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 "ib_isert.h"
  33
  34 #define ISERT_MAX_CONN          8
  35 #define ISER_MAX_RX_CQ_LEN      (ISERT_QP_MAX_RECV_DTOS * ISERT_MAX_CONN)
  36 #define ISER_MAX_TX_CQ_LEN      (ISERT_QP_MAX_REQ_DTOS  * ISERT_MAX_CONN)
  37 #define ISER_MAX_CQ_LEN         (ISER_MAX_RX_CQ_LEN + ISER_MAX_TX_CQ_LEN + \
  38                                  ISERT_MAX_CONN)
  39
  40 static int isert_debug_level;
  41 module_param_named(debug_level, isert_debug_level, int, 0644);
  42 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0 (default:0)");
  43
  44 static DEFINE_MUTEX(device_list_mutex);
  45 static LIST_HEAD(device_list);
  46 static struct workqueue_struct *isert_comp_wq;
  47 static struct workqueue_struct *isert_release_wq;
  48
  49 static void
  50 isert_unmap_cmd(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn);
  51 static int
  52 isert_map_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
  53                struct isert_rdma_wr *wr);
  54 static void
  55 isert_unreg_rdma(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn);
  56 static int
  57 isert_reg_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
  58                struct isert_rdma_wr *wr);
  59 static int
  60 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd);
  61 static int
  62 isert_rdma_post_recvl(struct isert_conn *isert_conn);
  63 static int
  64 isert_rdma_accept(struct isert_conn *isert_conn);
  65 struct rdma_cm_id *isert_setup_id(struct isert_np *isert_np);
  66
  67 static void isert_release_work(struct work_struct *work);
  68
  69 static inline bool
  70 isert_prot_cmd(struct isert_conn *conn, struct se_cmd *cmd)
  71 {
  72         return (conn->pi_support &&
  73                 cmd->prot_op != TARGET_PROT_NORMAL);
  74 }
  75
  76
  77 static void
  78 isert_qp_event_callback(struct ib_event *e, void *context)
  79 {
  80         struct isert_conn *isert_conn = context;
  81
  82         isert_err("%s (%d): conn %p\n",
  83                   ib_event_msg(e->event), e->event, isert_conn);
  84
  85         switch (e->event) {
  86         case IB_EVENT_COMM_EST:
  87                 rdma_notify(isert_conn->cm_id, IB_EVENT_COMM_EST);
  88                 break;
  89         case IB_EVENT_QP_LAST_WQE_REACHED:
  90                 isert_warn("Reached TX IB_EVENT_QP_LAST_WQE_REACHED\n");
  91                 break;
  92         default:
  93                 break;
  94         }
  95 }
  96
  97 static struct isert_comp *
  98 isert_comp_get(struct isert_conn *isert_conn)
  99 {
 100         struct isert_device *device = isert_conn->device;
 101         struct isert_comp *comp;
 102         int i, min = 0;
 103
 104         mutex_lock(&device_list_mutex);
 105         for (i = 0; i < device->comps_used; i++)
 106                 if (device->comps[i].active_qps <
 107                     device->comps[min].active_qps)
 108                         min = i;
 109         comp = &device->comps[min];
 110         comp->active_qps++;
 111         mutex_unlock(&device_list_mutex);
 112
 113         isert_info("conn %p, using comp %p min_index: %d\n",
 114                    isert_conn, comp, min);
 115
 116         return comp;
 117 }
 118
 119 static void
 120 isert_comp_put(struct isert_comp *comp)
 121 {
 122         mutex_lock(&device_list_mutex);
 123         comp->active_qps--;
 124         mutex_unlock(&device_list_mutex);
 125 }
 126
 127 static struct ib_qp *
 128 isert_create_qp(struct isert_conn *isert_conn,
 129                 struct isert_comp *comp,
 130                 struct rdma_cm_id *cma_id)
 131 {
 132         struct isert_device *device = isert_conn->device;
 133         struct ib_qp_init_attr attr;
 134         int ret;
 135
 136         memset(&attr, 0, sizeof(struct ib_qp_init_attr));
 137         attr.event_handler = isert_qp_event_callback;
 138         attr.qp_context = isert_conn;
 139         attr.send_cq = comp->cq;
 140         attr.recv_cq = comp->cq;
 141         attr.cap.max_send_wr = ISERT_QP_MAX_REQ_DTOS;
 142         attr.cap.max_recv_wr = ISERT_QP_MAX_RECV_DTOS + 1;
 143         attr.cap.max_send_sge = device->ib_device->attrs.max_sge;
 144         isert_conn->max_sge = min(device->ib_device->attrs.max_sge,
 145                                   device->ib_device->attrs.max_sge_rd);
 146         attr.cap.max_recv_sge = 1;
 147         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 148         attr.qp_type = IB_QPT_RC;
 149         if (device->pi_capable)
 150                 attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
 151
 152         ret = rdma_create_qp(cma_id, device->pd, &attr);
 153         if (ret) {
 154                 isert_err("rdma_create_qp failed for cma_id %d\n", ret);
 155                 return ERR_PTR(ret);
 156         }
 157
 158         return cma_id->qp;
 159 }
 160
 161 static int
 162 isert_conn_setup_qp(struct isert_conn *isert_conn, struct rdma_cm_id *cma_id)
 163 {
 164         struct isert_comp *comp;
 165         int ret;
 166
 167         comp = isert_comp_get(isert_conn);
 168         isert_conn->qp = isert_create_qp(isert_conn, comp, cma_id);
 169         if (IS_ERR(isert_conn->qp)) {
 170                 ret = PTR_ERR(isert_conn->qp);
 171                 goto err;
 172         }
 173
 174         return 0;
 175 err:
 176         isert_comp_put(comp);
 177         return ret;
 178 }
 179
 180 static void
 181 isert_cq_event_callback(struct ib_event *e, void *context)
 182 {
 183         isert_dbg("event: %d\n", e->event);
 184 }
 185
 186 static int
 187 isert_alloc_rx_descriptors(struct isert_conn *isert_conn)
 188 {
 189         struct isert_device *device = isert_conn->device;
 190         struct ib_device *ib_dev = device->ib_device;
 191         struct iser_rx_desc *rx_desc;
 192         struct ib_sge *rx_sg;
 193         u64 dma_addr;
 194         int i, j;
 195
 196         isert_conn->rx_descs = kzalloc(ISERT_QP_MAX_RECV_DTOS *
 197                                 sizeof(struct iser_rx_desc), GFP_KERNEL);
 198         if (!isert_conn->rx_descs)
 199                 goto fail;
 200
 201         rx_desc = isert_conn->rx_descs;
 202
 203         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 204                 dma_addr = ib_dma_map_single(ib_dev, (void *)rx_desc,
 205                                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 206                 if (ib_dma_mapping_error(ib_dev, dma_addr))
 207                         goto dma_map_fail;
 208
 209                 rx_desc->dma_addr = dma_addr;
 210
 211                 rx_sg = &rx_desc->rx_sg;
 212                 rx_sg->addr = rx_desc->dma_addr;
 213                 rx_sg->length = ISER_RX_PAYLOAD_SIZE;
 214                 rx_sg->lkey = device->pd->local_dma_lkey;
 215         }
 216
 217         return 0;
 218
 219 dma_map_fail:
 220         rx_desc = isert_conn->rx_descs;
 221         for (j = 0; j < i; j++, rx_desc++) {
 222                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 223                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 224         }
 225         kfree(isert_conn->rx_descs);
 226         isert_conn->rx_descs = NULL;
 227 fail:
 228         isert_err("conn %p failed to allocate rx descriptors\n", isert_conn);
 229
 230         return -ENOMEM;
 231 }
 232
 233 static void
 234 isert_free_rx_descriptors(struct isert_conn *isert_conn)
 235 {
 236         struct ib_device *ib_dev = isert_conn->device->ib_device;
 237         struct iser_rx_desc *rx_desc;
 238         int i;
 239
 240         if (!isert_conn->rx_descs)
 241                 return;
 242
 243         rx_desc = isert_conn->rx_descs;
 244         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 245                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 246                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 247         }
 248
 249         kfree(isert_conn->rx_descs);
 250         isert_conn->rx_descs = NULL;
 251 }
 252
 253 static void isert_cq_work(struct work_struct *);
 254 static void isert_cq_callback(struct ib_cq *, void *);
 255
 256 static void
 257 isert_free_comps(struct isert_device *device)
 258 {
 259         int i;
 260
 261         for (i = 0; i < device->comps_used; i++) {
 262                 struct isert_comp *comp = &device->comps[i];
 263
 264                 if (comp->cq) {
 265                         cancel_work_sync(&comp->work);
 266                         ib_destroy_cq(comp->cq);
 267                 }
 268         }
 269         kfree(device->comps);
 270 }
 271
 272 static int
 273 isert_alloc_comps(struct isert_device *device)
 274 {
 275         int i, max_cqe, ret = 0;
 276
 277         device->comps_used = min(ISERT_MAX_CQ, min_t(int, num_online_cpus(),
 278                                  device->ib_device->num_comp_vectors));
 279
 280         isert_info("Using %d CQs, %s supports %d vectors support "
 281                    "Fast registration %d pi_capable %d\n",
 282                    device->comps_used, device->ib_device->name,
 283                    device->ib_device->num_comp_vectors, device->use_fastreg,
 284                    device->pi_capable);
 285
 286         device->comps = kcalloc(device->comps_used, sizeof(struct isert_comp),
 287                                 GFP_KERNEL);
 288         if (!device->comps) {
 289                 isert_err("Unable to allocate completion contexts\n");
 290                 return -ENOMEM;
 291         }
 292
 293         max_cqe = min(ISER_MAX_CQ_LEN, device->ib_device->attrs.max_cqe);
 294
 295         for (i = 0; i < device->comps_used; i++) {
 296                 struct ib_cq_init_attr cq_attr = {};
 297                 struct isert_comp *comp = &device->comps[i];
 298
 299                 comp->device = device;
 300                 INIT_WORK(&comp->work, isert_cq_work);
 301                 cq_attr.cqe = max_cqe;
 302                 cq_attr.comp_vector = i;
 303                 comp->cq = ib_create_cq(device->ib_device,
 304                                         isert_cq_callback,
 305                                         isert_cq_event_callback,
 306                                         (void *)comp,
 307                                         &cq_attr);
 308                 if (IS_ERR(comp->cq)) {
 309                         isert_err("Unable to allocate cq\n");
 310                         ret = PTR_ERR(comp->cq);
 311                         comp->cq = NULL;
 312                         goto out_cq;
 313                 }
 314
 315                 ret = ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP);
 316                 if (ret)
 317                         goto out_cq;
 318         }
 319
 320         return 0;
 321 out_cq:
 322         isert_free_comps(device);
 323         return ret;
 324 }
 325
 326 static int
 327 isert_create_device_ib_res(struct isert_device *device)
 328 {
 329         struct ib_device *ib_dev = device->ib_device;
 330         int ret;
 331
 332         isert_dbg("devattr->max_sge: %d\n", ib_dev->attrs.max_sge);
 333         isert_dbg("devattr->max_sge_rd: %d\n", ib_dev->attrs.max_sge_rd);
 334
 335         /* asign function handlers */
 336         if (ib_dev->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS &&
 337             ib_dev->attrs.device_cap_flags & IB_DEVICE_SIGNATURE_HANDOVER) {
 338                 device->use_fastreg = 1;
 339                 device->reg_rdma_mem = isert_reg_rdma;
 340                 device->unreg_rdma_mem = isert_unreg_rdma;
 341         } else {
 342                 device->use_fastreg = 0;
 343                 device->reg_rdma_mem = isert_map_rdma;
 344                 device->unreg_rdma_mem = isert_unmap_cmd;
 345         }
 346
 347         ret = isert_alloc_comps(device);
 348         if (ret)
 349                 return ret;
 350
 351         device->pd = ib_alloc_pd(ib_dev);
 352         if (IS_ERR(device->pd)) {
 353                 ret = PTR_ERR(device->pd);
 354                 isert_err("failed to allocate pd, device %p, ret=%d\n",
 355                           device, ret);
 356                 goto out_cq;
 357         }
 358
 359         /* Check signature cap */
 360         device->pi_capable = ib_dev->attrs.device_cap_flags &
 361                              IB_DEVICE_SIGNATURE_HANDOVER ? true : false;
 362
 363         return 0;
 364
 365 out_cq:
 366         isert_free_comps(device);
 367         return ret;
 368 }
 369
 370 static void
 371 isert_free_device_ib_res(struct isert_device *device)
 372 {
 373         isert_info("device %p\n", device);
 374
 375         ib_dealloc_pd(device->pd);
 376         isert_free_comps(device);
 377 }
 378
 379 static void
 380 isert_device_put(struct isert_device *device)
 381 {
 382         mutex_lock(&device_list_mutex);
 383         device->refcount--;
 384         isert_info("device %p refcount %d\n", device, device->refcount);
 385         if (!device->refcount) {
 386                 isert_free_device_ib_res(device);
 387                 list_del(&device->dev_node);
 388                 kfree(device);
 389         }
 390         mutex_unlock(&device_list_mutex);
 391 }
 392
 393 static struct isert_device *
 394 isert_device_get(struct rdma_cm_id *cma_id)
 395 {
 396         struct isert_device *device;
 397         int ret;
 398
 399         mutex_lock(&device_list_mutex);
 400         list_for_each_entry(device, &device_list, dev_node) {
 401                 if (device->ib_device->node_guid == cma_id->device->node_guid) {
 402                         device->refcount++;
 403                         isert_info("Found iser device %p refcount %d\n",
 404                                    device, device->refcount);
 405                         mutex_unlock(&device_list_mutex);
 406                         return device;
 407                 }
 408         }
 409
 410         device = kzalloc(sizeof(struct isert_device), GFP_KERNEL);
 411         if (!device) {
 412                 mutex_unlock(&device_list_mutex);
 413                 return ERR_PTR(-ENOMEM);
 414         }
 415
 416         INIT_LIST_HEAD(&device->dev_node);
 417
 418         device->ib_device = cma_id->device;
 419         ret = isert_create_device_ib_res(device);
 420         if (ret) {
 421                 kfree(device);
 422                 mutex_unlock(&device_list_mutex);
 423                 return ERR_PTR(ret);
 424         }
 425
 426         device->refcount++;
 427         list_add_tail(&device->dev_node, &device_list);
 428         isert_info("Created a new iser device %p refcount %d\n",
 429                    device, device->refcount);
 430         mutex_unlock(&device_list_mutex);
 431
 432         return device;
 433 }
 434
 435 static void
 436 isert_conn_free_fastreg_pool(struct isert_conn *isert_conn)
 437 {
 438         struct fast_reg_descriptor *fr_desc, *tmp;
 439         int i = 0;
 440
 441         if (list_empty(&isert_conn->fr_pool))
 442                 return;
 443
 444         isert_info("Freeing conn %p fastreg pool", isert_conn);
 445
 446         list_for_each_entry_safe(fr_desc, tmp,
 447                                  &isert_conn->fr_pool, list) {
 448                 list_del(&fr_desc->list);
 449                 ib_dereg_mr(fr_desc->data_mr);
 450                 if (fr_desc->pi_ctx) {
 451                         ib_dereg_mr(fr_desc->pi_ctx->prot_mr);
 452                         ib_dereg_mr(fr_desc->pi_ctx->sig_mr);
 453                         kfree(fr_desc->pi_ctx);
 454                 }
 455                 kfree(fr_desc);
 456                 ++i;
 457         }
 458
 459         if (i < isert_conn->fr_pool_size)
 460                 isert_warn("Pool still has %d regions registered\n",
 461                         isert_conn->fr_pool_size - i);
 462 }
 463
 464 static int
 465 isert_create_pi_ctx(struct fast_reg_descriptor *desc,
 466                     struct ib_device *device,
 467                     struct ib_pd *pd)
 468 {
 469         struct pi_context *pi_ctx;
 470         int ret;
 471
 472         pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
 473         if (!pi_ctx) {
 474                 isert_err("Failed to allocate pi context\n");
 475                 return -ENOMEM;
 476         }
 477
 478         pi_ctx->prot_mr = ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG,
 479                                       ISCSI_ISER_SG_TABLESIZE);
 480         if (IS_ERR(pi_ctx->prot_mr)) {
 481                 isert_err("Failed to allocate prot frmr err=%ld\n",
 482                           PTR_ERR(pi_ctx->prot_mr));
 483                 ret = PTR_ERR(pi_ctx->prot_mr);
 484                 goto err_pi_ctx;
 485         }
 486         desc->ind |= ISERT_PROT_KEY_VALID;
 487
 488         pi_ctx->sig_mr = ib_alloc_mr(pd, IB_MR_TYPE_SIGNATURE, 2);
 489         if (IS_ERR(pi_ctx->sig_mr)) {
 490                 isert_err("Failed to allocate signature enabled mr err=%ld\n",
 491                           PTR_ERR(pi_ctx->sig_mr));
 492                 ret = PTR_ERR(pi_ctx->sig_mr);
 493                 goto err_prot_mr;
 494         }
 495
 496         desc->pi_ctx = pi_ctx;
 497         desc->ind |= ISERT_SIG_KEY_VALID;
 498         desc->ind &= ~ISERT_PROTECTED;
 499
 500         return 0;
 501
 502 err_prot_mr:
 503         ib_dereg_mr(pi_ctx->prot_mr);
 504 err_pi_ctx:
 505         kfree(pi_ctx);
 506
 507         return ret;
 508 }
 509
 510 static int
 511 isert_create_fr_desc(struct ib_device *ib_device, struct ib_pd *pd,
 512                      struct fast_reg_descriptor *fr_desc)
 513 {
 514         fr_desc->data_mr = ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG,
 515                                        ISCSI_ISER_SG_TABLESIZE);
 516         if (IS_ERR(fr_desc->data_mr)) {
 517                 isert_err("Failed to allocate data frmr err=%ld\n",
 518                           PTR_ERR(fr_desc->data_mr));
 519                 return PTR_ERR(fr_desc->data_mr);
 520         }
 521         fr_desc->ind |= ISERT_DATA_KEY_VALID;
 522
 523         isert_dbg("Created fr_desc %p\n", fr_desc);
 524
 525         return 0;
 526 }
 527
 528 static int
 529 isert_conn_create_fastreg_pool(struct isert_conn *isert_conn)
 530 {
 531         struct fast_reg_descriptor *fr_desc;
 532         struct isert_device *device = isert_conn->device;
 533         struct se_session *se_sess = isert_conn->conn->sess->se_sess;
 534         struct se_node_acl *se_nacl = se_sess->se_node_acl;
 535         int i, ret, tag_num;
 536         /*
 537          * Setup the number of FRMRs based upon the number of tags
 538          * available to session in iscsi_target_locate_portal().
 539          */
 540         tag_num = max_t(u32, ISCSIT_MIN_TAGS, se_nacl->queue_depth);
 541         tag_num = (tag_num * 2) + ISCSIT_EXTRA_TAGS;
 542
 543         isert_conn->fr_pool_size = 0;
 544         for (i = 0; i < tag_num; i++) {
 545                 fr_desc = kzalloc(sizeof(*fr_desc), GFP_KERNEL);
 546                 if (!fr_desc) {
 547                         isert_err("Failed to allocate fast_reg descriptor\n");
 548                         ret = -ENOMEM;
 549                         goto err;
 550                 }
 551
 552                 ret = isert_create_fr_desc(device->ib_device,
 553                                            device->pd, fr_desc);
 554                 if (ret) {
 555                         isert_err("Failed to create fastreg descriptor err=%d\n",
 556                                ret);
 557                         kfree(fr_desc);
 558                         goto err;
 559                 }
 560
 561                 list_add_tail(&fr_desc->list, &isert_conn->fr_pool);
 562                 isert_conn->fr_pool_size++;
 563         }
 564
 565         isert_dbg("Creating conn %p fastreg pool size=%d",
 566                  isert_conn, isert_conn->fr_pool_size);
 567
 568         return 0;
 569
 570 err:
 571         isert_conn_free_fastreg_pool(isert_conn);
 572         return ret;
 573 }
 574
 575 static void
 576 isert_init_conn(struct isert_conn *isert_conn)
 577 {
 578         isert_conn->state = ISER_CONN_INIT;
 579         INIT_LIST_HEAD(&isert_conn->node);
 580         init_completion(&isert_conn->login_comp);
 581         init_completion(&isert_conn->login_req_comp);
 582         init_completion(&isert_conn->wait);
 583         kref_init(&isert_conn->kref);
 584         mutex_init(&isert_conn->mutex);
 585         spin_lock_init(&isert_conn->pool_lock);
 586         INIT_LIST_HEAD(&isert_conn->fr_pool);
 587         INIT_WORK(&isert_conn->release_work, isert_release_work);
 588 }
 589
 590 static void
 591 isert_free_login_buf(struct isert_conn *isert_conn)
 592 {
 593         struct ib_device *ib_dev = isert_conn->device->ib_device;
 594
 595         ib_dma_unmap_single(ib_dev, isert_conn->login_rsp_dma,
 596                             ISER_RX_LOGIN_SIZE, DMA_TO_DEVICE);
 597         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 598                             ISCSI_DEF_MAX_RECV_SEG_LEN,
 599                             DMA_FROM_DEVICE);
 600         kfree(isert_conn->login_buf);
 601 }
 602
 603 static int
 604 isert_alloc_login_buf(struct isert_conn *isert_conn,
 605                       struct ib_device *ib_dev)
 606 {
 607         int ret;
 608
 609         isert_conn->login_buf = kzalloc(ISCSI_DEF_MAX_RECV_SEG_LEN +
 610                                         ISER_RX_LOGIN_SIZE, GFP_KERNEL);
 611         if (!isert_conn->login_buf) {
 612                 isert_err("Unable to allocate isert_conn->login_buf\n");
 613                 return -ENOMEM;
 614         }
 615
 616         isert_conn->login_req_buf = isert_conn->login_buf;
 617         isert_conn->login_rsp_buf = isert_conn->login_buf +
 618                                     ISCSI_DEF_MAX_RECV_SEG_LEN;
 619
 620         isert_dbg("Set login_buf: %p login_req_buf: %p login_rsp_buf: %p\n",
 621                  isert_conn->login_buf, isert_conn->login_req_buf,
 622                  isert_conn->login_rsp_buf);
 623
 624         isert_conn->login_req_dma = ib_dma_map_single(ib_dev,
 625                                 (void *)isert_conn->login_req_buf,
 626                                 ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_FROM_DEVICE);
 627
 628         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_req_dma);
 629         if (ret) {
 630                 isert_err("login_req_dma mapping error: %d\n", ret);
 631                 isert_conn->login_req_dma = 0;
 632                 goto out_login_buf;
 633         }
 634
 635         isert_conn->login_rsp_dma = ib_dma_map_single(ib_dev,
 636                                         (void *)isert_conn->login_rsp_buf,
 637                                         ISER_RX_LOGIN_SIZE, DMA_TO_DEVICE);
 638
 639         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_rsp_dma);
 640         if (ret) {
 641                 isert_err("login_rsp_dma mapping error: %d\n", ret);
 642                 isert_conn->login_rsp_dma = 0;
 643                 goto out_req_dma_map;
 644         }
 645
 646         return 0;
 647
 648 out_req_dma_map:
 649         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 650                             ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_FROM_DEVICE);
 651 out_login_buf:
 652         kfree(isert_conn->login_buf);
 653         return ret;
 654 }
 655
 656 static void
 657 isert_set_nego_params(struct isert_conn *isert_conn,
 658                       struct rdma_conn_param *param)
 659 {
 660         struct ib_device_attr *attr = &isert_conn->device->ib_device->attrs;
 661
 662         /* Set max inflight RDMA READ requests */
 663         isert_conn->initiator_depth = min_t(u8, param->initiator_depth,
 664                                 attr->max_qp_init_rd_atom);
 665         isert_dbg("Using initiator_depth: %u\n", isert_conn->initiator_depth);
 666
 667         if (param->private_data) {
 668                 u8 flags = *(u8 *)param->private_data;
 669
 670                 /*
 671                  * use remote invalidation if the both initiator
 672                  * and the HCA support it
 673                  */
 674                 isert_conn->snd_w_inv = !(flags & ISER_SEND_W_INV_NOT_SUP) &&
 675                                           (attr->device_cap_flags &
 676                                            IB_DEVICE_MEM_MGT_EXTENSIONS);
 677                 if (isert_conn->snd_w_inv)
 678                         isert_info("Using remote invalidation\n");
 679         }
 680 }
 681
 682 static int
 683 isert_connect_request(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 684 {
 685         struct isert_np *isert_np = cma_id->context;
 686         struct iscsi_np *np = isert_np->np;
 687         struct isert_conn *isert_conn;
 688         struct isert_device *device;
 689         int ret = 0;
 690
 691         spin_lock_bh(&np->np_thread_lock);
 692         if (!np->enabled) {
 693                 spin_unlock_bh(&np->np_thread_lock);
 694                 isert_dbg("iscsi_np is not enabled, reject connect request\n");
 695                 return rdma_reject(cma_id, NULL, 0);
 696         }
 697         spin_unlock_bh(&np->np_thread_lock);
 698
 699         isert_dbg("cma_id: %p, portal: %p\n",
 700                  cma_id, cma_id->context);
 701
 702         isert_conn = kzalloc(sizeof(struct isert_conn), GFP_KERNEL);
 703         if (!isert_conn)
 704                 return -ENOMEM;
 705
 706         isert_init_conn(isert_conn);
 707         isert_conn->cm_id = cma_id;
 708
 709         ret = isert_alloc_login_buf(isert_conn, cma_id->device);
 710         if (ret)
 711                 goto out;
 712
 713         device = isert_device_get(cma_id);
 714         if (IS_ERR(device)) {
 715                 ret = PTR_ERR(device);
 716                 goto out_rsp_dma_map;
 717         }
 718         isert_conn->device = device;
 719
 720         isert_set_nego_params(isert_conn, &event->param.conn);
 721
 722         ret = isert_conn_setup_qp(isert_conn, cma_id);
 723         if (ret)
 724                 goto out_conn_dev;
 725
 726         ret = isert_rdma_post_recvl(isert_conn);
 727         if (ret)
 728                 goto out_conn_dev;
 729
 730         ret = isert_rdma_accept(isert_conn);
 731         if (ret)
 732                 goto out_conn_dev;
 733
 734         mutex_lock(&isert_np->mutex);
 735         list_add_tail(&isert_conn->node, &isert_np->accepted);
 736         mutex_unlock(&isert_np->mutex);
 737
 738         return 0;
 739
 740 out_conn_dev:
 741         isert_device_put(device);
 742 out_rsp_dma_map:
 743         isert_free_login_buf(isert_conn);
 744 out:
 745         kfree(isert_conn);
 746         rdma_reject(cma_id, NULL, 0);
 747         return ret;
 748 }
 749
 750 static void
 751 isert_connect_release(struct isert_conn *isert_conn)
 752 {
 753         struct isert_device *device = isert_conn->device;
 754
 755         isert_dbg("conn %p\n", isert_conn);
 756
 757         BUG_ON(!device);
 758
 759         if (device->use_fastreg)
 760                 isert_conn_free_fastreg_pool(isert_conn);
 761
 762         isert_free_rx_descriptors(isert_conn);
 763         if (isert_conn->cm_id)
 764                 rdma_destroy_id(isert_conn->cm_id);
 765
 766         if (isert_conn->qp) {
 767                 struct isert_comp *comp = isert_conn->qp->recv_cq->cq_context;
 768
 769                 isert_comp_put(comp);
 770                 ib_destroy_qp(isert_conn->qp);
 771         }
 772
 773         if (isert_conn->login_buf)
 774                 isert_free_login_buf(isert_conn);
 775
 776         isert_device_put(device);
 777
 778         kfree(isert_conn);
 779 }
 780
 781 static void
 782 isert_connected_handler(struct rdma_cm_id *cma_id)
 783 {
 784         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 785         struct isert_np *isert_np = cma_id->context;
 786
 787         isert_info("conn %p\n", isert_conn);
 788
 789         mutex_lock(&isert_conn->mutex);
 790         isert_conn->state = ISER_CONN_UP;
 791         kref_get(&isert_conn->kref);
 792         mutex_unlock(&isert_conn->mutex);
 793
 794         mutex_lock(&isert_np->mutex);
 795         list_move_tail(&isert_conn->node, &isert_np->pending);
 796         mutex_unlock(&isert_np->mutex);
 797
 798         isert_info("np %p: Allow accept_np to continue\n", isert_np);
 799         up(&isert_np->sem);
 800 }
 801
 802 static void
 803 isert_release_kref(struct kref *kref)
 804 {
 805         struct isert_conn *isert_conn = container_of(kref,
 806                                 struct isert_conn, kref);
 807
 808         isert_info("conn %p final kref %s/%d\n", isert_conn, current->comm,
 809                    current->pid);
 810
 811         isert_connect_release(isert_conn);
 812 }
 813
 814 static void
 815 isert_put_conn(struct isert_conn *isert_conn)
 816 {
 817         kref_put(&isert_conn->kref, isert_release_kref);
 818 }
 819
 820 /**
 821  * isert_conn_terminate() - Initiate connection termination
 822  * @isert_conn: isert connection struct
 823  *
 824  * Notes:
 825  * In case the connection state is FULL_FEATURE, move state
 826  * to TEMINATING and start teardown sequence (rdma_disconnect).
 827  * In case the connection state is UP, complete flush as well.
 828  *
 829  * This routine must be called with mutex held. Thus it is
 830  * safe to call multiple times.
 831  */
 832 static void
 833 isert_conn_terminate(struct isert_conn *isert_conn)
 834 {
 835         int err;
 836
 837         switch (isert_conn->state) {
 838         case ISER_CONN_TERMINATING:
 839                 break;
 840         case ISER_CONN_UP:
 841         case ISER_CONN_FULL_FEATURE: /* FALLTHRU */
 842                 isert_info("Terminating conn %p state %d\n",
 843                            isert_conn, isert_conn->state);
 844                 isert_conn->state = ISER_CONN_TERMINATING;
 845                 err = rdma_disconnect(isert_conn->cm_id);
 846                 if (err)
 847                         isert_warn("Failed rdma_disconnect isert_conn %p\n",
 848                                    isert_conn);
 849                 break;
 850         default:
 851                 isert_warn("conn %p teminating in state %d\n",
 852                            isert_conn, isert_conn->state);
 853         }
 854 }
 855
 856 static int
 857 isert_np_cma_handler(struct isert_np *isert_np,
 858                      enum rdma_cm_event_type event)
 859 {
 860         isert_dbg("%s (%d): isert np %p\n",
 861                   rdma_event_msg(event), event, isert_np);
 862
 863         switch (event) {
 864         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 865                 isert_np->cm_id = NULL;
 866                 break;
 867         case RDMA_CM_EVENT_ADDR_CHANGE:
 868                 isert_np->cm_id = isert_setup_id(isert_np);
 869                 if (IS_ERR(isert_np->cm_id)) {
 870                         isert_err("isert np %p setup id failed: %ld\n",
 871                                   isert_np, PTR_ERR(isert_np->cm_id));
 872                         isert_np->cm_id = NULL;
 873                 }
 874                 break;
 875         default:
 876                 isert_err("isert np %p Unexpected event %d\n",
 877                           isert_np, event);
 878         }
 879
 880         return -1;
 881 }
 882
 883 static int
 884 isert_disconnected_handler(struct rdma_cm_id *cma_id,
 885                            enum rdma_cm_event_type event)
 886 {
 887         struct isert_np *isert_np = cma_id->context;
 888         struct isert_conn *isert_conn;
 889         bool terminating = false;
 890
 891         if (isert_np->cm_id == cma_id)
 892                 return isert_np_cma_handler(cma_id->context, event);
 893
 894         isert_conn = cma_id->qp->qp_context;
 895
 896         mutex_lock(&isert_conn->mutex);
 897         terminating = (isert_conn->state == ISER_CONN_TERMINATING);
 898         isert_conn_terminate(isert_conn);
 899         mutex_unlock(&isert_conn->mutex);
 900
 901         isert_info("conn %p completing wait\n", isert_conn);
 902         complete(&isert_conn->wait);
 903
 904         if (terminating)
 905                 goto out;
 906
 907         mutex_lock(&isert_np->mutex);
 908         if (!list_empty(&isert_conn->node)) {
 909                 list_del_init(&isert_conn->node);
 910                 isert_put_conn(isert_conn);
 911                 queue_work(isert_release_wq, &isert_conn->release_work);
 912         }
 913         mutex_unlock(&isert_np->mutex);
 914
 915 out:
 916         return 0;
 917 }
 918
 919 static int
 920 isert_connect_error(struct rdma_cm_id *cma_id)
 921 {
 922         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 923
 924         list_del_init(&isert_conn->node);
 925         isert_conn->cm_id = NULL;
 926         isert_put_conn(isert_conn);
 927
 928         return -1;
 929 }
 930
 931 static int
 932 isert_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 933 {
 934         int ret = 0;
 935
 936         isert_info("%s (%d): status %d id %p np %p\n",
 937                    rdma_event_msg(event->event), event->event,
 938                    event->status, cma_id, cma_id->context);
 939
 940         switch (event->event) {
 941         case RDMA_CM_EVENT_CONNECT_REQUEST:
 942                 ret = isert_connect_request(cma_id, event);
 943                 if (ret)
 944                         isert_err("failed handle connect request %d\n", ret);
 945                 break;
 946         case RDMA_CM_EVENT_ESTABLISHED:
 947                 isert_connected_handler(cma_id);
 948                 break;
 949         case RDMA_CM_EVENT_ADDR_CHANGE:    /* FALLTHRU */
 950         case RDMA_CM_EVENT_DISCONNECTED:   /* FALLTHRU */
 951         case RDMA_CM_EVENT_DEVICE_REMOVAL: /* FALLTHRU */
 952         case RDMA_CM_EVENT_TIMEWAIT_EXIT:  /* FALLTHRU */
 953                 ret = isert_disconnected_handler(cma_id, event->event);
 954                 break;
 955         case RDMA_CM_EVENT_REJECTED:       /* FALLTHRU */
 956         case RDMA_CM_EVENT_UNREACHABLE:    /* FALLTHRU */
 957         case RDMA_CM_EVENT_CONNECT_ERROR:
 958                 ret = isert_connect_error(cma_id);
 959                 break;
 960         default:
 961                 isert_err("Unhandled RDMA CMA event: %d\n", event->event);
 962                 break;
 963         }
 964
 965         return ret;
 966 }
 967
 968 static int
 969 isert_post_recvm(struct isert_conn *isert_conn, u32 count)
 970 {
 971         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 972         int i, ret;
 973         struct iser_rx_desc *rx_desc;
 974
 975         for (rx_wr = isert_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
 976                 rx_desc = &isert_conn->rx_descs[i];
 977                 rx_wr->wr_id = (uintptr_t)rx_desc;
 978                 rx_wr->sg_list = &rx_desc->rx_sg;
 979                 rx_wr->num_sge = 1;
 980                 rx_wr->next = rx_wr + 1;
 981         }
 982         rx_wr--;
 983         rx_wr->next = NULL; /* mark end of work requests list */
 984
 985         isert_conn->post_recv_buf_count += count;
 986         ret = ib_post_recv(isert_conn->qp, isert_conn->rx_wr,
 987                            &rx_wr_failed);
 988         if (ret) {
 989                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
 990                 isert_conn->post_recv_buf_count -= count;
 991         }
 992
 993         return ret;
 994 }
 995
 996 static int
 997 isert_post_recv(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc)
 998 {
 999         struct ib_recv_wr *rx_wr_failed, rx_wr;
1000         int ret;
1001
1002         rx_wr.wr_id = (uintptr_t)rx_desc;
1003         rx_wr.sg_list = &rx_desc->rx_sg;
1004         rx_wr.num_sge = 1;
1005         rx_wr.next = NULL;
1006
1007         isert_conn->post_recv_buf_count++;
1008         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_failed);
1009         if (ret) {
1010                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
1011                 isert_conn->post_recv_buf_count--;
1012         }
1013
1014         return ret;
1015 }
1016
1017 static int
1018 isert_post_send(struct isert_conn *isert_conn, struct iser_tx_desc *tx_desc)
1019 {
1020         struct ib_device *ib_dev = isert_conn->cm_id->device;
1021         struct ib_send_wr send_wr, *send_wr_failed;
1022         int ret;
1023
1024         ib_dma_sync_single_for_device(ib_dev, tx_desc->dma_addr,
1025                                       ISER_HEADERS_LEN, DMA_TO_DEVICE);
1026
1027         send_wr.next    = NULL;
1028         send_wr.wr_id   = (uintptr_t)tx_desc;
1029         send_wr.sg_list = tx_desc->tx_sg;
1030         send_wr.num_sge = tx_desc->num_sge;
1031         send_wr.opcode  = IB_WR_SEND;
1032         send_wr.send_flags = IB_SEND_SIGNALED;
1033
1034         ret = ib_post_send(isert_conn->qp, &send_wr, &send_wr_failed);
1035         if (ret)
1036                 isert_err("ib_post_send() failed, ret: %d\n", ret);
1037
1038         return ret;
1039 }
1040
1041 static void
1042 isert_create_send_desc(struct isert_conn *isert_conn,
1043                        struct isert_cmd *isert_cmd,
1044                        struct iser_tx_desc *tx_desc)
1045 {
1046         struct isert_device *device = isert_conn->device;
1047         struct ib_device *ib_dev = device->ib_device;
1048
1049         ib_dma_sync_single_for_cpu(ib_dev, tx_desc->dma_addr,
1050                                    ISER_HEADERS_LEN, DMA_TO_DEVICE);
1051
1052         memset(&tx_desc->iser_header, 0, sizeof(struct iser_ctrl));
1053         tx_desc->iser_header.flags = ISCSI_CTRL;
1054
1055         tx_desc->num_sge = 1;
1056         tx_desc->isert_cmd = isert_cmd;
1057
1058         if (tx_desc->tx_sg[0].lkey != device->pd->local_dma_lkey) {
1059                 tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
1060                 isert_dbg("tx_desc %p lkey mismatch, fixing\n", tx_desc);
1061         }
1062 }
1063
1064 static int
1065 isert_init_tx_hdrs(struct isert_conn *isert_conn,
1066                    struct iser_tx_desc *tx_desc)
1067 {
1068         struct isert_device *device = isert_conn->device;
1069         struct ib_device *ib_dev = device->ib_device;
1070         u64 dma_addr;
1071
1072         dma_addr = ib_dma_map_single(ib_dev, (void *)tx_desc,
1073                         ISER_HEADERS_LEN, DMA_TO_DEVICE);
1074         if (ib_dma_mapping_error(ib_dev, dma_addr)) {
1075                 isert_err("ib_dma_mapping_error() failed\n");
1076                 return -ENOMEM;
1077         }
1078
1079         tx_desc->dma_addr = dma_addr;
1080         tx_desc->tx_sg[0].addr  = tx_desc->dma_addr;
1081         tx_desc->tx_sg[0].length = ISER_HEADERS_LEN;
1082         tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
1083
1084         isert_dbg("Setup tx_sg[0].addr: 0x%llx length: %u lkey: 0x%x\n",
1085                   tx_desc->tx_sg[0].addr, tx_desc->tx_sg[0].length,
1086                   tx_desc->tx_sg[0].lkey);
1087
1088         return 0;
1089 }
1090
1091 static void
1092 isert_init_send_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1093                    struct ib_send_wr *send_wr)
1094 {
1095         struct iser_tx_desc *tx_desc = &isert_cmd->tx_desc;
1096
1097         isert_cmd->rdma_wr.iser_ib_op = ISER_IB_SEND;
1098         send_wr->wr_id = (uintptr_t)&isert_cmd->tx_desc;
1099
1100         if (isert_conn->snd_w_inv && isert_cmd->inv_rkey) {
1101                 send_wr->opcode  = IB_WR_SEND_WITH_INV;
1102                 send_wr->ex.invalidate_rkey = isert_cmd->inv_rkey;
1103         } else {
1104                 send_wr->opcode = IB_WR_SEND;
1105         }
1106
1107         send_wr->sg_list = &tx_desc->tx_sg[0];
1108         send_wr->num_sge = isert_cmd->tx_desc.num_sge;
1109         send_wr->send_flags = IB_SEND_SIGNALED;
1110 }
1111
1112 static int
1113 isert_rdma_post_recvl(struct isert_conn *isert_conn)
1114 {
1115         struct ib_recv_wr rx_wr, *rx_wr_fail;
1116         struct ib_sge sge;
1117         int ret;
1118
1119         memset(&sge, 0, sizeof(struct ib_sge));
1120         sge.addr = isert_conn->login_req_dma;
1121         sge.length = ISER_RX_LOGIN_SIZE;
1122         sge.lkey = isert_conn->device->pd->local_dma_lkey;
1123
1124         isert_dbg("Setup sge: addr: %llx length: %d 0x%08x\n",
1125                 sge.addr, sge.length, sge.lkey);
1126
1127         memset(&rx_wr, 0, sizeof(struct ib_recv_wr));
1128         rx_wr.wr_id = (uintptr_t)isert_conn->login_req_buf;
1129         rx_wr.sg_list = &sge;
1130         rx_wr.num_sge = 1;
1131
1132         isert_conn->post_recv_buf_count++;
1133         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_fail);
1134         if (ret) {
1135                 isert_err("ib_post_recv() failed: %d\n", ret);
1136                 isert_conn->post_recv_buf_count--;
1137         }
1138
1139         return ret;
1140 }
1141
1142 static int
1143 isert_put_login_tx(struct iscsi_conn *conn, struct iscsi_login *login,
1144                    u32 length)
1145 {
1146         struct isert_conn *isert_conn = conn->context;
1147         struct isert_device *device = isert_conn->device;
1148         struct ib_device *ib_dev = device->ib_device;
1149         struct iser_tx_desc *tx_desc = &isert_conn->login_tx_desc;
1150         int ret;
1151
1152         isert_create_send_desc(isert_conn, NULL, tx_desc);
1153
1154         memcpy(&tx_desc->iscsi_header, &login->rsp[0],
1155                sizeof(struct iscsi_hdr));
1156
1157         isert_init_tx_hdrs(isert_conn, tx_desc);
1158
1159         if (length > 0) {
1160                 struct ib_sge *tx_dsg = &tx_desc->tx_sg[1];
1161
1162                 ib_dma_sync_single_for_cpu(ib_dev, isert_conn->login_rsp_dma,
1163                                            length, DMA_TO_DEVICE);
1164
1165                 memcpy(isert_conn->login_rsp_buf, login->rsp_buf, length);
1166
1167                 ib_dma_sync_single_for_device(ib_dev, isert_conn->login_rsp_dma,
1168                                               length, DMA_TO_DEVICE);
1169
1170                 tx_dsg->addr    = isert_conn->login_rsp_dma;
1171                 tx_dsg->length  = length;
1172                 tx_dsg->lkey    = isert_conn->device->pd->local_dma_lkey;
1173                 tx_desc->num_sge = 2;
1174         }
1175         if (!login->login_failed) {
1176                 if (login->login_complete) {
1177                         if (!conn->sess->sess_ops->SessionType &&
1178                             isert_conn->device->use_fastreg) {
1179                                 ret = isert_conn_create_fastreg_pool(isert_conn);
1180                                 if (ret) {
1181                                         isert_err("Conn: %p failed to create"
1182                                                " fastreg pool\n", isert_conn);
1183                                         return ret;
1184                                 }
1185                         }
1186
1187                         ret = isert_alloc_rx_descriptors(isert_conn);
1188                         if (ret)
1189                                 return ret;
1190
1191                         ret = isert_post_recvm(isert_conn,
1192                                                ISERT_QP_MAX_RECV_DTOS);
1193                         if (ret)
1194                                 return ret;
1195
1196                         /* Now we are in FULL_FEATURE phase */
1197                         mutex_lock(&isert_conn->mutex);
1198                         isert_conn->state = ISER_CONN_FULL_FEATURE;
1199                         mutex_unlock(&isert_conn->mutex);
1200                         goto post_send;
1201                 }
1202
1203                 ret = isert_rdma_post_recvl(isert_conn);
1204                 if (ret)
1205                         return ret;
1206         }
1207 post_send:
1208         ret = isert_post_send(isert_conn, tx_desc);
1209         if (ret)
1210                 return ret;
1211
1212         return 0;
1213 }
1214
1215 static void
1216 isert_rx_login_req(struct isert_conn *isert_conn)
1217 {
1218         struct iser_rx_desc *rx_desc = (void *)isert_conn->login_req_buf;
1219         int rx_buflen = isert_conn->login_req_len;
1220         struct iscsi_conn *conn = isert_conn->conn;
1221         struct iscsi_login *login = conn->conn_login;
1222         int size;
1223
1224         isert_info("conn %p\n", isert_conn);
1225
1226         WARN_ON_ONCE(!login);
1227
1228         if (login->first_request) {
1229                 struct iscsi_login_req *login_req =
1230                         (struct iscsi_login_req *)&rx_desc->iscsi_header;
1231                 /*
1232                  * Setup the initial iscsi_login values from the leading
1233                  * login request PDU.
1234                  */
1235                 login->leading_connection = (!login_req->tsih) ? 1 : 0;
1236                 login->current_stage =
1237                         (login_req->flags & ISCSI_FLAG_LOGIN_CURRENT_STAGE_MASK)
1238                          >> 2;
1239                 login->version_min      = login_req->min_version;
1240                 login->version_max      = login_req->max_version;
1241                 memcpy(login->isid, login_req->isid, 6);
1242                 login->cmd_sn           = be32_to_cpu(login_req->cmdsn);
1243                 login->init_task_tag    = login_req->itt;
1244                 login->initial_exp_statsn = be32_to_cpu(login_req->exp_statsn);
1245                 login->cid              = be16_to_cpu(login_req->cid);
1246                 login->tsih             = be16_to_cpu(login_req->tsih);
1247         }
1248
1249         memcpy(&login->req[0], (void *)&rx_desc->iscsi_header, ISCSI_HDR_LEN);
1250
1251         size = min(rx_buflen, MAX_KEY_VALUE_PAIRS);
1252         isert_dbg("Using login payload size: %d, rx_buflen: %d "
1253                   "MAX_KEY_VALUE_PAIRS: %d\n", size, rx_buflen,
1254                   MAX_KEY_VALUE_PAIRS);
1255         memcpy(login->req_buf, &rx_desc->data[0], size);
1256
1257         if (login->first_request) {
1258                 complete(&isert_conn->login_comp);
1259                 return;
1260         }
1261         schedule_delayed_work(&conn->login_work, 0);
1262 }
1263
1264 static struct iscsi_cmd
1265 *isert_allocate_cmd(struct iscsi_conn *conn, struct iser_rx_desc *rx_desc)
1266 {
1267         struct isert_conn *isert_conn = conn->context;
1268         struct isert_cmd *isert_cmd;
1269         struct iscsi_cmd *cmd;
1270
1271         cmd = iscsit_allocate_cmd(conn, TASK_INTERRUPTIBLE);
1272         if (!cmd) {
1273                 isert_err("Unable to allocate iscsi_cmd + isert_cmd\n");
1274                 return NULL;
1275         }
1276         isert_cmd = iscsit_priv_cmd(cmd);
1277         isert_cmd->conn = isert_conn;
1278         isert_cmd->iscsi_cmd = cmd;
1279         isert_cmd->rx_desc = rx_desc;
1280
1281         return cmd;
1282 }
1283
1284 static int
1285 isert_handle_scsi_cmd(struct isert_conn *isert_conn,
1286                       struct isert_cmd *isert_cmd, struct iscsi_cmd *cmd,
1287                       struct iser_rx_desc *rx_desc, unsigned char *buf)
1288 {
1289         struct iscsi_conn *conn = isert_conn->conn;
1290         struct iscsi_scsi_req *hdr = (struct iscsi_scsi_req *)buf;
1291         int imm_data, imm_data_len, unsol_data, sg_nents, rc;
1292         bool dump_payload = false;
1293         unsigned int data_len;
1294
1295         rc = iscsit_setup_scsi_cmd(conn, cmd, buf);
1296         if (rc < 0)
1297                 return rc;
1298
1299         imm_data = cmd->immediate_data;
1300         imm_data_len = cmd->first_burst_len;
1301         unsol_data = cmd->unsolicited_data;
1302         data_len = cmd->se_cmd.data_length;
1303
1304         if (imm_data && imm_data_len == data_len)
1305                 cmd->se_cmd.se_cmd_flags |= SCF_PASSTHROUGH_SG_TO_MEM_NOALLOC;
1306         rc = iscsit_process_scsi_cmd(conn, cmd, hdr);
1307         if (rc < 0) {
1308                 return 0;
1309         } else if (rc > 0) {
1310                 dump_payload = true;
1311                 goto sequence_cmd;
1312         }
1313
1314         if (!imm_data)
1315                 return 0;
1316
1317         if (imm_data_len != data_len) {
1318                 sg_nents = max(1UL, DIV_ROUND_UP(imm_data_len, PAGE_SIZE));
1319                 sg_copy_from_buffer(cmd->se_cmd.t_data_sg, sg_nents,
1320                                     &rx_desc->data[0], imm_data_len);
1321                 isert_dbg("Copy Immediate sg_nents: %u imm_data_len: %d\n",
1322                           sg_nents, imm_data_len);
1323         } else {
1324                 sg_init_table(&isert_cmd->sg, 1);
1325                 cmd->se_cmd.t_data_sg = &isert_cmd->sg;
1326                 cmd->se_cmd.t_data_nents = 1;
1327                 sg_set_buf(&isert_cmd->sg, &rx_desc->data[0], imm_data_len);
1328                 isert_dbg("Transfer Immediate imm_data_len: %d\n",
1329                           imm_data_len);
1330         }
1331
1332         cmd->write_data_done += imm_data_len;
1333
1334         if (cmd->write_data_done == cmd->se_cmd.data_length) {
1335                 spin_lock_bh(&cmd->istate_lock);
1336                 cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1337                 cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1338                 spin_unlock_bh(&cmd->istate_lock);
1339         }
1340
1341 sequence_cmd:
1342         rc = iscsit_sequence_cmd(conn, cmd, buf, hdr->cmdsn);
1343
1344         if (!rc && dump_payload == false && unsol_data)
1345                 iscsit_set_unsoliticed_dataout(cmd);
1346         else if (dump_payload && imm_data)
1347                 target_put_sess_cmd(&cmd->se_cmd);
1348
1349         return 0;
1350 }
1351
1352 static int
1353 isert_handle_iscsi_dataout(struct isert_conn *isert_conn,
1354                            struct iser_rx_desc *rx_desc, unsigned char *buf)
1355 {
1356         struct scatterlist *sg_start;
1357         struct iscsi_conn *conn = isert_conn->conn;
1358         struct iscsi_cmd *cmd = NULL;
1359         struct iscsi_data *hdr = (struct iscsi_data *)buf;
1360         u32 unsol_data_len = ntoh24(hdr->dlength);
1361         int rc, sg_nents, sg_off, page_off;
1362
1363         rc = iscsit_check_dataout_hdr(conn, buf, &cmd);
1364         if (rc < 0)
1365                 return rc;
1366         else if (!cmd)
1367                 return 0;
1368         /*
1369          * FIXME: Unexpected unsolicited_data out
1370          */
1371         if (!cmd->unsolicited_data) {
1372                 isert_err("Received unexpected solicited data payload\n");
1373                 dump_stack();
1374                 return -1;
1375         }
1376
1377         isert_dbg("Unsolicited DataOut unsol_data_len: %u, "
1378                   "write_data_done: %u, data_length: %u\n",
1379                   unsol_data_len,  cmd->write_data_done,
1380                   cmd->se_cmd.data_length);
1381
1382         sg_off = cmd->write_data_done / PAGE_SIZE;
1383         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
1384         sg_nents = max(1UL, DIV_ROUND_UP(unsol_data_len, PAGE_SIZE));
1385         page_off = cmd->write_data_done % PAGE_SIZE;
1386         /*
1387          * FIXME: Non page-aligned unsolicited_data out
1388          */
1389         if (page_off) {
1390                 isert_err("unexpected non-page aligned data payload\n");
1391                 dump_stack();
1392                 return -1;
1393         }
1394         isert_dbg("Copying DataOut: sg_start: %p, sg_off: %u "
1395                   "sg_nents: %u from %p %u\n", sg_start, sg_off,
1396                   sg_nents, &rx_desc->data[0], unsol_data_len);
1397
1398         sg_copy_from_buffer(sg_start, sg_nents, &rx_desc->data[0],
1399                             unsol_data_len);
1400
1401         rc = iscsit_check_dataout_payload(cmd, hdr, false);
1402         if (rc < 0)
1403                 return rc;
1404
1405         /*
1406          * multiple data-outs on the same command can arrive -
1407          * so post the buffer before hand
1408          */
1409         rc = isert_post_recv(isert_conn, rx_desc);
1410         if (rc) {
1411                 isert_err("ib_post_recv failed with %d\n", rc);
1412                 return rc;
1413         }
1414         return 0;
1415 }
1416
1417 static int
1418 isert_handle_nop_out(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1419                      struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1420                      unsigned char *buf)
1421 {
1422         struct iscsi_conn *conn = isert_conn->conn;
1423         struct iscsi_nopout *hdr = (struct iscsi_nopout *)buf;
1424         int rc;
1425
1426         rc = iscsit_setup_nop_out(conn, cmd, hdr);
1427         if (rc < 0)
1428                 return rc;
1429         /*
1430          * FIXME: Add support for NOPOUT payload using unsolicited RDMA payload
1431          */
1432
1433         return iscsit_process_nop_out(conn, cmd, hdr);
1434 }
1435
1436 static int
1437 isert_handle_text_cmd(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1438                       struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1439                       struct iscsi_text *hdr)
1440 {
1441         struct iscsi_conn *conn = isert_conn->conn;
1442         u32 payload_length = ntoh24(hdr->dlength);
1443         int rc;
1444         unsigned char *text_in = NULL;
1445
1446         rc = iscsit_setup_text_cmd(conn, cmd, hdr);
1447         if (rc < 0)
1448                 return rc;
1449
1450         if (payload_length) {
1451                 text_in = kzalloc(payload_length, GFP_KERNEL);
1452                 if (!text_in) {
1453                         isert_err("Unable to allocate text_in of payload_length: %u\n",
1454                                   payload_length);
1455                         return -ENOMEM;
1456                 }
1457         }
1458         cmd->text_in_ptr = text_in;
1459
1460         memcpy(cmd->text_in_ptr, &rx_desc->data[0], payload_length);
1461
1462         return iscsit_process_text_cmd(conn, cmd, hdr);
1463 }
1464
1465 static int
1466 isert_rx_opcode(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc,
1467                 uint32_t read_stag, uint64_t read_va,
1468                 uint32_t write_stag, uint64_t write_va)
1469 {
1470         struct iscsi_hdr *hdr = &rx_desc->iscsi_header;
1471         struct iscsi_conn *conn = isert_conn->conn;
1472         struct iscsi_cmd *cmd;
1473         struct isert_cmd *isert_cmd;
1474         int ret = -EINVAL;
1475         u8 opcode = (hdr->opcode & ISCSI_OPCODE_MASK);
1476
1477         if (conn->sess->sess_ops->SessionType &&
1478            (!(opcode & ISCSI_OP_TEXT) || !(opcode & ISCSI_OP_LOGOUT))) {
1479                 isert_err("Got illegal opcode: 0x%02x in SessionType=Discovery,"
1480                           " ignoring\n", opcode);
1481                 return 0;
1482         }
1483
1484         switch (opcode) {
1485         case ISCSI_OP_SCSI_CMD:
1486                 cmd = isert_allocate_cmd(conn, rx_desc);
1487                 if (!cmd)
1488                         break;
1489
1490                 isert_cmd = iscsit_priv_cmd(cmd);
1491                 isert_cmd->read_stag = read_stag;
1492                 isert_cmd->read_va = read_va;
1493                 isert_cmd->write_stag = write_stag;
1494                 isert_cmd->write_va = write_va;
1495                 isert_cmd->inv_rkey = read_stag ? read_stag : write_stag;
1496
1497                 ret = isert_handle_scsi_cmd(isert_conn, isert_cmd, cmd,
1498                                         rx_desc, (unsigned char *)hdr);
1499                 break;
1500         case ISCSI_OP_NOOP_OUT:
1501                 cmd = isert_allocate_cmd(conn, rx_desc);
1502                 if (!cmd)
1503                         break;
1504
1505                 isert_cmd = iscsit_priv_cmd(cmd);
1506                 ret = isert_handle_nop_out(isert_conn, isert_cmd, cmd,
1507                                            rx_desc, (unsigned char *)hdr);
1508                 break;
1509         case ISCSI_OP_SCSI_DATA_OUT:
1510                 ret = isert_handle_iscsi_dataout(isert_conn, rx_desc,
1511                                                 (unsigned char *)hdr);
1512                 break;
1513         case ISCSI_OP_SCSI_TMFUNC:
1514                 cmd = isert_allocate_cmd(conn, rx_desc);
1515                 if (!cmd)
1516                         break;
1517
1518                 ret = iscsit_handle_task_mgt_cmd(conn, cmd,
1519                                                 (unsigned char *)hdr);
1520                 break;
1521         case ISCSI_OP_LOGOUT:
1522                 cmd = isert_allocate_cmd(conn, rx_desc);
1523                 if (!cmd)
1524                         break;
1525
1526                 ret = iscsit_handle_logout_cmd(conn, cmd, (unsigned char *)hdr);
1527                 break;
1528         case ISCSI_OP_TEXT:
1529                 if (be32_to_cpu(hdr->ttt) != 0xFFFFFFFF)
1530                         cmd = iscsit_find_cmd_from_itt(conn, hdr->itt);
1531                 else
1532                         cmd = isert_allocate_cmd(conn, rx_desc);
1533
1534                 if (!cmd)
1535                         break;
1536
1537                 isert_cmd = iscsit_priv_cmd(cmd);
1538                 ret = isert_handle_text_cmd(isert_conn, isert_cmd, cmd,
1539                                             rx_desc, (struct iscsi_text *)hdr);
1540                 break;
1541         default:
1542                 isert_err("Got unknown iSCSI OpCode: 0x%02x\n", opcode);
1543                 dump_stack();
1544                 break;
1545         }
1546
1547         return ret;
1548 }
1549
1550 static void
1551 isert_rx_do_work(struct iser_rx_desc *rx_desc, struct isert_conn *isert_conn)
1552 {
1553         struct iser_ctrl *iser_ctrl = &rx_desc->iser_header;
1554         uint64_t read_va = 0, write_va = 0;
1555         uint32_t read_stag = 0, write_stag = 0;
1556
1557         switch (iser_ctrl->flags & 0xF0) {
1558         case ISCSI_CTRL:
1559                 if (iser_ctrl->flags & ISER_RSV) {
1560                         read_stag = be32_to_cpu(iser_ctrl->read_stag);
1561                         read_va = be64_to_cpu(iser_ctrl->read_va);
1562                         isert_dbg("ISER_RSV: read_stag: 0x%x read_va: 0x%llx\n",
1563                                   read_stag, (unsigned long long)read_va);
1564                 }
1565                 if (iser_ctrl->flags & ISER_WSV) {
1566                         write_stag = be32_to_cpu(iser_ctrl->write_stag);
1567                         write_va = be64_to_cpu(iser_ctrl->write_va);
1568                         isert_dbg("ISER_WSV: write_stag: 0x%x write_va: 0x%llx\n",
1569                                   write_stag, (unsigned long long)write_va);
1570                 }
1571
1572                 isert_dbg("ISER ISCSI_CTRL PDU\n");
1573                 break;
1574         case ISER_HELLO:
1575                 isert_err("iSER Hello message\n");
1576                 break;
1577         default:
1578                 isert_warn("Unknown iSER hdr flags: 0x%02x\n", iser_ctrl->flags);
1579                 break;
1580         }
1581
1582         isert_rx_opcode(isert_conn, rx_desc,
1583                         read_stag, read_va, write_stag, write_va);
1584 }
1585
1586 static void
1587 isert_rcv_completion(struct iser_rx_desc *desc,
1588                      struct isert_conn *isert_conn,
1589                      u32 xfer_len)
1590 {
1591         struct ib_device *ib_dev = isert_conn->cm_id->device;
1592         struct iscsi_hdr *hdr;
1593         u64 rx_dma;
1594         int rx_buflen;
1595
1596         if ((char *)desc == isert_conn->login_req_buf) {
1597                 rx_dma = isert_conn->login_req_dma;
1598                 rx_buflen = ISER_RX_LOGIN_SIZE;
1599                 isert_dbg("login_buf: Using rx_dma: 0x%llx, rx_buflen: %d\n",
1600                          rx_dma, rx_buflen);
1601         } else {
1602                 rx_dma = desc->dma_addr;
1603                 rx_buflen = ISER_RX_PAYLOAD_SIZE;
1604                 isert_dbg("req_buf: Using rx_dma: 0x%llx, rx_buflen: %d\n",
1605                          rx_dma, rx_buflen);
1606         }
1607
1608         ib_dma_sync_single_for_cpu(ib_dev, rx_dma, rx_buflen, DMA_FROM_DEVICE);
1609
1610         hdr = &desc->iscsi_header;
1611         isert_dbg("iSCSI opcode: 0x%02x, ITT: 0x%08x, flags: 0x%02x dlen: %d\n",
1612                  hdr->opcode, hdr->itt, hdr->flags,
1613                  (int)(xfer_len - ISER_HEADERS_LEN));
1614
1615         if ((char *)desc == isert_conn->login_req_buf) {
1616                 isert_conn->login_req_len = xfer_len - ISER_HEADERS_LEN;
1617                 if (isert_conn->conn) {
1618                         struct iscsi_login *login = isert_conn->conn->conn_login;
1619
1620                         if (login && !login->first_request)
1621                                 isert_rx_login_req(isert_conn);
1622                 }
1623                 mutex_lock(&isert_conn->mutex);
1624                 complete(&isert_conn->login_req_comp);
1625                 mutex_unlock(&isert_conn->mutex);
1626         } else {
1627                 isert_rx_do_work(desc, isert_conn);
1628         }
1629
1630         ib_dma_sync_single_for_device(ib_dev, rx_dma, rx_buflen,
1631                                       DMA_FROM_DEVICE);
1632
1633         isert_conn->post_recv_buf_count--;
1634 }
1635
1636 static int
1637 isert_map_data_buf(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1638                    struct scatterlist *sg, u32 nents, u32 length, u32 offset,
1639                    enum iser_ib_op_code op, struct isert_data_buf *data)
1640 {
1641         struct ib_device *ib_dev = isert_conn->cm_id->device;
1642
1643         data->dma_dir = op == ISER_IB_RDMA_WRITE ?
1644                               DMA_TO_DEVICE : DMA_FROM_DEVICE;
1645
1646         data->len = length - offset;
1647         data->offset = offset;
1648         data->sg_off = data->offset / PAGE_SIZE;
1649
1650         data->sg = &sg[data->sg_off];
1651         data->nents = min_t(unsigned int, nents - data->sg_off,
1652                                           ISCSI_ISER_SG_TABLESIZE);
1653         data->len = min_t(unsigned int, data->len, ISCSI_ISER_SG_TABLESIZE *
1654                                         PAGE_SIZE);
1655
1656         data->dma_nents = ib_dma_map_sg(ib_dev, data->sg, data->nents,
1657                                         data->dma_dir);
1658         if (unlikely(!data->dma_nents)) {
1659                 isert_err("Cmd: unable to dma map SGs %p\n", sg);
1660                 return -EINVAL;
1661         }
1662
1663         isert_dbg("Mapped cmd: %p count: %u sg: %p sg_nents: %u rdma_len %d\n",
1664                   isert_cmd, data->dma_nents, data->sg, data->nents, data->len);
1665
1666         return 0;
1667 }
1668
1669 static void
1670 isert_unmap_data_buf(struct isert_conn *isert_conn, struct isert_data_buf *data)
1671 {
1672         struct ib_device *ib_dev = isert_conn->cm_id->device;
1673
1674         ib_dma_unmap_sg(ib_dev, data->sg, data->nents, data->dma_dir);
1675         memset(data, 0, sizeof(*data));
1676 }
1677
1678
1679
1680 static void
1681 isert_unmap_cmd(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn)
1682 {
1683         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1684
1685         isert_dbg("Cmd %p\n", isert_cmd);
1686
1687         if (wr->data.sg) {
1688                 isert_dbg("Cmd %p unmap_sg op\n", isert_cmd);
1689                 isert_unmap_data_buf(isert_conn, &wr->data);
1690         }
1691
1692         if (wr->rdma_wr) {
1693                 isert_dbg("Cmd %p free send_wr\n", isert_cmd);
1694                 kfree(wr->rdma_wr);
1695                 wr->rdma_wr = NULL;
1696         }
1697
1698         if (wr->ib_sge) {
1699                 isert_dbg("Cmd %p free ib_sge\n", isert_cmd);
1700                 kfree(wr->ib_sge);
1701                 wr->ib_sge = NULL;
1702         }
1703 }
1704
1705 static void
1706 isert_unreg_rdma(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn)
1707 {
1708         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1709
1710         isert_dbg("Cmd %p\n", isert_cmd);
1711
1712         if (wr->fr_desc) {
1713                 isert_dbg("Cmd %p free fr_desc %p\n", isert_cmd, wr->fr_desc);
1714                 if (wr->fr_desc->ind & ISERT_PROTECTED) {
1715                         isert_unmap_data_buf(isert_conn, &wr->prot);
1716                         wr->fr_desc->ind &= ~ISERT_PROTECTED;
1717                 }
1718                 spin_lock_bh(&isert_conn->pool_lock);
1719                 list_add_tail(&wr->fr_desc->list, &isert_conn->fr_pool);
1720                 spin_unlock_bh(&isert_conn->pool_lock);
1721                 wr->fr_desc = NULL;
1722         }
1723
1724         if (wr->data.sg) {
1725                 isert_dbg("Cmd %p unmap_sg op\n", isert_cmd);
1726                 isert_unmap_data_buf(isert_conn, &wr->data);
1727         }
1728
1729         wr->ib_sge = NULL;
1730         wr->rdma_wr = NULL;
1731 }
1732
1733 static void
1734 isert_put_cmd(struct isert_cmd *isert_cmd, bool comp_err)
1735 {
1736         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1737         struct isert_conn *isert_conn = isert_cmd->conn;
1738         struct iscsi_conn *conn = isert_conn->conn;
1739         struct isert_device *device = isert_conn->device;
1740         struct iscsi_text_rsp *hdr;
1741
1742         isert_dbg("Cmd %p\n", isert_cmd);
1743
1744         switch (cmd->iscsi_opcode) {
1745         case ISCSI_OP_SCSI_CMD:
1746                 spin_lock_bh(&conn->cmd_lock);
1747                 if (!list_empty(&cmd->i_conn_node))
1748                         list_del_init(&cmd->i_conn_node);
1749                 spin_unlock_bh(&conn->cmd_lock);
1750
1751                 if (cmd->data_direction == DMA_TO_DEVICE) {
1752                         iscsit_stop_dataout_timer(cmd);
1753                         /*
1754                          * Check for special case during comp_err where
1755                          * WRITE_PENDING has been handed off from core,
1756                          * but requires an extra target_put_sess_cmd()
1757                          * before transport_generic_free_cmd() below.
1758                          */
1759                         if (comp_err &&
1760                             cmd->se_cmd.t_state == TRANSPORT_WRITE_PENDING) {
1761                                 struct se_cmd *se_cmd = &cmd->se_cmd;
1762
1763                                 target_put_sess_cmd(se_cmd);
1764                         }
1765                 }
1766
1767                 device->unreg_rdma_mem(isert_cmd, isert_conn);
1768                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1769                 break;
1770         case ISCSI_OP_SCSI_TMFUNC:
1771                 spin_lock_bh(&conn->cmd_lock);
1772                 if (!list_empty(&cmd->i_conn_node))
1773                         list_del_init(&cmd->i_conn_node);
1774                 spin_unlock_bh(&conn->cmd_lock);
1775
1776                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1777                 break;
1778         case ISCSI_OP_REJECT:
1779         case ISCSI_OP_NOOP_OUT:
1780         case ISCSI_OP_TEXT:
1781                 hdr = (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
1782                 /* If the continue bit is on, keep the command alive */
1783                 if (hdr->flags & ISCSI_FLAG_TEXT_CONTINUE)
1784                         break;
1785
1786                 spin_lock_bh(&conn->cmd_lock);
1787                 if (!list_empty(&cmd->i_conn_node))
1788                         list_del_init(&cmd->i_conn_node);
1789                 spin_unlock_bh(&conn->cmd_lock);
1790
1791                 /*
1792                  * Handle special case for REJECT when iscsi_add_reject*() has
1793                  * overwritten the original iscsi_opcode assignment, and the
1794                  * associated cmd->se_cmd needs to be released.
1795                  */
1796                 if (cmd->se_cmd.se_tfo != NULL) {
1797                         isert_dbg("Calling transport_generic_free_cmd for 0x%02x\n",
1798                                  cmd->iscsi_opcode);
1799                         transport_generic_free_cmd(&cmd->se_cmd, 0);
1800                         break;
1801                 }
1802                 /*
1803                  * Fall-through
1804                  */
1805         default:
1806                 iscsit_release_cmd(cmd);
1807                 break;
1808         }
1809 }
1810
1811 static void
1812 isert_unmap_tx_desc(struct iser_tx_desc *tx_desc, struct ib_device *ib_dev)
1813 {
1814         if (tx_desc->dma_addr != 0) {
1815                 isert_dbg("unmap single for tx_desc->dma_addr\n");
1816                 ib_dma_unmap_single(ib_dev, tx_desc->dma_addr,
1817                                     ISER_HEADERS_LEN, DMA_TO_DEVICE);
1818                 tx_desc->dma_addr = 0;
1819         }
1820 }
1821
1822 static void
1823 isert_completion_put(struct iser_tx_desc *tx_desc, struct isert_cmd *isert_cmd,
1824                      struct ib_device *ib_dev, bool comp_err)
1825 {
1826         if (isert_cmd->pdu_buf_dma != 0) {
1827                 isert_dbg("unmap single for isert_cmd->pdu_buf_dma\n");
1828                 ib_dma_unmap_single(ib_dev, isert_cmd->pdu_buf_dma,
1829                                     isert_cmd->pdu_buf_len, DMA_TO_DEVICE);
1830                 isert_cmd->pdu_buf_dma = 0;
1831         }
1832
1833         isert_unmap_tx_desc(tx_desc, ib_dev);
1834         isert_put_cmd(isert_cmd, comp_err);
1835 }
1836
1837 static int
1838 isert_check_pi_status(struct se_cmd *se_cmd, struct ib_mr *sig_mr)
1839 {
1840         struct ib_mr_status mr_status;
1841         int ret;
1842
1843         ret = ib_check_mr_status(sig_mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
1844         if (ret) {
1845                 isert_err("ib_check_mr_status failed, ret %d\n", ret);
1846                 goto fail_mr_status;
1847         }
1848
1849         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1850                 u64 sec_offset_err;
1851                 u32 block_size = se_cmd->se_dev->dev_attrib.block_size + 8;
1852
1853                 switch (mr_status.sig_err.err_type) {
1854                 case IB_SIG_BAD_GUARD:
1855                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_GUARD_CHECK_FAILED;
1856                         break;
1857                 case IB_SIG_BAD_REFTAG:
1858                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_REF_TAG_CHECK_FAILED;
1859                         break;
1860                 case IB_SIG_BAD_APPTAG:
1861                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_APP_TAG_CHECK_FAILED;
1862                         break;
1863                 }
1864                 sec_offset_err = mr_status.sig_err.sig_err_offset;
1865                 do_div(sec_offset_err, block_size);
1866                 se_cmd->bad_sector = sec_offset_err + se_cmd->t_task_lba;
1867
1868                 isert_err("PI error found type %d at sector 0x%llx "
1869                           "expected 0x%x vs actual 0x%x\n",
1870                           mr_status.sig_err.err_type,
1871                           (unsigned long long)se_cmd->bad_sector,
1872                           mr_status.sig_err.expected,
1873                           mr_status.sig_err.actual);
1874                 ret = 1;
1875         }
1876
1877 fail_mr_status:
1878         return ret;
1879 }
1880
1881 static void
1882 isert_completion_rdma_write(struct iser_tx_desc *tx_desc,
1883                             struct isert_cmd *isert_cmd)
1884 {
1885         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1886         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1887         struct se_cmd *se_cmd = &cmd->se_cmd;
1888         struct isert_conn *isert_conn = isert_cmd->conn;
1889         struct isert_device *device = isert_conn->device;
1890         int ret = 0;
1891
1892         if (wr->fr_desc && wr->fr_desc->ind & ISERT_PROTECTED) {
1893                 ret = isert_check_pi_status(se_cmd,
1894                                             wr->fr_desc->pi_ctx->sig_mr);
1895                 wr->fr_desc->ind &= ~ISERT_PROTECTED;
1896         }
1897
1898         device->unreg_rdma_mem(isert_cmd, isert_conn);
1899         wr->rdma_wr_num = 0;
1900         if (ret)
1901                 transport_send_check_condition_and_sense(se_cmd,
1902                                                          se_cmd->pi_err, 0);
1903         else
1904                 isert_put_response(isert_conn->conn, cmd);
1905 }
1906
1907 static void
1908 isert_completion_rdma_read(struct iser_tx_desc *tx_desc,
1909                            struct isert_cmd *isert_cmd)
1910 {
1911         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1912         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1913         struct se_cmd *se_cmd = &cmd->se_cmd;
1914         struct isert_conn *isert_conn = isert_cmd->conn;
1915         struct isert_device *device = isert_conn->device;
1916         int ret = 0;
1917
1918         if (wr->fr_desc && wr->fr_desc->ind & ISERT_PROTECTED) {
1919                 ret = isert_check_pi_status(se_cmd,
1920                                             wr->fr_desc->pi_ctx->sig_mr);
1921                 wr->fr_desc->ind &= ~ISERT_PROTECTED;
1922         }
1923
1924         iscsit_stop_dataout_timer(cmd);
1925         device->unreg_rdma_mem(isert_cmd, isert_conn);
1926         cmd->write_data_done = wr->data.len;
1927         wr->rdma_wr_num = 0;
1928
1929         isert_dbg("Cmd: %p RDMA_READ comp calling execute_cmd\n", isert_cmd);
1930         spin_lock_bh(&cmd->istate_lock);
1931         cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1932         cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1933         spin_unlock_bh(&cmd->istate_lock);
1934
1935         if (ret) {
1936                 target_put_sess_cmd(se_cmd);
1937                 transport_send_check_condition_and_sense(se_cmd,
1938                                                          se_cmd->pi_err, 0);
1939         } else {
1940                 target_execute_cmd(se_cmd);
1941         }
1942 }
1943
1944 static void
1945 isert_do_control_comp(struct work_struct *work)
1946 {
1947         struct isert_cmd *isert_cmd = container_of(work,
1948                         struct isert_cmd, comp_work);
1949         struct isert_conn *isert_conn = isert_cmd->conn;
1950         struct ib_device *ib_dev = isert_conn->cm_id->device;
1951         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1952
1953         isert_dbg("Cmd %p i_state %d\n", isert_cmd, cmd->i_state);
1954
1955         switch (cmd->i_state) {
1956         case ISTATE_SEND_TASKMGTRSP:
1957                 iscsit_tmr_post_handler(cmd, cmd->conn);
1958         case ISTATE_SEND_REJECT:   /* FALLTHRU */
1959         case ISTATE_SEND_TEXTRSP:  /* FALLTHRU */
1960                 cmd->i_state = ISTATE_SENT_STATUS;
1961                 isert_completion_put(&isert_cmd->tx_desc, isert_cmd,
1962                                      ib_dev, false);
1963                 break;
1964         case ISTATE_SEND_LOGOUTRSP:
1965                 iscsit_logout_post_handler(cmd, cmd->conn);
1966                 break;
1967         default:
1968                 isert_err("Unknown i_state %d\n", cmd->i_state);
1969                 dump_stack();
1970                 break;
1971         }
1972 }
1973
1974 static void
1975 isert_response_completion(struct iser_tx_desc *tx_desc,
1976                           struct isert_cmd *isert_cmd,
1977                           struct isert_conn *isert_conn,
1978                           struct ib_device *ib_dev)
1979 {
1980         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1981
1982         if (cmd->i_state == ISTATE_SEND_TASKMGTRSP ||
1983             cmd->i_state == ISTATE_SEND_LOGOUTRSP ||
1984             cmd->i_state == ISTATE_SEND_REJECT ||
1985             cmd->i_state == ISTATE_SEND_TEXTRSP) {
1986                 isert_unmap_tx_desc(tx_desc, ib_dev);
1987
1988                 INIT_WORK(&isert_cmd->comp_work, isert_do_control_comp);
1989                 queue_work(isert_comp_wq, &isert_cmd->comp_work);
1990                 return;
1991         }
1992
1993         cmd->i_state = ISTATE_SENT_STATUS;
1994         isert_completion_put(tx_desc, isert_cmd, ib_dev, false);
1995 }
1996
1997 static void
1998 isert_snd_completion(struct iser_tx_desc *tx_desc,
1999                       struct isert_conn *isert_conn)
2000 {
2001         struct ib_device *ib_dev = isert_conn->cm_id->device;
2002         struct isert_cmd *isert_cmd = tx_desc->isert_cmd;
2003         struct isert_rdma_wr *wr;
2004
2005         if (!isert_cmd) {
2006                 isert_unmap_tx_desc(tx_desc, ib_dev);
2007                 return;
2008         }
2009         wr = &isert_cmd->rdma_wr;
2010
2011         isert_dbg("Cmd %p iser_ib_op %d\n", isert_cmd, wr->iser_ib_op);
2012
2013         switch (wr->iser_ib_op) {
2014         case ISER_IB_SEND:
2015                 isert_response_completion(tx_desc, isert_cmd,
2016                                           isert_conn, ib_dev);
2017                 break;
2018         case ISER_IB_RDMA_WRITE:
2019                 isert_completion_rdma_write(tx_desc, isert_cmd);
2020                 break;
2021         case ISER_IB_RDMA_READ:
2022                 isert_completion_rdma_read(tx_desc, isert_cmd);
2023                 break;
2024         default:
2025                 isert_err("Unknown wr->iser_ib_op: 0x%x\n", wr->iser_ib_op);
2026                 dump_stack();
2027                 break;
2028         }
2029 }
2030
2031 /**
2032  * is_isert_tx_desc() - Indicate if the completion wr_id
2033  *     is a TX descriptor or not.
2034  * @isert_conn: iser connection
2035  * @wr_id: completion WR identifier
2036  *
2037  * Since we cannot rely on wc opcode in FLUSH errors
2038  * we must work around it by checking if the wr_id address
2039  * falls in the iser connection rx_descs buffer. If so
2040  * it is an RX descriptor, otherwize it is a TX.
2041  */
2042 static inline bool
2043 is_isert_tx_desc(struct isert_conn *isert_conn, void *wr_id)
2044 {
2045         void *start = isert_conn->rx_descs;
2046         int len = ISERT_QP_MAX_RECV_DTOS * sizeof(*isert_conn->rx_descs);
2047
2048         if (wr_id >= start && wr_id < start + len)
2049                 return false;
2050
2051         return true;
2052 }
2053
2054 static void
2055 isert_cq_comp_err(struct isert_conn *isert_conn, struct ib_wc *wc)
2056 {
2057         if (wc->wr_id == ISER_BEACON_WRID) {
2058                 isert_info("conn %p completing wait_comp_err\n",
2059                            isert_conn);
2060                 complete(&isert_conn->wait_comp_err);
2061         } else if (is_isert_tx_desc(isert_conn, (void *)(uintptr_t)wc->wr_id)) {
2062                 struct ib_device *ib_dev = isert_conn->cm_id->device;
2063                 struct isert_cmd *isert_cmd;
2064                 struct iser_tx_desc *desc;
2065
2066                 desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
2067                 isert_cmd = desc->isert_cmd;
2068                 if (!isert_cmd)
2069                         isert_unmap_tx_desc(desc, ib_dev);
2070                 else
2071                         isert_completion_put(desc, isert_cmd, ib_dev, true);
2072         } else {
2073                 isert_conn->post_recv_buf_count--;
2074                 if (!isert_conn->post_recv_buf_count)
2075                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
2076         }
2077 }
2078
2079 static void
2080 isert_handle_wc(struct ib_wc *wc)
2081 {
2082         struct isert_conn *isert_conn;
2083         struct iser_tx_desc *tx_desc;
2084         struct iser_rx_desc *rx_desc;
2085
2086         isert_conn = wc->qp->qp_context;
2087         if (likely(wc->status == IB_WC_SUCCESS)) {
2088                 if (wc->opcode == IB_WC_RECV) {
2089                         rx_desc = (struct iser_rx_desc *)(uintptr_t)wc->wr_id;
2090                         isert_rcv_completion(rx_desc, isert_conn, wc->byte_len);
2091                 } else {
2092                         tx_desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
2093                         isert_snd_completion(tx_desc, isert_conn);
2094                 }
2095         } else {
2096                 if (wc->status != IB_WC_WR_FLUSH_ERR)
2097                         isert_err("%s (%d): wr id %llx vend_err %x\n",
2098                                   ib_wc_status_msg(wc->status), wc->status,
2099                                   wc->wr_id, wc->vendor_err);
2100                 else
2101                         isert_dbg("%s (%d): wr id %llx\n",
2102                                   ib_wc_status_msg(wc->status), wc->status,
2103                                   wc->wr_id);
2104
2105                 if (wc->wr_id != ISER_FASTREG_LI_WRID)
2106                         isert_cq_comp_err(isert_conn, wc);
2107         }
2108 }
2109
2110 static void
2111 isert_cq_work(struct work_struct *work)
2112 {
2113         enum { isert_poll_budget = 65536 };
2114         struct isert_comp *comp = container_of(work, struct isert_comp,
2115                                                work);
2116         struct ib_wc *const wcs = comp->wcs;
2117         int i, n, completed = 0;
2118
2119         while ((n = ib_poll_cq(comp->cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
2120                 for (i = 0; i < n; i++)
2121                         isert_handle_wc(&wcs[i]);
2122
2123                 completed += n;
2124                 if (completed >= isert_poll_budget)
2125                         break;
2126         }
2127
2128         ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP);
2129 }
2130
2131 static void
2132 isert_cq_callback(struct ib_cq *cq, void *context)
2133 {
2134         struct isert_comp *comp = context;
2135
2136         queue_work(isert_comp_wq, &comp->work);
2137 }
2138
2139 static int
2140 isert_post_response(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd)
2141 {
2142         struct ib_send_wr *wr_failed;
2143         int ret;
2144
2145         ret = isert_post_recv(isert_conn, isert_cmd->rx_desc);
2146         if (ret) {
2147                 isert_err("ib_post_recv failed with %d\n", ret);
2148                 return ret;
2149         }
2150
2151         ret = ib_post_send(isert_conn->qp, &isert_cmd->tx_desc.send_wr,
2152                            &wr_failed);
2153         if (ret) {
2154                 isert_err("ib_post_send failed with %d\n", ret);
2155                 return ret;
2156         }
2157         return ret;
2158 }
2159
2160 static int
2161 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2162 {
2163         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2164         struct isert_conn *isert_conn = conn->context;
2165         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2166         struct iscsi_scsi_rsp *hdr = (struct iscsi_scsi_rsp *)
2167                                 &isert_cmd->tx_desc.iscsi_header;
2168
2169         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2170         iscsit_build_rsp_pdu(cmd, conn, true, hdr);
2171         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2172         /*
2173          * Attach SENSE DATA payload to iSCSI Response PDU
2174          */
2175         if (cmd->se_cmd.sense_buffer &&
2176             ((cmd->se_cmd.se_cmd_flags & SCF_TRANSPORT_TASK_SENSE) ||
2177             (cmd->se_cmd.se_cmd_flags & SCF_EMULATED_TASK_SENSE))) {
2178                 struct isert_device *device = isert_conn->device;
2179                 struct ib_device *ib_dev = device->ib_device;
2180                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2181                 u32 padding, pdu_len;
2182
2183                 put_unaligned_be16(cmd->se_cmd.scsi_sense_length,
2184                                    cmd->sense_buffer);
2185                 cmd->se_cmd.scsi_sense_length += sizeof(__be16);
2186
2187                 padding = -(cmd->se_cmd.scsi_sense_length) & 3;
2188                 hton24(hdr->dlength, (u32)cmd->se_cmd.scsi_sense_length);
2189                 pdu_len = cmd->se_cmd.scsi_sense_length + padding;
2190
2191                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2192                                 (void *)cmd->sense_buffer, pdu_len,
2193                                 DMA_TO_DEVICE);
2194
2195                 isert_cmd->pdu_buf_len = pdu_len;
2196                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2197                 tx_dsg->length  = pdu_len;
2198                 tx_dsg->lkey    = device->pd->local_dma_lkey;
2199                 isert_cmd->tx_desc.num_sge = 2;
2200         }
2201
2202         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2203
2204         isert_dbg("Posting SCSI Response\n");
2205
2206         return isert_post_response(isert_conn, isert_cmd);
2207 }
2208
2209 static void
2210 isert_aborted_task(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2211 {
2212         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2213         struct isert_conn *isert_conn = conn->context;
2214         struct isert_device *device = isert_conn->device;
2215
2216         spin_lock_bh(&conn->cmd_lock);
2217         if (!list_empty(&cmd->i_conn_node))
2218                 list_del_init(&cmd->i_conn_node);
2219         spin_unlock_bh(&conn->cmd_lock);
2220
2221         if (cmd->data_direction == DMA_TO_DEVICE)
2222                 iscsit_stop_dataout_timer(cmd);
2223
2224         device->unreg_rdma_mem(isert_cmd, isert_conn);
2225 }
2226
2227 static enum target_prot_op
2228 isert_get_sup_prot_ops(struct iscsi_conn *conn)
2229 {
2230         struct isert_conn *isert_conn = conn->context;
2231         struct isert_device *device = isert_conn->device;
2232
2233         if (conn->tpg->tpg_attrib.t10_pi) {
2234                 if (device->pi_capable) {
2235                         isert_info("conn %p PI offload enabled\n", isert_conn);
2236                         isert_conn->pi_support = true;
2237                         return TARGET_PROT_ALL;
2238                 }
2239         }
2240
2241         isert_info("conn %p PI offload disabled\n", isert_conn);
2242         isert_conn->pi_support = false;
2243
2244         return TARGET_PROT_NORMAL;
2245 }
2246
2247 static int
2248 isert_put_nopin(struct iscsi_cmd *cmd, struct iscsi_conn *conn,
2249                 bool nopout_response)
2250 {
2251         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2252         struct isert_conn *isert_conn = conn->context;
2253         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2254
2255         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2256         iscsit_build_nopin_rsp(cmd, conn, (struct iscsi_nopin *)
2257                                &isert_cmd->tx_desc.iscsi_header,
2258                                nopout_response);
2259         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2260         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2261
2262         isert_dbg("conn %p Posting NOPIN Response\n", isert_conn);
2263
2264         return isert_post_response(isert_conn, isert_cmd);
2265 }
2266
2267 static int
2268 isert_put_logout_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2269 {
2270         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2271         struct isert_conn *isert_conn = conn->context;
2272         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2273
2274         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2275         iscsit_build_logout_rsp(cmd, conn, (struct iscsi_logout_rsp *)
2276                                 &isert_cmd->tx_desc.iscsi_header);
2277         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2278         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2279
2280         isert_dbg("conn %p Posting Logout Response\n", isert_conn);
2281
2282         return isert_post_response(isert_conn, isert_cmd);
2283 }
2284
2285 static int
2286 isert_put_tm_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2287 {
2288         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2289         struct isert_conn *isert_conn = conn->context;
2290         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2291
2292         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2293         iscsit_build_task_mgt_rsp(cmd, conn, (struct iscsi_tm_rsp *)
2294                                   &isert_cmd->tx_desc.iscsi_header);
2295         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2296         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2297
2298         isert_dbg("conn %p Posting Task Management Response\n", isert_conn);
2299
2300         return isert_post_response(isert_conn, isert_cmd);
2301 }
2302
2303 static int
2304 isert_put_reject(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2305 {
2306         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2307         struct isert_conn *isert_conn = conn->context;
2308         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2309         struct isert_device *device = isert_conn->device;
2310         struct ib_device *ib_dev = device->ib_device;
2311         struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2312         struct iscsi_reject *hdr =
2313                 (struct iscsi_reject *)&isert_cmd->tx_desc.iscsi_header;
2314
2315         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2316         iscsit_build_reject(cmd, conn, hdr);
2317         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2318
2319         hton24(hdr->dlength, ISCSI_HDR_LEN);
2320         isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2321                         (void *)cmd->buf_ptr, ISCSI_HDR_LEN,
2322                         DMA_TO_DEVICE);
2323         isert_cmd->pdu_buf_len = ISCSI_HDR_LEN;
2324         tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2325         tx_dsg->length  = ISCSI_HDR_LEN;
2326         tx_dsg->lkey    = device->pd->local_dma_lkey;
2327         isert_cmd->tx_desc.num_sge = 2;
2328
2329         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2330
2331         isert_dbg("conn %p Posting Reject\n", isert_conn);
2332
2333         return isert_post_response(isert_conn, isert_cmd);
2334 }
2335
2336 static int
2337 isert_put_text_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2338 {
2339         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2340         struct isert_conn *isert_conn = conn->context;
2341         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2342         struct iscsi_text_rsp *hdr =
2343                 (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
2344         u32 txt_rsp_len;
2345         int rc;
2346
2347         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2348         rc = iscsit_build_text_rsp(cmd, conn, hdr, ISCSI_INFINIBAND);
2349         if (rc < 0)
2350                 return rc;
2351
2352         txt_rsp_len = rc;
2353         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2354
2355         if (txt_rsp_len) {
2356                 struct isert_device *device = isert_conn->device;
2357                 struct ib_device *ib_dev = device->ib_device;
2358                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2359                 void *txt_rsp_buf = cmd->buf_ptr;
2360
2361                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2362                                 txt_rsp_buf, txt_rsp_len, DMA_TO_DEVICE);
2363
2364                 isert_cmd->pdu_buf_len = txt_rsp_len;
2365                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2366                 tx_dsg->length  = txt_rsp_len;
2367                 tx_dsg->lkey    = device->pd->local_dma_lkey;
2368                 isert_cmd->tx_desc.num_sge = 2;
2369         }
2370         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2371
2372         isert_dbg("conn %p Text Response\n", isert_conn);
2373
2374         return isert_post_response(isert_conn, isert_cmd);
2375 }
2376
2377 static int
2378 isert_build_rdma_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
2379                     struct ib_sge *ib_sge, struct ib_rdma_wr *rdma_wr,
2380                     u32 data_left, u32 offset)
2381 {
2382         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
2383         struct scatterlist *sg_start, *tmp_sg;
2384         struct isert_device *device = isert_conn->device;
2385         struct ib_device *ib_dev = device->ib_device;
2386         u32 sg_off, page_off;
2387         int i = 0, sg_nents;
2388
2389         sg_off = offset / PAGE_SIZE;
2390         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
2391         sg_nents = min(cmd->se_cmd.t_data_nents - sg_off, isert_conn->max_sge);
2392         page_off = offset % PAGE_SIZE;
2393
2394         rdma_wr->wr.sg_list = ib_sge;
2395         rdma_wr->wr.wr_id = (uintptr_t)&isert_cmd->tx_desc;
2396         /*
2397          * Perform mapping of TCM scatterlist memory ib_sge dma_addr.
2398          */
2399         for_each_sg(sg_start, tmp_sg, sg_nents, i) {
2400                 isert_dbg("RDMA from SGL dma_addr: 0x%llx dma_len: %u, "
2401                           "page_off: %u\n",
2402                           (unsigned long long)tmp_sg->dma_address,
2403                           tmp_sg->length, page_off);
2404
2405                 ib_sge->addr = ib_sg_dma_address(ib_dev, tmp_sg) + page_off;
2406                 ib_sge->length = min_t(u32, data_left,
2407                                 ib_sg_dma_len(ib_dev, tmp_sg) - page_off);
2408                 ib_sge->lkey = device->pd->local_dma_lkey;
2409
2410                 isert_dbg("RDMA ib_sge: addr: 0x%llx  length: %u lkey: %x\n",
2411                           ib_sge->addr, ib_sge->length, ib_sge->lkey);
2412                 page_off = 0;
2413                 data_left -= ib_sge->length;
2414                 if (!data_left)
2415                         break;
2416                 ib_sge++;
2417                 isert_dbg("Incrementing ib_sge pointer to %p\n", ib_sge);
2418         }
2419
2420         rdma_wr->wr.num_sge = ++i;
2421         isert_dbg("Set outgoing sg_list: %p num_sg: %u from TCM SGLs\n",
2422                   rdma_wr->wr.sg_list, rdma_wr->wr.num_sge);
2423
2424         return rdma_wr->wr.num_sge;
2425 }
2426
2427 static int
2428 isert_map_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2429                struct isert_rdma_wr *wr)
2430 {
2431         struct se_cmd *se_cmd = &cmd->se_cmd;
2432         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2433         struct isert_conn *isert_conn = conn->context;
2434         struct isert_data_buf *data = &wr->data;
2435         struct ib_rdma_wr *rdma_wr;
2436         struct ib_sge *ib_sge;
2437         u32 offset, data_len, data_left, rdma_write_max, va_offset = 0;
2438         int ret = 0, i, ib_sge_cnt;
2439
2440         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2441
2442         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2443         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2444                                  se_cmd->t_data_nents, se_cmd->data_length,
2445                                  offset, wr->iser_ib_op, &wr->data);
2446         if (ret)
2447                 return ret;
2448
2449         data_left = data->len;
2450         offset = data->offset;
2451
2452         ib_sge = kzalloc(sizeof(struct ib_sge) * data->nents, GFP_KERNEL);
2453         if (!ib_sge) {
2454                 isert_warn("Unable to allocate ib_sge\n");
2455                 ret = -ENOMEM;
2456                 goto unmap_cmd;
2457         }
2458         wr->ib_sge = ib_sge;
2459
2460         wr->rdma_wr_num = DIV_ROUND_UP(data->nents, isert_conn->max_sge);
2461         wr->rdma_wr = kzalloc(sizeof(struct ib_rdma_wr) * wr->rdma_wr_num,
2462                                 GFP_KERNEL);
2463         if (!wr->rdma_wr) {
2464                 isert_dbg("Unable to allocate wr->rdma_wr\n");
2465                 ret = -ENOMEM;
2466                 goto unmap_cmd;
2467         }
2468
2469         wr->isert_cmd = isert_cmd;
2470         rdma_write_max = isert_conn->max_sge * PAGE_SIZE;
2471
2472         for (i = 0; i < wr->rdma_wr_num; i++) {
2473                 rdma_wr = &isert_cmd->rdma_wr.rdma_wr[i];
2474                 data_len = min(data_left, rdma_write_max);
2475
2476                 rdma_wr->wr.send_flags = 0;
2477                 if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2478                         rdma_wr->wr.opcode = IB_WR_RDMA_WRITE;
2479                         rdma_wr->remote_addr = isert_cmd->read_va + offset;
2480                         rdma_wr->rkey = isert_cmd->read_stag;
2481                         if (i + 1 == wr->rdma_wr_num)
2482                                 rdma_wr->wr.next = &isert_cmd->tx_desc.send_wr;
2483                         else
2484                                 rdma_wr->wr.next = &wr->rdma_wr[i + 1].wr;
2485                 } else {
2486                         rdma_wr->wr.opcode = IB_WR_RDMA_READ;
2487                         rdma_wr->remote_addr = isert_cmd->write_va + va_offset;
2488                         rdma_wr->rkey = isert_cmd->write_stag;
2489                         if (i + 1 == wr->rdma_wr_num)
2490                                 rdma_wr->wr.send_flags = IB_SEND_SIGNALED;
2491                         else
2492                                 rdma_wr->wr.next = &wr->rdma_wr[i + 1].wr;
2493                 }
2494
2495                 ib_sge_cnt = isert_build_rdma_wr(isert_conn, isert_cmd, ib_sge,
2496                                         rdma_wr, data_len, offset);
2497                 ib_sge += ib_sge_cnt;
2498
2499                 offset += data_len;
2500                 va_offset += data_len;
2501                 data_left -= data_len;
2502         }
2503
2504         return 0;
2505 unmap_cmd:
2506         isert_unmap_data_buf(isert_conn, data);
2507
2508         return ret;
2509 }
2510
2511 static inline void
2512 isert_inv_rkey(struct ib_send_wr *inv_wr, struct ib_mr *mr)
2513 {
2514         u32 rkey;
2515
2516         memset(inv_wr, 0, sizeof(*inv_wr));
2517         inv_wr->wr_id = ISER_FASTREG_LI_WRID;
2518         inv_wr->opcode = IB_WR_LOCAL_INV;
2519         inv_wr->ex.invalidate_rkey = mr->rkey;
2520
2521         /* Bump the key */
2522         rkey = ib_inc_rkey(mr->rkey);
2523         ib_update_fast_reg_key(mr, rkey);
2524 }
2525
2526 static int
2527 isert_fast_reg_mr(struct isert_conn *isert_conn,
2528                   struct fast_reg_descriptor *fr_desc,
2529                   struct isert_data_buf *mem,
2530                   enum isert_indicator ind,
2531                   struct ib_sge *sge)
2532 {
2533         struct isert_device *device = isert_conn->device;
2534         struct ib_device *ib_dev = device->ib_device;
2535         struct ib_mr *mr;
2536         struct ib_reg_wr reg_wr;
2537         struct ib_send_wr inv_wr, *bad_wr, *wr = NULL;
2538         int ret, n;
2539
2540         if (mem->dma_nents == 1) {
2541                 sge->lkey = device->pd->local_dma_lkey;
2542                 sge->addr = ib_sg_dma_address(ib_dev, &mem->sg[0]);
2543                 sge->length = ib_sg_dma_len(ib_dev, &mem->sg[0]);
2544                 isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2545                          sge->addr, sge->length, sge->lkey);
2546                 return 0;
2547         }
2548
2549         if (ind == ISERT_DATA_KEY_VALID)
2550                 /* Registering data buffer */
2551                 mr = fr_desc->data_mr;
2552         else
2553                 /* Registering protection buffer */
2554                 mr = fr_desc->pi_ctx->prot_mr;
2555
2556         if (!(fr_desc->ind & ind)) {
2557                 isert_inv_rkey(&inv_wr, mr);
2558                 wr = &inv_wr;
2559         }
2560
2561         n = ib_map_mr_sg(mr, mem->sg, mem->nents, PAGE_SIZE);
2562         if (unlikely(n != mem->nents)) {
2563                 isert_err("failed to map mr sg (%d/%d)\n",
2564                          n, mem->nents);
2565                 return n < 0 ? n : -EINVAL;
2566         }
2567
2568         isert_dbg("Use fr_desc %p sg_nents %d offset %u\n",
2569                   fr_desc, mem->nents, mem->offset);
2570
2571         reg_wr.wr.next = NULL;
2572         reg_wr.wr.opcode = IB_WR_REG_MR;
2573         reg_wr.wr.wr_id = ISER_FASTREG_LI_WRID;
2574         reg_wr.wr.send_flags = 0;
2575         reg_wr.wr.num_sge = 0;
2576         reg_wr.mr = mr;
2577         reg_wr.key = mr->lkey;
2578         reg_wr.access = IB_ACCESS_LOCAL_WRITE;
2579
2580         if (!wr)
2581                 wr = &reg_wr.wr;
2582         else
2583                 wr->next = &reg_wr.wr;
2584
2585         ret = ib_post_send(isert_conn->qp, wr, &bad_wr);
2586         if (ret) {
2587                 isert_err("fast registration failed, ret:%d\n", ret);
2588                 return ret;
2589         }
2590         fr_desc->ind &= ~ind;
2591
2592         sge->lkey = mr->lkey;
2593         sge->addr = mr->iova;
2594         sge->length = mr->length;
2595
2596         isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2597                   sge->addr, sge->length, sge->lkey);
2598
2599         return ret;
2600 }
2601
2602 static inline void
2603 isert_set_dif_domain(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs,
2604                      struct ib_sig_domain *domain)
2605 {
2606         domain->sig_type = IB_SIG_TYPE_T10_DIF;
2607         domain->sig.dif.bg_type = IB_T10DIF_CRC;
2608         domain->sig.dif.pi_interval = se_cmd->se_dev->dev_attrib.block_size;
2609         domain->sig.dif.ref_tag = se_cmd->reftag_seed;
2610         /*
2611          * At the moment we hard code those, but if in the future
2612          * the target core would like to use it, we will take it
2613          * from se_cmd.
2614          */
2615         domain->sig.dif.apptag_check_mask = 0xffff;
2616         domain->sig.dif.app_escape = true;
2617         domain->sig.dif.ref_escape = true;
2618         if (se_cmd->prot_type == TARGET_DIF_TYPE1_PROT ||
2619             se_cmd->prot_type == TARGET_DIF_TYPE2_PROT)
2620                 domain->sig.dif.ref_remap = true;
2621 };
2622
2623 static int
2624 isert_set_sig_attrs(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs)
2625 {
2626         switch (se_cmd->prot_op) {
2627         case TARGET_PROT_DIN_INSERT:
2628         case TARGET_PROT_DOUT_STRIP:
2629                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
2630                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2631                 break;
2632         case TARGET_PROT_DOUT_INSERT:
2633         case TARGET_PROT_DIN_STRIP:
2634                 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
2635                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2636                 break;
2637         case TARGET_PROT_DIN_PASS:
2638         case TARGET_PROT_DOUT_PASS:
2639                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2640                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2641                 break;
2642         default:
2643                 isert_err("Unsupported PI operation %d\n", se_cmd->prot_op);
2644                 return -EINVAL;
2645         }
2646
2647         return 0;
2648 }
2649
2650 static inline u8
2651 isert_set_prot_checks(u8 prot_checks)
2652 {
2653         return (prot_checks & TARGET_DIF_CHECK_GUARD  ? 0xc0 : 0) |
2654                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x30 : 0) |
2655                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x0f : 0);
2656 }
2657
2658 static int
2659 isert_reg_sig_mr(struct isert_conn *isert_conn,
2660                  struct se_cmd *se_cmd,
2661                  struct isert_rdma_wr *rdma_wr,
2662                  struct fast_reg_descriptor *fr_desc)
2663 {
2664         struct ib_sig_handover_wr sig_wr;
2665         struct ib_send_wr inv_wr, *bad_wr, *wr = NULL;
2666         struct pi_context *pi_ctx = fr_desc->pi_ctx;
2667         struct ib_sig_attrs sig_attrs;
2668         int ret;
2669
2670         memset(&sig_attrs, 0, sizeof(sig_attrs));
2671         ret = isert_set_sig_attrs(se_cmd, &sig_attrs);
2672         if (ret)
2673                 goto err;
2674
2675         sig_attrs.check_mask = isert_set_prot_checks(se_cmd->prot_checks);
2676
2677         if (!(fr_desc->ind & ISERT_SIG_KEY_VALID)) {
2678                 isert_inv_rkey(&inv_wr, pi_ctx->sig_mr);
2679                 wr = &inv_wr;
2680         }
2681
2682         memset(&sig_wr, 0, sizeof(sig_wr));
2683         sig_wr.wr.opcode = IB_WR_REG_SIG_MR;
2684         sig_wr.wr.wr_id = ISER_FASTREG_LI_WRID;
2685         sig_wr.wr.sg_list = &rdma_wr->ib_sg[DATA];
2686         sig_wr.wr.num_sge = 1;
2687         sig_wr.access_flags = IB_ACCESS_LOCAL_WRITE;
2688         sig_wr.sig_attrs = &sig_attrs;
2689         sig_wr.sig_mr = pi_ctx->sig_mr;
2690         if (se_cmd->t_prot_sg)
2691                 sig_wr.prot = &rdma_wr->ib_sg[PROT];
2692
2693         if (!wr)
2694                 wr = &sig_wr.wr;
2695         else
2696                 wr->next = &sig_wr.wr;
2697
2698         ret = ib_post_send(isert_conn->qp, wr, &bad_wr);
2699         if (ret) {
2700                 isert_err("fast registration failed, ret:%d\n", ret);
2701                 goto err;
2702         }
2703         fr_desc->ind &= ~ISERT_SIG_KEY_VALID;
2704
2705         rdma_wr->ib_sg[SIG].lkey = pi_ctx->sig_mr->lkey;
2706         rdma_wr->ib_sg[SIG].addr = 0;
2707         rdma_wr->ib_sg[SIG].length = se_cmd->data_length;
2708         if (se_cmd->prot_op != TARGET_PROT_DIN_STRIP &&
2709             se_cmd->prot_op != TARGET_PROT_DOUT_INSERT)
2710                 /*
2711                  * We have protection guards on the wire
2712                  * so we need to set a larget transfer
2713                  */
2714                 rdma_wr->ib_sg[SIG].length += se_cmd->prot_length;
2715
2716         isert_dbg("sig_sge: addr: 0x%llx  length: %u lkey: %x\n",
2717                   rdma_wr->ib_sg[SIG].addr, rdma_wr->ib_sg[SIG].length,
2718                   rdma_wr->ib_sg[SIG].lkey);
2719 err:
2720         return ret;
2721 }
2722
2723 static int
2724 isert_handle_prot_cmd(struct isert_conn *isert_conn,
2725                       struct isert_cmd *isert_cmd,
2726                       struct isert_rdma_wr *wr)
2727 {
2728         struct isert_device *device = isert_conn->device;
2729         struct se_cmd *se_cmd = &isert_cmd->iscsi_cmd->se_cmd;
2730         int ret;
2731
2732         if (!wr->fr_desc->pi_ctx) {
2733                 ret = isert_create_pi_ctx(wr->fr_desc,
2734                                           device->ib_device,
2735                                           device->pd);
2736                 if (ret) {
2737                         isert_err("conn %p failed to allocate pi_ctx\n",
2738                                   isert_conn);
2739                         return ret;
2740                 }
2741         }
2742
2743         if (se_cmd->t_prot_sg) {
2744                 ret = isert_map_data_buf(isert_conn, isert_cmd,
2745                                          se_cmd->t_prot_sg,
2746                                          se_cmd->t_prot_nents,
2747                                          se_cmd->prot_length,
2748                                          0, wr->iser_ib_op, &wr->prot);
2749                 if (ret) {
2750                         isert_err("conn %p failed to map protection buffer\n",
2751                                   isert_conn);
2752                         return ret;
2753                 }
2754
2755                 memset(&wr->ib_sg[PROT], 0, sizeof(wr->ib_sg[PROT]));
2756                 ret = isert_fast_reg_mr(isert_conn, wr->fr_desc, &wr->prot,
2757                                         ISERT_PROT_KEY_VALID, &wr->ib_sg[PROT]);
2758                 if (ret) {
2759                         isert_err("conn %p failed to fast reg mr\n",
2760                                   isert_conn);
2761                         goto unmap_prot_cmd;
2762                 }
2763         }
2764
2765         ret = isert_reg_sig_mr(isert_conn, se_cmd, wr, wr->fr_desc);
2766         if (ret) {
2767                 isert_err("conn %p failed to fast reg mr\n",
2768                           isert_conn);
2769                 goto unmap_prot_cmd;
2770         }
2771         wr->fr_desc->ind |= ISERT_PROTECTED;
2772
2773         return 0;
2774
2775 unmap_prot_cmd:
2776         if (se_cmd->t_prot_sg)
2777                 isert_unmap_data_buf(isert_conn, &wr->prot);
2778
2779         return ret;
2780 }
2781
2782 static int
2783 isert_reg_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2784                struct isert_rdma_wr *wr)
2785 {
2786         struct se_cmd *se_cmd = &cmd->se_cmd;
2787         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2788         struct isert_conn *isert_conn = conn->context;
2789         struct fast_reg_descriptor *fr_desc = NULL;
2790         struct ib_rdma_wr *rdma_wr;
2791         struct ib_sge *ib_sg;
2792         u32 offset;
2793         int ret = 0;
2794         unsigned long flags;
2795
2796         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2797
2798         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2799         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2800                                  se_cmd->t_data_nents, se_cmd->data_length,
2801                                  offset, wr->iser_ib_op, &wr->data);
2802         if (ret)
2803                 return ret;
2804
2805         if (wr->data.dma_nents != 1 || isert_prot_cmd(isert_conn, se_cmd)) {
2806                 spin_lock_irqsave(&isert_conn->pool_lock, flags);
2807                 fr_desc = list_first_entry(&isert_conn->fr_pool,
2808                                            struct fast_reg_descriptor, list);
2809                 list_del(&fr_desc->list);
2810                 spin_unlock_irqrestore(&isert_conn->pool_lock, flags);
2811                 wr->fr_desc = fr_desc;
2812         }
2813
2814         ret = isert_fast_reg_mr(isert_conn, fr_desc, &wr->data,
2815                                 ISERT_DATA_KEY_VALID, &wr->ib_sg[DATA]);
2816         if (ret)
2817                 goto unmap_cmd;
2818
2819         if (isert_prot_cmd(isert_conn, se_cmd)) {
2820                 ret = isert_handle_prot_cmd(isert_conn, isert_cmd, wr);
2821                 if (ret)
2822                         goto unmap_cmd;
2823
2824                 ib_sg = &wr->ib_sg[SIG];
2825         } else {
2826                 ib_sg = &wr->ib_sg[DATA];
2827         }
2828
2829         memcpy(&wr->s_ib_sge, ib_sg, sizeof(*ib_sg));
2830         wr->ib_sge = &wr->s_ib_sge;
2831         wr->rdma_wr_num = 1;
2832         memset(&wr->s_rdma_wr, 0, sizeof(wr->s_rdma_wr));
2833         wr->rdma_wr = &wr->s_rdma_wr;
2834         wr->isert_cmd = isert_cmd;
2835
2836         rdma_wr = &isert_cmd->rdma_wr.s_rdma_wr;
2837         rdma_wr->wr.sg_list = &wr->s_ib_sge;
2838         rdma_wr->wr.num_sge = 1;
2839         rdma_wr->wr.wr_id = (uintptr_t)&isert_cmd->tx_desc;
2840         if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2841                 rdma_wr->wr.opcode = IB_WR_RDMA_WRITE;
2842                 rdma_wr->remote_addr = isert_cmd->read_va;
2843                 rdma_wr->rkey = isert_cmd->read_stag;
2844                 rdma_wr->wr.send_flags = !isert_prot_cmd(isert_conn, se_cmd) ?
2845                                       0 : IB_SEND_SIGNALED;
2846         } else {
2847                 rdma_wr->wr.opcode = IB_WR_RDMA_READ;
2848                 rdma_wr->remote_addr = isert_cmd->write_va;
2849                 rdma_wr->rkey = isert_cmd->write_stag;
2850                 rdma_wr->wr.send_flags = IB_SEND_SIGNALED;
2851         }
2852
2853         return 0;
2854
2855 unmap_cmd:
2856         if (fr_desc) {
2857                 spin_lock_irqsave(&isert_conn->pool_lock, flags);
2858                 list_add_tail(&fr_desc->list, &isert_conn->fr_pool);
2859                 spin_unlock_irqrestore(&isert_conn->pool_lock, flags);
2860         }
2861         isert_unmap_data_buf(isert_conn, &wr->data);
2862
2863         return ret;
2864 }
2865
2866 static int
2867 isert_put_datain(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2868 {
2869         struct se_cmd *se_cmd = &cmd->se_cmd;
2870         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2871         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2872         struct isert_conn *isert_conn = conn->context;
2873         struct isert_device *device = isert_conn->device;
2874         struct ib_send_wr *wr_failed;
2875         int rc;
2876
2877         isert_dbg("Cmd: %p RDMA_WRITE data_length: %u\n",
2878                  isert_cmd, se_cmd->data_length);
2879
2880         wr->iser_ib_op = ISER_IB_RDMA_WRITE;
2881         rc = device->reg_rdma_mem(conn, cmd, wr);
2882         if (rc) {
2883                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2884                 return rc;
2885         }
2886
2887         if (!isert_prot_cmd(isert_conn, se_cmd)) {
2888                 /*
2889                  * Build isert_conn->tx_desc for iSCSI response PDU and attach
2890                  */
2891                 isert_create_send_desc(isert_conn, isert_cmd,
2892                                        &isert_cmd->tx_desc);
2893                 iscsit_build_rsp_pdu(cmd, conn, true, (struct iscsi_scsi_rsp *)
2894                                      &isert_cmd->tx_desc.iscsi_header);
2895                 isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2896                 isert_init_send_wr(isert_conn, isert_cmd,
2897                                    &isert_cmd->tx_desc.send_wr);
2898                 isert_cmd->rdma_wr.s_rdma_wr.wr.next = &isert_cmd->tx_desc.send_wr;
2899                 wr->rdma_wr_num += 1;
2900
2901                 rc = isert_post_recv(isert_conn, isert_cmd->rx_desc);
2902                 if (rc) {
2903                         isert_err("ib_post_recv failed with %d\n", rc);
2904                         return rc;
2905                 }
2906         }
2907
2908         rc = ib_post_send(isert_conn->qp, &wr->rdma_wr->wr, &wr_failed);
2909         if (rc)
2910                 isert_warn("ib_post_send() failed for IB_WR_RDMA_WRITE\n");
2911
2912         if (!isert_prot_cmd(isert_conn, se_cmd))
2913                 isert_dbg("Cmd: %p posted RDMA_WRITE + Response for iSER Data "
2914                          "READ\n", isert_cmd);
2915         else
2916                 isert_dbg("Cmd: %p posted RDMA_WRITE for iSER Data READ\n",
2917                          isert_cmd);
2918
2919         return 1;
2920 }
2921
2922 static int
2923 isert_get_dataout(struct iscsi_conn *conn, struct iscsi_cmd *cmd, bool recovery)
2924 {
2925         struct se_cmd *se_cmd = &cmd->se_cmd;
2926         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2927         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2928         struct isert_conn *isert_conn = conn->context;
2929         struct isert_device *device = isert_conn->device;
2930         struct ib_send_wr *wr_failed;
2931         int rc;
2932
2933         isert_dbg("Cmd: %p RDMA_READ data_length: %u write_data_done: %u\n",
2934                  isert_cmd, se_cmd->data_length, cmd->write_data_done);
2935         wr->iser_ib_op = ISER_IB_RDMA_READ;
2936         rc = device->reg_rdma_mem(conn, cmd, wr);
2937         if (rc) {
2938                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2939                 return rc;
2940         }
2941
2942         rc = ib_post_send(isert_conn->qp, &wr->rdma_wr->wr, &wr_failed);
2943         if (rc)
2944                 isert_warn("ib_post_send() failed for IB_WR_RDMA_READ\n");
2945
2946         isert_dbg("Cmd: %p posted RDMA_READ memory for ISER Data WRITE\n",
2947                  isert_cmd);
2948
2949         return 0;
2950 }
2951
2952 static int
2953 isert_immediate_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2954 {
2955         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2956         int ret = 0;
2957
2958         switch (state) {
2959         case ISTATE_REMOVE:
2960                 spin_lock_bh(&conn->cmd_lock);
2961                 list_del_init(&cmd->i_conn_node);
2962                 spin_unlock_bh(&conn->cmd_lock);
2963                 isert_put_cmd(isert_cmd, true);
2964                 break;
2965         case ISTATE_SEND_NOPIN_WANT_RESPONSE:
2966                 ret = isert_put_nopin(cmd, conn, false);
2967                 break;
2968         default:
2969                 isert_err("Unknown immediate state: 0x%02x\n", state);
2970                 ret = -EINVAL;
2971                 break;
2972         }
2973
2974         return ret;
2975 }
2976
2977 static int
2978 isert_response_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2979 {
2980         struct isert_conn *isert_conn = conn->context;
2981         int ret;
2982
2983         switch (state) {
2984         case ISTATE_SEND_LOGOUTRSP:
2985                 ret = isert_put_logout_rsp(cmd, conn);
2986                 if (!ret)
2987                         isert_conn->logout_posted = true;
2988                 break;
2989         case ISTATE_SEND_NOPIN:
2990                 ret = isert_put_nopin(cmd, conn, true);
2991                 break;
2992         case ISTATE_SEND_TASKMGTRSP:
2993                 ret = isert_put_tm_rsp(cmd, conn);
2994                 break;
2995         case ISTATE_SEND_REJECT:
2996                 ret = isert_put_reject(cmd, conn);
2997                 break;
2998         case ISTATE_SEND_TEXTRSP:
2999                 ret = isert_put_text_rsp(cmd, conn);
3000                 break;
3001         case ISTATE_SEND_STATUS:
3002                 /*
3003                  * Special case for sending non GOOD SCSI status from TX thread
3004                  * context during pre se_cmd excecution failure.
3005                  */
3006                 ret = isert_put_response(conn, cmd);
3007                 break;
3008         default:
3009                 isert_err("Unknown response state: 0x%02x\n", state);
3010                 ret = -EINVAL;
3011                 break;
3012         }
3013
3014         return ret;
3015 }
3016
3017 struct rdma_cm_id *
3018 isert_setup_id(struct isert_np *isert_np)
3019 {
3020         struct iscsi_np *np = isert_np->np;
3021         struct rdma_cm_id *id;
3022         struct sockaddr *sa;
3023         int ret;
3024
3025         sa = (struct sockaddr *)&np->np_sockaddr;
3026         isert_dbg("ksockaddr: %p, sa: %p\n", &np->np_sockaddr, sa);
3027
3028         id = rdma_create_id(&init_net, isert_cma_handler, isert_np,
3029                             RDMA_PS_TCP, IB_QPT_RC);
3030         if (IS_ERR(id)) {
3031                 isert_err("rdma_create_id() failed: %ld\n", PTR_ERR(id));
3032                 ret = PTR_ERR(id);
3033                 goto out;
3034         }
3035         isert_dbg("id %p context %p\n", id, id->context);
3036
3037         ret = rdma_bind_addr(id, sa);
3038         if (ret) {
3039                 isert_err("rdma_bind_addr() failed: %d\n", ret);
3040                 goto out_id;
3041         }
3042
3043         ret = rdma_listen(id, 0);
3044         if (ret) {
3045                 isert_err("rdma_listen() failed: %d\n", ret);
3046                 goto out_id;
3047         }
3048
3049         return id;
3050 out_id:
3051         rdma_destroy_id(id);
3052 out:
3053         return ERR_PTR(ret);
3054 }
3055
3056 static int
3057 isert_setup_np(struct iscsi_np *np,
3058                struct sockaddr_storage *ksockaddr)
3059 {
3060         struct isert_np *isert_np;
3061         struct rdma_cm_id *isert_lid;
3062         int ret;
3063
3064         isert_np = kzalloc(sizeof(struct isert_np), GFP_KERNEL);
3065         if (!isert_np) {
3066                 isert_err("Unable to allocate struct isert_np\n");
3067                 return -ENOMEM;
3068         }
3069         sema_init(&isert_np->sem, 0);
3070         mutex_init(&isert_np->mutex);
3071         INIT_LIST_HEAD(&isert_np->accepted);
3072         INIT_LIST_HEAD(&isert_np->pending);
3073         isert_np->np = np;
3074
3075         /*
3076          * Setup the np->np_sockaddr from the passed sockaddr setup
3077          * in iscsi_target_configfs.c code..
3078          */
3079         memcpy(&np->np_sockaddr, ksockaddr,
3080                sizeof(struct sockaddr_storage));
3081
3082         isert_lid = isert_setup_id(isert_np);
3083         if (IS_ERR(isert_lid)) {
3084                 ret = PTR_ERR(isert_lid);
3085                 goto out;
3086         }
3087
3088         isert_np->cm_id = isert_lid;
3089         np->np_context = isert_np;
3090
3091         return 0;
3092
3093 out:
3094         kfree(isert_np);
3095
3096         return ret;
3097 }
3098
3099 static int
3100 isert_rdma_accept(struct isert_conn *isert_conn)
3101 {
3102         struct rdma_cm_id *cm_id = isert_conn->cm_id;
3103         struct rdma_conn_param cp;
3104         int ret;
3105         struct iser_cm_hdr rsp_hdr;
3106
3107         memset(&cp, 0, sizeof(struct rdma_conn_param));
3108         cp.initiator_depth = isert_conn->initiator_depth;
3109         cp.retry_count = 7;
3110         cp.rnr_retry_count = 7;
3111
3112         memset(&rsp_hdr, 0, sizeof(rsp_hdr));
3113         rsp_hdr.flags = ISERT_ZBVA_NOT_USED;
3114         if (!isert_conn->snd_w_inv)
3115                 rsp_hdr.flags = rsp_hdr.flags | ISERT_SEND_W_INV_NOT_USED;
3116         cp.private_data = (void *)&rsp_hdr;
3117         cp.private_data_len = sizeof(rsp_hdr);
3118
3119         ret = rdma_accept(cm_id, &cp);
3120         if (ret) {
3121                 isert_err("rdma_accept() failed with: %d\n", ret);
3122                 return ret;
3123         }
3124
3125         return 0;
3126 }
3127
3128 static int
3129 isert_get_login_rx(struct iscsi_conn *conn, struct iscsi_login *login)
3130 {
3131         struct isert_conn *isert_conn = conn->context;
3132         int ret;
3133
3134         isert_info("before login_req comp conn: %p\n", isert_conn);
3135         ret = wait_for_completion_interruptible(&isert_conn->login_req_comp);
3136         if (ret) {
3137                 isert_err("isert_conn %p interrupted before got login req\n",
3138                           isert_conn);
3139                 return ret;
3140         }
3141         reinit_completion(&isert_conn->login_req_comp);
3142
3143         /*
3144          * For login requests after the first PDU, isert_rx_login_req() will
3145          * kick schedule_delayed_work(&conn->login_work) as the packet is
3146          * received, which turns this callback from iscsi_target_do_login_rx()
3147          * into a NOP.
3148          */
3149         if (!login->first_request)
3150                 return 0;
3151
3152         isert_rx_login_req(isert_conn);
3153
3154         isert_info("before login_comp conn: %p\n", conn);
3155         ret = wait_for_completion_interruptible(&isert_conn->login_comp);
3156         if (ret)
3157                 return ret;
3158
3159         isert_info("processing login->req: %p\n", login->req);
3160
3161         return 0;
3162 }
3163
3164 static void
3165 isert_set_conn_info(struct iscsi_np *np, struct iscsi_conn *conn,
3166                     struct isert_conn *isert_conn)
3167 {
3168         struct rdma_cm_id *cm_id = isert_conn->cm_id;
3169         struct rdma_route *cm_route = &cm_id->route;
3170
3171         conn->login_family = np->np_sockaddr.ss_family;
3172
3173         conn->login_sockaddr = cm_route->addr.dst_addr;
3174         conn->local_sockaddr = cm_route->addr.src_addr;
3175 }
3176
3177 static int
3178 isert_accept_np(struct iscsi_np *np, struct iscsi_conn *conn)
3179 {
3180         struct isert_np *isert_np = np->np_context;
3181         struct isert_conn *isert_conn;
3182         int ret;
3183
3184 accept_wait:
3185         ret = down_interruptible(&isert_np->sem);
3186         if (ret)
3187                 return -ENODEV;
3188
3189         spin_lock_bh(&np->np_thread_lock);
3190         if (np->np_thread_state >= ISCSI_NP_THREAD_RESET) {
3191                 spin_unlock_bh(&np->np_thread_lock);
3192                 isert_dbg("np_thread_state %d\n",
3193                          np->np_thread_state);
3194                 /**
3195                  * No point in stalling here when np_thread
3196                  * is in state RESET/SHUTDOWN/EXIT - bail
3197                  **/
3198                 return -ENODEV;
3199         }
3200         spin_unlock_bh(&np->np_thread_lock);
3201
3202         mutex_lock(&isert_np->mutex);
3203         if (list_empty(&isert_np->pending)) {
3204                 mutex_unlock(&isert_np->mutex);
3205                 goto accept_wait;
3206         }
3207         isert_conn = list_first_entry(&isert_np->pending,
3208                         struct isert_conn, node);
3209         list_del_init(&isert_conn->node);
3210         mutex_unlock(&isert_np->mutex);
3211
3212         conn->context = isert_conn;
3213         isert_conn->conn = conn;
3214
3215         isert_set_conn_info(np, conn, isert_conn);
3216
3217         isert_dbg("Processing isert_conn: %p\n", isert_conn);
3218
3219         return 0;
3220 }
3221
3222 static void
3223 isert_free_np(struct iscsi_np *np)
3224 {
3225         struct isert_np *isert_np = np->np_context;
3226         struct isert_conn *isert_conn, *n;
3227
3228         if (isert_np->cm_id)
3229                 rdma_destroy_id(isert_np->cm_id);
3230
3231         /*
3232          * FIXME: At this point we don't have a good way to insure
3233          * that at this point we don't have hanging connections that
3234          * completed RDMA establishment but didn't start iscsi login
3235          * process. So work-around this by cleaning up what ever piled
3236          * up in accepted and pending lists.
3237          */
3238         mutex_lock(&isert_np->mutex);
3239         if (!list_empty(&isert_np->pending)) {
3240                 isert_info("Still have isert pending connections\n");
3241                 list_for_each_entry_safe(isert_conn, n,
3242                                          &isert_np->pending,
3243                                          node) {
3244                         isert_info("cleaning isert_conn %p state (%d)\n",
3245                                    isert_conn, isert_conn->state);
3246                         isert_connect_release(isert_conn);
3247                 }
3248         }
3249
3250         if (!list_empty(&isert_np->accepted)) {
3251                 isert_info("Still have isert accepted connections\n");
3252                 list_for_each_entry_safe(isert_conn, n,
3253                                          &isert_np->accepted,
3254                                          node) {
3255                         isert_info("cleaning isert_conn %p state (%d)\n",
3256                                    isert_conn, isert_conn->state);
3257                         isert_connect_release(isert_conn);
3258                 }
3259         }
3260         mutex_unlock(&isert_np->mutex);
3261
3262         np->np_context = NULL;
3263         kfree(isert_np);
3264 }
3265
3266 static void isert_release_work(struct work_struct *work)
3267 {
3268         struct isert_conn *isert_conn = container_of(work,
3269                                                      struct isert_conn,
3270                                                      release_work);
3271
3272         isert_info("Starting release conn %p\n", isert_conn);
3273
3274         wait_for_completion(&isert_conn->wait);
3275
3276         mutex_lock(&isert_conn->mutex);
3277         isert_conn->state = ISER_CONN_DOWN;
3278         mutex_unlock(&isert_conn->mutex);
3279
3280         isert_info("Destroying conn %p\n", isert_conn);
3281         isert_put_conn(isert_conn);
3282 }
3283
3284 static void
3285 isert_wait4logout(struct isert_conn *isert_conn)
3286 {
3287         struct iscsi_conn *conn = isert_conn->conn;
3288
3289         isert_info("conn %p\n", isert_conn);
3290
3291         if (isert_conn->logout_posted) {
3292                 isert_info("conn %p wait for conn_logout_comp\n", isert_conn);
3293                 wait_for_completion_timeout(&conn->conn_logout_comp,
3294                                             SECONDS_FOR_LOGOUT_COMP * HZ);
3295         }
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 static void
3310 isert_wait4flush(struct isert_conn *isert_conn)
3311 {
3312         struct ib_recv_wr *bad_wr;
3313
3314         isert_info("conn %p\n", isert_conn);
3315
3316         init_completion(&isert_conn->wait_comp_err);
3317         isert_conn->beacon.wr_id = ISER_BEACON_WRID;
3318         /* post an indication that all flush errors were consumed */
3319         if (ib_post_recv(isert_conn->qp, &isert_conn->beacon, &bad_wr)) {
3320                 isert_err("conn %p failed to post beacon", isert_conn);
3321                 return;
3322         }
3323
3324         wait_for_completion(&isert_conn->wait_comp_err);
3325 }
3326
3327 /**
3328  * isert_put_unsol_pending_cmds() - Drop commands waiting for
3329  *     unsolicitate dataout
3330  * @conn:    iscsi connection
3331  *
3332  * We might still have commands that are waiting for unsolicited
3333  * dataouts messages. We must put the extra reference on those
3334  * before blocking on the target_wait_for_session_cmds
3335  */
3336 static void
3337 isert_put_unsol_pending_cmds(struct iscsi_conn *conn)
3338 {
3339         struct iscsi_cmd *cmd, *tmp;
3340         static LIST_HEAD(drop_cmd_list);
3341
3342         spin_lock_bh(&conn->cmd_lock);
3343         list_for_each_entry_safe(cmd, tmp, &conn->conn_cmd_list, i_conn_node) {
3344                 if ((cmd->cmd_flags & ICF_NON_IMMEDIATE_UNSOLICITED_DATA) &&
3345                     (cmd->write_data_done < conn->sess->sess_ops->FirstBurstLength) &&
3346                     (cmd->write_data_done < cmd->se_cmd.data_length))
3347                         list_move_tail(&cmd->i_conn_node, &drop_cmd_list);
3348         }
3349         spin_unlock_bh(&conn->cmd_lock);
3350
3351         list_for_each_entry_safe(cmd, tmp, &drop_cmd_list, i_conn_node) {
3352                 list_del_init(&cmd->i_conn_node);
3353                 if (cmd->i_state != ISTATE_REMOVE) {
3354                         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
3355
3356                         isert_info("conn %p dropping cmd %p\n", conn, cmd);
3357                         isert_put_cmd(isert_cmd, true);
3358                 }
3359         }
3360 }
3361
3362 static void isert_wait_conn(struct iscsi_conn *conn)
3363 {
3364         struct isert_conn *isert_conn = conn->context;
3365
3366         isert_info("Starting conn %p\n", isert_conn);
3367
3368         mutex_lock(&isert_conn->mutex);
3369         /*
3370          * Only wait for wait_comp_err if the isert_conn made it
3371          * into full feature phase..
3372          */
3373         if (isert_conn->state == ISER_CONN_INIT) {
3374                 mutex_unlock(&isert_conn->mutex);
3375                 return;
3376         }
3377         isert_conn_terminate(isert_conn);
3378         mutex_unlock(&isert_conn->mutex);
3379
3380         isert_wait4flush(isert_conn);
3381         isert_put_unsol_pending_cmds(conn);
3382         isert_wait4cmds(conn);
3383         isert_wait4logout(isert_conn);
3384
3385         queue_work(isert_release_wq, &isert_conn->release_work);
3386 }
3387
3388 static void isert_free_conn(struct iscsi_conn *conn)
3389 {
3390         struct isert_conn *isert_conn = conn->context;
3391
3392         isert_wait4flush(isert_conn);
3393         isert_put_conn(isert_conn);
3394 }
3395
3396 static struct iscsit_transport iser_target_transport = {
3397         .name                   = "IB/iSER",
3398         .transport_type         = ISCSI_INFINIBAND,
3399         .priv_size              = sizeof(struct isert_cmd),
3400         .owner                  = THIS_MODULE,
3401         .iscsit_setup_np        = isert_setup_np,
3402         .iscsit_accept_np       = isert_accept_np,
3403         .iscsit_free_np         = isert_free_np,
3404         .iscsit_wait_conn       = isert_wait_conn,
3405         .iscsit_free_conn       = isert_free_conn,
3406         .iscsit_get_login_rx    = isert_get_login_rx,
3407         .iscsit_put_login_tx    = isert_put_login_tx,
3408         .iscsit_immediate_queue = isert_immediate_queue,
3409         .iscsit_response_queue  = isert_response_queue,
3410         .iscsit_get_dataout     = isert_get_dataout,
3411         .iscsit_queue_data_in   = isert_put_datain,
3412         .iscsit_queue_status    = isert_put_response,
3413         .iscsit_aborted_task    = isert_aborted_task,
3414         .iscsit_get_sup_prot_ops = isert_get_sup_prot_ops,
3415 };
3416
3417 static int __init isert_init(void)
3418 {
3419         int ret;
3420
3421         isert_comp_wq = alloc_workqueue("isert_comp_wq",
3422                                         WQ_UNBOUND | WQ_HIGHPRI, 0);
3423         if (!isert_comp_wq) {
3424                 isert_err("Unable to allocate isert_comp_wq\n");
3425                 ret = -ENOMEM;
3426                 return -ENOMEM;
3427         }
3428
3429         isert_release_wq = alloc_workqueue("isert_release_wq", WQ_UNBOUND,
3430                                         WQ_UNBOUND_MAX_ACTIVE);
3431         if (!isert_release_wq) {
3432                 isert_err("Unable to allocate isert_release_wq\n");
3433                 ret = -ENOMEM;
3434                 goto destroy_comp_wq;
3435         }
3436
3437         iscsit_register_transport(&iser_target_transport);
3438         isert_info("iSER_TARGET[0] - Loaded iser_target_transport\n");
3439
3440         return 0;
3441
3442 destroy_comp_wq:
3443         destroy_workqueue(isert_comp_wq);
3444
3445         return ret;
3446 }
3447
3448 static void __exit isert_exit(void)
3449 {
3450         flush_scheduled_work();
3451         destroy_workqueue(isert_release_wq);
3452         destroy_workqueue(isert_comp_wq);
3453         iscsit_unregister_transport(&iser_target_transport);
3454         isert_info("iSER_TARGET[0] - Released iser_target_transport\n");
3455 }
3456
3457 MODULE_DESCRIPTION("iSER-Target for mainline target infrastructure");
3458 MODULE_VERSION("1.0");
3459 MODULE_AUTHOR("nab@Linux-iSCSI.org");
3460 MODULE_LICENSE("GPL");
3461
3462 module_init(isert_init);
3463 module_exit(isert_exit);