP_KERNEL);
 567                 if (!fr_desc) {
 568                         isert_err("Failed to allocate fast_reg descriptor\n");
 569                         ret = -ENOMEM;
 570                         goto err;
 571                 }
 572
 573                 ret = isert_create_fr_desc(device->ib_device,
 574                                            device->pd, fr_desc);
 575                 if (ret) {
 576                         isert_err("Failed to create fastreg descriptor err=%d\n",
 577                                ret);
 578                         kfree(fr_desc);
 579                         goto err;
 580                 }
 581
 582                 list_add_tail(&fr_desc->list, &isert_conn->fr_pool);
 583                 isert_conn->fr_pool_size++;
 584         }
 585
 586         isert_dbg("Creating conn %p fastreg pool size=%d",
 587                  isert_conn, isert_conn->fr_pool_size);
 588
 589         return 0;
 590
 591 err:
 592         isert_conn_free_fastreg_pool(isert_conn);
 593         return ret;
 594 }
 595
 596 static void
 597 isert_init_conn(struct isert_conn *isert_conn)
 598 {
 599         isert_conn->state = ISER_CONN_INIT;
 600         INIT_LIST_HEAD(&isert_conn->node);
 601         init_completion(&isert_conn->login_comp);
 602         init_completion(&isert_conn->login_req_comp);
 603         init_completion(&isert_conn->wait);
 604         kref_init(&isert_conn->kref);
 605         mutex_init(&isert_conn->mutex);
 606         spin_lock_init(&isert_conn->pool_lock);
 607         INIT_LIST_HEAD(&isert_conn->fr_pool);
 608         INIT_WORK(&isert_conn->release_work, isert_release_work);
 609 }
 610
 611 static void
 612 isert_free_login_buf(struct isert_conn *isert_conn)
 613 {
 614         struct ib_device *ib_dev = isert_conn->device->ib_device;
 615
 616         ib_dma_unmap_single(ib_dev, isert_conn->login_rsp_dma,
 617                             ISER_RX_LOGIN_SIZE, DMA_TO_DEVICE);
 618         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 619                             ISCSI_DEF_MAX_RECV_SEG_LEN,
 620                             DMA_FROM_DEVICE);
 621         kfree(isert_conn->login_buf);
 622 }
 623
 624 static int
 625 isert_alloc_login_buf(struct isert_conn *isert_conn,
 626                       struct ib_device *ib_dev)
 627 {
 628         int ret;
 629
 630         isert_conn->login_buf = kzalloc(ISCSI_DEF_MAX_RECV_SEG_LEN +
 631                                         ISER_RX_LOGIN_SIZE, GFP_KERNEL);
 632         if (!isert_conn->login_buf) {
 633                 isert_err("Unable to allocate isert_conn->login_buf\n");
 634                 return -ENOMEM;
 635         }
 636
 637         isert_conn->login_req_buf = isert_conn->login_buf;
 638         isert_conn->login_rsp_buf = isert_conn->login_buf +
 639                                     ISCSI_DEF_MAX_RECV_SEG_LEN;
 640
 641         isert_dbg("Set login_buf: %p login_req_buf: %p login_rsp_buf: %p\n",
 642                  isert_conn->login_buf, isert_conn->login_req_buf,
 643                  isert_conn->login_rsp_buf);
 644
 645         isert_conn->login_req_dma = ib_dma_map_single(ib_dev,
 646                                 (void *)isert_conn->login_req_buf,
 647                                 ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_FROM_DEVICE);
 648
 649         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_req_dma);
 650         if (ret) {
 651                 isert_err("login_req_dma mapping error: %d\n", ret);
 652                 isert_conn->login_req_dma = 0;
 653                 goto out_login_buf;
 654         }
 655
 656         isert_conn->login_rsp_dma = ib_dma_map_single(ib_dev,
 657                                         (void *)isert_conn->login_rsp_buf,
 658                                         ISER_RX_LOGIN_SIZE, DMA_TO_DEVICE);
 659
 660         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_rsp_dma);
 661         if (ret) {
 662                 isert_err("login_rsp_dma mapping error: %d\n", ret);
 663                 isert_conn->login_rsp_dma = 0;
 664                 goto out_req_dma_map;
 665         }
 666
 667         return 0;
 668
 669 out_req_dma_map:
 670         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 671                             ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_FROM_DEVICE);
 672 out_login_buf:
 673         kfree(isert_conn->login_buf);
 674         return ret;
 675 }
 676
 677 static int
 678 isert_connect_request(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 679 {
 680         struct isert_np *isert_np = cma_id->context;
 681         struct iscsi_np *np = isert_np->np;
 682         struct isert_conn *isert_conn;
 683         struct isert_device *device;
 684         int ret = 0;
 685
 686         spin_lock_bh(&np->np_thread_lock);
 687         if (!np->enabled) {
 688                 spin_unlock_bh(&np->np_thread_lock);
 689                 isert_dbg("iscsi_np is not enabled, reject connect request\n");
 690                 return rdma_reject(cma_id, NULL, 0);
 691         }
 692         spin_unlock_bh(&np->np_thread_lock);
 693
 694         isert_dbg("cma_id: %p, portal: %p\n",
 695                  cma_id, cma_id->context);
 696
 697         isert_conn = kzalloc(sizeof(struct isert_conn), GFP_KERNEL);
 698         if (!isert_conn)
 699                 return -ENOMEM;
 700
 701         isert_init_conn(isert_conn);
 702         isert_conn->cm_id = cma_id;
 703
 704         ret = isert_alloc_login_buf(isert_conn, cma_id->device);
 705         if (ret)
 706                 goto out;
 707
 708         device = isert_device_get(cma_id);
 709         if (IS_ERR(device)) {
 710                 ret = PTR_ERR(device);
 711                 goto out_rsp_dma_map;
 712         }
 713         isert_conn->device = device;
 714
 715         /* Set max inflight RDMA READ requests */
 716         isert_conn->initiator_depth = min_t(u8,
 717                                 event->param.conn.initiator_depth,
 718                                 device->dev_attr.max_qp_init_rd_atom);
 719         isert_dbg("Using initiator_depth: %u\n", isert_conn->initiator_depth);
 720
 721         ret = isert_conn_setup_qp(isert_conn, cma_id);
 722         if (ret)
 723                 goto out_conn_dev;
 724
 725         ret = isert_rdma_post_recvl(isert_conn);
 726         if (ret)
 727                 goto out_conn_dev;
 728
 729         ret = isert_rdma_accept(isert_conn);
 730         if (ret)
 731                 goto out_conn_dev;
 732
 733         mutex_lock(&isert_np->mutex);
 734         list_add_tail(&isert_conn->node, &isert_np->accepted);
 735         mutex_unlock(&isert_np->mutex);
 736
 737         return 0;
 738
 739 out_conn_dev:
 740         isert_device_put(device);
 741 out_rsp_dma_map:
 742         isert_free_login_buf(isert_conn);
 743 out:
 744         kfree(isert_conn);
 745         rdma_reject(cma_id, NULL, 0);
 746         return ret;
 747 }
 748
 749 static void
 750 isert_connect_release(struct isert_conn *isert_conn)
 751 {
 752         struct isert_device *device = isert_conn->device;
 753
 754         isert_dbg("conn %p\n", isert_conn);
 755
 756         BUG_ON(!device);
 757
 758         if (device->use_fastreg)
 759                 isert_conn_free_fastreg_pool(isert_conn);
 760
 761         isert_free_rx_descriptors(isert_conn);
 762         if (isert_conn->cm_id)
 763                 rdma_destroy_id(isert_conn->cm_id);
 764
 765         if (isert_conn->qp) {
 766                 struct isert_comp *comp = isert_conn->qp->recv_cq->cq_context;
 767
 768                 isert_comp_put(comp);
 769                 ib_destroy_qp(isert_conn->qp);
 770         }
 771
 772         if (isert_conn->login_buf)
 773                 isert_free_login_buf(isert_conn);
 774
 775         isert_device_put(device);
 776
 777         kfree(isert_conn);
 778 }
 779
 780 static void
 781 isert_connected_handler(struct rdma_cm_id *cma_id)
 782 {
 783         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 784         struct isert_np *isert_np = cma_id->context;
 785
 786         isert_info("conn %p\n", isert_conn);
 787
 788         mutex_lock(&isert_conn->mutex);
 789         isert_conn->state = ISER_CONN_UP;
 790         kref_get(&isert_conn->kref);
 791         mutex_unlock(&isert_conn->mutex);
 792
 793         mutex_lock(&isert_np->mutex);
 794         list_move_tail(&isert_conn->node, &isert_np->pending);
 795         mutex_unlock(&isert_np->mutex);
 796
 797         isert_info("np %p: Allow accept_np to continue\n", isert_np);
 798         up(&isert_np->sem);
 799 }
 800
 801 static void
 802 isert_release_kref(struct kref *kref)
 803 {
 804         struct isert_conn *isert_conn = container_of(kref,
 805                                 struct isert_conn, kref);
 806
 807         isert_info("conn %p final kref %s/%d\n", isert_conn, current->comm,
 808                    current->pid);
 809
 810         isert_connect_release(isert_conn);
 811 }
 812
 813 static void
 814 isert_put_conn(struct isert_conn *isert_conn)
 815 {
 816         kref_put(&isert_conn->kref, isert_release_kref);
 817 }
 818
 819 static void
 820 isert_handle_unbound_conn(struct isert_conn *isert_conn)
 821 {
 822         struct isert_np *isert_np = isert_conn->cm_id->context;
 823
 824         mutex_lock(&isert_np->mutex);
 825         if (!list_empty(&isert_conn->node)) {
 826                 /*
 827                  * This means iscsi doesn't know this connection
 828                  * so schedule a cleanup ourselves
 829                  */
 830                 list_del_init(&isert_conn->node);
 831                 isert_put_conn(isert_conn);
 832                 complete(&isert_conn->wait);
 833                 queue_work(isert_release_wq, &isert_conn->release_work);
 834         }
 835         mutex_unlock(&isert_np->mutex);
 836 }
 837
 838 /**
 839  * isert_conn_terminate() - Initiate connection termination
 840  * @isert_conn: isert connection struct
 841  *
 842  * Notes:
 843  * In case the connection state is BOUND, move state
 844  * to TEMINATING and start teardown sequence (rdma_disconnect).
 845  * In case the connection state is UP, complete flush as well.
 846  *
 847  * This routine must be called with mutex held. Thus it is
 848  * safe to call multiple times.
 849  */
 850 static void
 851 isert_conn_terminate(struct isert_conn *isert_conn)
 852 {
 853         int err;
 854
 855         if (isert_conn->state >= ISER_CONN_TERMINATING)
 856                 return;
 857
 858         isert_info("Terminating conn %p state %d\n",
 859                    isert_conn, isert_conn->state);
 860         isert_conn->state = ISER_CONN_TERMINATING;
 861         err = rdma_disconnect(isert_conn->cm_id);
 862         if (err)
 863                 isert_warn("Failed rdma_disconnect isert_conn %p\n",
 864                            isert_conn);
 865
 866         isert_info("conn %p completing wait\n", isert_conn);
 867         complete(&isert_conn->wait);
 868 }
 869
 870 static int
 871 isert_np_cma_handler(struct isert_np *isert_np,
 872                      enum rdma_cm_event_type event)
 873 {
 874         isert_dbg("%s (%d): isert np %p\n",
 875                   rdma_event_msg(event), event, isert_np);
 876
 877         switch (event) {
 878         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 879                 isert_np->cm_id = NULL;
 880                 break;
 881         case RDMA_CM_EVENT_ADDR_CHANGE:
 882                 isert_np->cm_id = isert_setup_id(isert_np);
 883                 if (IS_ERR(isert_np->cm_id)) {
 884                         isert_err("isert np %p setup id failed: %ld\n",
 885                                   isert_np, PTR_ERR(isert_np->cm_id));
 886                         isert_np->cm_id = NULL;
 887                 }
 888                 break;
 889         default:
 890                 isert_err("isert np %p Unexpected event %d\n",
 891                           isert_np, event);
 892         }
 893
 894         return -1;
 895 }
 896
 897 static int
 898 isert_disconnected_handler(struct rdma_cm_id *cma_id,
 899                            enum rdma_cm_event_type event)
 900 {
 901         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 902
 903         mutex_lock(&isert_conn->mutex);
 904         switch (isert_conn->state) {
 905         case ISER_CONN_TERMINATING:
 906                 break;
 907         case ISER_CONN_UP:
 908                 isert_conn_terminate(isert_conn);
 909                 isert_wait4flush(isert_conn);
 910                 isert_handle_unbound_conn(isert_conn);
 911                 break;
 912         case ISER_CONN_BOUND:
 913         case ISER_CONN_FULL_FEATURE: /* FALLTHRU */
 914                 iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
 915                 break;
 916         default:
 917                 isert_warn("conn %p teminating in state %d\n",
 918                            isert_conn, isert_conn->state);
 919         }
 920         mutex_unlock(&isert_conn->mutex);
 921
 922         return 0;
 923 }
 924
 925 static int
 926 isert_connect_error(struct rdma_cm_id *cma_id)
 927 {
 928         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 929
 930         list_del_init(&isert_conn->node);
 931         isert_conn->cm_id = NULL;
 932         isert_put_conn(isert_conn);
 933
 934         return -1;
 935 }
 936
 937 static int
 938 isert_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 939 {
 940         struct isert_np *isert_np = cma_id->context;
 941         int ret = 0;
 942
 943         isert_info("%s (%d): status %d id %p np %p\n",
 944                    rdma_event_msg(event->event), event->event,
 945                    event->status, cma_id, cma_id->context);
 946
 947         if (isert_np->cm_id == cma_id)
 948                 return isert_np_cma_handler(cma_id->context, event->event);
 949
 950         switch (event->event) {
 951         case RDMA_CM_EVENT_CONNECT_REQUEST:
 952                 ret = isert_connect_request(cma_id, event);
 953                 if (ret)
 954                         isert_err("failed handle connect request %d\n", ret);
 955                 break;
 956         case RDMA_CM_EVENT_ESTABLISHED:
 957                 isert_connected_handler(cma_id);
 958                 break;
 959         case RDMA_CM_EVENT_ADDR_CHANGE:    /* FALLTHRU */
 960         case RDMA_CM_EVENT_DISCONNECTED:   /* FALLTHRU */
 961         case RDMA_CM_EVENT_DEVICE_REMOVAL: /* FALLTHRU */
 962         case RDMA_CM_EVENT_TIMEWAIT_EXIT:  /* FALLTHRU */
 963                 ret = isert_disconnected_handler(cma_id, event->event);
 964                 break;
 965         case RDMA_CM_EVENT_REJECTED:       /* FALLTHRU */
 966         case RDMA_CM_EVENT_UNREACHABLE:    /* FALLTHRU */
 967         case RDMA_CM_EVENT_CONNECT_ERROR:
 968                 ret = isert_connect_error(cma_id);
 969                 break;
 970         default:
 971                 isert_err("Unhandled RDMA CMA event: %d\n", event->event);
 972                 break;
 973         }
 974
 975         return ret;
 976 }
 977
 978 static int
 979 isert_post_recvm(struct isert_conn *isert_conn, u32 count)
 980 {
 981         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 982         int i, ret;
 983         struct iser_rx_desc *rx_desc;
 984
 985         for (rx_wr = isert_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
 986                 rx_desc = &isert_conn->rx_descs[i];
 987                 rx_wr->wr_id = (uintptr_t)rx_desc;
 988                 rx_wr->sg_list = &rx_desc->rx_sg;
 989                 rx_wr->num_sge = 1;
 990                 rx_wr->next = rx_wr + 1;
 991         }
 992         rx_wr--;
 993         rx_wr->next = NULL; /* mark end of work requests list */
 994
 995         ret = ib_post_recv(isert_conn->qp, isert_conn->rx_wr,
 996                            &rx_wr_failed);
 997         if (ret)
 998                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
 999
1000         return ret;
1001 }
1002
1003 static int
1004 isert_post_recv(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc)
1005 {
1006         struct ib_recv_wr *rx_wr_failed, rx_wr;
1007         int ret;
1008
1009         rx_wr.wr_id = (uintptr_t)rx_desc;
1010         rx_wr.sg_list = &rx_desc->rx_sg;
1011         rx_wr.num_sge = 1;
1012         rx_wr.next = NULL;
1013
1014         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_failed);
1015         if (ret)
1016                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
1017
1018         return ret;
1019 }
1020
1021 static int
1022 isert_post_send(struct isert_conn *isert_conn, struct iser_tx_desc *tx_desc)
1023 {
1024         struct ib_device *ib_dev = isert_conn->cm_id->device;
1025         struct ib_send_wr send_wr, *send_wr_failed;
1026         int ret;
1027
1028         ib_dma_sync_single_for_device(ib_dev, tx_desc->dma_addr,
1029                                       ISER_HEADERS_LEN, DMA_TO_DEVICE);
1030
1031         send_wr.next    = NULL;
1032         send_wr.wr_id   = (uintptr_t)tx_desc;
1033         send_wr.sg_list = tx_desc->tx_sg;
1034         send_wr.num_sge = tx_desc->num_sge;
1035         send_wr.opcode  = IB_WR_SEND;
1036         send_wr.send_flags = IB_SEND_SIGNALED;
1037
1038         ret = ib_post_send(isert_conn->qp, &send_wr, &send_wr_failed);
1039         if (ret)
1040                 isert_err("ib_post_send() failed, ret: %d\n", ret);
1041
1042         return ret;
1043 }
1044
1045 static void
1046 isert_create_send_desc(struct isert_conn *isert_conn,
1047                        struct isert_cmd *isert_cmd,
1048                        struct iser_tx_desc *tx_desc)
1049 {
1050         struct isert_device *device = isert_conn->device;
1051         struct ib_device *ib_dev = device->ib_device;
1052
1053         ib_dma_sync_single_for_cpu(ib_dev, tx_desc->dma_addr,
1054                                    ISER_HEADERS_LEN, DMA_TO_DEVICE);
1055
1056         memset(&tx_desc->iser_header, 0, sizeof(struct iser_hdr));
1057         tx_desc->iser_header.flags = ISER_VER;
1058
1059         tx_desc->num_sge = 1;
1060         tx_desc->isert_cmd = isert_cmd;
1061
1062         if (tx_desc->tx_sg[0].lkey != device->pd->local_dma_lkey) {
1063                 tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
1064                 isert_dbg("tx_desc %p lkey mismatch, fixing\n", tx_desc);
1065         }
1066 }
1067
1068 static int
1069 isert_init_tx_hdrs(struct isert_conn *isert_conn,
1070                    struct iser_tx_desc *tx_desc)
1071 {
1072         struct isert_device *device = isert_conn->device;
1073         struct ib_device *ib_dev = device->ib_device;
1074         u64 dma_addr;
1075
1076         dma_addr = ib_dma_map_single(ib_dev, (void *)tx_desc,
1077                         ISER_HEADERS_LEN, DMA_TO_DEVICE);
1078         if (ib_dma_mapping_error(ib_dev, dma_addr)) {
1079                 isert_err("ib_dma_mapping_error() failed\n");
1080                 return -ENOMEM;
1081         }
1082
1083         tx_desc->dma_addr = dma_addr;
1084         tx_desc->tx_sg[0].addr  = tx_desc->dma_addr;
1085         tx_desc->tx_sg[0].length = ISER_HEADERS_LEN;
1086         tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
1087
1088         isert_dbg("Setup tx_sg[0].addr: 0x%llx length: %u lkey: 0x%x\n",
1089                   tx_desc->tx_sg[0].addr, tx_desc->tx_sg[0].length,
1090                   tx_desc->tx_sg[0].lkey);
1091
1092         return 0;
1093 }
1094
1095 static void
1096 isert_init_send_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1097                    struct ib_send_wr *send_wr)
1098 {
1099         struct iser_tx_desc *tx_desc = &isert_cmd->tx_desc;
1100
1101         isert_cmd->rdma_wr.iser_ib_op = ISER_IB_SEND;
1102         send_wr->wr_id = (uintptr_t)&isert_cmd->tx_desc;
1103         send_wr->opcode = IB_WR_SEND;
1104         send_wr->sg_list = &tx_desc->tx_sg[0];
1105         send_wr->num_sge = isert_cmd->tx_desc.num_sge;
1106         send_wr->send_flags = IB_SEND_SIGNALED;
1107 }
1108
1109 static int
1110 isert_rdma_post_recvl(struct isert_conn *isert_conn)
1111 {
1112         struct ib_recv_wr rx_wr, *rx_wr_fail;
1113         struct ib_sge sge;
1114         int ret;
1115
1116         memset(&sge, 0, sizeof(struct ib_sge));
1117         sge.addr = isert_conn->login_req_dma;
1118         sge.length = ISER_RX_LOGIN_SIZE;
1119         sge.lkey = isert_conn->device->pd->local_dma_lkey;
1120
1121         isert_dbg("Setup sge: addr: %llx length: %d 0x%08x\n",
1122                 sge.addr, sge.length, sge.lkey);
1123
1124         memset(&rx_wr, 0, sizeof(struct ib_recv_wr));
1125         rx_wr.wr_id = (uintptr_t)isert_conn->login_req_buf;
1126         rx_wr.sg_list = &sge;
1127         rx_wr.num_sge = 1;
1128
1129         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_fail);
1130         if (ret)
1131                 isert_err("ib_post_recv() failed: %d\n", ret);
1132
1133         return ret;
1134 }
1135
1136 static int
1137 isert_put_login_tx(struct iscsi_conn *conn, struct iscsi_login *login,
1138                    u32 length)
1139 {
1140         struct isert_conn *isert_conn = conn->context;
1141         struct isert_device *device = isert_conn->device;
1142         struct ib_device *ib_dev = device->ib_device;
1143         struct iser_tx_desc *tx_desc = &isert_conn->login_tx_desc;
1144         int ret;
1145
1146         isert_create_send_desc(isert_conn, NULL, tx_desc);
1147
1148         memcpy(&tx_desc->iscsi_header, &login->rsp[0],
1149                sizeof(struct iscsi_hdr));
1150
1151         isert_init_tx_hdrs(isert_conn, tx_desc);
1152
1153         if (length > 0) {
1154                 struct ib_sge *tx_dsg = &tx_desc->tx_sg[1];
1155
1156                 ib_dma_sync_single_for_cpu(ib_dev, isert_conn->login_rsp_dma,
1157                                            length, DMA_TO_DEVICE);
1158
1159                 memcpy(isert_conn->login_rsp_buf, login->rsp_buf, length);
1160
1161                 ib_dma_sync_single_for_device(ib_dev, isert_conn->login_rsp_dma,
1162                                               length, DMA_TO_DEVICE);
1163
1164                 tx_dsg->addr    = isert_conn->login_rsp_dma;
1165                 tx_dsg->length  = length;
1166                 tx_dsg->lkey    = isert_conn->device->pd->local_dma_lkey;
1167                 tx_desc->num_sge = 2;
1168         }
1169         if (!login->login_failed) {
1170                 if (login->login_complete) {
1171                         if (!conn->sess->sess_ops->SessionType &&
1172                             isert_conn->device->use_fastreg) {
1173                                 ret = isert_conn_create_fastreg_pool(isert_conn);
1174                                 if (ret) {
1175                                         isert_err("Conn: %p failed to create"
1176                                                " fastreg pool\n", isert_conn);
1177                                         return ret;
1178                                 }
1179                         }
1180
1181                         ret = isert_alloc_rx_descriptors(isert_conn);
1182                         if (ret)
1183                                 return ret;
1184
1185                         ret = isert_post_recvm(isert_conn,
1186                                                ISERT_QP_MAX_RECV_DTOS);
1187                         if (ret)
1188                                 return ret;
1189
1190                         /* Now we are in FULL_FEATURE phase */
1191                         mutex_lock(&isert_conn->mutex);
1192                         isert_conn->state = ISER_CONN_FULL_FEATURE;
1193                         mutex_unlock(&isert_conn->mutex);
1194                         goto post_send;
1195                 }
1196
1197                 ret = isert_rdma_post_recvl(isert_conn);
1198                 if (ret)
1199                         return ret;
1200         }
1201 post_send:
1202         ret = isert_post_send(isert_conn, tx_desc);
1203         if (ret)
1204                 return ret;
1205
1206         return 0;
1207 }
1208
1209 static void
1210 isert_rx_login_req(struct isert_conn *isert_conn)
1211 {
1212         struct iser_rx_desc *rx_desc = (void *)isert_conn->login_req_buf;
1213         int rx_buflen = isert_conn->login_req_len;
1214         struct iscsi_conn *conn = isert_conn->conn;
1215         struct iscsi_login *login = conn->conn_login;
1216         int size;
1217
1218         isert_info("conn %p\n", isert_conn);
1219
1220         WARN_ON_ONCE(!login);
1221
1222         if (login->first_request) {
1223                 struct iscsi_login_req *login_req =
1224                         (struct iscsi_login_req *)&rx_desc->iscsi_header;
1225                 /*
1226                  * Setup the initial iscsi_login values from the leading
1227                  * login request PDU.
1228                  */
1229                 login->leading_connection = (!login_req->tsih) ? 1 : 0;
1230                 login->current_stage =
1231                         (login_req->flags & ISCSI_FLAG_LOGIN_CURRENT_STAGE_MASK)
1232                          >> 2;
1233                 login->version_min      = login_req->min_version;
1234                 login->version_max      = login_req->max_version;
1235                 memcpy(login->isid, login_req->isid, 6);
1236                 login->cmd_sn           = be32_to_cpu(login_req->cmdsn);
1237                 login->init_task_tag    = login_req->itt;
1238                 login->initial_exp_statsn = be32_to_cpu(login_req->exp_statsn);
1239                 login->cid              = be16_to_cpu(login_req->cid);
1240                 login->tsih             = be16_to_cpu(login_req->tsih);
1241         }
1242
1243         memcpy(&login->req[0], (void *)&rx_desc->iscsi_header, ISCSI_HDR_LEN);
1244
1245         size = min(rx_buflen, MAX_KEY_VALUE_PAIRS);
1246         isert_dbg("Using login payload size: %d, rx_buflen: %d "
1247                   "MAX_KEY_VALUE_PAIRS: %d\n", size, rx_buflen,
1248                   MAX_KEY_VALUE_PAIRS);
1249         memcpy(login->req_buf, &rx_desc->data[0], size);
1250
1251         if (login->first_request) {
1252                 complete(&isert_conn->login_comp);
1253                 return;
1254         }
1255         schedule_delayed_work(&conn->login_work, 0);
1256 }
1257
1258 static struct iscsi_cmd
1259 *isert_allocate_cmd(struct iscsi_conn *conn, struct iser_rx_desc *rx_desc)
1260 {
1261         struct isert_conn *isert_conn = conn->context;
1262         struct isert_cmd *isert_cmd;
1263         struct iscsi_cmd *cmd;
1264
1265         cmd = iscsit_allocate_cmd(conn, TASK_INTERRUPTIBLE);
1266         if (!cmd) {
1267                 isert_err("Unable to allocate iscsi_cmd + isert_cmd\n");
1268                 return NULL;
1269         }
1270         isert_cmd = iscsit_priv_cmd(cmd);
1271         isert_cmd->conn = isert_conn;
1272         isert_cmd->iscsi_cmd = cmd;
1273         isert_cmd->rx_desc = rx_desc;
1274
1275         return cmd;
1276 }
1277
1278 static int
1279 isert_handle_scsi_cmd(struct isert_conn *isert_conn,
1280                       struct isert_cmd *isert_cmd, struct iscsi_cmd *cmd,
1281                       struct iser_rx_desc *rx_desc, unsigned char *buf)
1282 {
1283         struct iscsi_conn *conn = isert_conn->conn;
1284         struct iscsi_scsi_req *hdr = (struct iscsi_scsi_req *)buf;
1285         int imm_data, imm_data_len, unsol_data, sg_nents, rc;
1286         bool dump_payload = false;
1287         unsigned int data_len;
1288
1289         rc = iscsit_setup_scsi_cmd(conn, cmd, buf);
1290         if (rc < 0)
1291                 return rc;
1292
1293         imm_data = cmd->immediate_data;
1294         imm_data_len = cmd->first_burst_len;
1295         unsol_data = cmd->unsolicited_data;
1296         data_len = cmd->se_cmd.data_length;
1297
1298         if (imm_data && imm_data_len == data_len)
1299                 cmd->se_cmd.se_cmd_flags |= SCF_PASSTHROUGH_SG_TO_MEM_NOALLOC;
1300         rc = iscsit_process_scsi_cmd(conn, cmd, hdr);
1301         if (rc < 0) {
1302                 return 0;
1303         } else if (rc > 0) {
1304                 dump_payload = true;
1305                 goto sequence_cmd;
1306         }
1307
1308         if (!imm_data)
1309                 return 0;
1310
1311         if (imm_data_len != data_len) {
1312                 sg_nents = max(1UL, DIV_ROUND_UP(imm_data_len, PAGE_SIZE));
1313                 sg_copy_from_buffer(cmd->se_cmd.t_data_sg, sg_nents,
1314                                     &rx_desc->data[0], imm_data_len);
1315                 isert_dbg("Copy Immediate sg_nents: %u imm_data_len: %d\n",
1316                           sg_nents, imm_data_len);
1317         } else {
1318                 sg_init_table(&isert_cmd->sg, 1);
1319                 cmd->se_cmd.t_data_sg = &isert_cmd->sg;
1320                 cmd->se_cmd.t_data_nents = 1;
1321                 sg_set_buf(&isert_cmd->sg, &rx_desc->data[0], imm_data_len);
1322                 isert_dbg("Transfer Immediate imm_data_len: %d\n",
1323                           imm_data_len);
1324         }
1325
1326         cmd->write_data_done += imm_data_len;
1327
1328         if (cmd->write_data_done == cmd->se_cmd.data_length) {
1329                 spin_lock_bh(&cmd->istate_lock);
1330                 cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1331                 cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1332                 spin_unlock_bh(&cmd->istate_lock);
1333         }
1334
1335 sequence_cmd:
1336         rc = iscsit_sequence_cmd(conn, cmd, buf, hdr->cmdsn);
1337
1338         if (!rc && dump_payload == false && unsol_data)
1339                 iscsit_set_unsoliticed_dataout(cmd);
1340         else if (dump_payload && imm_data)
1341                 target_put_sess_cmd(&cmd->se_cmd);
1342
1343         return 0;
1344 }
1345
1346 static int
1347 isert_handle_iscsi_dataout(struct isert_conn *isert_conn,
1348                            struct iser_rx_desc *rx_desc, unsigned char *buf)
1349 {
1350         struct scatterlist *sg_start;
1351         struct iscsi_conn *conn = isert_conn->conn;
1352         struct iscsi_cmd *cmd = NULL;
1353         struct iscsi_data *hdr = (struct iscsi_data *)buf;
1354         u32 unsol_data_len = ntoh24(hdr->dlength);
1355         int rc, sg_nents, sg_off, page_off;
1356
1357         rc = iscsit_check_dataout_hdr(conn, buf, &cmd);
1358         if (rc < 0)
1359                 return rc;
1360         else if (!cmd)
1361                 return 0;
1362         /*
1363          * FIXME: Unexpected unsolicited_data out
1364          */
1365         if (!cmd->unsolicited_data) {
1366                 isert_err("Received unexpected solicited data payload\n");
1367                 dump_stack();
1368                 return -1;
1369         }
1370
1371         isert_dbg("Unsolicited DataOut unsol_data_len: %u, "
1372                   "write_data_done: %u, data_length: %u\n",
1373                   unsol_data_len,  cmd->write_data_done,
1374                   cmd->se_cmd.data_length);
1375
1376         sg_off = cmd->write_data_done / PAGE_SIZE;
1377         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
1378         sg_nents = max(1UL, DIV_ROUND_UP(unsol_data_len, PAGE_SIZE));
1379         page_off = cmd->write_data_done % PAGE_SIZE;
1380         /*
1381          * FIXME: Non page-aligned unsolicited_data out
1382          */
1383         if (page_off) {
1384                 isert_err("unexpected non-page aligned data payload\n");
1385                 dump_stack();
1386                 return -1;
1387         }
1388         isert_dbg("Copying DataOut: sg_start: %p, sg_off: %u "
1389                   "sg_nents: %u from %p %u\n", sg_start, sg_off,
1390                   sg_nents, &rx_desc->data[0], unsol_data_len);
1391
1392         sg_copy_from_buffer(sg_start, sg_nents, &rx_desc->data[0],
1393                             unsol_data_len);
1394
1395         rc = iscsit_check_dataout_payload(cmd, hdr, false);
1396         if (rc < 0)
1397                 return rc;
1398
1399         /*
1400          * multiple data-outs on the same command can arrive -
1401          * so post the buffer before hand
1402          */
1403         rc = isert_post_recv(isert_conn, rx_desc);
1404         if (rc) {
1405                 isert_err("ib_post_recv failed with %d\n", rc);
1406                 return rc;
1407         }
1408         return 0;
1409 }
1410
1411 static int
1412 isert_handle_nop_out(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1413                      struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1414                      unsigned char *buf)
1415 {
1416         struct iscsi_conn *conn = isert_conn->conn;
1417         struct iscsi_nopout *hdr = (struct iscsi_nopout *)buf;
1418         int rc;
1419
1420         rc = iscsit_setup_nop_out(conn, cmd, hdr);
1421         if (rc < 0)
1422                 return rc;
1423         /*
1424          * FIXME: Add support for NOPOUT payload using unsolicited RDMA payload
1425          */
1426
1427         return iscsit_process_nop_out(conn, cmd, hdr);
1428 }
1429
1430 static int
1431 isert_handle_text_cmd(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1432                       struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1433                       struct iscsi_text *hdr)
1434 {
1435         struct iscsi_conn *conn = isert_conn->conn;
1436         u32 payload_length = ntoh24(hdr->dlength);
1437         int rc;
1438         unsigned char *text_in = NULL;
1439
1440         rc = iscsit_setup_text_cmd(conn, cmd, hdr);
1441         if (rc < 0)
1442                 return rc;
1443
1444         if (payload_length) {
1445                 text_in = kzalloc(payload_length, GFP_KERNEL);
1446                 if (!text_in) {
1447                         isert_err("Unable to allocate text_in of payload_length: %u\n",
1448                                   payload_length);
1449                         return -ENOMEM;
1450                 }
1451         }
1452         cmd->text_in_ptr = text_in;
1453
1454         memcpy(cmd->text_in_ptr, &rx_desc->data[0], payload_length);
1455
1456         return iscsit_process_text_cmd(conn, cmd, hdr);
1457 }
1458
1459 static int
1460 isert_rx_opcode(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc,
1461                 uint32_t read_stag, uint64_t read_va,
1462                 uint32_t write_stag, uint64_t write_va)
1463 {
1464         struct iscsi_hdr *hdr = &rx_desc->iscsi_header;
1465         struct iscsi_conn *conn = isert_conn->conn;
1466         struct iscsi_cmd *cmd;
1467         struct isert_cmd *isert_cmd;
1468         int ret = -EINVAL;
1469         u8 opcode = (hdr->opcode & ISCSI_OPCODE_MASK);
1470
1471         if (conn->sess->sess_ops->SessionType &&
1472            (!(opcode & ISCSI_OP_TEXT) || !(opcode & ISCSI_OP_LOGOUT))) {
1473                 isert_err("Got illegal opcode: 0x%02x in SessionType=Discovery,"
1474                           " ignoring\n", opcode);
1475                 return 0;
1476         }
1477
1478         switch (opcode) {
1479         case ISCSI_OP_SCSI_CMD:
1480                 cmd = isert_allocate_cmd(conn, rx_desc);
1481                 if (!cmd)
1482                         break;
1483
1484                 isert_cmd = iscsit_priv_cmd(cmd);
1485                 isert_cmd->read_stag = read_stag;
1486                 isert_cmd->read_va = read_va;
1487                 isert_cmd->write_stag = write_stag;
1488                 isert_cmd->write_va = write_va;
1489
1490                 ret = isert_handle_scsi_cmd(isert_conn, isert_cmd, cmd,
1491                                         rx_desc, (unsigned char *)hdr);
1492                 break;
1493         case ISCSI_OP_NOOP_OUT:
1494                 cmd = isert_allocate_cmd(conn, rx_desc);
1495                 if (!cmd)
1496                         break;
1497
1498                 isert_cmd = iscsit_priv_cmd(cmd);
1499                 ret = isert_handle_nop_out(isert_conn, isert_cmd, cmd,
1500                                            rx_desc, (unsigned char *)hdr);
1501                 break;
1502         case ISCSI_OP_SCSI_DATA_OUT:
1503                 ret = isert_handle_iscsi_dataout(isert_conn, rx_desc,
1504                                                 (unsigned char *)hdr);
1505                 break;
1506         case ISCSI_OP_SCSI_TMFUNC:
1507                 cmd = isert_allocate_cmd(conn, rx_desc);
1508                 if (!cmd)
1509                         break;
1510
1511                 ret = iscsit_handle_task_mgt_cmd(conn, cmd,
1512                                                 (unsigned char *)hdr);
1513                 break;
1514         case ISCSI_OP_LOGOUT:
1515                 cmd = isert_allocate_cmd(conn, rx_desc);
1516                 if (!cmd)
1517                         break;
1518
1519                 ret = iscsit_handle_logout_cmd(conn, cmd, (unsigned char *)hdr);
1520                 break;
1521         case ISCSI_OP_TEXT:
1522                 if (be32_to_cpu(hdr->ttt) != 0xFFFFFFFF)
1523                         cmd = iscsit_find_cmd_from_itt(conn, hdr->itt);
1524                 else
1525                         cmd = isert_allocate_cmd(conn, rx_desc);
1526
1527                 if (!cmd)
1528                         break;
1529
1530                 isert_cmd = iscsit_priv_cmd(cmd);
1531                 ret = isert_handle_text_cmd(isert_conn, isert_cmd, cmd,
1532                                             rx_desc, (struct iscsi_text *)hdr);
1533                 break;
1534         default:
1535                 isert_err("Got unknown iSCSI OpCode: 0x%02x\n", opcode);
1536                 dump_stack();
1537                 break;
1538         }
1539
1540         return ret;
1541 }
1542
1543 static void
1544 isert_rx_do_work(struct iser_rx_desc *rx_desc, struct isert_conn *isert_conn)
1545 {
1546         struct iser_hdr *iser_hdr = &rx_desc->iser_header;
1547         uint64_t read_va = 0, write_va = 0;
1548         uint32_t read_stag = 0, write_stag = 0;
1549
1550         switch (iser_hdr->flags & 0xF0) {
1551         case ISCSI_CTRL:
1552                 if (iser_hdr->flags & ISER_RSV) {
1553                         read_stag = be32_to_cpu(iser_hdr->read_stag);
1554                         read_va = be64_to_cpu(iser_hdr->read_va);
1555                         isert_dbg("ISER_RSV: read_stag: 0x%x read_va: 0x%llx\n",
1556                                   read_stag, (unsigned long long)read_va);
1557                 }
1558                 if (iser_hdr->flags & ISER_WSV) {
1559                         write_stag = be32_to_cpu(iser_hdr->write_stag);
1560                         write_va = be64_to_cpu(iser_hdr->write_va);
1561                         isert_dbg("ISER_WSV: write_stag: 0x%x write_va: 0x%llx\n",
1562                                   write_stag, (unsigned long long)write_va);
1563                 }
1564
1565                 isert_dbg("ISER ISCSI_CTRL PDU\n");
1566                 break;
1567         case ISER_HELLO:
1568                 isert_err("iSER Hello message\n");
1569                 break;
1570         default:
1571                 isert_warn("Unknown iSER hdr flags: 0x%02x\n", iser_hdr->flags);
1572                 break;
1573         }
1574
1575         isert_rx_opcode(isert_conn, rx_desc,
1576                         read_stag, read_va, write_stag, write_va);
1577 }
1578
1579 static void
1580 isert_rcv_completion(struct iser_rx_desc *desc,
1581                      struct isert_conn *isert_conn,
1582                      u32 xfer_len)
1583 {
1584         struct ib_device *ib_dev = isert_conn->device->ib_device;
1585         struct iscsi_hdr *hdr;
1586         u64 rx_dma;
1587         int rx_buflen;
1588
1589         if ((char *)desc == isert_conn->login_req_buf) {
1590                 rx_dma = isert_conn->login_req_dma;
1591                 rx_buflen = ISER_RX_LOGIN_SIZE;
1592                 isert_dbg("login_buf: Using rx_dma: 0x%llx, rx_buflen: %d\n",
1593                          rx_dma, rx_buflen);
1594         } else {
1595                 rx_dma = desc->dma_addr;
1596                 rx_buflen = ISER_RX_PAYLOAD_SIZE;
1597                 isert_dbg("req_buf: Using rx_dma: 0x%llx, rx_buflen: %d\n",
1598                          rx_dma, rx_buflen);
1599         }
1600
1601         ib_dma_sync_single_for_cpu(ib_dev, rx_dma, rx_buflen, DMA_FROM_DEVICE);
1602
1603         hdr = &desc->iscsi_header;
1604         isert_dbg("iSCSI opcode: 0x%02x, ITT: 0x%08x, flags: 0x%02x dlen: %d\n",
1605                  hdr->opcode, hdr->itt, hdr->flags,
1606                  (int)(xfer_len - ISER_HEADERS_LEN));
1607
1608         if ((char *)desc == isert_conn->login_req_buf) {
1609                 isert_conn->login_req_len = xfer_len - ISER_HEADERS_LEN;
1610                 if (isert_conn->conn) {
1611                         struct iscsi_login *login = isert_conn->conn->conn_login;
1612
1613                         if (login && !login->first_request)
1614                                 isert_rx_login_req(isert_conn);
1615                 }
1616                 mutex_lock(&isert_conn->mutex);
1617                 complete(&isert_conn->login_req_comp);
1618                 mutex_unlock(&isert_conn->mutex);
1619         } else {
1620                 isert_rx_do_work(desc, isert_conn);
1621         }
1622
1623         ib_dma_sync_single_for_device(ib_dev, rx_dma, rx_buflen,
1624                                       DMA_FROM_DEVICE);
1625
1626 }
1627
1628 static int
1629 isert_map_data_buf(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1630                    struct scatterlist *sg, u32 nents, u32 length, u32 offset,
1631                    enum iser_ib_op_code op, struct isert_data_buf *data)
1632 {
1633         struct ib_device *ib_dev = isert_conn->cm_id->device;
1634
1635         data->dma_dir = op == ISER_IB_RDMA_WRITE ?
1636                               DMA_TO_DEVICE : DMA_FROM_DEVICE;
1637
1638         data->len = length - offset;
1639         data->offset = offset;
1640         data->sg_off = data->offset / PAGE_SIZE;
1641
1642         data->sg = &sg[data->sg_off];
1643         data->nents = min_t(unsigned int, nents - data->sg_off,
1644                                           ISCSI_ISER_SG_TABLESIZE);
1645         data->len = min_t(unsigned int, data->len, ISCSI_ISER_SG_TABLESIZE *
1646                                         PAGE_SIZE);
1647
1648         data->dma_nents = ib_dma_map_sg(ib_dev, data->sg, data->nents,
1649                                         data->dma_dir);
1650         if (unlikely(!data->dma_nents)) {
1651                 isert_err("Cmd: unable to dma map SGs %p\n", sg);
1652                 return -EINVAL;
1653         }
1654
1655         isert_dbg("Mapped cmd: %p count: %u sg: %p sg_nents: %u rdma_len %d\n",
1656                   isert_cmd, data->dma_nents, data->sg, data->nents, data->len);
1657
1658         return 0;
1659 }
1660
1661 static void
1662 isert_unmap_data_buf(struct isert_conn *isert_conn, struct isert_data_buf *data)
1663 {
1664         struct ib_device *ib_dev = isert_conn->cm_id->device;
1665
1666         ib_dma_unmap_sg(ib_dev, data->sg, data->nents, data->dma_dir);
1667         memset(data, 0, sizeof(*data));
1668 }
1669
1670
1671
1672 static void
1673 isert_unmap_cmd(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn)
1674 {
1675         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1676
1677         isert_dbg("Cmd %p\n", isert_cmd);
1678
1679         if (wr->data.sg) {
1680                 isert_dbg("Cmd %p unmap_sg op\n", isert_cmd);
1681                 isert_unmap_data_buf(isert_conn, &wr->data);
1682         }
1683
1684         if (wr->rdma_wr) {
1685                 isert_dbg("Cmd %p free send_wr\n", isert_cmd);
1686                 kfree(wr->rdma_wr);
1687                 wr->rdma_wr = NULL;
1688         }
1689
1690         if (wr->ib_sge) {
1691                 isert_dbg("Cmd %p free ib_sge\n", isert_cmd);
1692                 kfree(wr->ib_sge);
1693                 wr->ib_sge = NULL;
1694         }
1695 }
1696
1697 static void
1698 isert_unreg_rdma(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn)
1699 {
1700         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1701
1702         isert_dbg("Cmd %p\n", isert_cmd);
1703
1704         if (wr->fr_desc) {
1705                 isert_dbg("Cmd %p free fr_desc %p\n", isert_cmd, wr->fr_desc);
1706                 if (wr->fr_desc->ind & ISERT_PROTECTED) {
1707                         isert_unmap_data_buf(isert_conn, &wr->prot);
1708                         wr->fr_desc->ind &= ~ISERT_PROTECTED;
1709                 }
1710                 spin_lock_bh(&isert_conn->pool_lock);
1711                 list_add_tail(&wr->fr_desc->list, &isert_conn->fr_pool);
1712                 spin_unlock_bh(&isert_conn->pool_lock);
1713                 wr->fr_desc = NULL;
1714         }
1715
1716         if (wr->data.sg) {
1717                 isert_dbg("Cmd %p unmap_sg op\n", isert_cmd);
1718                 isert_unmap_data_buf(isert_conn, &wr->data);
1719         }
1720
1721         wr->ib_sge = NULL;
1722         wr->rdma_wr = NULL;
1723 }
1724
1725 static void
1726 isert_put_cmd(struct isert_cmd *isert_cmd, bool comp_err)
1727 {
1728         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1729         struct isert_conn *isert_conn = isert_cmd->conn;
1730         struct iscsi_conn *conn = isert_conn->conn;
1731         struct isert_device *device = isert_conn->device;
1732         struct iscsi_text_rsp *hdr;
1733
1734         isert_dbg("Cmd %p\n", isert_cmd);
1735
1736         switch (cmd->iscsi_opcode) {
1737         case ISCSI_OP_SCSI_CMD:
1738                 spin_lock_bh(&conn->cmd_lock);
1739                 if (!list_empty(&cmd->i_conn_node))
1740                         list_del_init(&cmd->i_conn_node);
1741                 spin_unlock_bh(&conn->cmd_lock);
1742
1743                 if (cmd->data_direction == DMA_TO_DEVICE) {
1744                         iscsit_stop_dataout_timer(cmd);
1745                         /*
1746                          * Check for special case during comp_err where
1747                          * WRITE_PENDING has been handed off from core,
1748                          * but requires an extra target_put_sess_cmd()
1749                          * before transport_generic_free_cmd() below.
1750                          */
1751                         if (comp_err &&
1752                             cmd->se_cmd.t_state == TRANSPORT_WRITE_PENDING) {
1753                                 struct se_cmd *se_cmd = &cmd->se_cmd;
1754
1755                                 target_put_sess_cmd(se_cmd);
1756                         }
1757                 }
1758
1759                 device->unreg_rdma_mem(isert_cmd, isert_conn);
1760                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1761                 break;
1762         case ISCSI_OP_SCSI_TMFUNC:
1763                 spin_lock_bh(&conn->cmd_lock);
1764                 if (!list_empty(&cmd->i_conn_node))
1765                         list_del_init(&cmd->i_conn_node);
1766                 spin_unlock_bh(&conn->cmd_lock);
1767
1768                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1769                 break;
1770         case ISCSI_OP_REJECT:
1771         case ISCSI_OP_NOOP_OUT:
1772         case ISCSI_OP_TEXT:
1773                 hdr = (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
1774                 /* If the continue bit is on, keep the command alive */
1775                 if (hdr->flags & ISCSI_FLAG_TEXT_CONTINUE)
1776                         break;
1777
1778                 spin_lock_bh(&conn->cmd_lock);
1779                 if (!list_empty(&cmd->i_conn_node))
1780                         list_del_init(&cmd->i_conn_node);
1781                 spin_unlock_bh(&conn->cmd_lock);
1782
1783                 /*
1784                  * Handle special case for REJECT when iscsi_add_reject*() has
1785                  * overwritten the original iscsi_opcode assignment, and the
1786                  * associated cmd->se_cmd needs to be released.
1787                  */
1788                 if (cmd->se_cmd.se_tfo != NULL) {
1789                         isert_dbg("Calling transport_generic_free_cmd for 0x%02x\n",
1790                                  cmd->iscsi_opcode);
1791                         transport_generic_free_cmd(&cmd->se_cmd, 0);
1792                         break;
1793                 }
1794                 /*
1795                  * Fall-through
1796                  */
1797         default:
1798                 iscsit_release_cmd(cmd);
1799                 break;
1800         }
1801 }
1802
1803 static void
1804 isert_unmap_tx_desc(struct iser_tx_desc *tx_desc, struct ib_device *ib_dev)
1805 {
1806         if (tx_desc->dma_addr != 0) {
1807                 isert_dbg("unmap single for tx_desc->dma_addr\n");
1808                 ib_dma_unmap_single(ib_dev, tx_desc->dma_addr,
1809                                     ISER_HEADERS_LEN, DMA_TO_DEVICE);
1810                 tx_desc->dma_addr = 0;
1811         }
1812 }
1813
1814 static void
1815 isert_completion_put(struct iser_tx_desc *tx_desc, struct isert_cmd *isert_cmd,
1816                      struct ib_device *ib_dev, bool comp_err)
1817 {
1818         if (isert_cmd->pdu_buf_dma != 0) {
1819                 isert_dbg("unmap single for isert_cmd->pdu_buf_dma\n");
1820                 ib_dma_unmap_single(ib_dev, isert_cmd->pdu_buf_dma,
1821                                     isert_cmd->pdu_buf_len, DMA_TO_DEVICE);
1822                 isert_cmd->pdu_buf_dma = 0;
1823         }
1824
1825         isert_unmap_tx_desc(tx_desc, ib_dev);
1826         isert_put_cmd(isert_cmd, comp_err);
1827 }
1828
1829 static int
1830 isert_check_pi_status(struct se_cmd *se_cmd, struct ib_mr *sig_mr)
1831 {
1832         struct ib_mr_status mr_status;
1833         int ret;
1834
1835         ret = ib_check_mr_status(sig_mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
1836         if (ret) {
1837                 isert_err("ib_check_mr_status failed, ret %d\n", ret);
1838                 goto fail_mr_status;
1839         }
1840
1841         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1842                 u64 sec_offset_err;
1843                 u32 block_size = se_cmd->se_dev->dev_attrib.block_size + 8;
1844
1845                 switch (mr_status.sig_err.err_type) {
1846                 case IB_SIG_BAD_GUARD:
1847                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_GUARD_CHECK_FAILED;
1848                         break;
1849                 case IB_SIG_BAD_REFTAG:
1850                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_REF_TAG_CHECK_FAILED;
1851                         break;
1852                 case IB_SIG_BAD_APPTAG:
1853                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_APP_TAG_CHECK_FAILED;
1854                         break;
1855                 }
1856                 sec_offset_err = mr_status.sig_err.sig_err_offset;
1857                 do_div(sec_offset_err, block_size);
1858                 se_cmd->bad_sector = sec_offset_err + se_cmd->t_task_lba;
1859
1860                 isert_err("PI error found type %d at sector 0x%llx "
1861                           "expected 0x%x vs actual 0x%x\n",
1862                           mr_status.sig_err.err_type,
1863                           (unsigned long long)se_cmd->bad_sector,
1864                           mr_status.sig_err.expected,
1865                           mr_status.sig_err.actual);
1866                 ret = 1;
1867         }
1868
1869 fail_mr_status:
1870         return ret;
1871 }
1872
1873 static void
1874 isert_completion_rdma_write(struct iser_tx_desc *tx_desc,
1875                             struct isert_cmd *isert_cmd)
1876 {
1877         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1878         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1879         struct se_cmd *se_cmd = &cmd->se_cmd;
1880         struct isert_conn *isert_conn = isert_cmd->conn;
1881         struct isert_device *device = isert_conn->device;
1882         int ret = 0;
1883
1884         if (wr->fr_desc && wr->fr_desc->ind & ISERT_PROTECTED) {
1885                 ret = isert_check_pi_status(se_cmd,
1886                                             wr->fr_desc->pi_ctx->sig_mr);
1887                 wr->fr_desc->ind &= ~ISERT_PROTECTED;
1888         }
1889
1890         device->unreg_rdma_mem(isert_cmd, isert_conn);
1891         wr->rdma_wr_num = 0;
1892         if (ret)
1893                 transport_send_check_condition_and_sense(se_cmd,
1894                                                          se_cmd->pi_err, 0);
1895         else
1896                 isert_put_response(isert_conn->conn, cmd);
1897 }
1898
1899 static void
1900 isert_completion_rdma_read(struct iser_tx_desc *tx_desc,
1901                            struct isert_cmd *isert_cmd)
1902 {
1903         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1904         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1905         struct se_cmd *se_cmd = &cmd->se_cmd;
1906         struct isert_conn *isert_conn = isert_cmd->conn;
1907         struct isert_device *device = isert_conn->device;
1908         int ret = 0;
1909
1910         if (wr->fr_desc && wr->fr_desc->ind & ISERT_PROTECTED) {
1911                 ret = isert_check_pi_status(se_cmd,
1912                                             wr->fr_desc->pi_ctx->sig_mr);
1913                 wr->fr_desc->ind &= ~ISERT_PROTECTED;
1914         }
1915
1916         iscsit_stop_dataout_timer(cmd);
1917         device->unreg_rdma_mem(isert_cmd, isert_conn);
1918         cmd->write_data_done = wr->data.len;
1919         wr->rdma_wr_num = 0;
1920
1921         isert_dbg("Cmd: %p RDMA_READ comp calling execute_cmd\n", isert_cmd);
1922         spin_lock_bh(&cmd->istate_lock);
1923         cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1924         cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1925         spin_unlock_bh(&cmd->istate_lock);
1926
1927         if (ret) {
1928                 target_put_sess_cmd(se_cmd);
1929                 transport_send_check_condition_and_sense(se_cmd,
1930                                                          se_cmd->pi_err, 0);
1931         } else {
1932                 target_execute_cmd(se_cmd);
1933         }
1934 }
1935
1936 static void
1937 isert_do_control_comp(struct work_struct *work)
1938 {
1939         struct isert_cmd *isert_cmd = container_of(work,
1940                         struct isert_cmd, comp_work);
1941         struct isert_conn *isert_conn = isert_cmd->conn;
1942         struct ib_device *ib_dev = isert_conn->cm_id->device;
1943         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1944
1945         isert_dbg("Cmd %p i_state %d\n", isert_cmd, cmd->i_state);
1946
1947         switch (cmd->i_state) {
1948         case ISTATE_SEND_TASKMGTRSP:
1949                 iscsit_tmr_post_handler(cmd, cmd->conn);
1950         case ISTATE_SEND_REJECT:   /* FALLTHRU */
1951         case ISTATE_SEND_TEXTRSP:  /* FALLTHRU */
1952                 cmd->i_state = ISTATE_SENT_STATUS;
1953                 isert_completion_put(&isert_cmd->tx_desc, isert_cmd,
1954                                      ib_dev, false);
1955                 break;
1956         case ISTATE_SEND_LOGOUTRSP:
1957                 iscsit_logout_post_handler(cmd, cmd->conn);
1958                 break;
1959         default:
1960                 isert_err("Unknown i_state %d\n", cmd->i_state);
1961                 dump_stack();
1962                 break;
1963         }
1964 }
1965
1966 static void
1967 isert_response_completion(struct iser_tx_desc *tx_desc,
1968                           struct isert_cmd *isert_cmd,
1969                           struct isert_conn *isert_conn,
1970                           struct ib_device *ib_dev)
1971 {
1972         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1973
1974         if (cmd->i_state == ISTATE_SEND_TASKMGTRSP ||
1975             cmd->i_state == ISTATE_SEND_LOGOUTRSP ||
1976             cmd->i_state == ISTATE_SEND_REJECT ||
1977             cmd->i_state == ISTATE_SEND_TEXTRSP) {
1978                 isert_unmap_tx_desc(tx_desc, ib_dev);
1979
1980                 INIT_WORK(&isert_cmd->comp_work, isert_do_control_comp);
1981                 queue_work(isert_comp_wq, &isert_cmd->comp_work);
1982                 return;
1983         }
1984
1985         cmd->i_state = ISTATE_SENT_STATUS;
1986         isert_completion_put(tx_desc, isert_cmd, ib_dev, false);
1987 }
1988
1989 static void
1990 isert_snd_completion(struct iser_tx_desc *tx_desc,
1991                       struct isert_conn *isert_conn)
1992 {
1993         struct ib_device *ib_dev = isert_conn->cm_id->device;
1994         struct isert_cmd *isert_cmd = tx_desc->isert_cmd;
1995         struct isert_rdma_wr *wr;
1996
1997         if (!isert_cmd) {
1998                 isert_unmap_tx_desc(tx_desc, ib_dev);
1999                 return;
2000         }
2001         wr = &isert_cmd->rdma_wr;
2002
2003         isert_dbg("Cmd %p iser_ib_op %d\n", isert_cmd, wr->iser_ib_op);
2004
2005         switch (wr->iser_ib_op) {
2006         case ISER_IB_SEND:
2007                 isert_response_completion(tx_desc, isert_cmd,
2008                                           isert_conn, ib_dev);
2009                 break;
2010         case ISER_IB_RDMA_WRITE:
2011                 isert_completion_rdma_write(tx_desc, isert_cmd);
2012                 break;
2013         case ISER_IB_RDMA_READ:
2014                 isert_completion_rdma_read(tx_desc, isert_cmd);
2015                 break;
2016         default:
2017                 isert_err("Unknown wr->iser_ib_op: 0x%x\n", wr->iser_ib_op);
2018                 dump_stack();
2019                 break;
2020         }
2021 }
2022
2023 /**
2024  * is_isert_tx_desc() - Indicate if the completion wr_id
2025  *     is a TX descriptor or not.
2026  * @isert_conn: iser connection
2027  * @wr_id: completion WR identifier
2028  *
2029  * Since we cannot rely on wc opcode in FLUSH errors
2030  * we must work around it by checking if the wr_id address
2031  * falls in the iser connection rx_descs buffer. If so
2032  * it is an RX descriptor, otherwize it is a TX.
2033  */
2034 static inline bool
2035 is_isert_tx_desc(struct isert_conn *isert_conn, void *wr_id)
2036 {
2037         void *start = isert_conn->rx_descs;
2038         int len = ISERT_QP_MAX_RECV_DTOS * sizeof(*isert_conn->rx_descs);
2039
2040         if ((wr_id >= start && wr_id < start + len) ||
2041             (wr_id == isert_conn->login_req_buf))
2042                 return false;
2043
2044         return true;
2045 }
2046
2047 static void
2048 isert_cq_comp_err(struct isert_conn *isert_conn, struct ib_wc *wc)
2049 {
2050         if (wc->wr_id == ISER_BEACON_WRID) {
2051                 isert_info("conn %p completing wait_comp_err\n",
2052                            isert_conn);
2053                 complete(&isert_conn->wait_comp_err);
2054         } else if (is_isert_tx_desc(isert_conn, (void *)(uintptr_t)wc->wr_id)) {
2055                 struct ib_device *ib_dev = isert_conn->cm_id->device;
2056                 struct isert_cmd *isert_cmd;
2057                 struct iser_tx_desc *desc;
2058
2059                 desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
2060                 isert_cmd = desc->isert_cmd;
2061                 if (!isert_cmd)
2062                         isert_unmap_tx_desc(desc, ib_dev);
2063                 else
2064                         isert_completion_put(desc, isert_cmd, ib_dev, true);
2065         }
2066 }
2067
2068 static void
2069 isert_handle_wc(struct ib_wc *wc)
2070 {
2071         struct isert_conn *isert_conn;
2072         struct iser_tx_desc *tx_desc;
2073         struct iser_rx_desc *rx_desc;
2074
2075         isert_conn = wc->qp->qp_context;
2076         if (likely(wc->status == IB_WC_SUCCESS)) {
2077                 if (wc->opcode == IB_WC_RECV) {
2078                         rx_desc = (struct iser_rx_desc *)(uintptr_t)wc->wr_id;
2079                         isert_rcv_completion(rx_desc, isert_conn, wc->byte_len);
2080                 } else {
2081                         tx_desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
2082                         isert_snd_completion(tx_desc, isert_conn);
2083                 }
2084         } else {
2085                 if (wc->status != IB_WC_WR_FLUSH_ERR)
2086                         isert_err("%s (%d): wr id %llx vend_err %x\n",
2087                                   ib_wc_status_msg(wc->status), wc->status,
2088                                   wc->wr_id, wc->vendor_err);
2089                 else
2090                         isert_dbg("%s (%d): wr id %llx\n",
2091                                   ib_wc_status_msg(wc->status), wc->status,
2092                                   wc->wr_id);
2093
2094                 if (wc->wr_id != ISER_FASTREG_LI_WRID)
2095                         isert_cq_comp_err(isert_conn, wc);
2096         }
2097 }
2098
2099 static void
2100 isert_cq_work(struct work_struct *work)
2101 {
2102         enum { isert_poll_budget = 65536 };
2103         struct isert_comp *comp = container_of(work, struct isert_comp,
2104                                                work);
2105         struct ib_wc *const wcs = comp->wcs;
2106         int i, n, completed = 0;
2107
2108         while ((n = ib_poll_cq(comp->cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
2109                 for (i = 0; i < n; i++)
2110                         isert_handle_wc(&wcs[i]);
2111
2112                 completed += n;
2113                 if (completed >= isert_poll_budget)
2114                         break;
2115         }
2116
2117         ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP);
2118 }
2119
2120 static void
2121 isert_cq_callback(struct ib_cq *cq, void *context)
2122 {
2123         struct isert_comp *comp = context;
2124
2125         queue_work(isert_comp_wq, &comp->work);
2126 }
2127
2128 static int
2129 isert_post_response(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd)
2130 {
2131         struct ib_send_wr *wr_failed;
2132         int ret;
2133
2134         ret = isert_post_recv(isert_conn, isert_cmd->rx_desc);
2135         if (ret) {
2136                 isert_err("ib_post_recv failed with %d\n", ret);
2137                 return ret;
2138         }
2139
2140         ret = ib_post_send(isert_conn->qp, &isert_cmd->tx_desc.send_wr,
2141                            &wr_failed);
2142         if (ret) {
2143                 isert_err("ib_post_send failed with %d\n", ret);
2144                 return ret;
2145         }
2146         return ret;
2147 }
2148
2149 static int
2150 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2151 {
2152         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2153         struct isert_conn *isert_conn = conn->context;
2154         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2155         struct iscsi_scsi_rsp *hdr = (struct iscsi_scsi_rsp *)
2156                                 &isert_cmd->tx_desc.iscsi_header;
2157
2158         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2159         iscsit_build_rsp_pdu(cmd, conn, true, hdr);
2160         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2161         /*
2162          * Attach SENSE DATA payload to iSCSI Response PDU
2163          */
2164         if (cmd->se_cmd.sense_buffer &&
2165             ((cmd->se_cmd.se_cmd_flags & SCF_TRANSPORT_TASK_SENSE) ||
2166             (cmd->se_cmd.se_cmd_flags & SCF_EMULATED_TASK_SENSE))) {
2167                 struct isert_device *device = isert_conn->device;
2168                 struct ib_device *ib_dev = device->ib_device;
2169                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2170                 u32 padding, pdu_len;
2171
2172                 put_unaligned_be16(cmd->se_cmd.scsi_sense_length,
2173                                    cmd->sense_buffer);
2174                 cmd->se_cmd.scsi_sense_length += sizeof(__be16);
2175
2176                 padding = -(cmd->se_cmd.scsi_sense_length) & 3;
2177                 hton24(hdr->dlength, (u32)cmd->se_cmd.scsi_sense_length);
2178                 pdu_len = cmd->se_cmd.scsi_sense_length + padding;
2179
2180                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2181                                 (void *)cmd->sense_buffer, pdu_len,
2182                                 DMA_TO_DEVICE);
2183
2184                 isert_cmd->pdu_buf_len = pdu_len;
2185                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2186                 tx_dsg->length  = pdu_len;
2187                 tx_dsg->lkey    = device->pd->local_dma_lkey;
2188                 isert_cmd->tx_desc.num_sge = 2;
2189         }
2190
2191         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2192
2193         isert_dbg("Posting SCSI Response\n");
2194
2195         return isert_post_response(isert_conn, isert_cmd);
2196 }
2197
2198 static void
2199 isert_aborted_task(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2200 {
2201         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2202         struct isert_conn *isert_conn = conn->context;
2203         struct isert_device *device = isert_conn->device;
2204
2205         spin_lock_bh(&conn->cmd_lock);
2206         if (!list_empty(&cmd->i_conn_node))
2207                 list_del_init(&cmd->i_conn_node);
2208         spin_unlock_bh(&conn->cmd_lock);
2209
2210         if (cmd->data_direction == DMA_TO_DEVICE)
2211                 iscsit_stop_dataout_timer(cmd);
2212
2213         device->unreg_rdma_mem(isert_cmd, isert_conn);
2214 }
2215
2216 static enum target_prot_op
2217 isert_get_sup_prot_ops(struct iscsi_conn *conn)
2218 {
2219         struct isert_conn *isert_conn = conn->context;
2220         struct isert_device *device = isert_conn->device;
2221
2222         if (conn->tpg->tpg_attrib.t10_pi) {
2223                 if (device->pi_capable) {
2224                         isert_info("conn %p PI offload enabled\n", isert_conn);
2225                         isert_conn->pi_support = true;
2226                         return TARGET_PROT_ALL;
2227                 }
2228         }
2229
2230         isert_info("conn %p PI offload disabled\n", isert_conn);
2231         isert_conn->pi_support = false;
2232
2233         return TARGET_PROT_NORMAL;
2234 }
2235
2236 static int
2237 isert_put_nopin(struct iscsi_cmd *cmd, struct iscsi_conn *conn,
2238                 bool nopout_response)
2239 {
2240         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2241         struct isert_conn *isert_conn = conn->context;
2242         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2243
2244         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2245         iscsit_build_nopin_rsp(cmd, conn, (struct iscsi_nopin *)
2246                                &isert_cmd->tx_desc.iscsi_header,
2247                                nopout_response);
2248         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2249         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2250
2251         isert_dbg("conn %p Posting NOPIN Response\n", isert_conn);
2252
2253         return isert_post_response(isert_conn, isert_cmd);
2254 }
2255
2256 static int
2257 isert_put_logout_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2258 {
2259         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2260         struct isert_conn *isert_conn = conn->context;
2261         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2262
2263         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2264         iscsit_build_logout_rsp(cmd, conn, (struct iscsi_logout_rsp *)
2265                                 &isert_cmd->tx_desc.iscsi_header);
2266         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2267         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2268
2269         isert_dbg("conn %p Posting Logout Response\n", isert_conn);
2270
2271         return isert_post_response(isert_conn, isert_cmd);
2272 }
2273
2274 static int
2275 isert_put_tm_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2276 {
2277         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2278         struct isert_conn *isert_conn = conn->context;
2279         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2280
2281         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2282         iscsit_build_task_mgt_rsp(cmd, conn, (struct iscsi_tm_rsp *)
2283                                   &isert_cmd->tx_desc.iscsi_header);
2284         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2285         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2286
2287         isert_dbg("conn %p Posting Task Management Response\n", isert_conn);
2288
2289         return isert_post_response(isert_conn, isert_cmd);
2290 }
2291
2292 static int
2293 isert_put_reject(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2294 {
2295         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2296         struct isert_conn *isert_conn = conn->context;
2297         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2298         struct isert_device *device = isert_conn->device;
2299         struct ib_device *ib_dev = device->ib_device;
2300         struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2301         struct iscsi_reject *hdr =
2302                 (struct iscsi_reject *)&isert_cmd->tx_desc.iscsi_header;
2303
2304         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2305         iscsit_build_reject(cmd, conn, hdr);
2306         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2307
2308         hton24(hdr->dlength, ISCSI_HDR_LEN);
2309         isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2310                         (void *)cmd->buf_ptr, ISCSI_HDR_LEN,
2311                         DMA_TO_DEVICE);
2312         isert_cmd->pdu_buf_len = ISCSI_HDR_LEN;
2313         tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2314         tx_dsg->length  = ISCSI_HDR_LEN;
2315         tx_dsg->lkey    = device->pd->local_dma_lkey;
2316         isert_cmd->tx_desc.num_sge = 2;
2317
2318         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2319
2320         isert_dbg("conn %p Posting Reject\n", isert_conn);
2321
2322         return isert_post_response(isert_conn, isert_cmd);
2323 }
2324
2325 static int
2326 isert_put_text_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2327 {
2328         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2329         struct isert_conn *isert_conn = conn->context;
2330         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2331         struct iscsi_text_rsp *hdr =
2332                 (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
2333         u32 txt_rsp_len;
2334         int rc;
2335
2336         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2337         rc = iscsit_build_text_rsp(cmd, conn, hdr, ISCSI_INFINIBAND);
2338         if (rc < 0)
2339                 return rc;
2340
2341         txt_rsp_len = rc;
2342         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2343
2344         if (txt_rsp_len) {
2345                 struct isert_device *device = isert_conn->device;
2346                 struct ib_device *ib_dev = device->ib_device;
2347                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2348                 void *txt_rsp_buf = cmd->buf_ptr;
2349
2350                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2351                                 txt_rsp_buf, txt_rsp_len, DMA_TO_DEVICE);
2352
2353                 isert_cmd->pdu_buf_len = txt_rsp_len;
2354                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2355                 tx_dsg->length  = txt_rsp_len;
2356                 tx_dsg->lkey    = device->pd->local_dma_lkey;
2357                 isert_cmd->tx_desc.num_sge = 2;
2358         }
2359         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2360
2361         isert_dbg("conn %p Text Response\n", isert_conn);
2362
2363         return isert_post_response(isert_conn, isert_cmd);
2364 }
2365
2366 static int
2367 isert_build_rdma_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
2368                     struct ib_sge *ib_sge, struct ib_rdma_wr *rdma_wr,
2369                     u32 data_left, u32 offset)
2370 {
2371         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
2372         struct scatterlist *sg_start, *tmp_sg;
2373         struct isert_device *device = isert_conn->device;
2374         struct ib_device *ib_dev = device->ib_device;
2375         u32 sg_off, page_off;
2376         int i = 0, sg_nents;
2377
2378         sg_off = offset / PAGE_SIZE;
2379         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
2380         sg_nents = min(cmd->se_cmd.t_data_nents - sg_off, isert_conn->max_sge);
2381         page_off = offset % PAGE_SIZE;
2382
2383         rdma_wr->wr.sg_list = ib_sge;
2384         rdma_wr->wr.wr_id = (uintptr_t)&isert_cmd->tx_desc;
2385         /*
2386          * Perform mapping of TCM scatterlist memory ib_sge dma_addr.
2387          */
2388         for_each_sg(sg_start, tmp_sg, sg_nents, i) {
2389                 isert_dbg("RDMA from SGL dma_addr: 0x%llx dma_len: %u, "
2390                           "page_off: %u\n",
2391                           (unsigned long long)tmp_sg->dma_address,
2392                           tmp_sg->length, page_off);
2393
2394                 ib_sge->addr = ib_sg_dma_address(ib_dev, tmp_sg) + page_off;
2395                 ib_sge->length = min_t(u32, data_left,
2396                                 ib_sg_dma_len(ib_dev, tmp_sg) - page_off);
2397                 ib_sge->lkey = device->pd->local_dma_lkey;
2398
2399                 isert_dbg("RDMA ib_sge: addr: 0x%llx  length: %u lkey: %x\n",
2400                           ib_sge->addr, ib_sge->length, ib_sge->lkey);
2401                 page_off = 0;
2402                 data_left -= ib_sge->length;
2403                 if (!data_left)
2404                         break;
2405                 ib_sge++;
2406                 isert_dbg("Incrementing ib_sge pointer to %p\n", ib_sge);
2407         }
2408
2409         rdma_wr->wr.num_sge = ++i;
2410         isert_dbg("Set outgoing sg_list: %p num_sg: %u from TCM SGLs\n",
2411                   rdma_wr->wr.sg_list, rdma_wr->wr.num_sge);
2412
2413         return rdma_wr->wr.num_sge;
2414 }
2415
2416 static int
2417 isert_map_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2418                struct isert_rdma_wr *wr)
2419 {
2420         struct se_cmd *se_cmd = &cmd->se_cmd;
2421         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2422         struct isert_conn *isert_conn = conn->context;
2423         struct isert_data_buf *data = &wr->data;
2424         struct ib_rdma_wr *rdma_wr;
2425         struct ib_sge *ib_sge;
2426         u32 offset, data_len, data_left, rdma_write_max, va_offset = 0;
2427         int ret = 0, i, ib_sge_cnt;
2428
2429         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2430
2431         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2432         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2433                                  se_cmd->t_data_nents, se_cmd->data_length,
2434                                  offset, wr->iser_ib_op, &wr->data);
2435         if (ret)
2436                 return ret;
2437
2438         data_left = data->len;
2439         offset = data->offset;
2440
2441         ib_sge = kzalloc(sizeof(struct ib_sge) * data->nents, GFP_KERNEL);
2442         if (!ib_sge) {
2443                 isert_warn("Unable to allocate ib_sge\n");
2444                 ret = -ENOMEM;
2445                 goto unmap_cmd;
2446         }
2447         wr->ib_sge = ib_sge;
2448
2449         wr->rdma_wr_num = DIV_ROUND_UP(data->nents, isert_conn->max_sge);
2450         wr->rdma_wr = kzalloc(sizeof(struct ib_rdma_wr) * wr->rdma_wr_num,
2451                                 GFP_KERNEL);
2452         if (!wr->rdma_wr) {
2453                 isert_dbg("Unable to allocate wr->rdma_wr\n");
2454                 ret = -ENOMEM;
2455                 goto unmap_cmd;
2456         }
2457
2458         wr->isert_cmd = isert_cmd;
2459         rdma_write_max = isert_conn->max_sge * PAGE_SIZE;
2460
2461         for (i = 0; i < wr->rdma_wr_num; i++) {
2462                 rdma_wr = &isert_cmd->rdma_wr.rdma_wr[i];
2463                 data_len = min(data_left, rdma_write_max);
2464
2465                 rdma_wr->wr.send_flags = 0;
2466                 if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2467                         rdma_wr->wr.opcode = IB_WR_RDMA_WRITE;
2468                         rdma_wr->remote_addr = isert_cmd->read_va + offset;
2469                         rdma_wr->rkey = isert_cmd->read_stag;
2470                         if (i + 1 == wr->rdma_wr_num)
2471                                 rdma_wr->wr.next = &isert_cmd->tx_desc.send_wr;
2472                         else
2473                                 rdma_wr->wr.next = &wr->rdma_wr[i + 1].wr;
2474                 } else {
2475                         rdma_wr->wr.opcode = IB_WR_RDMA_READ;
2476                         rdma_wr->remote_addr = isert_cmd->write_va + va_offset;
2477                         rdma_wr->rkey = isert_cmd->write_stag;
2478                         if (i + 1 == wr->rdma_wr_num)
2479                                 rdma_wr->wr.send_flags = IB_SEND_SIGNALED;
2480                         else
2481                                 rdma_wr->wr.next = &wr->rdma_wr[i + 1].wr;
2482                 }
2483
2484                 ib_sge_cnt = isert_build_rdma_wr(isert_conn, isert_cmd, ib_sge,
2485                                         rdma_wr, data_len, offset);
2486                 ib_sge += ib_sge_cnt;
2487
2488                 offset += data_len;
2489                 va_offset += data_len;
2490                 data_left -= data_len;
2491         }
2492
2493         return 0;
2494 unmap_cmd:
2495         isert_unmap_data_buf(isert_conn, data);
2496
2497         return ret;
2498 }
2499
2500 static inline void
2501 isert_inv_rkey(struct ib_send_wr *inv_wr, struct ib_mr *mr)
2502 {
2503         u32 rkey;
2504
2505         memset(inv_wr, 0, sizeof(*inv_wr));
2506         inv_wr->wr_id = ISER_FASTREG_LI_WRID;
2507         inv_wr->opcode = IB_WR_LOCAL_INV;
2508         inv_wr->ex.invalidate_rkey = mr->rkey;
2509
2510         /* Bump the key */
2511         rkey = ib_inc_rkey(mr->rkey);
2512         ib_update_fast_reg_key(mr, rkey);
2513 }
2514
2515 static int
2516 isert_fast_reg_mr(struct isert_conn *isert_conn,
2517                   struct fast_reg_descriptor *fr_desc,
2518                   struct isert_data_buf *mem,
2519                   enum isert_indicator ind,
2520                   struct ib_sge *sge)
2521 {
2522         struct isert_device *device = isert_conn->device;
2523         struct ib_device *ib_dev = device->ib_device;
2524         struct ib_mr *mr;
2525         struct ib_reg_wr reg_wr;
2526         struct ib_send_wr inv_wr, *bad_wr, *wr = NULL;
2527         int ret, n;
2528
2529         if (mem->dma_nents == 1) {
2530                 sge->lkey = device->pd->local_dma_lkey;
2531                 sge->addr = ib_sg_dma_address(ib_dev, &mem->sg[0]);
2532                 sge->length = ib_sg_dma_len(ib_dev, &mem->sg[0]);
2533                 isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2534                          sge->addr, sge->length, sge->lkey);
2535                 return 0;
2536         }
2537
2538         if (ind == ISERT_DATA_KEY_VALID)
2539                 /* Registering data buffer */
2540                 mr = fr_desc->data_mr;
2541         else
2542                 /* Registering protection buffer */
2543                 mr = fr_desc->pi_ctx->prot_mr;
2544
2545         if (!(fr_desc->ind & ind)) {
2546                 isert_inv_rkey(&inv_wr, mr);
2547                 wr = &inv_wr;
2548         }
2549
2550         n = ib_map_mr_sg(mr, mem->sg, mem->nents, PAGE_SIZE);
2551         if (unlikely(n != mem->nents)) {
2552                 isert_err("failed to map mr sg (%d/%d)\n",
2553                          n, mem->nents);
2554                 return n < 0 ? n : -EINVAL;
2555         }
2556
2557         isert_dbg("Use fr_desc %p sg_nents %d offset %u\n",
2558                   fr_desc, mem->nents, mem->offset);
2559
2560         reg_wr.wr.next = NULL;
2561         reg_wr.wr.opcode = IB_WR_REG_MR;
2562         reg_wr.wr.wr_id = ISER_FASTREG_LI_WRID;
2563         reg_wr.wr.send_flags = 0;
2564         reg_wr.wr.num_sge = 0;
2565         reg_wr.mr = mr;
2566         reg_wr.key = mr->lkey;
2567         reg_wr.access = IB_ACCESS_LOCAL_WRITE;
2568
2569         if (!wr)
2570                 wr = &reg_wr.wr;
2571         else
2572                 wr->next = &reg_wr.wr;
2573
2574         ret = ib_post_send(isert_conn->qp, wr, &bad_wr);
2575         if (ret) {
2576                 isert_err("fast registration failed, ret:%d\n", ret);
2577                 return ret;
2578         }
2579         fr_desc->ind &= ~ind;
2580
2581         sge->lkey = mr->lkey;
2582         sge->addr = mr->iova;
2583         sge->length = mr->length;
2584
2585         isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2586                   sge->addr, sge->length, sge->lkey);
2587
2588         return ret;
2589 }
2590
2591 static inline void
2592 isert_set_dif_domain(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs,
2593                      struct ib_sig_domain *domain)
2594 {
2595         domain->sig_type = IB_SIG_TYPE_T10_DIF;
2596         domain->sig.dif.bg_type = IB_T10DIF_CRC;
2597         domain->sig.dif.pi_interval = se_cmd->se_dev->dev_attrib.block_size;
2598         domain->sig.dif.ref_tag = se_cmd->reftag_seed;
2599         /*
2600          * At the moment we hard code those, but if in the future
2601          * the target core would like to use it, we will take it
2602          * from se_cmd.
2603          */
2604         domain->sig.dif.apptag_check_mask = 0xffff;
2605         domain->sig.dif.app_escape = true;
2606         domain->sig.dif.ref_escape = true;
2607         if (se_cmd->prot_type == TARGET_DIF_TYPE1_PROT ||
2608             se_cmd->prot_type == TARGET_DIF_TYPE2_PROT)
2609                 domain->sig.dif.ref_remap = true;
2610 };
2611
2612 static int
2613 isert_set_sig_attrs(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs)
2614 {
2615         switch (se_cmd->prot_op) {
2616         case TARGET_PROT_DIN_INSERT:
2617         case TARGET_PROT_DOUT_STRIP:
2618                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
2619                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2620                 break;
2621         case TARGET_PROT_DOUT_INSERT:
2622         case TARGET_PROT_DIN_STRIP:
2623                 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
2624                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2625                 break;
2626         case TARGET_PROT_DIN_PASS:
2627         case TARGET_PROT_DOUT_PASS:
2628                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2629                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2630                 break;
2631         default:
2632                 isert_err("Unsupported PI operation %d\n", se_cmd->prot_op);
2633                 return -EINVAL;
2634         }
2635
2636         return 0;
2637 }
2638
2639 static inline u8
2640 isert_set_prot_checks(u8 prot_checks)
2641 {
2642         return (prot_checks & TARGET_DIF_CHECK_GUARD  ? 0xc0 : 0) |
2643                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x30 : 0) |
2644                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x0f : 0);
2645 }
2646
2647 static int
2648 isert_reg_sig_mr(struct isert_conn *isert_conn,
2649                  struct se_cmd *se_cmd,
2650                  struct isert_rdma_wr *rdma_wr,
2651                  struct fast_reg_descriptor *fr_desc)
2652 {
2653         struct ib_sig_handover_wr sig_wr;
2654         struct ib_send_wr inv_wr, *bad_wr, *wr = NULL;
2655         struct pi_context *pi_ctx = fr_desc->pi_ctx;
2656         struct ib_sig_attrs sig_attrs;
2657         int ret;
2658
2659         memset(&sig_attrs, 0, sizeof(sig_attrs));
2660         ret = isert_set_sig_attrs(se_cmd, &sig_attrs);
2661         if (ret)
2662                 goto err;
2663
2664         sig_attrs.check_mask = isert_set_prot_checks(se_cmd->prot_checks);
2665
2666         if (!(fr_desc->ind & ISERT_SIG_KEY_VALID)) {
2667                 isert_inv_rkey(&inv_wr, pi_ctx->sig_mr);
2668                 wr = &inv_wr;
2669         }
2670
2671         memset(&sig_wr, 0, sizeof(sig_wr));
2672         sig_wr.wr.opcode = IB_WR_REG_SIG_MR;
2673         sig_wr.wr.wr_id = ISER_FASTREG_LI_WRID;
2674         sig_wr.wr.sg_list = &rdma_wr->ib_sg[DATA];
2675         sig_wr.wr.num_sge = 1;
2676         sig_wr.access_flags = IB_ACCESS_LOCAL_WRITE;
2677         sig_wr.sig_attrs = &sig_attrs;
2678         sig_wr.sig_mr = pi_ctx->sig_mr;
2679         if (se_cmd->t_prot_sg)
2680                 sig_wr.prot = &rdma_wr->ib_sg[PROT];
2681
2682         if (!wr)
2683                 wr = &sig_wr.wr;
2684         else
2685                 wr->next = &sig_wr.wr;
2686
2687         ret = ib_post_send(isert_conn->qp, wr, &bad_wr);
2688         if (ret) {
2689                 isert_err("fast registration failed, ret:%d\n", ret);
2690                 goto err;
2691         }
2692         fr_desc->ind &= ~ISERT_SIG_KEY_VALID;
2693
2694         rdma_wr->ib_sg[SIG].lkey = pi_ctx->sig_mr->lkey;
2695         rdma_wr->ib_sg[SIG].addr = 0;
2696         rdma_wr->ib_sg[SIG].length = se_cmd->data_length;
2697         if (se_cmd->prot_op != TARGET_PROT_DIN_STRIP &&
2698             se_cmd->prot_op != TARGET_PROT_DOUT_INSERT)
2699                 /*
2700                  * We have protection guards on the wire
2701                  * so we need to set a larget transfer
2702                  */
2703                 rdma_wr->ib_sg[SIG].length += se_cmd->prot_length;
2704
2705         isert_dbg("sig_sge: addr: 0x%llx  length: %u lkey: %x\n",
2706                   rdma_wr->ib_sg[SIG].addr, rdma_wr->ib_sg[SIG].length,
2707                   rdma_wr->ib_sg[SIG].lkey);
2708 err:
2709         return ret;
2710 }
2711
2712 static int
2713 isert_handle_prot_cmd(struct isert_conn *isert_conn,
2714                       struct isert_cmd *isert_cmd,
2715                       struct isert_rdma_wr *wr)
2716 {
2717         struct isert_device *device = isert_conn->device;
2718         struct se_cmd *se_cmd = &isert_cmd->iscsi_cmd->se_cmd;
2719         int ret;
2720
2721         if (!wr->fr_desc->pi_ctx) {
2722                 ret = isert_create_pi_ctx(wr->fr_desc,
2723                                           device->ib_device,
2724                                           device->pd);
2725                 if (ret) {
2726                         isert_err("conn %p failed to allocate pi_ctx\n",
2727                                   isert_conn);
2728                         return ret;
2729                 }
2730         }
2731
2732         if (se_cmd->t_prot_sg) {
2733                 ret = isert_map_data_buf(isert_conn, isert_cmd,
2734                                          se_cmd->t_prot_sg,
2735                                          se_cmd->t_prot_nents,
2736                                          se_cmd->prot_length,
2737                                          0, wr->iser_ib_op, &wr->prot);
2738                 if (ret) {
2739                         isert_err("conn %p failed to map protection buffer\n",
2740                                   isert_conn);
2741                         return ret;
2742                 }
2743
2744                 memset(&wr->ib_sg[PROT], 0, sizeof(wr->ib_sg[PROT]));
2745                 ret = isert_fast_reg_mr(isert_conn, wr->fr_desc, &wr->prot,
2746                                         ISERT_PROT_KEY_VALID, &wr->ib_sg[PROT]);
2747                 if (ret) {
2748                         isert_err("conn %p failed to fast reg mr\n",
2749                                   isert_conn);
2750                         goto unmap_prot_cmd;
2751                 }
2752         }
2753
2754         ret = isert_reg_sig_mr(isert_conn, se_cmd, wr, wr->fr_desc);
2755         if (ret) {
2756                 isert_err("conn %p failed to fast reg mr\n",
2757                           isert_conn);
2758                 goto unmap_prot_cmd;
2759         }
2760         wr->fr_desc->ind |= ISERT_PROTECTED;
2761
2762         return 0;
2763
2764 unmap_prot_cmd:
2765         if (se_cmd->t_prot_sg)
2766                 isert_unmap_data_buf(isert_conn, &wr->prot);
2767
2768         return ret;
2769 }
2770
2771 static int
2772 isert_reg_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2773                struct isert_rdma_wr *wr)
2774 {
2775         struct se_cmd *se_cmd = &cmd->se_cmd;
2776         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2777         struct isert_conn *isert_conn = conn->context;
2778         struct fast_reg_descriptor *fr_desc = NULL;
2779         struct ib_rdma_wr *rdma_wr;
2780         struct ib_sge *ib_sg;
2781         u32 offset;
2782         int ret = 0;
2783         unsigned long flags;
2784
2785         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2786
2787         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2788         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2789                                  se_cmd->t_data_nents, se_cmd->data_length,
2790                                  offset, wr->iser_ib_op, &wr->data);
2791         if (ret)
2792                 return ret;
2793
2794         if (wr->data.dma_nents != 1 || isert_prot_cmd(isert_conn, se_cmd)) {
2795                 spin_lock_irqsave(&isert_conn->pool_lock, flags);
2796                 fr_desc = list_first_entry(&isert_conn->fr_pool,
2797                                            struct fast_reg_descriptor, list);
2798                 list_del(&fr_desc->list);
2799                 spin_unlock_irqrestore(&isert_conn->pool_lock, flags);
2800                 wr->fr_desc = fr_desc;
2801         }
2802
2803         ret = isert_fast_reg_mr(isert_conn, fr_desc, &wr->data,
2804                                 ISERT_DATA_KEY_VALID, &wr->ib_sg[DATA]);
2805         if (ret)
2806                 goto unmap_cmd;
2807
2808         if (isert_prot_cmd(isert_conn, se_cmd)) {
2809                 ret = isert_handle_prot_cmd(isert_conn, isert_cmd, wr);
2810                 if (ret)
2811                         goto unmap_cmd;
2812
2813                 ib_sg = &wr->ib_sg[SIG];
2814         } else {
2815                 ib_sg = &wr->ib_sg[DATA];
2816         }
2817
2818         memcpy(&wr->s_ib_sge, ib_sg, sizeof(*ib_sg));
2819         wr->ib_sge = &wr->s_ib_sge;
2820         wr->rdma_wr_num = 1;
2821         memset(&wr->s_rdma_wr, 0, sizeof(wr->s_rdma_wr));
2822         wr->rdma_wr = &wr->s_rdma_wr;
2823         wr->isert_cmd = isert_cmd;
2824
2825         rdma_wr = &isert_cmd->rdma_wr.s_rdma_wr;
2826         rdma_wr->wr.sg_list = &wr->s_ib_sge;
2827         rdma_wr->wr.num_sge = 1;
2828         rdma_wr->wr.wr_id = (uintptr_t)&isert_cmd->tx_desc;
2829         if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2830                 rdma_wr->wr.opcode = IB_WR_RDMA_WRITE;
2831                 rdma_wr->remote_addr = isert_cmd->read_va;
2832                 rdma_wr->rkey = isert_cmd->read_stag;
2833                 rdma_wr->wr.send_flags = !isert_prot_cmd(isert_conn, se_cmd) ?
2834                                       0 : IB_SEND_SIGNALED;
2835         } else {
2836                 rdma_wr->wr.opcode = IB_WR_RDMA_READ;
2837                 rdma_wr->remote_addr = isert_cmd->write_va;
2838                 rdma_wr->rkey = isert_cmd->write_stag;
2839                 rdma_wr->wr.send_flags = IB_SEND_SIGNALED;
2840         }
2841
2842         return 0;
2843
2844 unmap_cmd:
2845         if (fr_desc) {
2846                 spin_lock_irqsave(&isert_conn->pool_lock, flags);
2847                 list_add_tail(&fr_desc->list, &isert_conn->fr_pool);
2848                 spin_unlock_irqrestore(&isert_conn->pool_lock, flags);
2849         }
2850         isert_unmap_data_buf(isert_conn, &wr->data);
2851
2852         return ret;
2853 }
2854
2855 static int
2856 isert_put_datain(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2857 {
2858         struct se_cmd *se_cmd = &cmd->se_cmd;
2859         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2860         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2861         struct isert_conn *isert_conn = conn->context;
2862         struct isert_device *device = isert_conn->device;
2863         struct ib_send_wr *wr_failed;
2864         int rc;
2865
2866         isert_dbg("Cmd: %p RDMA_WRITE data_length: %u\n",
2867                  isert_cmd, se_cmd->data_length);
2868
2869         wr->iser_ib_op = ISER_IB_RDMA_WRITE;
2870         rc = device->reg_rdma_mem(conn, cmd, wr);
2871         if (rc) {
2872                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2873                 return rc;
2874         }
2875
2876         if (!isert_prot_cmd(isert_conn, se_cmd)) {
2877                 /*
2878                  * Build isert_conn->tx_desc for iSCSI response PDU and attach
2879                  */
2880                 isert_create_send_desc(isert_conn, isert_cmd,
2881                                        &isert_cmd->tx_desc);
2882                 iscsit_build_rsp_pdu(cmd, conn, true, (struct iscsi_scsi_rsp *)
2883                                      &isert_cmd->tx_desc.iscsi_header);
2884                 isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2885                 isert_init_send_wr(isert_conn, isert_cmd,
2886                                    &isert_cmd->tx_desc.send_wr);
2887                 isert_cmd->rdma_wr.s_rdma_wr.wr.next = &isert_cmd->tx_desc.send_wr;
2888                 wr->rdma_wr_num += 1;
2889
2890                 rc = isert_post_recv(isert_conn, isert_cmd->rx_desc);
2891                 if (rc) {
2892                         isert_err("ib_post_recv failed with %d\n", rc);
2893                         return rc;
2894                 }
2895         }
2896
2897         rc = ib_post_send(isert_conn->qp, &wr->rdma_wr->wr, &wr_failed);
2898         if (rc)
2899                 isert_warn("ib_post_send() failed for IB_WR_RDMA_WRITE\n");
2900
2901         if (!isert_prot_cmd(isert_conn, se_cmd))
2902                 isert_dbg("Cmd: %p posted RDMA_WRITE + Response for iSER Data "
2903                          "READ\n", isert_cmd);
2904         else
2905                 isert_dbg("Cmd: %p posted RDMA_WRITE for iSER Data READ\n",
2906                          isert_cmd);
2907
2908         return 1;
2909 }
2910
2911 static int
2912 isert_get_dataout(struct iscsi_conn *conn, struct iscsi_cmd *cmd, bool recovery)
2913 {
2914         struct se_cmd *se_cmd = &cmd->se_cmd;
2915         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2916         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2917         struct isert_conn *isert_conn = conn->context;
2918         struct isert_device *device = isert_conn->device;
2919         struct ib_send_wr *wr_failed;
2920         int rc;
2921
2922         isert_dbg("Cmd: %p RDMA_READ data_length: %u write_data_done: %u\n",
2923                  isert_cmd, se_cmd->data_length, cmd->write_data_done);
2924         wr->iser_ib_op = ISER_IB_RDMA_READ;
2925         rc = device->reg_rdma_mem(conn, cmd, wr);
2926         if (rc) {
2927                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2928                 return rc;
2929         }
2930
2931         rc = ib_post_send(isert_conn->qp, &wr->rdma_wr->wr, &wr_failed);
2932         if (rc)
2933                 isert_warn("ib_post_send() failed for IB_WR_RDMA_READ\n");
2934
2935         isert_dbg("Cmd: %p posted RDMA_READ memory for ISER Data WRITE\n",
2936                  isert_cmd);
2937
2938         return 0;
2939 }
2940
2941 static int
2942 isert_immediate_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2943 {
2944         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2945         int ret = 0;
2946
2947         switch (state) {
2948         case ISTATE_REMOVE:
2949                 spin_lock_bh(&conn->cmd_lock);
2950                 list_del_init(&cmd->i_conn_node);
2951                 spin_unlock_bh(&conn->cmd_lock);
2952                 isert_put_cmd(isert_cmd, true);
2953                 break;
2954         case ISTATE_SEND_NOPIN_WANT_RESPONSE:
2955                 ret = isert_put_nopin(cmd, conn, false);
2956                 break;
2957         default:
2958                 isert_err("Unknown immediate state: 0x%02x\n", state);
2959                 ret = -EINVAL;
2960                 break;
2961         }
2962
2963         return ret;
2964 }
2965
2966 static int
2967 isert_response_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2968 {
2969         struct isert_conn *isert_conn = conn->context;
2970         int ret;
2971
2972         switch (state) {
2973         case ISTATE_SEND_LOGOUTRSP:
2974                 ret = isert_put_logout_rsp(cmd, conn);
2975                 if (!ret)
2976                         isert_conn->logout_posted = true;
2977                 break;
2978         case ISTATE_SEND_NOPIN:
2979                 ret = isert_put_nopin(cmd, conn, true);
2980                 break;
2981         case ISTATE_SEND_TASKMGTRSP:
2982                 ret = isert_put_tm_rsp(cmd, conn);
2983                 break;
2984         case ISTATE_SEND_REJECT:
2985                 ret = isert_put_reject(cmd, conn);
2986                 break;
2987         case ISTATE_SEND_TEXTRSP:
2988                 ret = isert_put_text_rsp(cmd, conn);
2989                 break;
2990         case ISTATE_SEND_STATUS:
2991                 /*
2992                  * Special case for sending non GOOD SCSI status from TX thread
2993                  * context during pre se_cmd excecution failure.
2994                  */
2995                 ret = isert_put_response(conn, cmd);
2996                 break;
2997         default:
2998                 isert_err("Unknown response state: 0x%02x\n", state);
2999                 ret = -EINVAL;
3000                 break;
3001         }
3002
3003         return ret;
3004 }
3005
3006 struct rdma_cm_id *
3007 isert_setup_id(struct isert_np *isert_np)
3008 {
3009         struct iscsi_np *np = isert_np->np;
3010         struct rdma_cm_id *id;
3011         struct sockaddr *sa;
3012         int ret;
3013
3014         sa = (struct sockaddr *)&np->np_sockaddr;
3015         isert_dbg("ksockaddr: %p, sa: %p\n", &np->np_sockaddr, sa);
3016
3017         id = rdma_create_id(&init_net, isert_cma_handler, isert_np,
3018                             RDMA_PS_TCP, IB_QPT_RC);
3019         if (IS_ERR(id)) {
3020                 isert_err("rdma_create_id() failed: %ld\n", PTR_ERR(id));
3021                 ret = PTR_ERR(id);
3022                 goto out;
3023         }
3024         isert_dbg("id %p context %p\n", id, id->context);
3025
3026         ret = rdma_bind_addr(id, sa);
3027         if (ret) {
3028                 isert_err("rdma_bind_addr() failed: %d\n", ret);
3029                 goto out_id;
3030         }
3031
3032         ret = rdma_listen(id, 0);
3033         if (ret) {
3034                 isert_err("rdma_listen() failed: %d\n", ret);
3035                 goto out_id;
3036         }
3037
3038         return id;
3039 out_id:
3040         rdma_destroy_id(id);
3041 out:
3042         return ERR_PTR(ret);
3043 }
3044
3045 static int
3046 isert_setup_np(struct iscsi_np *np,
3047                struct sockaddr_storage *ksockaddr)
3048 {
3049         struct isert_np *isert_np;
3050         struct rdma_cm_id *isert_lid;
3051         int ret;
3052
3053         isert_np = kzalloc(sizeof(struct isert_np), GFP_KERNEL);
3054         if (!isert_np) {
3055                 isert_err("Unable to allocate struct isert_np\n");
3056                 return -ENOMEM;
3057         }
3058         sema_init(&isert_np->sem, 0);
3059         mutex_init(&isert_np->mutex);
3060         INIT_LIST_HEAD(&isert_np->accepted);
3061         INIT_LIST_HEAD(&isert_np->pending);
3062         isert_np->np = np;
3063
3064         /*
3065          * Setup the np->np_sockaddr from the passed sockaddr setup
3066          * in iscsi_target_configfs.c code..
3067          */
3068         memcpy(&np->np_sockaddr, ksockaddr,
3069                sizeof(struct sockaddr_storage));
3070
3071         isert_lid = isert_setup_id(isert_np);
3072         if (IS_ERR(isert_lid)) {
3073                 ret = PTR_ERR(isert_lid);
3074                 goto out;
3075         }
3076
3077         isert_np->cm_id = isert_lid;
3078         np->np_context = isert_np;
3079
3080         return 0;
3081
3082 out:
3083         kfree(isert_np);
3084
3085         return ret;
3086 }
3087
3088 static int
3089 isert_rdma_accept(struct isert_conn *isert_conn)
3090 {
3091         struct rdma_cm_id *cm_id = isert_conn->cm_id;
3092         struct rdma_conn_param cp;
3093         int ret;
3094
3095         memset(&cp, 0, sizeof(struct rdma_conn_param));
3096         cp.initiator_depth = isert_conn->initiator_depth;
3097         cp.retry_count = 7;
3098         cp.rnr_retry_count = 7;
3099
3100         ret = rdma_accept(cm_id, &cp);
3101         if (ret) {
3102                 isert_err("rdma_accept() failed with: %d\n", ret);
3103                 return ret;
3104         }
3105
3106         return 0;
3107 }
3108
3109 static int
3110 isert_get_login_rx(struct iscsi_conn *conn, struct iscsi_login *login)
3111 {
3112         struct isert_conn *isert_conn = conn->context;
3113         int ret;
3114
3115         isert_info("before login_req comp conn: %p\n", isert_conn);
3116         ret = wait_for_completion_interruptible(&isert_conn->login_req_comp);
3117         if (ret) {
3118                 isert_err("isert_conn %p interrupted before got login req\n",
3119                           isert_conn);
3120                 return ret;
3121         }
3122         reinit_completion(&isert_conn->login_req_comp);
3123
3124         /*
3125          * For login requests after the first PDU, isert_rx_login_req() will
3126          * kick schedule_delayed_work(&conn->login_work) as the packet is
3127          * received, which turns this callback from iscsi_target_do_login_rx()
3128          * into a NOP.
3129          */
3130         if (!login->first_request)
3131                 return 0;
3132
3133         isert_rx_login_req(isert_conn);
3134
3135         isert_info("before login_comp conn: %p\n", conn);
3136         ret = wait_for_completion_interruptible(&isert_conn->login_comp);
3137         if (ret)
3138                 return ret;
3139
3140         isert_info("processing login->req: %p\n", login->req);
3141
3142         return 0;
3143 }
3144
3145 static void
3146 isert_set_conn_info(struct iscsi_np *np, struct iscsi_conn *conn,
3147                     struct isert_conn *isert_conn)
3148 {
3149         struct rdma_cm_id *cm_id = isert_conn->cm_id;
3150         struct rdma_route *cm_route = &cm_id->route;
3151
3152         conn->login_family = np->np_sockaddr.ss_family;
3153
3154         conn->login_sockaddr = cm_route->addr.dst_addr;
3155         conn->local_sockaddr = cm_route->addr.src_addr;
3156 }
3157
3158 static int
3159 isert_accept_np(struct iscsi_np *np, struct iscsi_conn *conn)
3160 {
3161         struct isert_np *isert_np = np->np_context;
3162         struct isert_conn *isert_conn;
3163         int ret;
3164
3165 accept_wait:
3166         ret = down_interruptible(&isert_np->sem);
3167         if (ret)
3168                 return -ENODEV;
3169
3170         spin_lock_bh(&np->np_thread_lock);
3171         if (np->np_thread_state >= ISCSI_NP_THREAD_RESET) {
3172                 spin_unlock_bh(&np->np_thread_lock);
3173                 isert_dbg("np_thread_state %d\n",
3174                          np->np_thread_state);
3175                 /**
3176                  * No point in stalling here when np_thread
3177                  * is in state RESET/SHUTDOWN/EXIT - bail
3178                  **/
3179                 return -ENODEV;
3180         }
3181         spin_unlock_bh(&np->np_thread_lock);
3182
3183         mutex_lock(&isert_np->mutex);
3184         if (list_empty(&isert_np->pending)) {
3185                 mutex_unlock(&isert_np->mutex);
3186                 goto accept_wait;
3187         }
3188         isert_conn = list_first_entry(&isert_np->pending,
3189                         struct isert_conn, node);
3190         list_del_init(&isert_conn->node);
3191         mutex_unlock(&isert_np->mutex);
3192
3193         conn->context = isert_conn;
3194         isert_conn->conn = conn;
3195         isert_conn->state = ISER_CONN_BOUND;
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_wait4flush(struct isert_conn *isert_conn)
3282 {
3283         struct ib_recv_wr *bad_wr;
3284
3285         isert_info("conn %p\n", isert_conn);
3286
3287         init_completion(&isert_conn->wait_comp_err);
3288         isert_conn->beacon.wr_id = ISER_BEACON_WRID;
3289         /* post an indication that all flush errors were consumed */
3290         if (ib_post_recv(isert_conn->qp, &isert_conn->beacon, &bad_wr)) {
3291                 isert_err("conn %p failed to post beacon", isert_conn);
3292                 return;
3293         }
3294
3295         wait_for_completion(&isert_conn->wait_comp_err);
3296 }
3297
3298 static void
3299 isert_wait4cmds(struct iscsi_conn *conn)
3300 {
3301         isert_info("iscsi_conn %p\n", conn);
3302
3303         if (conn->sess) {
3304                 target_sess_cmd_list_set_waiting(conn->sess->se_sess);
3305                 target_wait_for_sess_cmds(conn->sess->se_sess);
3306         }
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);