drivers/infiniband/ulp/isert/ib_isert.c

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