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         /*
 161          * FIXME: Use devattr.max_sge - 2 for max_send_sge as
 162          * work-around for RDMA_READs with ConnectX-2.
 163          *
 164          * Also, still make sure to have at least two SGEs for
 165          * outgoing control PDU responses.
 166          */
 167         attr.cap.max_send_sge = max(2, device->dev_attr.max_sge - 2);
 168         isert_conn->max_sge = attr.cap.max_send_sge;
 169
 170         attr.cap.max_recv_sge = 1;
 171         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 172         attr.qp_type = IB_QPT_RC;
 173         if (device->pi_capable)
 174                 attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
 175
 176         ret = rdma_create_qp(cma_id, device->pd, &attr);
 177         if (ret) {
 178                 isert_err("rdma_create_qp failed for cma_id %d\n", ret);
 179                 return ERR_PTR(ret);
 180         }
 181
 182         return cma_id->qp;
 183 }
 184
 185 static int
 186 isert_conn_setup_qp(struct isert_conn *isert_conn, struct rdma_cm_id *cma_id)
 187 {
 188         struct isert_comp *comp;
 189         int ret;
 190
 191         comp = isert_comp_get(isert_conn);
 192         isert_conn->qp = isert_create_qp(isert_conn, comp, cma_id);
 193         if (IS_ERR(isert_conn->qp)) {
 194                 ret = PTR_ERR(isert_conn->qp);
 195                 goto err;
 196         }
 197
 198         return 0;
 199 err:
 200         isert_comp_put(comp);
 201         return ret;
 202 }
 203
 204 static void
 205 isert_cq_event_callback(struct ib_event *e, void *context)
 206 {
 207         isert_dbg("event: %d\n", e->event);
 208 }
 209
 210 static int
 211 isert_alloc_rx_descriptors(struct isert_conn *isert_conn)
 212 {
 213         struct isert_device *device = isert_conn->device;
 214         struct ib_device *ib_dev = device->ib_device;
 215         struct iser_rx_desc *rx_desc;
 216         struct ib_sge *rx_sg;
 217         u64 dma_addr;
 218         int i, j;
 219
 220         isert_conn->rx_descs = kzalloc(ISERT_QP_MAX_RECV_DTOS *
 221                                 sizeof(struct iser_rx_desc), GFP_KERNEL);
 222         if (!isert_conn->rx_descs)
 223                 goto fail;
 224
 225         rx_desc = isert_conn->rx_descs;
 226
 227         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 228                 dma_addr = ib_dma_map_single(ib_dev, (void *)rx_desc,
 229                                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 230                 if (ib_dma_mapping_error(ib_dev, dma_addr))
 231                         goto dma_map_fail;
 232
 233                 rx_desc->dma_addr = dma_addr;
 234
 235                 rx_sg = &rx_desc->rx_sg;
 236                 rx_sg->addr = rx_desc->dma_addr;
 237                 rx_sg->length = ISER_RX_PAYLOAD_SIZE;
 238                 rx_sg->lkey = device->pd->local_dma_lkey;
 239         }
 240
 241         return 0;
 242
 243 dma_map_fail:
 244         rx_desc = isert_conn->rx_descs;
 245         for (j = 0; j < i; j++, rx_desc++) {
 246                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 247                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 248         }
 249         kfree(isert_conn->rx_descs);
 250         isert_conn->rx_descs = NULL;
 251 fail:
 252         isert_err("conn %p failed to allocate rx descriptors\n", isert_conn);
 253
 254         return -ENOMEM;
 255 }
 256
 257 static void
 258 isert_free_rx_descriptors(struct isert_conn *isert_conn)
 259 {
 260         struct ib_device *ib_dev = isert_conn->device->ib_device;
 261         struct iser_rx_desc *rx_desc;
 262         int i;
 263
 264         if (!isert_conn->rx_descs)
 265                 return;
 266
 267         rx_desc = isert_conn->rx_descs;
 268         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 269                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 270                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 271         }
 272
 273         kfree(isert_conn->rx_descs);
 274         isert_conn->rx_descs = NULL;
 275 }
 276
 277 static void isert_cq_work(struct work_struct *);
 278 static void isert_cq_callback(struct ib_cq *, void *);
 279
 280 static void
 281 isert_free_comps(struct isert_device *device)
 282 {
 283         int i;
 284
 285         for (i = 0; i < device->comps_used; i++) {
 286                 struct isert_comp *comp = &device->comps[i];
 287
 288                 if (comp->cq) {
 289                         cancel_work_sync(&comp->work);
 290                         ib_destroy_cq(comp->cq);
 291                 }
 292         }
 293         kfree(device->comps);
 294 }
 295
 296 static int
 297 isert_alloc_comps(struct isert_device *device,
 298                   struct ib_device_attr *attr)
 299 {
 300         int i, max_cqe, ret = 0;
 301
 302         device->comps_used = min(ISERT_MAX_CQ, min_t(int, num_online_cpus(),
 303                                  device->ib_device->num_comp_vectors));
 304
 305         isert_info("Using %d CQs, %s supports %d vectors support "
 306                    "Fast registration %d pi_capable %d\n",
 307                    device->comps_used, device->ib_device->name,
 308                    device->ib_device->num_comp_vectors, device->use_fastreg,
 309                    device->pi_capable);
 310
 311         device->comps = kcalloc(device->comps_used, sizeof(struct isert_comp),
 312                                 GFP_KERNEL);
 313         if (!device->comps) {
 314                 isert_err("Unable to allocate completion contexts\n");
 315                 return -ENOMEM;
 316         }
 317
 318         max_cqe = min(ISER_MAX_CQ_LEN, attr->max_cqe);
 319
 320         for (i = 0; i < device->comps_used; i++) {
 321                 struct ib_cq_init_attr cq_attr = {};
 322                 struct isert_comp *comp = &device->comps[i];
 323
 324                 comp->device = device;
 325                 INIT_WORK(&comp->work, isert_cq_work);
 326                 cq_attr.cqe = max_cqe;
 327                 cq_attr.comp_vector = i;
 328                 comp->cq = ib_create_cq(device->ib_device,
 329                                         isert_cq_callback,
 330                                         isert_cq_event_callback,
 331                                         (void *)comp,
 332                                         &cq_attr);
 333                 if (IS_ERR(comp->cq)) {
 334                         isert_err("Unable to allocate cq\n");
 335                         ret = PTR_ERR(comp->cq);
 336                         comp->cq = NULL;
 337                         goto out_cq;
 338                 }
 339
 340                 ret = ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP);
 341                 if (ret)
 342                         goto out_cq;
 343         }
 344
 345         return 0;
 346 out_cq:
 347         isert_free_comps(device);
 348         return ret;
 349 }
 350
 351 static int
 352 isert_create_device_ib_res(struct isert_device *device)
 353 {
 354         struct ib_device_attr *dev_attr;
 355         int ret;
 356
 357         dev_attr = &device->dev_attr;
 358         ret = isert_query_device(device->ib_device, dev_attr);
 359         if (ret)
 360                 return ret;
 361
 362         /* asign function handlers */
 363         if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS &&
 364             dev_attr->device_cap_flags & IB_DEVICE_SIGNATURE_HANDOVER) {
 365                 device->use_fastreg = 1;
 366                 device->reg_rdma_mem = isert_reg_rdma;
 367                 device->unreg_rdma_mem = isert_unreg_rdma;
 368         } else {
 369                 device->use_fastreg = 0;
 370                 device->reg_rdma_mem = isert_map_rdma;
 371                 device->unreg_rdma_mem = isert_unmap_cmd;
 372         }
 373
 374         ret = isert_alloc_comps(device, dev_attr);
 375         if (ret)
 376                 return ret;
 377
 378         device->pd = ib_alloc_pd(device->ib_device);
 379         if (IS_ERR(device->pd)) {
 380                 ret = PTR_ERR(device->pd);
 381                 isert_err("failed to allocate pd, device %p, ret=%d\n",
 382                           device, ret);
 383                 goto out_cq;
 384         }
 385
 386         /* Check signature cap */
 387         device->pi_capable = dev_attr->device_cap_flags &
 388                              IB_DEVICE_SIGNATURE_HANDOVER ? true : false;
 389
 390         return 0;
 391
 392 out_cq:
 393         isert_free_comps(device);
 394         return ret;
 395 }
 396
 397 static void
 398 isert_free_device_ib_res(struct isert_device *device)
 399 {
 400         isert_info("device %p\n", device);
 401
 402         ib_dealloc_pd(device->pd);
 403         isert_free_comps(device);
 404 }
 405
 406 static void
 407 isert_device_put(struct isert_device *device)
 408 {
 409         mutex_lock(&device_list_mutex);
 410         device->refcount--;
 411         isert_info("device %p refcount %d\n", device, device->refcount);
 412         if (!device->refcount) {
 413                 isert_free_device_ib_res(device);
 414                 list_del(&device->dev_node);
 415                 kfree(device);
 416         }
 417         mutex_unlock(&device_list_mutex);
 418 }
 419
 420 static struct isert_device *
 421 isert_device_get(struct rdma_cm_id *cma_id)
 422 {
 423         struct isert_device *device;
 424         int ret;
 425
 426         mutex_lock(&device_list_mutex);
 427         list_for_each_entry(device, &device_list, dev_node) {
 428                 if (device->ib_device->node_guid == cma_id->device->node_guid) {
 429                         device->refcount++;
 430                         isert_info("Found iser device %p refcount %d\n",
 431                                    device, device->refcount);
 432                         mutex_unlock(&device_list_mutex);
 433                         return device;
 434                 }
 435         }
 436
 437         device = kzalloc(sizeof(struct isert_device), GFP_KERNEL);
 438         if (!device) {
 439                 mutex_unlock(&device_list_mutex);
 440                 return ERR_PTR(-ENOMEM);
 441         }
 442
 443         INIT_LIST_HEAD(&device->dev_node);
 444
 445         device->ib_device = cma_id->device;
 446         ret = isert_create_device_ib_res(device);
 447         if (ret) {
 448                 kfree(device);
 449                 mutex_unlock(&device_list_mutex);
 450                 return ERR_PTR(ret);
 451         }
 452
 453         device->refcount++;
 454         list_add_tail(&device->dev_node, &device_list);
 455         isert_info("Created a new iser device %p refcount %d\n",
 456                    device, device->refcount);
 457         mutex_unlock(&device_list_mutex);
 458
 459         return device;
 460 }
 461
 462 static void
 463 isert_conn_free_fastreg_pool(struct isert_conn *isert_conn)
 464 {
 465         struct fast_reg_descriptor *fr_desc, *tmp;
 466         int i = 0;
 467
 468         if (list_empty(&isert_conn->fr_pool))
 469                 return;
 470
 471         isert_info("Freeing conn %p fastreg pool", isert_conn);
 472
 473         list_for_each_entry_safe(fr_desc, tmp,
 474                                  &isert_conn->fr_pool, list) {
 475                 list_del(&fr_desc->list);
 476                 ib_dereg_mr(fr_desc->data_mr);
 477                 if (fr_desc->pi_ctx) {
 478                         ib_dereg_mr(fr_desc->pi_ctx->prot_mr);
 479                         ib_dereg_mr(fr_desc->pi_ctx->sig_mr);
 480                         kfree(fr_desc->pi_ctx);
 481                 }
 482                 kfree(fr_desc);
 483                 ++i;
 484         }
 485
 486         if (i < isert_conn->fr_pool_size)
 487                 isert_warn("Pool still has %d regions registered\n",
 488                         isert_conn->fr_pool_size - i);
 489 }
 490
 491 static int
 492 isert_create_pi_ctx(struct fast_reg_descriptor *desc,
 493                     struct ib_device *device,
 494                     struct ib_pd *pd)
 495 {
 496         struct pi_context *pi_ctx;
 497         int ret;
 498
 499         pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
 500         if (!pi_ctx) {
 501                 isert_err("Failed to allocate pi context\n");
 502                 return -ENOMEM;
 503         }
 504
 505         pi_ctx->prot_mr = ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG,
 506                                       ISCSI_ISER_SG_TABLESIZE);
 507         if (IS_ERR(pi_ctx->prot_mr)) {
 508                 isert_err("Failed to allocate prot frmr err=%ld\n",
 509                           PTR_ERR(pi_ctx->prot_mr));
 510                 ret = PTR_ERR(pi_ctx->prot_mr);
 511                 goto err_pi_ctx;
 512         }
 513         desc->ind |= ISERT_PROT_KEY_VALID;
 514
 515         pi_ctx->sig_mr = ib_alloc_mr(pd, IB_MR_TYPE_SIGNATURE, 2);
 516         if (IS_ERR(pi_ctx->sig_mr)) {
 517                 isert_err("Failed to allocate signature enabled mr err=%ld\n",
 518                           PTR_ERR(pi_ctx->sig_mr));
 519                 ret = PTR_ERR(pi_ctx->sig_mr);
 520                 goto err_prot_mr;
 521         }
 522
 523         desc->pi_ctx = pi_ctx;
 524         desc->ind |= ISERT_SIG_KEY_VALID;
 525         desc->ind &= ~ISERT_PROTECTED;
 526
 527         return 0;
 528
 529 err_prot_mr:
 530         ib_dereg_mr(pi_ctx->prot_mr);
 531 err_pi_ctx:
 532         kfree(pi_ctx);
 533
 534         return ret;
 535 }
 536
 537 static int
 538 isert_create_fr_desc(struct ib_device *ib_device, struct ib_pd *pd,
 539                      struct fast_reg_descriptor *fr_desc)
 540 {
 541         fr_desc->data_mr = ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG,
 542                                        ISCSI_ISER_SG_TABLESIZE);
 543         if (IS_ERR(fr_desc->data_mr)) {
 544                 isert_err("Failed to allocate data frmr err=%ld\n",
 545                           PTR_ERR(fr_desc->data_mr));
 546                 return PTR_ERR(fr_desc->data_mr);
 547         }
 548         fr_desc->ind |= ISERT_DATA_KEY_VALID;
 549
 550         isert_dbg("Created fr_desc %p\n", fr_desc);
 551
 552         return 0;
 553 }
 554
 555 static int
 556 isert_conn_create_fastreg_pool(struct isert_conn *isert_conn)
 557 {
 558         struct fast_reg_descriptor *fr_desc;
 559         struct isert_device *device = isert_conn->device;
 560         struct se_session *se_sess = isert_conn->conn->sess->se_sess;
 561         struct se_node_acl *se_nacl = se_sess->se_node_acl;
 562         int i, ret, tag_num;
 563         /*
 564          * Setup the number of FRMRs based upon the number of tags
 565          * available to session in iscsi_target_locate_portal().
 566          */
 567         tag_num = max_t(u32, ISCSIT_MIN_TAGS, se_nacl->queue_depth);
 568         tag_num = (tag_num * 2) + ISCSIT_EXTRA_TAGS;
 569
 570         isert_conn->fr_pool_size = 0;
 571         for (i = 0; i < tag_num; i++) {
 572                 fr_desc = kzalloc(sizeof(*fr_desc), GFP_KERNEL);
 573                 if (!fr_desc) {
 574                         isert_err("Failed to allocate fast_reg descriptor\n");
 575                         ret = -ENOMEM;
 576                         goto err;
 577                 }
 578
 579                 ret = isert_create_fr_desc(device->ib_device,
 580                                            device->pd, fr_desc);
 581                 if (ret) {
 582                         isert_err("Failed to create fastreg descriptor err=%d\n",
 583                                ret);
 584                         kfree(fr_desc);
 585                         goto err;
 586                 }
 587
 588                 list_add_tail(&fr_desc->list, &isert_conn->fr_pool);
 589                 isert_conn->fr_pool_size++;
 590         }
 591
 592         isert_dbg("Creating conn %p fastreg pool size=%d",
 593                  isert_conn, isert_conn->fr_pool_size);
 594
 595         return 0;
 596
 597 err:
 598         isert_conn_free_fastreg_pool(isert_conn);
 599         return ret;
 600 }
 601
 602 static void
 603 isert_init_conn(struct isert_conn *isert_conn)
 604 {
 605         isert_conn->state = ISER_CONN_INIT;
 606         INIT_LIST_HEAD(&isert_conn->node);
 607         init_completion(&isert_conn->login_comp);
 608         init_completion(&isert_conn->login_req_comp);
 609         init_completion(&isert_conn->wait);
 610         kref_init(&isert_conn->kref);
 611         mutex_init(&isert_conn->mutex);
 612         spin_lock_init(&isert_conn->pool_lock);
 613         INIT_LIST_HEAD(&isert_conn->fr_pool);
 614         INIT_WORK(&isert_conn->release_work, isert_release_work);
 615 }
 616
 617 static void
 618 isert_free_login_buf(struct isert_conn *isert_conn)
 619 {
 620         struct ib_device *ib_dev = isert_conn->device->ib_device;
 621
 622         ib_dma_unmap_single(ib_dev, isert_conn->login_rsp_dma,
 623                             ISER_RX_LOGIN_SIZE, DMA_TO_DEVICE);
 624         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 625                             ISCSI_DEF_MAX_RECV_SEG_LEN,
 626                             DMA_FROM_DEVICE);
 627         kfree(isert_conn->login_buf);
 628 }
 629
 630 static int
 631 isert_alloc_login_buf(struct isert_conn *isert_conn,
 632                       struct ib_device *ib_dev)
 633 {
 634         int ret;
 635
 636         isert_conn->login_buf = kzalloc(ISCSI_DEF_MAX_RECV_SEG_LEN +
 637                                         ISER_RX_LOGIN_SIZE, GFP_KERNEL);
 638         if (!isert_conn->login_buf) {
 639                 isert_err("Unable to allocate isert_conn->login_buf\n");
 640                 return -ENOMEM;
 641         }
 642
 643         isert_conn->login_req_buf = isert_conn->login_buf;
 644         isert_conn->login_rsp_buf = isert_conn->login_buf +
 645                                     ISCSI_DEF_MAX_RECV_SEG_LEN;
 646
 647         isert_dbg("Set login_buf: %p login_req_buf: %p login_rsp_buf: %p\n",
 648                  isert_conn->login_buf, isert_conn->login_req_buf,
 649                  isert_conn->login_rsp_buf);
 650
 651         isert_conn->login_req_dma = ib_dma_map_single(ib_dev,
 652                                 (void *)isert_conn->login_req_buf,
 653                                 ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_FROM_DEVICE);
 654
 655         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_req_dma);
 656         if (ret) {
 657                 isert_err("login_req_dma mapping error: %d\n", ret);
 658                 isert_conn->login_req_dma = 0;
 659                 goto out_login_buf;
 660         }
 661
 662         isert_conn->login_rsp_dma = ib_dma_map_single(ib_dev,
 663                                         (void *)isert_conn->login_rsp_buf,
 664                                         ISER_RX_LOGIN_SIZE, DMA_TO_DEVICE);
 665
 666         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_rsp_dma);
 667         if (ret) {
 668                 isert_err("login_rsp_dma mapping error: %d\n", ret);
 669                 isert_conn->login_rsp_dma = 0;
 670                 goto out_req_dma_map;
 671         }
 672
 673         return 0;
 674
 675 out_req_dma_map:
 676         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 677                             ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_FROM_DEVICE);
 678 out_login_buf:
 679         kfree(isert_conn->login_buf);
 680         return ret;
 681 }
 682
 683 static int
 684 isert_connect_request(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 685 {
 686         struct isert_np *isert_np = cma_id->context;
 687         struct iscsi_np *np = isert_np->np;
 688         struct isert_conn *isert_conn;
 689         struct isert_device *device;
 690         int ret = 0;
 691
 692         spin_lock_bh(&np->np_thread_lock);
 693         if (!np->enabled) {
 694                 spin_unlock_bh(&np->np_thread_lock);
 695                 isert_dbg("iscsi_np is not enabled, reject connect request\n");
 696                 return rdma_reject(cma_id, NULL, 0);
 697         }
 698         spin_unlock_bh(&np->np_thread_lock);
 699
 700         isert_dbg("cma_id: %p, portal: %p\n",
 701                  cma_id, cma_id->context);
 702
 703         isert_conn = kzalloc(sizeof(struct isert_conn), GFP_KERNEL);
 704         if (!isert_conn)
 705                 return -ENOMEM;
 706
 707         isert_init_conn(isert_conn);
 708         isert_conn->cm_id = cma_id;
 709
 710         ret = isert_alloc_login_buf(isert_conn, cma_id->device);
 711         if (ret)
 712                 goto out;
 713
 714         device = isert_device_get(cma_id);
 715         if (IS_ERR(device)) {
 716                 ret = PTR_ERR(device);
 717                 goto out_rsp_dma_map;
 718         }
 719         isert_conn->device = device;
 720
 721         /* Set max inflight RDMA READ requests */
 722         isert_conn->initiator_depth = min_t(u8,
 723                                 event->param.conn.initiator_depth,
 724                                 device->dev_attr.max_qp_init_rd_atom);
 725         isert_dbg("Using initiator_depth: %u\n", isert_conn->initiator_depth);
 726
 727         ret = isert_conn_setup_qp(isert_conn, cma_id);
 728         if (ret)
 729                 goto out_conn_dev;
 730
 731         ret = isert_rdma_post_recvl(isert_conn);
 732         if (ret)
 733                 goto out_conn_dev;
 734
 735         ret = isert_rdma_accept(isert_conn);
 736         if (ret)
 737                 goto out_conn_dev;
 738
 739         mutex_lock(&isert_np->mutex);
 740         list_add_tail(&isert_conn->node, &isert_np->accepted);
 741         mutex_unlock(&isert_np->mutex);
 742
 743         return 0;
 744
 745 out_conn_dev:
 746         isert_device_put(device);
 747 out_rsp_dma_map:
 748         isert_free_login_buf(isert_conn);
 749 out:
 750         kfree(isert_conn);
 751         rdma_reject(cma_id, NULL, 0);
 752         return ret;
 753 }
 754
 755 static void
 756 isert_connect_release(struct isert_conn *isert_conn)
 757 {
 758         struct isert_device *device = isert_conn->device;
 759
 760         isert_dbg("conn %p\n", isert_conn);
 761
 762         BUG_ON(!device);
 763
 764         if (device->use_fastreg)
 765                 isert_conn_free_fastreg_pool(isert_conn);
 766
 767         isert_free_rx_descriptors(isert_conn);
 768         if (isert_conn->cm_id)
 769                 rdma_destroy_id(isert_conn->cm_id);
 770
 771         if (isert_conn->qp) {
 772                 struct isert_comp *comp = isert_conn->qp->recv_cq->cq_context;
 773
 774                 isert_comp_put(comp);
 775                 ib_destroy_qp(isert_conn->qp);
 776         }
 777
 778         if (isert_conn->login_buf)
 779                 isert_free_login_buf(isert_conn);
 780
 781         isert_device_put(device);
 782
 783         kfree(isert_conn);
 784 }
 785
 786 static void
 787 isert_connected_handler(struct rdma_cm_id *cma_id)
 788 {
 789         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 790         struct isert_np *isert_np = cma_id->context;
 791
 792         isert_info("conn %p\n", isert_conn);
 793
 794         mutex_lock(&isert_conn->mutex);
 795         isert_conn->state = ISER_CONN_UP;
 796         kref_get(&isert_conn->kref);
 797         mutex_unlock(&isert_conn->mutex);
 798
 799         mutex_lock(&isert_np->mutex);
 800         list_move_tail(&isert_conn->node, &isert_np->pending);
 801         mutex_unlock(&isert_np->mutex);
 802
 803         isert_info("np %p: Allow accept_np to continue\n", isert_np);
 804         up(&isert_np->sem);
 805 }
 806
 807 static void
 808 isert_release_kref(struct kref *kref)
 809 {
 810         struct isert_conn *isert_conn = container_of(kref,
 811                                 struct isert_conn, kref);
 812
 813         isert_info("conn %p final kref %s/%d\n", isert_conn, current->comm,
 814                    current->pid);
 815
 816         isert_connect_release(isert_conn);
 817 }
 818
 819 static void
 820 isert_put_conn(struct isert_conn *isert_conn)
 821 {
 822         kref_put(&isert_conn->kref, isert_release_kref);
 823 }
 824
 825 /**
 826  * isert_conn_terminate() - Initiate connection termination
 827  * @isert_conn: isert connection struct
 828  *
 829  * Notes:
 830  * In case the connection state is FULL_FEATURE, move state
 831  * to TEMINATING and start teardown sequence (rdma_disconnect).
 832  * In case the connection state is UP, complete flush as well.
 833  *
 834  * This routine must be called with mutex held. Thus it is
 835  * safe to call multiple times.
 836  */
 837 static void
 838 isert_conn_terminate(struct isert_conn *isert_conn)
 839 {
 840         int err;
 841
 842         switch (isert_conn->state) {
 843         case ISER_CONN_TERMINATING:
 844                 break;
 845         case ISER_CONN_UP:
 846         case ISER_CONN_FULL_FEATURE: /* FALLTHRU */
 847                 isert_info("Terminating conn %p state %d\n",
 848                            isert_conn, isert_conn->state);
 849                 isert_conn->state = ISER_CONN_TERMINATING;
 850                 err = rdma_disconnect(isert_conn->cm_id);
 851                 if (err)
 852                         isert_warn("Failed rdma_disconnect isert_conn %p\n",
 853                                    isert_conn);
 854                 break;
 855         default:
 856                 isert_warn("conn %p teminating in state %d\n",
 857                            isert_conn, isert_conn->state);
 858         }
 859 }
 860
 861 static int
 862 isert_np_cma_handler(struct isert_np *isert_np,
 863                      enum rdma_cm_event_type event)
 864 {
 865         isert_dbg("%s (%d): isert np %p\n",
 866                   rdma_event_msg(event), event, isert_np);
 867
 868         switch (event) {
 869         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 870                 isert_np->cm_id = NULL;
 871                 break;
 872         case RDMA_CM_EVENT_ADDR_CHANGE:
 873                 isert_np->cm_id = isert_setup_id(isert_np);
 874                 if (IS_ERR(isert_np->cm_id)) {
 875                         isert_err("isert np %p setup id failed: %ld\n",
 876                                   isert_np, PTR_ERR(isert_np->cm_id));
 877                         isert_np->cm_id = NULL;
 878                 }
 879                 break;
 880         default:
 881                 isert_err("isert np %p Unexpected event %d\n",
 882                           isert_np, event);
 883         }
 884
 885         return -1;
 886 }
 887
 888 static int
 889 isert_disconnected_handler(struct rdma_cm_id *cma_id,
 890                            enum rdma_cm_event_type event)
 891 {
 892         struct isert_np *isert_np = cma_id->context;
 893         struct isert_conn *isert_conn;
 894         bool terminating = false;
 895
 896         if (isert_np->cm_id == cma_id)
 897                 return isert_np_cma_handler(cma_id->context, event);
 898
 899         isert_conn = cma_id->qp->qp_context;
 900
 901         mutex_lock(&isert_conn->mutex);
 902         terminating = (isert_conn->state == ISER_CONN_TERMINATING);
 903         isert_conn_terminate(isert_conn);
 904         mutex_unlock(&isert_conn->mutex);
 905
 906         isert_info("conn %p completing wait\n", isert_conn);
 907         complete(&isert_conn->wait);
 908
 909         if (terminating)
 910                 goto out;
 911
 912         mutex_lock(&isert_np->mutex);
 913         if (!list_empty(&isert_conn->node)) {
 914                 list_del_init(&isert_conn->node);
 915                 isert_put_conn(isert_conn);
 916                 queue_work(isert_release_wq, &isert_conn->release_work);
 917         }
 918         mutex_unlock(&isert_np->mutex);
 919
 920 out:
 921         return 0;
 922 }
 923
 924 static int
 925 isert_connect_error(struct rdma_cm_id *cma_id)
 926 {
 927         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 928
 929         list_del_init(&isert_conn->node);
 930         isert_conn->cm_id = NULL;
 931         isert_put_conn(isert_conn);
 932
 933         return -1;
 934 }
 935
 936 static int
 937 isert_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 938 {
 939         int ret = 0;
 940
 941         isert_info("%s (%d): status %d id %p np %p\n",
 942                    rdma_event_msg(event->event), event->event,
 943                    event->status, cma_id, cma_id->context);
 944
 945         switch (event->event) {
 946         case RDMA_CM_EVENT_CONNECT_REQUEST:
 947                 ret = isert_connect_request(cma_id, event);
 948                 if (ret)
 949                         isert_err("failed handle connect request %d\n", ret);
 950                 break;
 951         case RDMA_CM_EVENT_ESTABLISHED:
 952                 isert_connected_handler(cma_id);
 953                 break;
 954         case RDMA_CM_EVENT_ADDR_CHANGE:    /* FALLTHRU */
 955         case RDMA_CM_EVENT_DISCONNECTED:   /* FALLTHRU */
 956         case RDMA_CM_EVENT_DEVICE_REMOVAL: /* FALLTHRU */
 957         case RDMA_CM_EVENT_TIMEWAIT_EXIT:  /* FALLTHRU */
 958                 ret = isert_disconnected_handler(cma_id, event->event);
 959                 break;
 960         case RDMA_CM_EVENT_REJECTED:       /* FALLTHRU */
 961         case RDMA_CM_EVENT_UNREACHABLE:    /* FALLTHRU */
 962         case RDMA_CM_EVENT_CONNECT_ERROR:
 963                 ret = isert_connect_error(cma_id);
 964                 break;
 965         default:
 966                 isert_err("Unhandled RDMA CMA event: %d\n", event->event);
 967                 break;
 968         }
 969
 970         return ret;
 971 }
 972
 973 static int
 974 isert_post_recvm(struct isert_conn *isert_conn, u32 count)
 975 {
 976         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 977         int i, ret;
 978         struct iser_rx_desc *rx_desc;
 979
 980         for (rx_wr = isert_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
 981                 rx_desc = &isert_conn->rx_descs[i];
 982                 rx_wr->wr_id = (uintptr_t)rx_desc;
 983                 rx_wr->sg_list = &rx_desc->rx_sg;
 984                 rx_wr->num_sge = 1;
 985                 rx_wr->next = rx_wr + 1;
 986         }
 987         rx_wr--;
 988         rx_wr->next = NULL; /* mark end of work requests list */
 989
 990         isert_conn->post_recv_buf_count += count;
 991         ret = ib_post_recv(isert_conn->qp, isert_conn->rx_wr,
 992                            &rx_wr_failed);
 993         if (ret) {
 994                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
 995                 isert_conn->post_recv_buf_count -= count;
 996         }
 997
 998         return ret;
 999 }
1000
1001 static int
1002 isert_post_recv(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc)
1003 {
1004         struct ib_recv_wr *rx_wr_failed, rx_wr;
1005         int ret;
1006
1007         rx_wr.wr_id = (uintptr_t)rx_desc;
1008         rx_wr.sg_list = &rx_desc->rx_sg;
1009         rx_wr.num_sge = 1;
1010         rx_wr.next = NULL;
1011
1012         isert_conn->post_recv_buf_count++;
1013         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_failed);
1014         if (ret) {
1015                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
1016                 isert_conn->post_recv_buf_count--;
1017         }
1018
1019         return ret;
1020 }
1021
1022 static int
1023 isert_post_send(struct isert_conn *isert_conn, struct iser_tx_desc *tx_desc)
1024 {
1025         struct ib_device *ib_dev = isert_conn->cm_id->device;
1026         struct ib_send_wr send_wr, *send_wr_failed;
1027         int ret;
1028
1029         ib_dma_sync_single_for_device(ib_dev, tx_desc->dma_addr,
1030                                       ISER_HEADERS_LEN, DMA_TO_DEVICE);
1031
1032         send_wr.next    = NULL;
1033         send_wr.wr_id   = (uintptr_t)tx_desc;
1034         send_wr.sg_list = tx_desc->tx_sg;
1035         send_wr.num_sge = tx_desc->num_sge;
1036         send_wr.opcode  = IB_WR_SEND;
1037         send_wr.send_flags = IB_SEND_SIGNALED;
1038
1039         ret = ib_post_send(isert_conn->qp, &send_wr, &send_wr_failed);
1040         if (ret)
1041                 isert_err("ib_post_send() failed, ret: %d\n", ret);
1042
1043         return ret;
1044 }
1045
1046 static void
1047 isert_create_send_desc(struct isert_conn *isert_conn,
1048                        struct isert_cmd *isert_cmd,
1049                        struct iser_tx_desc *tx_desc)
1050 {
1051         struct isert_device *device = isert_conn->device;
1052         struct ib_device *ib_dev = device->ib_device;
1053
1054         ib_dma_sync_single_for_cpu(ib_dev, tx_desc->dma_addr,
1055                                    ISER_HEADERS_LEN, DMA_TO_DEVICE);
1056
1057         memset(&tx_desc->iser_header, 0, sizeof(struct iser_hdr));
1058         tx_desc->iser_header.flags = ISER_VER;
1059
1060         tx_desc->num_sge = 1;
1061         tx_desc->isert_cmd = isert_cmd;
1062
1063         if (tx_desc->tx_sg[0].lkey != device->pd->local_dma_lkey) {
1064                 tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
1065                 isert_dbg("tx_desc %p lkey mismatch, fixing\n", tx_desc);
1066         }
1067 }
1068
1069 static int
1070 isert_init_tx_hdrs(struct isert_conn *isert_conn,
1071                    struct iser_tx_desc *tx_desc)
1072 {
1073         struct isert_device *device = isert_conn->device;
1074         struct ib_device *ib_dev = device->ib_device;
1075         u64 dma_addr;
1076
1077         dma_addr = ib_dma_map_single(ib_dev, (void *)tx_desc,
1078                         ISER_HEADERS_LEN, DMA_TO_DEVICE);
1079         if (ib_dma_mapping_error(ib_dev, dma_addr)) {
1080                 isert_err("ib_dma_mapping_error() failed\n");
1081                 return -ENOMEM;
1082         }
1083
1084         tx_desc->dma_addr = dma_addr;
1085         tx_desc->tx_sg[0].addr  = tx_desc->dma_addr;
1086         tx_desc->tx_sg[0].length = ISER_HEADERS_LEN;
1087         tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
1088
1089         isert_dbg("Setup tx_sg[0].addr: 0x%llx length: %u lkey: 0x%x\n",
1090                   tx_desc->tx_sg[0].addr, tx_desc->tx_sg[0].length,
1091                   tx_desc->tx_sg[0].lkey);
1092
1093         return 0;
1094 }
1095
1096 static void
1097 isert_init_send_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1098                    struct ib_send_wr *send_wr)
1099 {
1100         struct iser_tx_desc *tx_desc = &isert_cmd->tx_desc;
1101
1102         isert_cmd->rdma_wr.iser_ib_op = ISER_IB_SEND;
1103         send_wr->wr_id = (uintptr_t)&isert_cmd->tx_desc;
1104         send_wr->opcode = IB_WR_SEND;
1105         send_wr->sg_list = &tx_desc->tx_sg[0];
1106         send_wr->num_sge = isert_cmd->tx_desc.num_sge;
1107         send_wr->send_flags = IB_SEND_SIGNALED;
1108 }
1109
1110 static int
1111 isert_rdma_post_recvl(struct isert_conn *isert_conn)
1112 {
1113         struct ib_recv_wr rx_wr, *rx_wr_fail;
1114         struct ib_sge sge;
1115         int ret;
1116
1117         memset(&sge, 0, sizeof(struct ib_sge));
1118         sge.addr = isert_conn->login_req_dma;
1119         sge.length = ISER_RX_LOGIN_SIZE;
1120         sge.lkey = isert_conn->device->pd->local_dma_lkey;
1121
1122         isert_dbg("Setup sge: addr: %llx length: %d 0x%08x\n",
1123                 sge.addr, sge.length, sge.lkey);
1124
1125         memset(&rx_wr, 0, sizeof(struct ib_recv_wr));
1126         rx_wr.wr_id = (uintptr_t)isert_conn->login_req_buf;
1127         rx_wr.sg_list = &sge;
1128         rx_wr.num_sge = 1;
1129
1130         isert_conn->post_recv_buf_count++;
1131         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_fail);
1132         if (ret) {
1133                 isert_err("ib_post_recv() failed: %d\n", ret);
1134                 isert_conn->post_recv_buf_count--;
1135         }
1136
1137         return ret;
1138 }
1139
1140 static int
1141 isert_put_login_tx(struct iscsi_conn *conn, struct iscsi_login *login,
1142                    u32 length)
1143 {
1144         struct isert_conn *isert_conn = conn->context;
1145         struct isert_device *device = isert_conn->device;
1146         struct ib_device *ib_dev = device->ib_device;
1147         struct iser_tx_desc *tx_desc = &isert_conn->login_tx_desc;
1148         int ret;
1149
1150         isert_create_send_desc(isert_conn, NULL, tx_desc);
1151
1152         memcpy(&tx_desc->iscsi_header, &login->rsp[0],
1153                sizeof(struct iscsi_hdr));
1154
1155         isert_init_tx_hdrs(isert_conn, tx_desc);
1156
1157         if (length > 0) {
1158                 struct ib_sge *tx_dsg = &tx_desc->tx_sg[1];
1159
1160                 ib_dma_sync_single_for_cpu(ib_dev, isert_conn->login_rsp_dma,
1161                                            length, DMA_TO_DEVICE);
1162
1163                 memcpy(isert_conn->login_rsp_buf, login->rsp_buf, length);
1164
1165                 ib_dma_sync_single_for_device(ib_dev, isert_conn->login_rsp_dma,
1166                                               length, DMA_TO_DEVICE);
1167
1168                 tx_dsg->addr    = isert_conn->login_rsp_dma;
1169                 tx_dsg->length  = length;
1170                 tx_dsg->lkey    = isert_conn->device->pd->local_dma_lkey;
1171                 tx_desc->num_sge = 2;
1172         }
1173         if (!login->login_failed) {
1174                 if (login->login_complete) {
1175                         if (!conn->sess->sess_ops->SessionType &&
1176                             isert_conn->device->use_fastreg) {
1177                                 ret = isert_conn_create_fastreg_pool(isert_conn);
1178                                 if (ret) {
1179                                         isert_err("Conn: %p failed to create"
1180                                                " fastreg pool\n", isert_conn);
1181                                         return ret;
1182                                 }
1183                         }
1184
1185                         ret = isert_alloc_rx_descriptors(isert_conn);
1186                         if (ret)
1187                                 return ret;
1188
1189                         ret = isert_post_recvm(isert_conn,
1190                                                ISERT_QP_MAX_RECV_DTOS);
1191                         if (ret)
1192                                 return ret;
1193
1194                         /* Now we are in FULL_FEATURE phase */
1195                         mutex_lock(&isert_conn->mutex);
1196                         isert_conn->state = ISER_CONN_FULL_FEATURE;
1197                         mutex_unlock(&isert_conn->mutex);
1198                         goto post_send;
1199                 }
1200
1201                 ret = isert_rdma_post_recvl(isert_conn);
1202                 if (ret)
1203                         return ret;
1204         }
1205 post_send:
1206         ret = isert_post_send(isert_conn, tx_desc);
1207         if (ret)
1208                 return ret;
1209
1210         return 0;
1211 }
1212
1213 static void
1214 isert_rx_login_req(struct isert_conn *isert_conn)
1215 {
1216         struct iser_rx_desc *rx_desc = (void *)isert_conn->login_req_buf;
1217         int rx_buflen = isert_conn->login_req_len;
1218         struct iscsi_conn *conn = isert_conn->conn;
1219         struct iscsi_login *login = conn->conn_login;
1220         int size;
1221
1222         isert_info("conn %p\n", isert_conn);
1223
1224         WARN_ON_ONCE(!login);
1225
1226         if (login->first_request) {
1227                 struct iscsi_login_req *login_req =
1228                         (struct iscsi_login_req *)&rx_desc->iscsi_header;
1229                 /*
1230                  * Setup the initial iscsi_login values from the leading
1231                  * login request PDU.
1232                  */
1233                 login->leading_connection = (!login_req->tsih) ? 1 : 0;
1234                 login->current_stage =
1235                         (login_req->flags & ISCSI_FLAG_LOGIN_CURRENT_STAGE_MASK)
1236                          >> 2;
1237                 login->version_min      = login_req->min_version;
1238                 login->version_max      = login_req->max_version;
1239                 memcpy(login->isid, login_req->isid, 6);
1240                 login->cmd_sn           = be32_to_cpu(login_req->cmdsn);
1241                 login->init_task_tag    = login_req->itt;
1242                 login->initial_exp_statsn = be32_to_cpu(login_req->exp_statsn);
1243                 login->cid              = be16_to_cpu(login_req->cid);
1244                 login->tsih             = be16_to_cpu(login_req->tsih);
1245         }
1246
1247         memcpy(&login->req[0], (void *)&rx_desc->iscsi_header, ISCSI_HDR_LEN);
1248
1249         size = min(rx_buflen, MAX_KEY_VALUE_PAIRS);
1250         isert_dbg("Using login payload size: %d, rx_buflen: %d "
1251                   "MAX_KEY_VALUE_PAIRS: %d\n", size, rx_buflen,
1252                   MAX_KEY_VALUE_PAIRS);
1253         memcpy(login->req_buf, &rx_desc->data[0], size);
1254
1255         if (login->first_request) {
1256                 complete(&isert_conn->login_comp);
1257                 return;
1258         }
1259         schedule_delayed_work(&conn->login_work, 0);
1260 }
1261
1262 static struct iscsi_cmd
1263 *isert_allocate_cmd(struct iscsi_conn *conn, struct iser_rx_desc *rx_desc)
1264 {
1265         struct isert_conn *isert_conn = conn->context;
1266         struct isert_cmd *isert_cmd;
1267         struct iscsi_cmd *cmd;
1268
1269         cmd = iscsit_allocate_cmd(conn, TASK_INTERRUPTIBLE);
1270         if (!cmd) {
1271                 isert_err("Unable to allocate iscsi_cmd + isert_cmd\n");
1272                 return NULL;
1273         }
1274         isert_cmd = iscsit_priv_cmd(cmd);
1275         isert_cmd->conn = isert_conn;
1276         isert_cmd->iscsi_cmd = cmd;
1277         isert_cmd->rx_desc = rx_desc;
1278
1279         return cmd;
1280 }
1281
1282 static int
1283 isert_handle_scsi_cmd(struct isert_conn *isert_conn,
1284                       struct isert_cmd *isert_cmd, struct iscsi_cmd *cmd,
1285                       struct iser_rx_desc *rx_desc, unsigned char *buf)
1286 {
1287         struct iscsi_conn *conn = isert_conn->conn;
1288         struct iscsi_scsi_req *hdr = (struct iscsi_scsi_req *)buf;
1289         int imm_data, imm_data_len, unsol_data, sg_nents, rc;
1290         bool dump_payload = false;
1291         unsigned int data_len;
1292
1293         rc = iscsit_setup_scsi_cmd(conn, cmd, buf);
1294         if (rc < 0)
1295                 return rc;
1296
1297         imm_data = cmd->immediate_data;
1298         imm_data_len = cmd->first_burst_len;
1299         unsol_data = cmd->unsolicited_data;
1300         data_len = cmd->se_cmd.data_length;
1301
1302         if (imm_data && imm_data_len == data_len)
1303                 cmd->se_cmd.se_cmd_flags |= SCF_PASSTHROUGH_SG_TO_MEM_NOALLOC;
1304         rc = iscsit_process_scsi_cmd(conn, cmd, hdr);
1305         if (rc < 0) {
1306                 return 0;
1307         } else if (rc > 0) {
1308                 dump_payload = true;
1309                 goto sequence_cmd;
1310         }
1311
1312         if (!imm_data)
1313                 return 0;
1314
1315         if (imm_data_len != data_len) {
1316                 sg_nents = max(1UL, DIV_ROUND_UP(imm_data_len, PAGE_SIZE));
1317                 sg_copy_from_buffer(cmd->se_cmd.t_data_sg, sg_nents,
1318                                     &rx_desc->data[0], imm_data_len);
1319                 isert_dbg("Copy Immediate sg_nents: %u imm_data_len: %d\n",
1320                           sg_nents, imm_data_len);
1321         } else {
1322                 sg_init_table(&isert_cmd->sg, 1);
1323                 cmd->se_cmd.t_data_sg = &isert_cmd->sg;
1324                 cmd->se_cmd.t_data_nents = 1;
1325                 sg_set_buf(&isert_cmd->sg, &rx_desc->data[0], imm_data_len);
1326                 isert_dbg("Transfer Immediate imm_data_len: %d\n",
1327                           imm_data_len);
1328         }
1329
1330         cmd->write_data_done += imm_data_len;
1331
1332         if (cmd->write_data_done == cmd->se_cmd.data_length) {
1333                 spin_lock_bh(&cmd->istate_lock);
1334                 cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1335                 cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1336                 spin_unlock_bh(&cmd->istate_lock);
1337         }
1338
1339 sequence_cmd:
1340         rc = iscsit_sequence_cmd(conn, cmd, buf, hdr->cmdsn);
1341
1342         if (!rc && dump_payload == false && unsol_data)
1343                 iscsit_set_unsoliticed_dataout(cmd);
1344         else if (dump_payload && imm_data)
1345                 target_put_sess_cmd(&cmd->se_cmd);
1346
1347         return 0;
1348 }
1349
1350 static int
1351 isert_handle_iscsi_dataout(struct isert_conn *isert_conn,
1352                            struct iser_rx_desc *rx_desc, unsigned char *buf)
1353 {
1354         struct scatterlist *sg_start;
1355         struct iscsi_conn *conn = isert_conn->conn;
1356         struct iscsi_cmd *cmd = NULL;
1357         struct iscsi_data *hdr = (struct iscsi_data *)buf;
1358         u32 unsol_data_len = ntoh24(hdr->dlength);
1359         int rc, sg_nents, sg_off, page_off;
1360
1361         rc = iscsit_check_dataout_hdr(conn, buf, &cmd);
1362         if (rc < 0)
1363                 return rc;
1364         else if (!cmd)
1365                 return 0;
1366         /*
1367          * FIXME: Unexpected unsolicited_data out
1368          */
1369         if (!cmd->unsolicited_data) {
1370                 isert_err("Received unexpected solicited data payload\n");
1371                 dump_stack();
1372                 return -1;
1373         }
1374
1375         isert_dbg("Unsolicited DataOut unsol_data_len: %u, "
1376                   "write_data_done: %u, data_length: %u\n",
1377                   unsol_data_len,  cmd->write_data_done,
1378                   cmd->se_cmd.data_length);
1379
1380         sg_off = cmd->write_data_done / PAGE_SIZE;
1381         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
1382         sg_nents = max(1UL, DIV_ROUND_UP(unsol_data_len, PAGE_SIZE));
1383         page_off = cmd->write_data_done % PAGE_SIZE;
1384         /*
1385          * FIXME: Non page-aligned unsolicited_data out
1386          */
1387         if (page_off) {
1388                 isert_err("unexpected non-page aligned data payload\n");
1389                 dump_stack();
1390                 return -1;
1391         }
1392         isert_dbg("Copying DataOut: sg_start: %p, sg_off: %u "
1393                   "sg_nents: %u from %p %u\n", sg_start, sg_off,
1394                   sg_nents, &rx_desc->data[0], unsol_data_len);
1395
1396         sg_copy_from_buffer(sg_start, sg_nents, &rx_desc->data[0],
1397                             unsol_data_len);
1398
1399         rc = iscsit_check_dataout_payload(cmd, hdr, false);
1400         if (rc < 0)
1401                 return rc;
1402
1403         /*
1404          * multiple data-outs on the same command can arrive -
1405          * so post the buffer before hand
1406          */
1407         rc = isert_post_recv(isert_conn, rx_desc);
1408         if (rc) {
1409                 isert_err("ib_post_recv failed with %d\n", rc);
1410                 return rc;
1411         }
1412         return 0;
1413 }
1414
1415 static int
1416 isert_handle_nop_out(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1417                      struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1418                      unsigned char *buf)
1419 {
1420         struct iscsi_conn *conn = isert_conn->conn;
1421         struct iscsi_nopout *hdr = (struct iscsi_nopout *)buf;
1422         int rc;
1423
1424         rc = iscsit_setup_nop_out(conn, cmd, hdr);
1425         if (rc < 0)
1426                 return rc;
1427         /*
1428          * FIXME: Add support for NOPOUT payload using unsolicited RDMA payload
1429          */
1430
1431         return iscsit_process_nop_out(conn, cmd, hdr);
1432 }
1433
1434 static int
1435 isert_handle_text_cmd(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1436                       struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1437                       struct iscsi_text *hdr)
1438 {
1439         struct iscsi_conn *conn = isert_conn->conn;
1440         u32 payload_length = ntoh24(hdr->dlength);
1441         int rc;
1442         unsigned char *text_in = NULL;
1443
1444         rc = iscsit_setup_text_cmd(conn, cmd, hdr);
1445         if (rc < 0)
1446                 return rc;
1447
1448         if (payload_length) {
1449                 text_in = kzalloc(payload_length, GFP_KERNEL);
1450                 if (!text_in) {
1451                         isert_err("Unable to allocate text_in of payload_length: %u\n",
1452                                   payload_length);
1453                         return -ENOMEM;
1454                 }
1455         }
1456         cmd->text_in_ptr = text_in;
1457
1458         memcpy(cmd->text_in_ptr, &rx_desc->data[0], payload_length);
1459
1460         return iscsit_process_text_cmd(conn, cmd, hdr);
1461 }
1462
1463 static int
1464 isert_rx_opcode(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc,
1465                 uint32_t read_stag, uint64_t read_va,
1466                 uint32_t write_stag, uint64_t write_va)
1467 {
1468         struct iscsi_hdr *hdr = &rx_desc->iscsi_header;
1469         struct iscsi_conn *conn = isert_conn->conn;
1470         struct iscsi_cmd *cmd;
1471         struct isert_cmd *isert_cmd;
1472         int ret = -EINVAL;
1473         u8 opcode = (hdr->opcode & ISCSI_OPCODE_MASK);
1474
1475         if (conn->sess->sess_ops->SessionType &&
1476            (!(opcode & ISCSI_OP_TEXT) || !(opcode & ISCSI_OP_LOGOUT))) {
1477                 isert_err("Got illegal opcode: 0x%02x in SessionType=Discovery,"
1478                           " ignoring\n", opcode);
1479                 return 0;
1480         }
1481
1482         switch (opcode) {
1483         case ISCSI_OP_SCSI_CMD:
1484                 cmd = isert_allocate_cmd(conn, rx_desc);
1485                 if (!cmd)
1486                         break;
1487
1488                 isert_cmd = iscsit_priv_cmd(cmd);
1489                 isert_cmd->read_stag = read_stag;
1490                 isert_cmd->read_va = read_va;
1491                 isert_cmd->write_stag = write_stag;
1492                 isert_cmd->write_va = write_va;
1493
1494                 ret = isert_handle_scsi_cmd(isert_conn, isert_cmd, cmd,
1495                                         rx_desc, (unsigned char *)hdr);
1496                 break;
1497         case ISCSI_OP_NOOP_OUT:
1498                 cmd = isert_allocate_cmd(conn, rx_desc);
1499                 if (!cmd)
1500                         break;
1501
1502                 isert_cmd = iscsit_priv_cmd(cmd);
1503                 ret = isert_handle_nop_out(isert_conn, isert_cmd, cmd,
1504                                            rx_desc, (unsigned char *)hdr);
1505                 break;
1506         case ISCSI_OP_SCSI_DATA_OUT:
1507                 ret = isert_handle_iscsi_dataout(isert_conn, rx_desc,
1508                                                 (unsigned char *)hdr);
1509                 break;
1510         case ISCSI_OP_SCSI_TMFUNC:
1511                 cmd = isert_allocate_cmd(conn, rx_desc);
1512                 if (!cmd)
1513                         break;
1514
1515                 ret = iscsit_handle_task_mgt_cmd(conn, cmd,
1516                                                 (unsigned char *)hdr);
1517                 break;
1518         case ISCSI_OP_LOGOUT:
1519                 cmd = isert_allocate_cmd(conn, rx_desc);
1520                 if (!cmd)
1521                         break;
1522
1523                 ret = iscsit_handle_logout_cmd(conn, cmd, (unsigned char *)hdr);
1524                 break;
1525         case ISCSI_OP_TEXT:
1526                 if (be32_to_cpu(hdr->ttt) != 0xFFFFFFFF)
1527                         cmd = iscsit_find_cmd_from_itt(conn, hdr->itt);
1528                 else
1529                         cmd = isert_allocate_cmd(conn, rx_desc);
1530
1531                 if (!cmd)
1532                         break;
1533
1534                 isert_cmd = iscsit_priv_cmd(cmd);
1535                 ret = isert_handle_text_cmd(isert_conn, isert_cmd, cmd,
1536                                             rx_desc, (struct iscsi_text *)hdr);
1537                 break;
1538         default:
1539                 isert_err("Got unknown iSCSI OpCode: 0x%02x\n", opcode);
1540                 dump_stack();
1541                 break;
1542         }
1543
1544         return ret;
1545 }
1546
1547 static void
1548 isert_rx_do_work(struct iser_rx_desc *rx_desc, struct isert_conn *isert_conn)
1549 {
1550         struct iser_hdr *iser_hdr = &rx_desc->iser_header;
1551         uint64_t read_va = 0, write_va = 0;
1552         uint32_t read_stag = 0, write_stag = 0;
1553
1554         switch (iser_hdr->flags & 0xF0) {
1555         case ISCSI_CTRL:
1556                 if (iser_hdr->flags & ISER_RSV) {
1557                         read_stag = be32_to_cpu(iser_hdr->read_stag);
1558                         read_va = be64_to_cpu(iser_hdr->read_va);
1559                         isert_dbg("ISER_RSV: read_stag: 0x%x read_va: 0x%llx\n",
1560                                   read_stag, (unsigned long long)read_va);
1561                 }
1562                 if (iser_hdr->flags & ISER_WSV) {
1563                         write_stag = be32_to_cpu(iser_hdr->write_stag);
1564                         write_va = be64_to_cpu(iser_hdr->write_va);
1565                         isert_dbg("ISER_WSV: write_stag: 0x%x write_va: 0x%llx\n",
1566                                   write_stag, (unsigned long long)write_va);
1567                 }
1568
1569                 isert_dbg("ISER ISCSI_CTRL PDU\n");
1570                 break;
1571         case ISER_HELLO:
1572                 isert_err("iSER Hello message\n");
1573                 break;
1574         default:
1575                 isert_warn("Unknown iSER hdr flags: 0x%02x\n", iser_hdr->flags);
1576                 break;
1577         }
1578
1579         isert_rx_opcode(isert_conn, rx_desc,
1580                         read_stag, read_va, write_stag, write_va);
1581 }
1582
1583 static void
1584 isert_rcv_completion(struct iser_rx_desc *desc,
1585                      struct isert_conn *isert_conn,
1586                      u32 xfer_len)
1587 {
1588         struct ib_device *ib_dev = isert_conn->cm_id->device;
1589         struct iscsi_hdr *hdr;
1590         u64 rx_dma;
1591         int rx_buflen;
1592
1593         if ((char *)desc == isert_conn->login_req_buf) {
1594                 rx_dma = isert_conn->login_req_dma;
1595                 rx_buflen = ISER_RX_LOGIN_SIZE;
1596                 isert_dbg("login_buf: Using rx_dma: 0x%llx, rx_buflen: %d\n",
1597                          rx_dma, rx_buflen);
1598         } else {
1599                 rx_dma = desc->dma_addr;
1600                 rx_buflen = ISER_RX_PAYLOAD_SIZE;
1601                 isert_dbg("req_buf: Using rx_dma: 0x%llx, rx_buflen: %d\n",
1602                          rx_dma, rx_buflen);
1603         }
1604
1605         ib_dma_sync_single_for_cpu(ib_dev, rx_dma, rx_buflen, DMA_FROM_DEVICE);
1606
1607         hdr = &desc->iscsi_header;
1608         isert_dbg("iSCSI opcode: 0x%02x, ITT: 0x%08x, flags: 0x%02x dlen: %d\n",
1609                  hdr->opcode, hdr->itt, hdr->flags,
1610                  (int)(xfer_len - ISER_HEADERS_LEN));
1611
1612         if ((char *)desc == isert_conn->login_req_buf) {
1613                 isert_conn->login_req_len = xfer_len - ISER_HEADERS_LEN;
1614                 if (isert_conn->conn) {
1615                         struct iscsi_login *login = isert_conn->conn->conn_login;
1616
1617                         if (login && !login->first_request)
1618                                 isert_rx_login_req(isert_conn);
1619                 }
1620                 mutex_lock(&isert_conn->mutex);
1621                 complete(&isert_conn->login_req_comp);
1622                 mutex_unlock(&isert_conn->mutex);
1623         } else {
1624                 isert_rx_do_work(desc, isert_conn);
1625         }
1626
1627         ib_dma_sync_single_for_device(ib_dev, rx_dma, rx_buflen,
1628                                       DMA_FROM_DEVICE);
1629
1630         isert_conn->post_recv_buf_count--;
1631 }
1632
1633 static int
1634 isert_map_data_buf(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1635                    struct scatterlist *sg, u32 nents, u32 length, u32 offset,
1636                    enum iser_ib_op_code op, struct isert_data_buf *data)
1637 {
1638         struct ib_device *ib_dev = isert_conn->cm_id->device;
1639
1640         data->dma_dir = op == ISER_IB_RDMA_WRITE ?
1641                               DMA_TO_DEVICE : DMA_FROM_DEVICE;
1642
1643         data->len = length - offset;
1644         data->offset = offset;
1645         data->sg_off = data->offset / PAGE_SIZE;
1646
1647         data->sg = &sg[data->sg_off];
1648         data->nents = min_t(unsigned int, nents - data->sg_off,
1649                                           ISCSI_ISER_SG_TABLESIZE);
1650         data->len = min_t(unsigned int, data->len, ISCSI_ISER_SG_TABLESIZE *
1651                                         PAGE_SIZE);
1652
1653         data->dma_nents = ib_dma_map_sg(ib_dev, data->sg, data->nents,
1654                                         data->dma_dir);
1655         if (unlikely(!data->dma_nents)) {
1656                 isert_err("Cmd: unable to dma map SGs %p\n", sg);
1657                 return -EINVAL;
1658         }
1659
1660         isert_dbg("Mapped cmd: %p count: %u sg: %p sg_nents: %u rdma_len %d\n",
1661                   isert_cmd, data->dma_nents, data->sg, data->nents, data->len);
1662
1663         return 0;
1664 }
1665
1666 static void
1667 isert_unmap_data_buf(struct isert_conn *isert_conn, struct isert_data_buf *data)
1668 {
1669         struct ib_device *ib_dev = isert_conn->cm_id->device;
1670
1671         ib_dma_unmap_sg(ib_dev, data->sg, data->nents, data->dma_dir);
1672         memset(data, 0, sizeof(*data));
1673 }
1674
1675
1676
1677 static void
1678 isert_unmap_cmd(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn)
1679 {
1680         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1681
1682         isert_dbg("Cmd %p\n", isert_cmd);
1683
1684         if (wr->data.sg) {
1685                 isert_dbg("Cmd %p unmap_sg op\n", isert_cmd);
1686                 isert_unmap_data_buf(isert_conn, &wr->data);
1687         }
1688
1689         if (wr->rdma_wr) {
1690                 isert_dbg("Cmd %p free send_wr\n", isert_cmd);
1691                 kfree(wr->rdma_wr);
1692                 wr->rdma_wr = NULL;
1693         }
1694
1695         if (wr->ib_sge) {
1696                 isert_dbg("Cmd %p free ib_sge\n", isert_cmd);
1697                 kfree(wr->ib_sge);
1698                 wr->ib_sge = NULL;
1699         }
1700 }
1701
1702 static void
1703 isert_unreg_rdma(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn)
1704 {
1705         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1706
1707         isert_dbg("Cmd %p\n", isert_cmd);
1708
1709         if (wr->fr_desc) {
1710                 isert_dbg("Cmd %p free fr_desc %p\n", isert_cmd, wr->fr_desc);
1711                 if (wr->fr_desc->ind & ISERT_PROTECTED) {
1712                         isert_unmap_data_buf(isert_conn, &wr->prot);
1713                         wr->fr_desc->ind &= ~ISERT_PROTECTED;
1714                 }
1715                 spin_lock_bh(&isert_conn->pool_lock);
1716                 list_add_tail(&wr->fr_desc->list, &isert_conn->fr_pool);
1717                 spin_unlock_bh(&isert_conn->pool_lock);
1718                 wr->fr_desc = NULL;
1719         }
1720
1721         if (wr->data.sg) {
1722                 isert_dbg("Cmd %p unmap_sg op\n", isert_cmd);
1723                 isert_unmap_data_buf(isert_conn, &wr->data);
1724         }
1725
1726         wr->ib_sge = NULL;
1727         wr->rdma_wr = NULL;
1728 }
1729
1730 static void
1731 isert_put_cmd(struct isert_cmd *isert_cmd, bool comp_err)
1732 {
1733         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1734         struct isert_conn *isert_conn = isert_cmd->conn;
1735         struct iscsi_conn *conn = isert_conn->conn;
1736         struct isert_device *device = isert_conn->device;
1737         struct iscsi_text_rsp *hdr;
1738
1739         isert_dbg("Cmd %p\n", isert_cmd);
1740
1741         switch (cmd->iscsi_opcode) {
1742         case ISCSI_OP_SCSI_CMD:
1743                 spin_lock_bh(&conn->cmd_lock);
1744                 if (!list_empty(&cmd->i_conn_node))
1745                         list_del_init(&cmd->i_conn_node);
1746                 spin_unlock_bh(&conn->cmd_lock);
1747
1748                 if (cmd->data_direction == DMA_TO_DEVICE) {
1749                         iscsit_stop_dataout_timer(cmd);
1750                         /*
1751                          * Check for special case during comp_err where
1752                          * WRITE_PENDING has been handed off from core,
1753                          * but requires an extra target_put_sess_cmd()
1754                          * before transport_generic_free_cmd() below.
1755                          */
1756                         if (comp_err &&
1757                             cmd->se_cmd.t_state == TRANSPORT_WRITE_PENDING) {
1758                                 struct se_cmd *se_cmd = &cmd->se_cmd;
1759
1760                                 target_put_sess_cmd(se_cmd);
1761                         }
1762                 }
1763
1764                 device->unreg_rdma_mem(isert_cmd, isert_conn);
1765                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1766                 break;
1767         case ISCSI_OP_SCSI_TMFUNC:
1768                 spin_lock_bh(&conn->cmd_lock);
1769                 if (!list_empty(&cmd->i_conn_node))
1770                         list_del_init(&cmd->i_conn_node);
1771                 spin_unlock_bh(&conn->cmd_lock);
1772
1773                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1774                 break;
1775         case ISCSI_OP_REJECT:
1776         case ISCSI_OP_NOOP_OUT:
1777         case ISCSI_OP_TEXT:
1778                 hdr = (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
1779                 /* If the continue bit is on, keep the command alive */
1780                 if (hdr->flags & ISCSI_FLAG_TEXT_CONTINUE)
1781                         break;
1782
1783                 spin_lock_bh(&conn->cmd_lock);
1784                 if (!list_empty(&cmd->i_conn_node))
1785                         list_del_init(&cmd->i_conn_node);
1786                 spin_unlock_bh(&conn->cmd_lock);
1787
1788                 /*
1789                  * Handle special case for REJECT when iscsi_add_reject*() has
1790                  * overwritten the original iscsi_opcode assignment, and the
1791                  * associated cmd->se_cmd needs to be released.
1792                  */
1793                 if (cmd->se_cmd.se_tfo != NULL) {
1794                         isert_dbg("Calling transport_generic_free_cmd for 0x%02x\n",
1795                                  cmd->iscsi_opcode);
1796                         transport_generic_free_cmd(&cmd->se_cmd, 0);
1797                         break;
1798                 }
1799                 /*
1800                  * Fall-through
1801                  */
1802         default:
1803                 iscsit_release_cmd(cmd);
1804                 break;
1805         }
1806 }
1807
1808 static void
1809 isert_unmap_tx_desc(struct iser_tx_desc *tx_desc, struct ib_device *ib_dev)
1810 {
1811         if (tx_desc->dma_addr != 0) {
1812                 isert_dbg("unmap single for tx_desc->dma_addr\n");
1813                 ib_dma_unmap_single(ib_dev, tx_desc->dma_addr,
1814                                     ISER_HEADERS_LEN, DMA_TO_DEVICE);
1815                 tx_desc->dma_addr = 0;
1816         }
1817 }
1818
1819 static void
1820 isert_completion_put(struct iser_tx_desc *tx_desc, struct isert_cmd *isert_cmd,
1821                      struct ib_device *ib_dev, bool comp_err)
1822 {
1823         if (isert_cmd->pdu_buf_dma != 0) {
1824                 isert_dbg("unmap single for isert_cmd->pdu_buf_dma\n");
1825                 ib_dma_unmap_single(ib_dev, isert_cmd->pdu_buf_dma,
1826                                     isert_cmd->pdu_buf_len, DMA_TO_DEVICE);
1827                 isert_cmd->pdu_buf_dma = 0;
1828         }
1829
1830         isert_unmap_tx_desc(tx_desc, ib_dev);
1831         isert_put_cmd(isert_cmd, comp_err);
1832 }
1833
1834 static int
1835 isert_check_pi_status(struct se_cmd *se_cmd, struct ib_mr *sig_mr)
1836 {
1837         struct ib_mr_status mr_status;
1838         int ret;
1839
1840         ret = ib_check_mr_status(sig_mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
1841         if (ret) {
1842                 isert_err("ib_check_mr_status failed, ret %d\n", ret);
1843                 goto fail_mr_status;
1844         }
1845
1846         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1847                 u64 sec_offset_err;
1848                 u32 block_size = se_cmd->se_dev->dev_attrib.block_size + 8;
1849
1850                 switch (mr_status.sig_err.err_type) {
1851                 case IB_SIG_BAD_GUARD:
1852                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_GUARD_CHECK_FAILED;
1853                         break;
1854                 case IB_SIG_BAD_REFTAG:
1855                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_REF_TAG_CHECK_FAILED;
1856                         break;
1857                 case IB_SIG_BAD_APPTAG:
1858                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_APP_TAG_CHECK_FAILED;
1859                         break;
1860                 }
1861                 sec_offset_err = mr_status.sig_err.sig_err_offset;
1862                 do_div(sec_offset_err, block_size);
1863                 se_cmd->bad_sector = sec_offset_err + se_cmd->t_task_lba;
1864
1865                 isert_err("PI error found type %d at sector 0x%llx "
1866                           "expected 0x%x vs actual 0x%x\n",
1867                           mr_status.sig_err.err_type,
1868                           (unsigned long long)se_cmd->bad_sector,
1869                           mr_status.sig_err.expected,
1870                           mr_status.sig_err.actual);
1871                 ret = 1;
1872         }
1873
1874 fail_mr_status:
1875         return ret;
1876 }
1877
1878 static void
1879 isert_completion_rdma_write(struct iser_tx_desc *tx_desc,
1880                             struct isert_cmd *isert_cmd)
1881 {
1882         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1883         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1884         struct se_cmd *se_cmd = &cmd->se_cmd;
1885         struct isert_conn *isert_conn = isert_cmd->conn;
1886         struct isert_device *device = isert_conn->device;
1887         int ret = 0;
1888
1889         if (wr->fr_desc && wr->fr_desc->ind & ISERT_PROTECTED) {
1890                 ret = isert_check_pi_status(se_cmd,
1891                                             wr->fr_desc->pi_ctx->sig_mr);
1892                 wr->fr_desc->ind &= ~ISERT_PROTECTED;
1893         }
1894
1895         device->unreg_rdma_mem(isert_cmd, isert_conn);
1896         wr->rdma_wr_num = 0;
1897         if (ret)
1898                 transport_send_check_condition_and_sense(se_cmd,
1899                                                          se_cmd->pi_err, 0);
1900         else
1901                 isert_put_response(isert_conn->conn, cmd);
1902 }
1903
1904 static void
1905 isert_completion_rdma_read(struct iser_tx_desc *tx_desc,
1906                            struct isert_cmd *isert_cmd)
1907 {
1908         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1909         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1910         struct se_cmd *se_cmd = &cmd->se_cmd;
1911         struct isert_conn *isert_conn = isert_cmd->conn;
1912         struct isert_device *device = isert_conn->device;
1913         int ret = 0;
1914
1915         if (wr->fr_desc && wr->fr_desc->ind & ISERT_PROTECTED) {
1916                 ret = isert_check_pi_status(se_cmd,
1917                                             wr->fr_desc->pi_ctx->sig_mr);
1918                 wr->fr_desc->ind &= ~ISERT_PROTECTED;
1919         }
1920
1921         iscsit_stop_dataout_timer(cmd);
1922         device->unreg_rdma_mem(isert_cmd, isert_conn);
1923         cmd->write_data_done = wr->data.len;
1924         wr->rdma_wr_num = 0;
1925
1926         isert_dbg("Cmd: %p RDMA_READ comp calling execute_cmd\n", isert_cmd);
1927         spin_lock_bh(&cmd->istate_lock);
1928         cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1929         cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1930         spin_unlock_bh(&cmd->istate_lock);
1931
1932         if (ret) {
1933                 target_put_sess_cmd(se_cmd);
1934                 transport_send_check_condition_and_sense(se_cmd,
1935                                                          se_cmd->pi_err, 0);
1936         } else {
1937                 target_execute_cmd(se_cmd);
1938         }
1939 }
1940
1941 static void
1942 isert_do_control_comp(struct work_struct *work)
1943 {
1944         struct isert_cmd *isert_cmd = container_of(work,
1945                         struct isert_cmd, comp_work);
1946         struct isert_conn *isert_conn = isert_cmd->conn;
1947         struct ib_device *ib_dev = isert_conn->cm_id->device;
1948         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1949
1950         isert_dbg("Cmd %p i_state %d\n", isert_cmd, cmd->i_state);
1951
1952         switch (cmd->i_state) {
1953         case ISTATE_SEND_TASKMGTRSP:
1954                 iscsit_tmr_post_handler(cmd, cmd->conn);
1955         case ISTATE_SEND_REJECT:   /* FALLTHRU */
1956         case ISTATE_SEND_TEXTRSP:  /* FALLTHRU */
1957                 cmd->i_state = ISTATE_SENT_STATUS;
1958                 isert_completion_put(&isert_cmd->tx_desc, isert_cmd,
1959                                      ib_dev, false);
1960                 break;
1961         case ISTATE_SEND_LOGOUTRSP:
1962                 iscsit_logout_post_handler(cmd, cmd->conn);
1963                 break;
1964         default:
1965                 isert_err("Unknown i_state %d\n", cmd->i_state);
1966                 dump_stack();
1967                 break;
1968         }
1969 }
1970
1971 static void
1972 isert_response_completion(struct iser_tx_desc *tx_desc,
1973                           struct isert_cmd *isert_cmd,
1974                           struct isert_conn *isert_conn,
1975                           struct ib_device *ib_dev)
1976 {
1977         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1978
1979         if (cmd->i_state == ISTATE_SEND_TASKMGTRSP ||
1980             cmd->i_state == ISTATE_SEND_LOGOUTRSP ||
1981             cmd->i_state == ISTATE_SEND_REJECT ||
1982             cmd->i_state == ISTATE_SEND_TEXTRSP) {
1983                 isert_unmap_tx_desc(tx_desc, ib_dev);
1984
1985                 INIT_WORK(&isert_cmd->comp_work, isert_do_control_comp);
1986                 queue_work(isert_comp_wq, &isert_cmd->comp_work);
1987                 return;
1988         }
1989
1990         cmd->i_state = ISTATE_SENT_STATUS;
1991         isert_completion_put(tx_desc, isert_cmd, ib_dev, false);
1992 }
1993
1994 static void
1995 isert_snd_completion(struct iser_tx_desc *tx_desc,
1996                       struct isert_conn *isert_conn)
1997 {
1998         struct ib_device *ib_dev = isert_conn->cm_id->device;
1999         struct isert_cmd *isert_cmd = tx_desc->isert_cmd;
2000         struct isert_rdma_wr *wr;
2001
2002         if (!isert_cmd) {
2003                 isert_unmap_tx_desc(tx_desc, ib_dev);
2004                 return;
2005         }
2006         wr = &isert_cmd->rdma_wr;
2007
2008         isert_dbg("Cmd %p iser_ib_op %d\n", isert_cmd, wr->iser_ib_op);
2009
2010         switch (wr->iser_ib_op) {
2011         case ISER_IB_SEND:
2012                 isert_response_completion(tx_desc, isert_cmd,
2013                                           isert_conn, ib_dev);
2014                 break;
2015         case ISER_IB_RDMA_WRITE:
2016                 isert_completion_rdma_write(tx_desc, isert_cmd);
2017                 break;
2018         case ISER_IB_RDMA_READ:
2019                 isert_completion_rdma_read(tx_desc, isert_cmd);
2020                 break;
2021         default:
2022                 isert_err("Unknown wr->iser_ib_op: 0x%x\n", wr->iser_ib_op);
2023                 dump_stack();
2024                 break;
2025         }
2026 }
2027
2028 /**
2029  * is_isert_tx_desc() - Indicate if the completion wr_id
2030  *     is a TX descriptor or not.
2031  * @isert_conn: iser connection
2032  * @wr_id: completion WR identifier
2033  *
2034  * Since we cannot rely on wc opcode in FLUSH errors
2035  * we must work around it by checking if the wr_id address
2036  * falls in the iser connection rx_descs buffer. If so
2037  * it is an RX descriptor, otherwize it is a TX.
2038  */
2039 static inline bool
2040 is_isert_tx_desc(struct isert_conn *isert_conn, void *wr_id)
2041 {
2042         void *start = isert_conn->rx_descs;
2043         int len = ISERT_QP_MAX_RECV_DTOS * sizeof(*isert_conn->rx_descs);
2044
2045         if (wr_id >= start && wr_id < start + len)
2046                 return false;
2047
2048         return true;
2049 }
2050
2051 static void
2052 isert_cq_comp_err(struct isert_conn *isert_conn, struct ib_wc *wc)
2053 {
2054         if (wc->wr_id == ISER_BEACON_WRID) {
2055                 isert_info("conn %p completing wait_comp_err\n",
2056                            isert_conn);
2057                 complete(&isert_conn->wait_comp_err);
2058         } else if (is_isert_tx_desc(isert_conn, (void *)(uintptr_t)wc->wr_id)) {
2059                 struct ib_device *ib_dev = isert_conn->cm_id->device;
2060                 struct isert_cmd *isert_cmd;
2061                 struct iser_tx_desc *desc;
2062
2063                 desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
2064                 isert_cmd = desc->isert_cmd;
2065                 if (!isert_cmd)
2066                         isert_unmap_tx_desc(desc, ib_dev);
2067                 else
2068                         isert_completion_put(desc, isert_cmd, ib_dev, true);
2069         } else {
2070                 isert_conn->post_recv_buf_count--;
2071                 if (!isert_conn->post_recv_buf_count)
2072                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
2073         }
2074 }
2075
2076 static void
2077 isert_handle_wc(struct ib_wc *wc)
2078 {
2079         struct isert_conn *isert_conn;
2080         struct iser_tx_desc *tx_desc;
2081         struct iser_rx_desc *rx_desc;
2082
2083         isert_conn = wc->qp->qp_context;
2084         if (likely(wc->status == IB_WC_SUCCESS)) {
2085                 if (wc->opcode == IB_WC_RECV) {
2086                         rx_desc = (struct iser_rx_desc *)(uintptr_t)wc->wr_id;
2087                         isert_rcv_completion(rx_desc, isert_conn, wc->byte_len);
2088                 } else {
2089                         tx_desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
2090                         isert_snd_completion(tx_desc, isert_conn);
2091                 }
2092         } else {
2093                 if (wc->status != IB_WC_WR_FLUSH_ERR)
2094                         isert_err("%s (%d): wr id %llx vend_err %x\n",
2095                                   ib_wc_status_msg(wc->status), wc->status,
2096                                   wc->wr_id, wc->vendor_err);
2097                 else
2098                         isert_dbg("%s (%d): wr id %llx\n",
2099                                   ib_wc_status_msg(wc->status), wc->status,
2100                                   wc->wr_id);
2101
2102                 if (wc->wr_id != ISER_FASTREG_LI_WRID)
2103                         isert_cq_comp_err(isert_conn, wc);
2104         }
2105 }
2106
2107 static void
2108 isert_cq_work(struct work_struct *work)
2109 {
2110         enum { isert_poll_budget = 65536 };
2111         struct isert_comp *comp = container_of(work, struct isert_comp,
2112                                                work);
2113         struct ib_wc *const wcs = comp->wcs;
2114         int i, n, completed = 0;
2115
2116         while ((n = ib_poll_cq(comp->cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
2117                 for (i = 0; i < n; i++)
2118                         isert_handle_wc(&wcs[i]);
2119
2120                 completed += n;
2121                 if (completed >= isert_poll_budget)
2122                         break;
2123         }
2124
2125         ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP);
2126 }
2127
2128 static void
2129 isert_cq_callback(struct ib_cq *cq, void *context)
2130 {
2131         struct isert_comp *comp = context;
2132
2133         queue_work(isert_comp_wq, &comp->work);
2134 }
2135
2136 static int
2137 isert_post_response(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd)
2138 {
2139         struct ib_send_wr *wr_failed;
2140         int ret;
2141
2142         ret = isert_post_recv(isert_conn, isert_cmd->rx_desc);
2143         if (ret) {
2144                 isert_err("ib_post_recv failed with %d\n", ret);
2145                 return ret;
2146         }
2147
2148         ret = ib_post_send(isert_conn->qp, &isert_cmd->tx_desc.send_wr,
2149                            &wr_failed);
2150         if (ret) {
2151                 isert_err("ib_post_send failed with %d\n", ret);
2152                 return ret;
2153         }
2154         return ret;
2155 }
2156
2157 static int
2158 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2159 {
2160         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2161         struct isert_conn *isert_conn = conn->context;
2162         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2163         struct iscsi_scsi_rsp *hdr = (struct iscsi_scsi_rsp *)
2164                                 &isert_cmd->tx_desc.iscsi_header;
2165
2166         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2167         iscsit_build_rsp_pdu(cmd, conn, true, hdr);
2168         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2169         /*
2170          * Attach SENSE DATA payload to iSCSI Response PDU
2171          */
2172         if (cmd->se_cmd.sense_buffer &&
2173             ((cmd->se_cmd.se_cmd_flags & SCF_TRANSPORT_TASK_SENSE) ||
2174             (cmd->se_cmd.se_cmd_flags & SCF_EMULATED_TASK_SENSE))) {
2175                 struct isert_device *device = isert_conn->device;
2176                 struct ib_device *ib_dev = device->ib_device;
2177                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2178                 u32 padding, pdu_len;
2179
2180                 put_unaligned_be16(cmd->se_cmd.scsi_sense_length,
2181                                    cmd->sense_buffer);
2182                 cmd->se_cmd.scsi_sense_length += sizeof(__be16);
2183
2184                 padding = -(cmd->se_cmd.scsi_sense_length) & 3;
2185                 hton24(hdr->dlength, (u32)cmd->se_cmd.scsi_sense_length);
2186                 pdu_len = cmd->se_cmd.scsi_sense_length + padding;
2187
2188                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2189                                 (void *)cmd->sense_buffer, pdu_len,
2190                                 DMA_TO_DEVICE);
2191
2192                 isert_cmd->pdu_buf_len = pdu_len;
2193                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2194                 tx_dsg->length  = pdu_len;
2195                 tx_dsg->lkey    = device->pd->local_dma_lkey;
2196                 isert_cmd->tx_desc.num_sge = 2;
2197         }
2198
2199         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2200
2201         isert_dbg("Posting SCSI Response\n");
2202
2203         return isert_post_response(isert_conn, isert_cmd);
2204 }
2205
2206 static void
2207 isert_aborted_task(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2208 {
2209         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2210         struct isert_conn *isert_conn = conn->context;
2211         struct isert_device *device = isert_conn->device;
2212
2213         spin_lock_bh(&conn->cmd_lock);
2214         if (!list_empty(&cmd->i_conn_node))
2215                 list_del_init(&cmd->i_conn_node);
2216         spin_unlock_bh(&conn->cmd_lock);
2217
2218         if (cmd->data_direction == DMA_TO_DEVICE)
2219                 iscsit_stop_dataout_timer(cmd);
2220
2221         device->unreg_rdma_mem(isert_cmd, isert_conn);
2222 }
2223
2224 static enum target_prot_op
2225 isert_get_sup_prot_ops(struct iscsi_conn *conn)
2226 {
2227         struct isert_conn *isert_conn = conn->context;
2228         struct isert_device *device = isert_conn->device;
2229
2230         if (conn->tpg->tpg_attrib.t10_pi) {
2231                 if (device->pi_capable) {
2232                         isert_info("conn %p PI offload enabled\n", isert_conn);
2233                         isert_conn->pi_support = true;
2234                         return TARGET_PROT_ALL;
2235                 }
2236         }
2237
2238         isert_info("conn %p PI offload disabled\n", isert_conn);
2239         isert_conn->pi_support = false;
2240
2241         return TARGET_PROT_NORMAL;
2242 }
2243
2244 static int
2245 isert_put_nopin(struct iscsi_cmd *cmd, struct iscsi_conn *conn,
2246                 bool nopout_response)
2247 {
2248         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2249         struct isert_conn *isert_conn = conn->context;
2250         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2251
2252         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2253         iscsit_build_nopin_rsp(cmd, conn, (struct iscsi_nopin *)
2254                                &isert_cmd->tx_desc.iscsi_header,
2255                                nopout_response);
2256         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2257         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2258
2259         isert_dbg("conn %p Posting NOPIN Response\n", isert_conn);
2260
2261         return isert_post_response(isert_conn, isert_cmd);
2262 }
2263
2264 static int
2265 isert_put_logout_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2266 {
2267         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2268         struct isert_conn *isert_conn = conn->context;
2269         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2270
2271         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2272         iscsit_build_logout_rsp(cmd, conn, (struct iscsi_logout_rsp *)
2273                                 &isert_cmd->tx_desc.iscsi_header);
2274         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2275         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2276
2277         isert_dbg("conn %p Posting Logout Response\n", isert_conn);
2278
2279         return isert_post_response(isert_conn, isert_cmd);
2280 }
2281
2282 static int
2283 isert_put_tm_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2284 {
2285         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2286         struct isert_conn *isert_conn = conn->context;
2287         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2288
2289         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2290         iscsit_build_task_mgt_rsp(cmd, conn, (struct iscsi_tm_rsp *)
2291                                   &isert_cmd->tx_desc.iscsi_header);
2292         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2293         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2294
2295         isert_dbg("conn %p Posting Task Management Response\n", isert_conn);
2296
2297         return isert_post_response(isert_conn, isert_cmd);
2298 }
2299
2300 static int
2301 isert_put_reject(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2302 {
2303         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2304         struct isert_conn *isert_conn = conn->context;
2305         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2306         struct isert_device *device = isert_conn->device;
2307         struct ib_device *ib_dev = device->ib_device;
2308         struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2309         struct iscsi_reject *hdr =
2310                 (struct iscsi_reject *)&isert_cmd->tx_desc.iscsi_header;
2311
2312         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2313         iscsit_build_reject(cmd, conn, hdr);
2314         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2315
2316         hton24(hdr->dlength, ISCSI_HDR_LEN);
2317         isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2318                         (void *)cmd->buf_ptr, ISCSI_HDR_LEN,
2319                         DMA_TO_DEVICE);
2320         isert_cmd->pdu_buf_len = ISCSI_HDR_LEN;
2321         tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2322         tx_dsg->length  = ISCSI_HDR_LEN;
2323         tx_dsg->lkey    = device->pd->local_dma_lkey;
2324         isert_cmd->tx_desc.num_sge = 2;
2325
2326         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2327
2328         isert_dbg("conn %p Posting Reject\n", isert_conn);
2329
2330         return isert_post_response(isert_conn, isert_cmd);
2331 }
2332
2333 static int
2334 isert_put_text_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2335 {
2336         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2337         struct isert_conn *isert_conn = conn->context;
2338         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2339         struct iscsi_text_rsp *hdr =
2340                 (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
2341         u32 txt_rsp_len;
2342         int rc;
2343
2344         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2345         rc = iscsit_build_text_rsp(cmd, conn, hdr, ISCSI_INFINIBAND);
2346         if (rc < 0)
2347                 return rc;
2348
2349         txt_rsp_len = rc;
2350         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2351
2352         if (txt_rsp_len) {
2353                 struct isert_device *device = isert_conn->device;
2354                 struct ib_device *ib_dev = device->ib_device;
2355                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2356                 void *txt_rsp_buf = cmd->buf_ptr;
2357
2358                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2359                                 txt_rsp_buf, txt_rsp_len, DMA_TO_DEVICE);
2360
2361                 isert_cmd->pdu_buf_len = txt_rsp_len;
2362                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2363                 tx_dsg->length  = txt_rsp_len;
2364                 tx_dsg->lkey    = device->pd->local_dma_lkey;
2365                 isert_cmd->tx_desc.num_sge = 2;
2366         }
2367         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2368
2369         isert_dbg("conn %p Text Response\n", isert_conn);
2370
2371         return isert_post_response(isert_conn, isert_cmd);
2372 }
2373
2374 static int
2375 isert_build_rdma_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
2376                     struct ib_sge *ib_sge, struct ib_rdma_wr *rdma_wr,
2377                     u32 data_left, u32 offset)
2378 {
2379         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
2380         struct scatterlist *sg_start, *tmp_sg;
2381         struct isert_device *device = isert_conn->device;
2382         struct ib_device *ib_dev = device->ib_device;
2383         u32 sg_off, page_off;
2384         int i = 0, sg_nents;
2385
2386         sg_off = offset / PAGE_SIZE;
2387         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
2388         sg_nents = min(cmd->se_cmd.t_data_nents - sg_off, isert_conn->max_sge);
2389         page_off = offset % PAGE_SIZE;
2390
2391         rdma_wr->wr.sg_list = ib_sge;
2392         rdma_wr->wr.wr_id = (uintptr_t)&isert_cmd->tx_desc;
2393         /*
2394          * Perform mapping of TCM scatterlist memory ib_sge dma_addr.
2395          */
2396         for_each_sg(sg_start, tmp_sg, sg_nents, i) {
2397                 isert_dbg("RDMA from SGL dma_addr: 0x%llx dma_len: %u, "
2398                           "page_off: %u\n",
2399                           (unsigned long long)tmp_sg->dma_address,
2400                           tmp_sg->length, page_off);
2401
2402                 ib_sge->addr = ib_sg_dma_address(ib_dev, tmp_sg) + page_off;
2403                 ib_sge->length = min_t(u32, data_left,
2404                                 ib_sg_dma_len(ib_dev, tmp_sg) - page_off);
2405                 ib_sge->lkey = device->pd->local_dma_lkey;
2406
2407                 isert_dbg("RDMA ib_sge: addr: 0x%llx  length: %u lkey: %x\n",
2408                           ib_sge->addr, ib_sge->length, ib_sge->lkey);
2409                 page_off = 0;
2410                 data_left -= ib_sge->length;
2411                 if (!data_left)
2412                         break;
2413                 ib_sge++;
2414                 isert_dbg("Incrementing ib_sge pointer to %p\n", ib_sge);
2415         }
2416
2417         rdma_wr->wr.num_sge = ++i;
2418         isert_dbg("Set outgoing sg_list: %p num_sg: %u from TCM SGLs\n",
2419                   rdma_wr->wr.sg_list, rdma_wr->wr.num_sge);
2420
2421         return rdma_wr->wr.num_sge;
2422 }
2423
2424 static int
2425 isert_map_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2426                struct isert_rdma_wr *wr)
2427 {
2428         struct se_cmd *se_cmd = &cmd->se_cmd;
2429         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2430         struct isert_conn *isert_conn = conn->context;
2431         struct isert_data_buf *data = &wr->data;
2432         struct ib_rdma_wr *rdma_wr;
2433         struct ib_sge *ib_sge;
2434         u32 offset, data_len, data_left, rdma_write_max, va_offset = 0;
2435         int ret = 0, i, ib_sge_cnt;
2436
2437         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2438
2439         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2440         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2441                                  se_cmd->t_data_nents, se_cmd->data_length,
2442                                  offset, wr->iser_ib_op, &wr->data);
2443         if (ret)
2444                 return ret;
2445
2446         data_left = data->len;
2447         offset = data->offset;
2448
2449         ib_sge = kzalloc(sizeof(struct ib_sge) * data->nents, GFP_KERNEL);
2450         if (!ib_sge) {
2451                 isert_warn("Unable to allocate ib_sge\n");
2452                 ret = -ENOMEM;
2453                 goto unmap_cmd;
2454         }
2455         wr->ib_sge = ib_sge;
2456
2457         wr->rdma_wr_num = DIV_ROUND_UP(data->nents, isert_conn->max_sge);
2458         wr->rdma_wr = kzalloc(sizeof(struct ib_rdma_wr) * wr->rdma_wr_num,
2459                                 GFP_KERNEL);
2460         if (!wr->rdma_wr) {
2461                 isert_dbg("Unable to allocate wr->rdma_wr\n");
2462                 ret = -ENOMEM;
2463                 goto unmap_cmd;
2464         }
2465
2466         wr->isert_cmd = isert_cmd;
2467         rdma_write_max = isert_conn->max_sge * PAGE_SIZE;
2468
2469         for (i = 0; i < wr->rdma_wr_num; i++) {
2470                 rdma_wr = &isert_cmd->rdma_wr.rdma_wr[i];
2471                 data_len = min(data_left, rdma_write_max);
2472
2473                 rdma_wr->wr.send_flags = 0;
2474                 if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2475                         rdma_wr->wr.opcode = IB_WR_RDMA_WRITE;
2476                         rdma_wr->remote_addr = isert_cmd->read_va + offset;
2477                         rdma_wr->rkey = isert_cmd->read_stag;
2478                         if (i + 1 == wr->rdma_wr_num)
2479                                 rdma_wr->wr.next = &isert_cmd->tx_desc.send_wr;
2480                         else
2481                                 rdma_wr->wr.next = &wr->rdma_wr[i + 1].wr;
2482                 } else {
2483                         rdma_wr->wr.opcode = IB_WR_RDMA_READ;
2484                         rdma_wr->remote_addr = isert_cmd->write_va + va_offset;
2485                         rdma_wr->rkey = isert_cmd->write_stag;
2486                         if (i + 1 == wr->rdma_wr_num)
2487                                 rdma_wr->wr.send_flags = IB_SEND_SIGNALED;
2488                         else
2489                                 rdma_wr->wr.next = &wr->rdma_wr[i + 1].wr;
2490                 }
2491
2492                 ib_sge_cnt = isert_build_rdma_wr(isert_conn, isert_cmd, ib_sge,
2493                                         rdma_wr, data_len, offset);
2494                 ib_sge += ib_sge_cnt;
2495
2496                 offset += data_len;
2497                 va_offset += data_len;
2498                 data_left -= data_len;
2499         }
2500
2501         return 0;
2502 unmap_cmd:
2503         isert_unmap_data_buf(isert_conn, data);
2504
2505         return ret;
2506 }
2507
2508 static inline void
2509 isert_inv_rkey(struct ib_send_wr *inv_wr, struct ib_mr *mr)
2510 {
2511         u32 rkey;
2512
2513         memset(inv_wr, 0, sizeof(*inv_wr));
2514         inv_wr->wr_id = ISER_FASTREG_LI_WRID;
2515         inv_wr->opcode = IB_WR_LOCAL_INV;
2516         inv_wr->ex.invalidate_rkey = mr->rkey;
2517
2518         /* Bump the key */
2519         rkey = ib_inc_rkey(mr->rkey);
2520         ib_update_fast_reg_key(mr, rkey);
2521 }
2522
2523 static int
2524 isert_fast_reg_mr(struct isert_conn *isert_conn,
2525                   struct fast_reg_descriptor *fr_desc,
2526                   struct isert_data_buf *mem,
2527                   enum isert_indicator ind,
2528                   struct ib_sge *sge)
2529 {
2530         struct isert_device *device = isert_conn->device;
2531         struct ib_device *ib_dev = device->ib_device;
2532         struct ib_mr *mr;
2533         struct ib_reg_wr reg_wr;
2534         struct ib_send_wr inv_wr, *bad_wr, *wr = NULL;
2535         int ret, n;
2536
2537         if (mem->dma_nents == 1) {
2538                 sge->lkey = device->pd->local_dma_lkey;
2539                 sge->addr = ib_sg_dma_address(ib_dev, &mem->sg[0]);
2540                 sge->length = ib_sg_dma_len(ib_dev, &mem->sg[0]);
2541                 isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2542                          sge->addr, sge->length, sge->lkey);
2543                 return 0;
2544         }
2545
2546         if (ind == ISERT_DATA_KEY_VALID)
2547                 /* Registering data buffer */
2548                 mr = fr_desc->data_mr;
2549         else
2550                 /* Registering protection buffer */
2551                 mr = fr_desc->pi_ctx->prot_mr;
2552
2553         if (!(fr_desc->ind & ind)) {
2554                 isert_inv_rkey(&inv_wr, mr);
2555                 wr = &inv_wr;
2556         }
2557
2558         n = ib_map_mr_sg(mr, mem->sg, mem->nents, PAGE_SIZE);
2559         if (unlikely(n != mem->nents)) {
2560                 isert_err("failed to map mr sg (%d/%d)\n",
2561                          n, mem->nents);
2562                 return n < 0 ? n : -EINVAL;
2563         }
2564
2565         isert_dbg("Use fr_desc %p sg_nents %d offset %u\n",
2566                   fr_desc, mem->nents, mem->offset);
2567
2568         reg_wr.wr.next = NULL;
2569         reg_wr.wr.opcode = IB_WR_REG_MR;
2570         reg_wr.wr.wr_id = ISER_FASTREG_LI_WRID;
2571         reg_wr.wr.send_flags = 0;
2572         reg_wr.wr.num_sge = 0;
2573         reg_wr.mr = mr;
2574         reg_wr.key = mr->lkey;
2575         reg_wr.access = IB_ACCESS_LOCAL_WRITE;
2576
2577         if (!wr)
2578                 wr = &reg_wr.wr;
2579         else
2580                 wr->next = &reg_wr.wr;
2581
2582         ret = ib_post_send(isert_conn->qp, wr, &bad_wr);
2583         if (ret) {
2584                 isert_err("fast registration failed, ret:%d\n", ret);
2585                 return ret;
2586         }
2587         fr_desc->ind &= ~ind;
2588
2589         sge->lkey = mr->lkey;
2590         sge->addr = mr->iova;
2591         sge->length = mr->length;
2592
2593         isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2594                   sge->addr, sge->length, sge->lkey);
2595
2596         return ret;
2597 }
2598
2599 static inline void
2600 isert_set_dif_domain(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs,
2601                      struct ib_sig_domain *domain)
2602 {
2603         domain->sig_type = IB_SIG_TYPE_T10_DIF;
2604         domain->sig.dif.bg_type = IB_T10DIF_CRC;
2605         domain->sig.dif.pi_interval = se_cmd->se_dev->dev_attrib.block_size;
2606         domain->sig.dif.ref_tag = se_cmd->reftag_seed;
2607         /*
2608          * At the moment we hard code those, but if in the future
2609          * the target core would like to use it, we will take it
2610          * from se_cmd.
2611          */
2612         domain->sig.dif.apptag_check_mask = 0xffff;
2613         domain->sig.dif.app_escape = true;
2614         domain->sig.dif.ref_escape = true;
2615         if (se_cmd->prot_type == TARGET_DIF_TYPE1_PROT ||
2616             se_cmd->prot_type == TARGET_DIF_TYPE2_PROT)
2617                 domain->sig.dif.ref_remap = true;
2618 };
2619
2620 static int
2621 isert_set_sig_attrs(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs)
2622 {
2623         switch (se_cmd->prot_op) {
2624         case TARGET_PROT_DIN_INSERT:
2625         case TARGET_PROT_DOUT_STRIP:
2626                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
2627                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2628                 break;
2629         case TARGET_PROT_DOUT_INSERT:
2630         case TARGET_PROT_DIN_STRIP:
2631                 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
2632                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2633                 break;
2634         case TARGET_PROT_DIN_PASS:
2635         case TARGET_PROT_DOUT_PASS:
2636                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2637                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2638                 break;
2639         default:
2640                 isert_err("Unsupported PI operation %d\n", se_cmd->prot_op);
2641                 return -EINVAL;
2642         }
2643
2644         return 0;
2645 }
2646
2647 static inline u8
2648 isert_set_prot_checks(u8 prot_checks)
2649 {
2650         return (prot_checks & TARGET_DIF_CHECK_GUARD  ? 0xc0 : 0) |
2651                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x30 : 0) |
2652                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x0f : 0);
2653 }
2654
2655 static int
2656 isert_reg_sig_mr(struct isert_conn *isert_conn,
2657                  struct se_cmd *se_cmd,
2658                  struct isert_rdma_wr *rdma_wr,
2659                  struct fast_reg_descriptor *fr_desc)
2660 {
2661         struct ib_sig_handover_wr sig_wr;
2662         struct ib_send_wr inv_wr, *bad_wr, *wr = NULL;
2663         struct pi_context *pi_ctx = fr_desc->pi_ctx;
2664         struct ib_sig_attrs sig_attrs;
2665         int ret;
2666
2667         memset(&sig_attrs, 0, sizeof(sig_attrs));
2668         ret = isert_set_sig_attrs(se_cmd, &sig_attrs);
2669         if (ret)
2670                 goto err;
2671
2672         sig_attrs.check_mask = isert_set_prot_checks(se_cmd->prot_checks);
2673
2674         if (!(fr_desc->ind & ISERT_SIG_KEY_VALID)) {
2675                 isert_inv_rkey(&inv_wr, pi_ctx->sig_mr);
2676                 wr = &inv_wr;
2677         }
2678
2679         memset(&sig_wr, 0, sizeof(sig_wr));
2680         sig_wr.wr.opcode = IB_WR_REG_SIG_MR;
2681         sig_wr.wr.wr_id = ISER_FASTREG_LI_WRID;
2682         sig_wr.wr.sg_list = &rdma_wr->ib_sg[DATA];
2683         sig_wr.wr.num_sge = 1;
2684         sig_wr.access_flags = IB_ACCESS_LOCAL_WRITE;
2685         sig_wr.sig_attrs = &sig_attrs;
2686         sig_wr.sig_mr = pi_ctx->sig_mr;
2687         if (se_cmd->t_prot_sg)
2688                 sig_wr.prot = &rdma_wr->ib_sg[PROT];
2689
2690         if (!wr)
2691                 wr = &sig_wr.wr;
2692         else
2693                 wr->next = &sig_wr.wr;
2694
2695         ret = ib_post_send(isert_conn->qp, wr, &bad_wr);
2696         if (ret) {
2697                 isert_err("fast registration failed, ret:%d\n", ret);
2698                 goto err;
2699         }
2700         fr_desc->ind &= ~ISERT_SIG_KEY_VALID;
2701
2702         rdma_wr->ib_sg[SIG].lkey = pi_ctx->sig_mr->lkey;
2703         rdma_wr->ib_sg[SIG].addr = 0;
2704         rdma_wr->ib_sg[SIG].length = se_cmd->data_length;
2705         if (se_cmd->prot_op != TARGET_PROT_DIN_STRIP &&
2706             se_cmd->prot_op != TARGET_PROT_DOUT_INSERT)
2707                 /*
2708                  * We have protection guards on the wire
2709                  * so we need to set a larget transfer
2710                  */
2711                 rdma_wr->ib_sg[SIG].length += se_cmd->prot_length;
2712
2713         isert_dbg("sig_sge: addr: 0x%llx  length: %u lkey: %x\n",
2714                   rdma_wr->ib_sg[SIG].addr, rdma_wr->ib_sg[SIG].length,
2715                   rdma_wr->ib_sg[SIG].lkey);
2716 err:
2717         return ret;
2718 }
2719
2720 static int
2721 isert_handle_prot_cmd(struct isert_conn *isert_conn,
2722                       struct isert_cmd *isert_cmd,
2723                       struct isert_rdma_wr *wr)
2724 {
2725         struct isert_device *device = isert_conn->device;
2726         struct se_cmd *se_cmd = &isert_cmd->iscsi_cmd->se_cmd;
2727         int ret;
2728
2729         if (!wr->fr_desc->pi_ctx) {
2730                 ret = isert_create_pi_ctx(wr->fr_desc,
2731                                           device->ib_device,
2732                                           device->pd);
2733                 if (ret) {
2734                         isert_err("conn %p failed to allocate pi_ctx\n",
2735                                   isert_conn);
2736                         return ret;
2737                 }
2738         }
2739
2740         if (se_cmd->t_prot_sg) {
2741                 ret = isert_map_data_buf(isert_conn, isert_cmd,
2742                                          se_cmd->t_prot_sg,
2743                                          se_cmd->t_prot_nents,
2744                                          se_cmd->prot_length,
2745                                          0, wr->iser_ib_op, &wr->prot);
2746                 if (ret) {
2747                         isert_err("conn %p failed to map protection buffer\n",
2748                                   isert_conn);
2749                         return ret;
2750                 }
2751
2752                 memset(&wr->ib_sg[PROT], 0, sizeof(wr->ib_sg[PROT]));
2753                 ret = isert_fast_reg_mr(isert_conn, wr->fr_desc, &wr->prot,
2754                                         ISERT_PROT_KEY_VALID, &wr->ib_sg[PROT]);
2755                 if (ret) {
2756                         isert_err("conn %p failed to fast reg mr\n",
2757                                   isert_conn);
2758                         goto unmap_prot_cmd;
2759                 }
2760         }
2761
2762         ret = isert_reg_sig_mr(isert_conn, se_cmd, wr, wr->fr_desc);
2763         if (ret) {
2764                 isert_err("conn %p failed to fast reg mr\n",
2765                           isert_conn);
2766                 goto unmap_prot_cmd;
2767         }
2768         wr->fr_desc->ind |= ISERT_PROTECTED;
2769
2770         return 0;
2771
2772 unmap_prot_cmd:
2773         if (se_cmd->t_prot_sg)
2774                 isert_unmap_data_buf(isert_conn, &wr->prot);
2775
2776         return ret;
2777 }
2778
2779 static int
2780 isert_reg_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2781                struct isert_rdma_wr *wr)
2782 {
2783         struct se_cmd *se_cmd = &cmd->se_cmd;
2784         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2785         struct isert_conn *isert_conn = conn->context;
2786         struct fast_reg_descriptor *fr_desc = NULL;
2787         struct ib_rdma_wr *rdma_wr;
2788         struct ib_sge *ib_sg;
2789         u32 offset;
2790         int ret = 0;
2791         unsigned long flags;
2792
2793         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2794
2795         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2796         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2797                                  se_cmd->t_data_nents, se_cmd->data_length,
2798                                  offset, wr->iser_ib_op, &wr->data);
2799         if (ret)
2800                 return ret;
2801
2802         if (wr->data.dma_nents != 1 || isert_prot_cmd(isert_conn, se_cmd)) {
2803                 spin_lock_irqsave(&isert_conn->pool_lock, flags);
2804                 fr_desc = list_first_entry(&isert_conn->fr_pool,
2805                                            struct fast_reg_descriptor, list);
2806                 list_del(&fr_desc->list);
2807                 spin_unlock_irqrestore(&isert_conn->pool_lock, flags);
2808                 wr->fr_desc = fr_desc;
2809         }
2810
2811         ret = isert_fast_reg_mr(isert_conn, fr_desc, &wr->data,
2812                                 ISERT_DATA_KEY_VALID, &wr->ib_sg[DATA]);
2813         if (ret)
2814                 goto unmap_cmd;
2815
2816         if (isert_prot_cmd(isert_conn, se_cmd)) {
2817                 ret = isert_handle_prot_cmd(isert_conn, isert_cmd, wr);
2818                 if (ret)
2819                         goto unmap_cmd;
2820
2821                 ib_sg = &wr->ib_sg[SIG];
2822         } else {
2823                 ib_sg = &wr->ib_sg[DATA];
2824         }
2825
2826         memcpy(&wr->s_ib_sge, ib_sg, sizeof(*ib_sg));
2827         wr->ib_sge = &wr->s_ib_sge;
2828         wr->rdma_wr_num = 1;
2829         memset(&wr->s_rdma_wr, 0, sizeof(wr->s_rdma_wr));
2830         wr->rdma_wr = &wr->s_rdma_wr;
2831         wr->isert_cmd = isert_cmd;
2832
2833         rdma_wr = &isert_cmd->rdma_wr.s_rdma_wr;
2834         rdma_wr->wr.sg_list = &wr->s_ib_sge;
2835         rdma_wr->wr.num_sge = 1;
2836         rdma_wr->wr.wr_id = (uintptr_t)&isert_cmd->tx_desc;
2837         if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2838                 rdma_wr->wr.opcode = IB_WR_RDMA_WRITE;
2839                 rdma_wr->remote_addr = isert_cmd->read_va;
2840                 rdma_wr->rkey = isert_cmd->read_stag;
2841                 rdma_wr->wr.send_flags = !isert_prot_cmd(isert_conn, se_cmd) ?
2842                                       0 : IB_SEND_SIGNALED;
2843         } else {
2844                 rdma_wr->wr.opcode = IB_WR_RDMA_READ;
2845                 rdma_wr->remote_addr = isert_cmd->write_va;
2846                 rdma_wr->rkey = isert_cmd->write_stag;
2847                 rdma_wr->wr.send_flags = IB_SEND_SIGNALED;
2848         }
2849
2850         return 0;
2851
2852 unmap_cmd:
2853         if (fr_desc) {
2854                 spin_lock_irqsave(&isert_conn->pool_lock, flags);
2855                 list_add_tail(&fr_desc->list, &isert_conn->fr_pool);
2856                 spin_unlock_irqrestore(&isert_conn->pool_lock, flags);
2857         }
2858         isert_unmap_data_buf(isert_conn, &wr->data);
2859
2860         return ret;
2861 }
2862
2863 static int
2864 isert_put_datain(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2865 {
2866         struct se_cmd *se_cmd = &cmd->se_cmd;
2867         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2868         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2869         struct isert_conn *isert_conn = conn->context;
2870         struct isert_device *device = isert_conn->device;
2871         struct ib_send_wr *wr_failed;
2872         int rc;
2873
2874         isert_dbg("Cmd: %p RDMA_WRITE data_length: %u\n",
2875                  isert_cmd, se_cmd->data_length);
2876
2877         wr->iser_ib_op = ISER_IB_RDMA_WRITE;
2878         rc = device->reg_rdma_mem(conn, cmd, wr);
2879         if (rc) {
2880                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2881                 return rc;
2882         }
2883
2884         if (!isert_prot_cmd(isert_conn, se_cmd)) {
2885                 /*
2886                  * Build isert_conn->tx_desc for iSCSI response PDU and attach
2887                  */
2888                 isert_create_send_desc(isert_conn, isert_cmd,
2889                                        &isert_cmd->tx_desc);
2890                 iscsit_build_rsp_pdu(cmd, conn, true, (struct iscsi_scsi_rsp *)
2891                                      &isert_cmd->tx_desc.iscsi_header);
2892                 isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2893                 isert_init_send_wr(isert_conn, isert_cmd,
2894                                    &isert_cmd->tx_desc.send_wr);
2895                 isert_cmd->rdma_wr.s_rdma_wr.wr.next = &isert_cmd->tx_desc.send_wr;
2896                 wr->rdma_wr_num += 1;
2897
2898                 rc = isert_post_recv(isert_conn, isert_cmd->rx_desc);
2899                 if (rc) {
2900                         isert_err("ib_post_recv failed with %d\n", rc);
2901                         return rc;
2902                 }
2903         }
2904
2905         rc = ib_post_send(isert_conn->qp, &wr->rdma_wr->wr, &wr_failed);
2906         if (rc)
2907                 isert_warn("ib_post_send() failed for IB_WR_RDMA_WRITE\n");
2908
2909         if (!isert_prot_cmd(isert_conn, se_cmd))
2910                 isert_dbg("Cmd: %p posted RDMA_WRITE + Response for iSER Data "
2911                          "READ\n", isert_cmd);
2912         else
2913                 isert_dbg("Cmd: %p posted RDMA_WRITE for iSER Data READ\n",
2914                          isert_cmd);
2915
2916         return 1;
2917 }
2918
2919 static int
2920 isert_get_dataout(struct iscsi_conn *conn, struct iscsi_cmd *cmd, bool recovery)
2921 {
2922         struct se_cmd *se_cmd = &cmd->se_cmd;
2923         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2924         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2925         struct isert_conn *isert_conn = conn->context;
2926         struct isert_device *device = isert_conn->device;
2927         struct ib_send_wr *wr_failed;
2928         int rc;
2929
2930         isert_dbg("Cmd: %p RDMA_READ data_length: %u write_data_done: %u\n",
2931                  isert_cmd, se_cmd->data_length, cmd->write_data_done);
2932         wr->iser_ib_op = ISER_IB_RDMA_READ;
2933         rc = device->reg_rdma_mem(conn, cmd, wr);
2934         if (rc) {
2935                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2936                 return rc;
2937         }
2938
2939         rc = ib_post_send(isert_conn->qp, &wr->rdma_wr->wr, &wr_failed);
2940         if (rc)
2941                 isert_warn("ib_post_send() failed for IB_WR_RDMA_READ\n");
2942
2943         isert_dbg("Cmd: %p posted RDMA_READ memory for ISER Data WRITE\n",
2944                  isert_cmd);
2945
2946         return 0;
2947 }
2948
2949 static int
2950 isert_immediate_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2951 {
2952         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2953         int ret = 0;
2954
2955         switch (state) {
2956         case ISTATE_REMOVE:
2957                 spin_lock_bh(&conn->cmd_lock);
2958                 list_del_init(&cmd->i_conn_node);
2959                 spin_unlock_bh(&conn->cmd_lock);
2960                 isert_put_cmd(isert_cmd, true);
2961                 break;
2962         case ISTATE_SEND_NOPIN_WANT_RESPONSE:
2963                 ret = isert_put_nopin(cmd, conn, false);
2964                 break;
2965         default:
2966                 isert_err("Unknown immediate state: 0x%02x\n", state);
2967                 ret = -EINVAL;
2968                 break;
2969         }
2970
2971         return ret;
2972 }
2973
2974 static int
2975 isert_response_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2976 {
2977         struct isert_conn *isert_conn = conn->context;
2978         int ret;
2979
2980         switch (state) {
2981         case ISTATE_SEND_LOGOUTRSP:
2982                 ret = isert_put_logout_rsp(cmd, conn);
2983                 if (!ret)
2984                         isert_conn->logout_posted = true;
2985                 break;
2986         case ISTATE_SEND_NOPIN:
2987                 ret = isert_put_nopin(cmd, conn, true);
2988                 break;
2989         case ISTATE_SEND_TASKMGTRSP:
2990                 ret = isert_put_tm_rsp(cmd, conn);
2991                 break;
2992         case ISTATE_SEND_REJECT:
2993                 ret = isert_put_reject(cmd, conn);
2994                 break;
2995         case ISTATE_SEND_TEXTRSP:
2996                 ret = isert_put_text_rsp(cmd, conn);
2997                 break;
2998         case ISTATE_SEND_STATUS:
2999                 /*
3000                  * Special case for sending non GOOD SCSI status from TX thread
3001                  * context during pre se_cmd excecution failure.
3002                  */
3003                 ret = isert_put_response(conn, cmd);
3004                 break;
3005         default:
3006                 isert_err("Unknown response state: 0x%02x\n", state);
3007                 ret = -EINVAL;
3008                 break;
3009         }
3010
3011         return ret;
3012 }
3013
3014 struct rdma_cm_id *
3015 isert_setup_id(struct isert_np *isert_np)
3016 {
3017         struct iscsi_np *np = isert_np->np;
3018         struct rdma_cm_id *id;
3019         struct sockaddr *sa;
3020         int ret;
3021
3022         sa = (struct sockaddr *)&np->np_sockaddr;
3023         isert_dbg("ksockaddr: %p, sa: %p\n", &np->np_sockaddr, sa);
3024
3025         id = rdma_create_id(&init_net, isert_cma_handler, isert_np,
3026                             RDMA_PS_TCP, IB_QPT_RC);
3027         if (IS_ERR(id)) {
3028                 isert_err("rdma_create_id() failed: %ld\n", PTR_ERR(id));
3029                 ret = PTR_ERR(id);
3030                 goto out;
3031         }
3032         isert_dbg("id %p context %p\n", id, id->context);
3033
3034         ret = rdma_bind_addr(id, sa);
3035         if (ret) {
3036                 isert_err("rdma_bind_addr() failed: %d\n", ret);
3037                 goto out_id;
3038         }
3039
3040         ret = rdma_listen(id, 0);
3041         if (ret) {
3042                 isert_err("rdma_listen() failed: %d\n", ret);
3043                 goto out_id;
3044         }
3045
3046         return id;
3047 out_id:
3048         rdma_destroy_id(id);
3049 out:
3050         return ERR_PTR(ret);
3051 }
3052
3053 static int
3054 isert_setup_np(struct iscsi_np *np,
3055                struct sockaddr_storage *ksockaddr)
3056 {
3057         struct isert_np *isert_np;
3058         struct rdma_cm_id *isert_lid;
3059         int ret;
3060
3061         isert_np = kzalloc(sizeof(struct isert_np), GFP_KERNEL);
3062         if (!isert_np) {
3063                 isert_err("Unable to allocate struct isert_np\n");
3064                 return -ENOMEM;
3065         }
3066         sema_init(&isert_np->sem, 0);
3067         mutex_init(&isert_np->mutex);
3068         INIT_LIST_HEAD(&isert_np->accepted);
3069         INIT_LIST_HEAD(&isert_np->pending);
3070         isert_np->np = np;
3071
3072         /*
3073          * Setup the np->np_sockaddr from the passed sockaddr setup
3074          * in iscsi_target_configfs.c code..
3075          */
3076         memcpy(&np->np_sockaddr, ksockaddr,
3077                sizeof(struct sockaddr_storage));
3078
3079         isert_lid = isert_setup_id(isert_np);
3080         if (IS_ERR(isert_lid)) {
3081                 ret = PTR_ERR(isert_lid);
3082                 goto out;
3083         }
3084
3085         isert_np->cm_id = isert_lid;
3086         np->np_context = isert_np;
3087
3088         return 0;
3089
3090 out:
3091         kfree(isert_np);
3092
3093         return ret;
3094 }
3095
3096 static int
3097 isert_rdma_accept(struct isert_conn *isert_conn)
3098 {
3099         struct rdma_cm_id *cm_id = isert_conn->cm_id;
3100         struct rdma_conn_param cp;
3101         int ret;
3102
3103         memset(&cp, 0, sizeof(struct rdma_conn_param));
3104         cp.initiator_depth = isert_conn->initiator_depth;
3105         cp.retry_count = 7;
3106         cp.rnr_retry_count = 7;
3107
3108         ret = rdma_accept(cm_id, &cp);
3109         if (ret) {
3110                 isert_err("rdma_accept() failed with: %d\n", ret);
3111                 return ret;
3112         }
3113
3114         return 0;
3115 }
3116
3117 static int
3118 isert_get_login_rx(struct iscsi_conn *conn, struct iscsi_login *login)
3119 {
3120         struct isert_conn *isert_conn = conn->context;
3121         int ret;
3122
3123         isert_info("before login_req comp conn: %p\n", isert_conn);
3124         ret = wait_for_completion_interruptible(&isert_conn->login_req_comp);
3125         if (ret) {
3126                 isert_err("isert_conn %p interrupted before got login req\n",
3127                           isert_conn);
3128                 return ret;
3129         }
3130         reinit_completion(&isert_conn->login_req_comp);
3131
3132         /*
3133          * For login requests after the first PDU, isert_rx_login_req() will
3134          * kick schedule_delayed_work(&conn->login_work) as the packet is
3135          * received, which turns this callback from iscsi_target_do_login_rx()
3136          * into a NOP.
3137          */
3138         if (!login->first_request)
3139                 return 0;
3140
3141         isert_rx_login_req(isert_conn);
3142
3143         isert_info("before login_comp conn: %p\n", conn);
3144         ret = wait_for_completion_interruptible(&isert_conn->login_comp);
3145         if (ret)
3146                 return ret;
3147
3148         isert_info("processing login->req: %p\n", login->req);
3149
3150         return 0;
3151 }
3152
3153 static void
3154 isert_set_conn_info(struct iscsi_np *np, struct iscsi_conn *conn,
3155                     struct isert_conn *isert_conn)
3156 {
3157         struct rdma_cm_id *cm_id = isert_conn->cm_id;
3158         struct rdma_route *cm_route = &cm_id->route;
3159
3160         conn->login_family = np->np_sockaddr.ss_family;
3161
3162         conn->login_sockaddr = cm_route->addr.dst_addr;
3163         conn->local_sockaddr = cm_route->addr.src_addr;
3164 }
3165
3166 static int
3167 isert_accept_np(struct iscsi_np *np, struct iscsi_conn *conn)
3168 {
3169         struct isert_np *isert_np = np->np_context;
3170         struct isert_conn *isert_conn;
3171         int ret;
3172
3173 accept_wait:
3174         ret = down_interruptible(&isert_np->sem);
3175         if (ret)
3176                 return -ENODEV;
3177
3178         spin_lock_bh(&np->np_thread_lock);
3179         if (np->np_thread_state >= ISCSI_NP_THREAD_RESET) {
3180                 spin_unlock_bh(&np->np_thread_lock);
3181                 isert_dbg("np_thread_state %d\n",
3182                          np->np_thread_state);
3183                 /**
3184                  * No point in stalling here when np_thread
3185                  * is in state RESET/SHUTDOWN/EXIT - bail
3186                  **/
3187                 return -ENODEV;
3188         }
3189         spin_unlock_bh(&np->np_thread_lock);
3190
3191         mutex_lock(&isert_np->mutex);
3192         if (list_empty(&isert_np->pending)) {
3193                 mutex_unlock(&isert_np->mutex);
3194                 goto accept_wait;
3195         }
3196         isert_conn = list_first_entry(&isert_np->pending,
3197                         struct isert_conn, node);
3198         list_del_init(&isert_conn->node);
3199         mutex_unlock(&isert_np->mutex);
3200
3201         conn->context = isert_conn;
3202         isert_conn->conn = conn;
3203
3204         isert_set_conn_info(np, conn, isert_conn);
3205
3206         isert_dbg("Processing isert_conn: %p\n", isert_conn);
3207
3208         return 0;
3209 }
3210
3211 static void
3212 isert_free_np(struct iscsi_np *np)
3213 {
3214         struct isert_np *isert_np = np->np_context;
3215         struct isert_conn *isert_conn, *n;
3216
3217         if (isert_np->cm_id)
3218                 rdma_destroy_id(isert_np->cm_id);
3219
3220         /*
3221          * FIXME: At this point we don't have a good way to insure
3222          * that at this point we don't have hanging connections that
3223          * completed RDMA establishment but didn't start iscsi login
3224          * process. So work-around this by cleaning up what ever piled
3225          * up in accepted and pending lists.
3226          */
3227         mutex_lock(&isert_np->mutex);
3228         if (!list_empty(&isert_np->pending)) {
3229                 isert_info("Still have isert pending connections\n");
3230                 list_for_each_entry_safe(isert_conn, n,
3231                                          &isert_np->pending,
3232                                          node) {
3233                         isert_info("cleaning isert_conn %p state (%d)\n",
3234                                    isert_conn, isert_conn->state);
3235                         isert_connect_release(isert_conn);
3236                 }
3237         }
3238
3239         if (!list_empty(&isert_np->accepted)) {
3240                 isert_info("Still have isert accepted connections\n");
3241                 list_for_each_entry_safe(isert_conn, n,
3242                                          &isert_np->accepted,
3243                                          node) {
3244                         isert_info("cleaning isert_conn %p state (%d)\n",
3245                                    isert_conn, isert_conn->state);
3246                         isert_connect_release(isert_conn);
3247                 }
3248         }
3249         mutex_unlock(&isert_np->mutex);
3250
3251         np->np_context = NULL;
3252         kfree(isert_np);
3253 }
3254
3255 static void isert_release_work(struct work_struct *work)
3256 {
3257         struct isert_conn *isert_conn = container_of(work,
3258                                                      struct isert_conn,
3259                                                      release_work);
3260
3261         isert_info("Starting release conn %p\n", isert_conn);
3262
3263         wait_for_completion(&isert_conn->wait);
3264
3265         mutex_lock(&isert_conn->mutex);
3266         isert_conn->state = ISER_CONN_DOWN;
3267         mutex_unlock(&isert_conn->mutex);
3268
3269         isert_info("Destroying conn %p\n", isert_conn);
3270         isert_put_conn(isert_conn);
3271 }
3272
3273 static void
3274 isert_wait4logout(struct isert_conn *isert_conn)
3275 {
3276         struct iscsi_conn *conn = isert_conn->conn;
3277
3278         isert_info("conn %p\n", isert_conn);
3279
3280         if (isert_conn->logout_posted) {
3281                 isert_info("conn %p wait for conn_logout_comp\n", isert_conn);
3282                 wait_for_completion_timeout(&conn->conn_logout_comp,
3283                                             SECONDS_FOR_LOGOUT_COMP * HZ);
3284         }
3285 }
3286
3287 static void
3288 isert_wait4cmds(struct iscsi_conn *conn)
3289 {
3290         isert_info("iscsi_conn %p\n", conn);
3291
3292         if (conn->sess) {
3293                 target_sess_cmd_list_set_waiting(conn->sess->se_sess);
3294                 target_wait_for_sess_cmds(conn->sess->se_sess);
3295         }
3296 }
3297
3298 static void
3299 isert_wait4flush(struct isert_conn *isert_conn)
3300 {
3301         struct ib_recv_wr *bad_wr;
3302
3303         isert_info("conn %p\n", isert_conn);
3304
3305         init_completion(&isert_conn->wait_comp_err);
3306         isert_conn->beacon.wr_id = ISER_BEACON_WRID;
3307         /* post an indication that all flush errors were consumed */
3308         if (ib_post_recv(isert_conn->qp, &isert_conn->beacon, &bad_wr)) {
3309                 isert_err("conn %p failed to post beacon", isert_conn);
3310                 return;
3311         }
3312
3313         wait_for_completion(&isert_conn->wait_comp_err);
3314 }
3315
3316 /**
3317  * isert_put_unsol_pending_cmds() - Drop commands waiting for
3318  *     unsolicitate dataout
3319  * @conn:    iscsi connection
3320  *
3321  * We might still have commands that are waiting for unsolicited
3322  * dataouts messages. We must put the extra reference on those
3323  * before blocking on the target_wait_for_session_cmds
3324  */
3325 static void
3326 isert_put_unsol_pending_cmds(struct iscsi_conn *conn)
3327 {
3328         struct iscsi_cmd *cmd, *tmp;
3329         static LIST_HEAD(drop_cmd_list);
3330
3331         spin_lock_bh(&conn->cmd_lock);
3332         list_for_each_entry_safe(cmd, tmp, &conn->conn_cmd_list, i_conn_node) {
3333                 if ((cmd->cmd_flags & ICF_NON_IMMEDIATE_UNSOLICITED_DATA) &&
3334                     (cmd->write_data_done < conn->sess->sess_ops->FirstBurstLength) &&
3335                     (cmd->write_data_done < cmd->se_cmd.data_length))
3336                         list_move_tail(&cmd->i_conn_node, &drop_cmd_list);
3337         }
3338         spin_unlock_bh(&conn->cmd_lock);
3339
3340         list_for_each_entry_safe(cmd, tmp, &drop_cmd_list, i_conn_node) {
3341                 list_del_init(&cmd->i_conn_node);
3342                 if (cmd->i_state != ISTATE_REMOVE) {
3343                         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
3344
3345                         isert_info("conn %p dropping cmd %p\n", conn, cmd);
3346                         isert_put_cmd(isert_cmd, true);
3347                 }
3348         }
3349 }
3350
3351 static void isert_wait_conn(struct iscsi_conn *conn)
3352 {
3353         struct isert_conn *isert_conn = conn->context;
3354
3355         isert_info("Starting conn %p\n", isert_conn);
3356
3357         mutex_lock(&isert_conn->mutex);
3358         /*
3359          * Only wait for wait_comp_err if the isert_conn made it
3360          * into full feature phase..
3361          */
3362         if (isert_conn->state == ISER_CONN_INIT) {
3363                 mutex_unlock(&isert_conn->mutex);
3364                 return;
3365         }
3366         isert_conn_terminate(isert_conn);
3367         mutex_unlock(&isert_conn->mutex);
3368
3369         isert_wait4flush(isert_conn);
3370         isert_put_unsol_pending_cmds(conn);
3371         isert_wait4cmds(conn);
3372         isert_wait4logout(isert_conn);
3373
3374         queue_work(isert_release_wq, &isert_conn->release_work);
3375 }
3376
3377 static void isert_free_conn(struct iscsi_conn *conn)
3378 {
3379         struct isert_conn *isert_conn = conn->context;
3380
3381         isert_wait4flush(isert_conn);
3382         isert_put_conn(isert_conn);
3383 }
3384
3385 static struct iscsit_transport iser_target_transport = {
3386         .name                   = "IB/iSER",
3387         .transport_type         = ISCSI_INFINIBAND,
3388         .priv_size              = sizeof(struct isert_cmd),
3389         .owner                  = THIS_MODULE,
3390         .iscsit_setup_np        = isert_setup_np,
3391         .iscsit_accept_np       = isert_accept_np,
3392         .iscsit_free_np         = isert_free_np,
3393         .iscsit_wait_conn       = isert_wait_conn,
3394         .iscsit_free_conn       = isert_free_conn,
3395         .iscsit_get_login_rx    = isert_get_login_rx,
3396         .iscsit_put_login_tx    = isert_put_login_tx,
3397         .iscsit_immediate_queue = isert_immediate_queue,
3398         .iscsit_response_queue  = isert_response_queue,
3399         .iscsit_get_dataout     = isert_get_dataout,
3400         .iscsit_queue_data_in   = isert_put_datain,
3401         .iscsit_queue_status    = isert_put_response,
3402         .iscsit_aborted_task    = isert_aborted_task,
3403         .iscsit_get_sup_prot_ops = isert_get_sup_prot_ops,
3404 };
3405
3406 static int __init isert_init(void)
3407 {
3408         int ret;
3409
3410         isert_comp_wq = alloc_workqueue("isert_comp_wq",
3411                                         WQ_UNBOUND | WQ_HIGHPRI, 0);
3412         if (!isert_comp_wq) {
3413                 isert_err("Unable to allocate isert_comp_wq\n");
3414                 ret = -ENOMEM;
3415                 return -ENOMEM;
3416         }
3417
3418         isert_release_wq = alloc_workqueue("isert_release_wq", WQ_UNBOUND,
3419                                         WQ_UNBOUND_MAX_ACTIVE);
3420         if (!isert_release_wq) {
3421                 isert_err("Unable to allocate isert_release_wq\n");
3422                 ret = -ENOMEM;
3423                 goto destroy_comp_wq;
3424         }
3425
3426         iscsit_register_transport(&iser_target_transport);
3427         isert_info("iSER_TARGET[0] - Loaded iser_target_transport\n");
3428
3429         return 0;
3430
3431 destroy_comp_wq:
3432         destroy_workqueue(isert_comp_wq);
3433
3434         return ret;
3435 }
3436
3437 static void __exit isert_exit(void)
3438 {
3439         flush_scheduled_work();
3440         destroy_workqueue(isert_release_wq);
3441         destroy_workqueue(isert_comp_wq);
3442         iscsit_unregister_transport(&iser_target_transport);
3443         isert_info("iSER_TARGET[0] - Released iser_target_transport\n");
3444 }
3445
3446 MODULE_DESCRIPTION("iSER-Target for mainline target infrastructure");
3447 MODULE_VERSION("1.0");
3448 MODULE_AUTHOR("nab@Linux-iSCSI.org");
3449 MODULE_LICENSE("GPL");
3450
3451 module_init(isert_init);
3452 module_exit(isert_exit);