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 inline bool
  69 isert_prot_cmd(struct isert_conn *conn, struct se_cmd *cmd)
  70 {
  71         return (conn->pi_support &&
  72                 cmd->prot_op != TARGET_PROT_NORMAL);
  73 }
  74
  75
  76 static void
  77 isert_qp_event_callback(struct ib_event *e, void *context)
  78 {
  79         struct isert_conn *isert_conn = context;
  80
  81         isert_err("conn %p event: %d\n", isert_conn, e->event);
  82         switch (e->event) {
  83         case IB_EVENT_COMM_EST:
  84                 rdma_notify(isert_conn->cm_id, IB_EVENT_COMM_EST);
  85                 break;
  86         case IB_EVENT_QP_LAST_WQE_REACHED:
  87                 isert_warn("Reached TX IB_EVENT_QP_LAST_WQE_REACHED\n");
  88                 break;
  89         default:
  90                 break;
  91         }
  92 }
  93
  94 static int
  95 isert_query_device(struct ib_device *ib_dev, struct ib_device_attr *devattr)
  96 {
  97         int ret;
  98
  99         ret = ib_query_device(ib_dev, devattr);
 100         if (ret) {
 101                 isert_err("ib_query_device() failed: %d\n", ret);
 102                 return ret;
 103         }
 104         isert_dbg("devattr->max_sge: %d\n", devattr->max_sge);
 105         isert_dbg("devattr->max_sge_rd: %d\n", devattr->max_sge_rd);
 106
 107         return 0;
 108 }
 109
 110 static struct isert_comp *
 111 isert_comp_get(struct isert_conn *isert_conn)
 112 {
 113         struct isert_device *device = isert_conn->device;
 114         struct isert_comp *comp;
 115         int i, min = 0;
 116
 117         mutex_lock(&device_list_mutex);
 118         for (i = 0; i < device->comps_used; i++)
 119                 if (device->comps[i].active_qps <
 120                     device->comps[min].active_qps)
 121                         min = i;
 122         comp = &device->comps[min];
 123         comp->active_qps++;
 124         mutex_unlock(&device_list_mutex);
 125
 126         isert_info("conn %p, using comp %p min_index: %d\n",
 127                    isert_conn, comp, min);
 128
 129         return comp;
 130 }
 131
 132 static void
 133 isert_comp_put(struct isert_comp *comp)
 134 {
 135         mutex_lock(&device_list_mutex);
 136         comp->active_qps--;
 137         mutex_unlock(&device_list_mutex);
 138 }
 139
 140 static struct ib_qp *
 141 isert_create_qp(struct isert_conn *isert_conn,
 142                 struct isert_comp *comp,
 143                 struct rdma_cm_id *cma_id)
 144 {
 145         struct isert_device *device = isert_conn->device;
 146         struct ib_qp_init_attr attr;
 147         int ret;
 148
 149         memset(&attr, 0, sizeof(struct ib_qp_init_attr));
 150         attr.event_handler = isert_qp_event_callback;
 151         attr.qp_context = isert_conn;
 152         attr.send_cq = comp->cq;
 153         attr.recv_cq = comp->cq;
 154         attr.cap.max_send_wr = ISERT_QP_MAX_REQ_DTOS;
 155         attr.cap.max_recv_wr = ISERT_QP_MAX_RECV_DTOS + 1;
 156         /*
 157          * FIXME: Use devattr.max_sge - 2 for max_send_sge as
 158          * work-around for RDMA_READs with ConnectX-2.
 159          *
 160          * Also, still make sure to have at least two SGEs for
 161          * outgoing control PDU responses.
 162          */
 163         attr.cap.max_send_sge = max(2, device->dev_attr.max_sge - 2);
 164         isert_conn->max_sge = attr.cap.max_send_sge;
 165
 166         attr.cap.max_recv_sge = 1;
 167         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 168         attr.qp_type = IB_QPT_RC;
 169         if (device->pi_capable)
 170                 attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
 171
 172         ret = rdma_create_qp(cma_id, device->pd, &attr);
 173         if (ret) {
 174                 isert_err("rdma_create_qp failed for cma_id %d\n", ret);
 175                 return ERR_PTR(ret);
 176         }
 177
 178         return cma_id->qp;
 179 }
 180
 181 static int
 182 isert_conn_setup_qp(struct isert_conn *isert_conn, struct rdma_cm_id *cma_id)
 183 {
 184         struct isert_comp *comp;
 185         int ret;
 186
 187         comp = isert_comp_get(isert_conn);
 188         isert_conn->qp = isert_create_qp(isert_conn, comp, cma_id);
 189         if (IS_ERR(isert_conn->qp)) {
 190                 ret = PTR_ERR(isert_conn->qp);
 191                 goto err;
 192         }
 193
 194         return 0;
 195 err:
 196         isert_comp_put(comp);
 197         return ret;
 198 }
 199
 200 static void
 201 isert_cq_event_callback(struct ib_event *e, void *context)
 202 {
 203         isert_dbg("event: %d\n", e->event);
 204 }
 205
 206 static int
 207 isert_alloc_rx_descriptors(struct isert_conn *isert_conn)
 208 {
 209         struct isert_device *device = isert_conn->device;
 210         struct ib_device *ib_dev = device->ib_device;
 211         struct iser_rx_desc *rx_desc;
 212         struct ib_sge *rx_sg;
 213         u64 dma_addr;
 214         int i, j;
 215
 216         isert_conn->rx_descs = kzalloc(ISERT_QP_MAX_RECV_DTOS *
 217                                 sizeof(struct iser_rx_desc), GFP_KERNEL);
 218         if (!isert_conn->rx_descs)
 219                 goto fail;
 220
 221         rx_desc = isert_conn->rx_descs;
 222
 223         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 224                 dma_addr = ib_dma_map_single(ib_dev, (void *)rx_desc,
 225                                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 226                 if (ib_dma_mapping_error(ib_dev, dma_addr))
 227                         goto dma_map_fail;
 228
 229                 rx_desc->dma_addr = dma_addr;
 230
 231                 rx_sg = &rx_desc->rx_sg;
 232                 rx_sg->addr = rx_desc->dma_addr;
 233                 rx_sg->length = ISER_RX_PAYLOAD_SIZE;
 234                 rx_sg->lkey = device->mr->lkey;
 235         }
 236
 237         isert_conn->rx_desc_head = 0;
 238
 239         return 0;
 240
 241 dma_map_fail:
 242         rx_desc = isert_conn->rx_descs;
 243         for (j = 0; j < i; j++, rx_desc++) {
 244                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 245                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 246         }
 247         kfree(isert_conn->rx_descs);
 248         isert_conn->rx_descs = NULL;
 249 fail:
 250         isert_err("conn %p failed to allocate rx descriptors\n", isert_conn);
 251
 252         return -ENOMEM;
 253 }
 254
 255 static void
 256 isert_free_rx_descriptors(struct isert_conn *isert_conn)
 257 {
 258         struct ib_device *ib_dev = isert_conn->device->ib_device;
 259         struct iser_rx_desc *rx_desc;
 260         int i;
 261
 262         if (!isert_conn->rx_descs)
 263                 return;
 264
 265         rx_desc = isert_conn->rx_descs;
 266         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 267                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 268                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 269         }
 270
 271         kfree(isert_conn->rx_descs);
 272         isert_conn->rx_descs = NULL;
 273 }
 274
 275 static void isert_cq_work(struct work_struct *);
 276 static void isert_cq_callback(struct ib_cq *, void *);
 277
 278 static void
 279 isert_free_comps(struct isert_device *device)
 280 {
 281         int i;
 282
 283         for (i = 0; i < device->comps_used; i++) {
 284                 struct isert_comp *comp = &device->comps[i];
 285
 286                 if (comp->cq) {
 287                         cancel_work_sync(&comp->work);
 288                         ib_destroy_cq(comp->cq);
 289                 }
 290         }
 291         kfree(device->comps);
 292 }
 293
 294 static int
 295 isert_alloc_comps(struct isert_device *device,
 296                   struct ib_device_attr *attr)
 297 {
 298         int i, max_cqe, ret = 0;
 299
 300         device->comps_used = min(ISERT_MAX_CQ, min_t(int, num_online_cpus(),
 301                                  device->ib_device->num_comp_vectors));
 302
 303         isert_info("Using %d CQs, %s supports %d vectors support "
 304                    "Fast registration %d pi_capable %d\n",
 305                    device->comps_used, device->ib_device->name,
 306                    device->ib_device->num_comp_vectors, device->use_fastreg,
 307                    device->pi_capable);
 308
 309         device->comps = kcalloc(device->comps_used, sizeof(struct isert_comp),
 310                                 GFP_KERNEL);
 311         if (!device->comps) {
 312                 isert_err("Unable to allocate completion contexts\n");
 313                 return -ENOMEM;
 314         }
 315
 316         max_cqe = min(ISER_MAX_CQ_LEN, attr->max_cqe);
 317
 318         for (i = 0; i < device->comps_used; i++) {
 319                 struct isert_comp *comp = &device->comps[i];
 320
 321                 comp->device = device;
 322                 INIT_WORK(&comp->work, isert_cq_work);
 323                 comp->cq = ib_create_cq(device->ib_device,
 324                                         isert_cq_callback,
 325                                         isert_cq_event_callback,
 326                                         (void *)comp,
 327                                         max_cqe, i);
 328                 if (IS_ERR(comp->cq)) {
 329                         isert_err("Unable to allocate cq\n");
 330                         ret = PTR_ERR(comp->cq);
 331                         comp->cq = NULL;
 332                         goto out_cq;
 333                 }
 334
 335                 ret = ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP);
 336                 if (ret)
 337                         goto out_cq;
 338         }
 339
 340         return 0;
 341 out_cq:
 342         isert_free_comps(device);
 343         return ret;
 344 }
 345
 346 static int
 347 isert_create_device_ib_res(struct isert_device *device)
 348 {
 349         struct ib_device_attr *dev_attr;
 350         int ret;
 351
 352         dev_attr = &device->dev_attr;
 353         ret = isert_query_device(device->ib_device, dev_attr);
 354         if (ret)
 355                 return ret;
 356
 357         /* asign function handlers */
 358         if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS &&
 359             dev_attr->device_cap_flags & IB_DEVICE_SIGNATURE_HANDOVER) {
 360                 device->use_fastreg = 1;
 361                 device->reg_rdma_mem = isert_reg_rdma;
 362                 device->unreg_rdma_mem = isert_unreg_rdma;
 363         } else {
 364                 device->use_fastreg = 0;
 365                 device->reg_rdma_mem = isert_map_rdma;
 366                 device->unreg_rdma_mem = isert_unmap_cmd;
 367         }
 368
 369         ret = isert_alloc_comps(device, dev_attr);
 370         if (ret)
 371                 return ret;
 372
 373         device->pd = ib_alloc_pd(device->ib_device);
 374         if (IS_ERR(device->pd)) {
 375                 ret = PTR_ERR(device->pd);
 376                 isert_err("failed to allocate pd, device %p, ret=%d\n",
 377                           device, ret);
 378                 goto out_cq;
 379         }
 380
 381         device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE);
 382         if (IS_ERR(device->mr)) {
 383                 ret = PTR_ERR(device->mr);
 384                 isert_err("failed to create dma mr, device %p, ret=%d\n",
 385                           device, ret);
 386                 goto out_mr;
 387         }
 388
 389         /* Check signature cap */
 390         device->pi_capable = dev_attr->device_cap_flags &
 391                              IB_DEVICE_SIGNATURE_HANDOVER ? true : false;
 392
 393         return 0;
 394
 395 out_mr:
 396         ib_dealloc_pd(device->pd);
 397 out_cq:
 398         isert_free_comps(device);
 399         return ret;
 400 }
 401
 402 static void
 403 isert_free_device_ib_res(struct isert_device *device)
 404 {
 405         isert_info("device %p\n", device);
 406
 407         ib_dereg_mr(device->mr);
 408         ib_dealloc_pd(device->pd);
 409         isert_free_comps(device);
 410 }
 411
 412 static void
 413 isert_device_put(struct isert_device *device)
 414 {
 415         mutex_lock(&device_list_mutex);
 416         device->refcount--;
 417         isert_info("device %p refcount %d\n", device, device->refcount);
 418         if (!device->refcount) {
 419                 isert_free_device_ib_res(device);
 420                 list_del(&device->dev_node);
 421                 kfree(device);
 422         }
 423         mutex_unlock(&device_list_mutex);
 424 }
 425
 426 static struct isert_device *
 427 isert_device_get(struct rdma_cm_id *cma_id)
 428 {
 429         struct isert_device *device;
 430         int ret;
 431
 432         mutex_lock(&device_list_mutex);
 433         list_for_each_entry(device, &device_list, dev_node) {
 434                 if (device->ib_device->node_guid == cma_id->device->node_guid) {
 435                         device->refcount++;
 436                         isert_info("Found iser device %p refcount %d\n",
 437                                    device, device->refcount);
 438                         mutex_unlock(&device_list_mutex);
 439                         return device;
 440                 }
 441         }
 442
 443         device = kzalloc(sizeof(struct isert_device), GFP_KERNEL);
 444         if (!device) {
 445                 mutex_unlock(&device_list_mutex);
 446                 return ERR_PTR(-ENOMEM);
 447         }
 448
 449         INIT_LIST_HEAD(&device->dev_node);
 450
 451         device->ib_device = cma_id->device;
 452         ret = isert_create_device_ib_res(device);
 453         if (ret) {
 454                 kfree(device);
 455                 mutex_unlock(&device_list_mutex);
 456                 return ERR_PTR(ret);
 457         }
 458
 459         device->refcount++;
 460         list_add_tail(&device->dev_node, &device_list);
 461         isert_info("Created a new iser device %p refcount %d\n",
 462                    device, device->refcount);
 463         mutex_unlock(&device_list_mutex);
 464
 465         return device;
 466 }
 467
 468 static void
 469 isert_conn_free_fastreg_pool(struct isert_conn *isert_conn)
 470 {
 471         struct fast_reg_descriptor *fr_desc, *tmp;
 472         int i = 0;
 473
 474         if (list_empty(&isert_conn->fr_pool))
 475                 return;
 476
 477         isert_info("Freeing conn %p fastreg pool", isert_conn);
 478
 479         list_for_each_entry_safe(fr_desc, tmp,
 480                                  &isert_conn->fr_pool, list) {
 481                 list_del(&fr_desc->list);
 482                 ib_free_fast_reg_page_list(fr_desc->data_frpl);
 483                 ib_dereg_mr(fr_desc->data_mr);
 484                 if (fr_desc->pi_ctx) {
 485                         ib_free_fast_reg_page_list(fr_desc->pi_ctx->prot_frpl);
 486                         ib_dereg_mr(fr_desc->pi_ctx->prot_mr);
 487                         ib_destroy_mr(fr_desc->pi_ctx->sig_mr);
 488                         kfree(fr_desc->pi_ctx);
 489                 }
 490                 kfree(fr_desc);
 491                 ++i;
 492         }
 493
 494         if (i < isert_conn->fr_pool_size)
 495                 isert_warn("Pool still has %d regions registered\n",
 496                         isert_conn->fr_pool_size - i);
 497 }
 498
 499 static int
 500 isert_create_pi_ctx(struct fast_reg_descriptor *desc,
 501                     struct ib_device *device,
 502                     struct ib_pd *pd)
 503 {
 504         struct ib_mr_init_attr mr_init_attr;
 505         struct pi_context *pi_ctx;
 506         int ret;
 507
 508         pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
 509         if (!pi_ctx) {
 510                 isert_err("Failed to allocate pi context\n");
 511                 return -ENOMEM;
 512         }
 513
 514         pi_ctx->prot_frpl = ib_alloc_fast_reg_page_list(device,
 515                                             ISCSI_ISER_SG_TABLESIZE);
 516         if (IS_ERR(pi_ctx->prot_frpl)) {
 517                 isert_err("Failed to allocate prot frpl err=%ld\n",
 518                           PTR_ERR(pi_ctx->prot_frpl));
 519                 ret = PTR_ERR(pi_ctx->prot_frpl);
 520                 goto err_pi_ctx;
 521         }
 522
 523         pi_ctx->prot_mr = ib_alloc_fast_reg_mr(pd, ISCSI_ISER_SG_TABLESIZE);
 524         if (IS_ERR(pi_ctx->prot_mr)) {
 525                 isert_err("Failed to allocate prot frmr err=%ld\n",
 526                           PTR_ERR(pi_ctx->prot_mr));
 527                 ret = PTR_ERR(pi_ctx->prot_mr);
 528                 goto err_prot_frpl;
 529         }
 530         desc->ind |= ISERT_PROT_KEY_VALID;
 531
 532         memset(&mr_init_attr, 0, sizeof(mr_init_attr));
 533         mr_init_attr.max_reg_descriptors = 2;
 534         mr_init_attr.flags |= IB_MR_SIGNATURE_EN;
 535         pi_ctx->sig_mr = ib_create_mr(pd, &mr_init_attr);
 536         if (IS_ERR(pi_ctx->sig_mr)) {
 537                 isert_err("Failed to allocate signature enabled mr err=%ld\n",
 538                           PTR_ERR(pi_ctx->sig_mr));
 539                 ret = PTR_ERR(pi_ctx->sig_mr);
 540                 goto err_prot_mr;
 541         }
 542
 543         desc->pi_ctx = pi_ctx;
 544         desc->ind |= ISERT_SIG_KEY_VALID;
 545         desc->ind &= ~ISERT_PROTECTED;
 546
 547         return 0;
 548
 549 err_prot_mr:
 550         ib_dereg_mr(pi_ctx->prot_mr);
 551 err_prot_frpl:
 552         ib_free_fast_reg_page_list(pi_ctx->prot_frpl);
 553 err_pi_ctx:
 554         kfree(pi_ctx);
 555
 556         return ret;
 557 }
 558
 559 static int
 560 isert_create_fr_desc(struct ib_device *ib_device, struct ib_pd *pd,
 561                      struct fast_reg_descriptor *fr_desc)
 562 {
 563         int ret;
 564
 565         fr_desc->data_frpl = ib_alloc_fast_reg_page_list(ib_device,
 566                                                          ISCSI_ISER_SG_TABLESIZE);
 567         if (IS_ERR(fr_desc->data_frpl)) {
 568                 isert_err("Failed to allocate data frpl err=%ld\n",
 569                           PTR_ERR(fr_desc->data_frpl));
 570                 return PTR_ERR(fr_desc->data_frpl);
 571         }
 572
 573         fr_desc->data_mr = ib_alloc_fast_reg_mr(pd, ISCSI_ISER_SG_TABLESIZE);
 574         if (IS_ERR(fr_desc->data_mr)) {
 575                 isert_err("Failed to allocate data frmr err=%ld\n",
 576                           PTR_ERR(fr_desc->data_mr));
 577                 ret = PTR_ERR(fr_desc->data_mr);
 578                 goto err_data_frpl;
 579         }
 580         fr_desc->ind |= ISERT_DATA_KEY_VALID;
 581
 582         isert_dbg("Created fr_desc %p\n", fr_desc);
 583
 584         return 0;
 585
 586 err_data_frpl:
 587         ib_free_fast_reg_page_list(fr_desc->data_frpl);
 588
 589         return ret;
 590 }
 591
 592 static int
 593 isert_conn_create_fastreg_pool(struct isert_conn *isert_conn)
 594 {
 595         struct fast_reg_descriptor *fr_desc;
 596         struct isert_device *device = isert_conn->device;
 597         struct se_session *se_sess = isert_conn->conn->sess->se_sess;
 598         struct se_node_acl *se_nacl = se_sess->se_node_acl;
 599         int i, ret, tag_num;
 600         /*
 601          * Setup the number of FRMRs based upon the number of tags
 602          * available to session in iscsi_target_locate_portal().
 603          */
 604         tag_num = max_t(u32, ISCSIT_MIN_TAGS, se_nacl->queue_depth);
 605         tag_num = (tag_num * 2) + ISCSIT_EXTRA_TAGS;
 606
 607         isert_conn->fr_pool_size = 0;
 608         for (i = 0; i < tag_num; i++) {
 609                 fr_desc = kzalloc(sizeof(*fr_desc), GFP_KERNEL);
 610                 if (!fr_desc) {
 611                         isert_err("Failed to allocate fast_reg descriptor\n");
 612                         ret = -ENOMEM;
 613                         goto err;
 614                 }
 615
 616                 ret = isert_create_fr_desc(device->ib_device,
 617                                            device->pd, fr_desc);
 618                 if (ret) {
 619                         isert_err("Failed to create fastreg descriptor err=%d\n",
 620                                ret);
 621                         kfree(fr_desc);
 622                         goto err;
 623                 }
 624
 625                 list_add_tail(&fr_desc->list, &isert_conn->fr_pool);
 626                 isert_conn->fr_pool_size++;
 627         }
 628
 629         isert_dbg("Creating conn %p fastreg pool size=%d",
 630                  isert_conn, isert_conn->fr_pool_size);
 631
 632         return 0;
 633
 634 err:
 635         isert_conn_free_fastreg_pool(isert_conn);
 636         return ret;
 637 }
 638
 639 static void
 640 isert_init_conn(struct isert_conn *isert_conn)
 641 {
 642         isert_conn->state = ISER_CONN_INIT;
 643         INIT_LIST_HEAD(&isert_conn->accept_node);
 644         init_completion(&isert_conn->login_comp);
 645         init_completion(&isert_conn->login_req_comp);
 646         init_completion(&isert_conn->wait);
 647         kref_init(&isert_conn->kref);
 648         mutex_init(&isert_conn->mutex);
 649         spin_lock_init(&isert_conn->pool_lock);
 650         INIT_LIST_HEAD(&isert_conn->fr_pool);
 651 }
 652
 653 static void
 654 isert_free_login_buf(struct isert_conn *isert_conn)
 655 {
 656         struct ib_device *ib_dev = isert_conn->device->ib_device;
 657
 658         ib_dma_unmap_single(ib_dev, isert_conn->login_rsp_dma,
 659                             ISER_RX_LOGIN_SIZE, DMA_TO_DEVICE);
 660         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 661                             ISCSI_DEF_MAX_RECV_SEG_LEN,
 662                             DMA_FROM_DEVICE);
 663         kfree(isert_conn->login_buf);
 664 }
 665
 666 static int
 667 isert_alloc_login_buf(struct isert_conn *isert_conn,
 668                       struct ib_device *ib_dev)
 669 {
 670         int ret;
 671
 672         isert_conn->login_buf = kzalloc(ISCSI_DEF_MAX_RECV_SEG_LEN +
 673                                         ISER_RX_LOGIN_SIZE, GFP_KERNEL);
 674         if (!isert_conn->login_buf) {
 675                 isert_err("Unable to allocate isert_conn->login_buf\n");
 676                 return -ENOMEM;
 677         }
 678
 679         isert_conn->login_req_buf = isert_conn->login_buf;
 680         isert_conn->login_rsp_buf = isert_conn->login_buf +
 681                                     ISCSI_DEF_MAX_RECV_SEG_LEN;
 682
 683         isert_dbg("Set login_buf: %p login_req_buf: %p login_rsp_buf: %p\n",
 684                  isert_conn->login_buf, isert_conn->login_req_buf,
 685                  isert_conn->login_rsp_buf);
 686
 687         isert_conn->login_req_dma = ib_dma_map_single(ib_dev,
 688                                 (void *)isert_conn->login_req_buf,
 689                                 ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_FROM_DEVICE);
 690
 691         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_req_dma);
 692         if (ret) {
 693                 isert_err("login_req_dma mapping error: %d\n", ret);
 694                 isert_conn->login_req_dma = 0;
 695                 goto out_login_buf;
 696         }
 697
 698         isert_conn->login_rsp_dma = ib_dma_map_single(ib_dev,
 699                                         (void *)isert_conn->login_rsp_buf,
 700                                         ISER_RX_LOGIN_SIZE, DMA_TO_DEVICE);
 701
 702         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_rsp_dma);
 703         if (ret) {
 704                 isert_err("login_rsp_dma mapping error: %d\n", ret);
 705                 isert_conn->login_rsp_dma = 0;
 706                 goto out_req_dma_map;
 707         }
 708
 709         return 0;
 710
 711 out_req_dma_map:
 712         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 713                             ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_FROM_DEVICE);
 714 out_login_buf:
 715         kfree(isert_conn->login_buf);
 716         return ret;
 717 }
 718
 719 static int
 720 isert_connect_request(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 721 {
 722         struct isert_np *isert_np = cma_id->context;
 723         struct iscsi_np *np = isert_np->np;
 724         struct isert_conn *isert_conn;
 725         struct isert_device *device;
 726         int ret = 0;
 727
 728         spin_lock_bh(&np->np_thread_lock);
 729         if (!np->enabled) {
 730                 spin_unlock_bh(&np->np_thread_lock);
 731                 isert_dbg("iscsi_np is not enabled, reject connect request\n");
 732                 return rdma_reject(cma_id, NULL, 0);
 733         }
 734         spin_unlock_bh(&np->np_thread_lock);
 735
 736         isert_dbg("cma_id: %p, portal: %p\n",
 737                  cma_id, cma_id->context);
 738
 739         isert_conn = kzalloc(sizeof(struct isert_conn), GFP_KERNEL);
 740         if (!isert_conn)
 741                 return -ENOMEM;
 742
 743         isert_init_conn(isert_conn);
 744         isert_conn->cm_id = cma_id;
 745
 746         ret = isert_alloc_login_buf(isert_conn, cma_id->device);
 747         if (ret)
 748                 goto out;
 749
 750         device = isert_device_get(cma_id);
 751         if (IS_ERR(device)) {
 752                 ret = PTR_ERR(device);
 753                 goto out_rsp_dma_map;
 754         }
 755         isert_conn->device = device;
 756
 757         /* Set max inflight RDMA READ requests */
 758         isert_conn->initiator_depth = min_t(u8,
 759                                 event->param.conn.initiator_depth,
 760                                 device->dev_attr.max_qp_init_rd_atom);
 761         isert_dbg("Using initiator_depth: %u\n", isert_conn->initiator_depth);
 762
 763         ret = isert_conn_setup_qp(isert_conn, cma_id);
 764         if (ret)
 765                 goto out_conn_dev;
 766
 767         ret = isert_rdma_post_recvl(isert_conn);
 768         if (ret)
 769                 goto out_conn_dev;
 770
 771         ret = isert_rdma_accept(isert_conn);
 772         if (ret)
 773                 goto out_conn_dev;
 774
 775         mutex_lock(&isert_np->np_accept_mutex);
 776         list_add_tail(&isert_conn->accept_node, &isert_np->np_accept_list);
 777         mutex_unlock(&isert_np->np_accept_mutex);
 778
 779         isert_info("np %p: Allow accept_np to continue\n", np);
 780         up(&isert_np->np_sem);
 781         return 0;
 782
 783 out_conn_dev:
 784         isert_device_put(device);
 785 out_rsp_dma_map:
 786         isert_free_login_buf(isert_conn);
 787 out:
 788         kfree(isert_conn);
 789         rdma_reject(cma_id, NULL, 0);
 790         return ret;
 791 }
 792
 793 static void
 794 isert_connect_release(struct isert_conn *isert_conn)
 795 {
 796         struct isert_device *device = isert_conn->device;
 797
 798         isert_dbg("conn %p\n", isert_conn);
 799
 800         BUG_ON(!device);
 801
 802         if (device->use_fastreg)
 803                 isert_conn_free_fastreg_pool(isert_conn);
 804
 805         isert_free_rx_descriptors(isert_conn);
 806         if (isert_conn->cm_id)
 807                 rdma_destroy_id(isert_conn->cm_id);
 808
 809         if (isert_conn->qp) {
 810                 struct isert_comp *comp = isert_conn->qp->recv_cq->cq_context;
 811
 812                 isert_comp_put(comp);
 813                 ib_destroy_qp(isert_conn->qp);
 814         }
 815
 816         if (isert_conn->login_buf)
 817                 isert_free_login_buf(isert_conn);
 818
 819         isert_device_put(device);
 820
 821         kfree(isert_conn);
 822 }
 823
 824 static void
 825 isert_connected_handler(struct rdma_cm_id *cma_id)
 826 {
 827         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 828
 829         isert_info("conn %p\n", isert_conn);
 830
 831         if (!kref_get_unless_zero(&isert_conn->kref)) {
 832                 isert_warn("conn %p connect_release is running\n", isert_conn);
 833                 return;
 834         }
 835
 836         mutex_lock(&isert_conn->mutex);
 837         if (isert_conn->state != ISER_CONN_FULL_FEATURE)
 838                 isert_conn->state = ISER_CONN_UP;
 839         mutex_unlock(&isert_conn->mutex);
 840 }
 841
 842 static void
 843 isert_release_kref(struct kref *kref)
 844 {
 845         struct isert_conn *isert_conn = container_of(kref,
 846                                 struct isert_conn, kref);
 847
 848         isert_info("conn %p final kref %s/%d\n", isert_conn, current->comm,
 849                    current->pid);
 850
 851         isert_connect_release(isert_conn);
 852 }
 853
 854 static void
 855 isert_put_conn(struct isert_conn *isert_conn)
 856 {
 857         kref_put(&isert_conn->kref, isert_release_kref);
 858 }
 859
 860 /**
 861  * isert_conn_terminate() - Initiate connection termination
 862  * @isert_conn: isert connection struct
 863  *
 864  * Notes:
 865  * In case the connection state is FULL_FEATURE, move state
 866  * to TEMINATING and start teardown sequence (rdma_disconnect).
 867  * In case the connection state is UP, complete flush as well.
 868  *
 869  * This routine must be called with mutex held. Thus it is
 870  * safe to call multiple times.
 871  */
 872 static void
 873 isert_conn_terminate(struct isert_conn *isert_conn)
 874 {
 875         int err;
 876
 877         switch (isert_conn->state) {
 878         case ISER_CONN_TERMINATING:
 879                 break;
 880         case ISER_CONN_UP:
 881         case ISER_CONN_FULL_FEATURE: /* FALLTHRU */
 882                 isert_info("Terminating conn %p state %d\n",
 883                            isert_conn, isert_conn->state);
 884                 isert_conn->state = ISER_CONN_TERMINATING;
 885                 err = rdma_disconnect(isert_conn->cm_id);
 886                 if (err)
 887                         isert_warn("Failed rdma_disconnect isert_conn %p\n",
 888                                    isert_conn);
 889                 break;
 890         default:
 891                 isert_warn("conn %p teminating in state %d\n",
 892                            isert_conn, isert_conn->state);
 893         }
 894 }
 895
 896 static int
 897 isert_np_cma_handler(struct isert_np *isert_np,
 898                      enum rdma_cm_event_type event)
 899 {
 900         isert_dbg("isert np %p, handling event %d\n", isert_np, event);
 901
 902         switch (event) {
 903         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 904                 isert_np->np_cm_id = NULL;
 905                 break;
 906         case RDMA_CM_EVENT_ADDR_CHANGE:
 907                 isert_np->np_cm_id = isert_setup_id(isert_np);
 908                 if (IS_ERR(isert_np->np_cm_id)) {
 909                         isert_err("isert np %p setup id failed: %ld\n",
 910                                   isert_np, PTR_ERR(isert_np->np_cm_id));
 911                         isert_np->np_cm_id = NULL;
 912                 }
 913                 break;
 914         default:
 915                 isert_err("isert np %p Unexpected event %d\n",
 916                           isert_np, event);
 917         }
 918
 919         return -1;
 920 }
 921
 922 static int
 923 isert_disconnected_handler(struct rdma_cm_id *cma_id,
 924                            enum rdma_cm_event_type event)
 925 {
 926         struct isert_np *isert_np = cma_id->context;
 927         struct isert_conn *isert_conn;
 928
 929         if (isert_np->np_cm_id == cma_id)
 930                 return isert_np_cma_handler(cma_id->context, event);
 931
 932         isert_conn = cma_id->qp->qp_context;
 933
 934         mutex_lock(&isert_conn->mutex);
 935         isert_conn_terminate(isert_conn);
 936         mutex_unlock(&isert_conn->mutex);
 937
 938         isert_info("conn %p completing wait\n", isert_conn);
 939         complete(&isert_conn->wait);
 940
 941         return 0;
 942 }
 943
 944 static int
 945 isert_connect_error(struct rdma_cm_id *cma_id)
 946 {
 947         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 948
 949         isert_conn->cm_id = NULL;
 950         isert_put_conn(isert_conn);
 951
 952         return -1;
 953 }
 954
 955 static int
 956 isert_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 957 {
 958         int ret = 0;
 959
 960         isert_info("event %d status %d id %p np %p\n", event->event,
 961                    event->status, cma_id, cma_id->context);
 962
 963         switch (event->event) {
 964         case RDMA_CM_EVENT_CONNECT_REQUEST:
 965                 ret = isert_connect_request(cma_id, event);
 966                 if (ret)
 967                         isert_err("failed handle connect request %d\n", ret);
 968                 break;
 969         case RDMA_CM_EVENT_ESTABLISHED:
 970                 isert_connected_handler(cma_id);
 971                 break;
 972         case RDMA_CM_EVENT_ADDR_CHANGE:    /* FALLTHRU */
 973         case RDMA_CM_EVENT_DISCONNECTED:   /* FALLTHRU */
 974         case RDMA_CM_EVENT_DEVICE_REMOVAL: /* FALLTHRU */
 975         case RDMA_CM_EVENT_TIMEWAIT_EXIT:  /* FALLTHRU */
 976                 ret = isert_disconnected_handler(cma_id, event->event);
 977                 break;
 978         case RDMA_CM_EVENT_REJECTED:       /* FALLTHRU */
 979         case RDMA_CM_EVENT_UNREACHABLE:    /* FALLTHRU */
 980         case RDMA_CM_EVENT_CONNECT_ERROR:
 981                 ret = isert_connect_error(cma_id);
 982                 break;
 983         default:
 984                 isert_err("Unhandled RDMA CMA event: %d\n", event->event);
 985                 break;
 986         }
 987
 988         return ret;
 989 }
 990
 991 static int
 992 isert_post_recv(struct isert_conn *isert_conn, u32 count)
 993 {
 994         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 995         int i, ret;
 996         unsigned int rx_head = isert_conn->rx_desc_head;
 997         struct iser_rx_desc *rx_desc;
 998
 999         for (rx_wr = isert_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
1000                 rx_desc         = &isert_conn->rx_descs[rx_head];
1001                 rx_wr->wr_id    = (uintptr_t)rx_desc;
1002                 rx_wr->sg_list  = &rx_desc->rx_sg;
1003                 rx_wr->num_sge  = 1;
1004                 rx_wr->next     = rx_wr + 1;
1005                 rx_head = (rx_head + 1) & (ISERT_QP_MAX_RECV_DTOS - 1);
1006         }
1007
1008         rx_wr--;
1009         rx_wr->next = NULL; /* mark end of work requests list */
1010
1011         isert_conn->post_recv_buf_count += count;
1012         ret = ib_post_recv(isert_conn->qp, isert_conn->rx_wr,
1013                                 &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 -= count;
1017         } else {
1018                 isert_dbg("Posted %d RX buffers\n", count);
1019                 isert_conn->rx_desc_head = rx_head;
1020         }
1021         return ret;
1022 }
1023
1024 static int
1025 isert_post_send(struct isert_conn *isert_conn, struct iser_tx_desc *tx_desc)
1026 {
1027         struct ib_device *ib_dev = isert_conn->cm_id->device;
1028         struct ib_send_wr send_wr, *send_wr_failed;
1029         int ret;
1030
1031         ib_dma_sync_single_for_device(ib_dev, tx_desc->dma_addr,
1032                                       ISER_HEADERS_LEN, DMA_TO_DEVICE);
1033
1034         send_wr.next    = NULL;
1035         send_wr.wr_id   = (uintptr_t)tx_desc;
1036         send_wr.sg_list = tx_desc->tx_sg;
1037         send_wr.num_sge = tx_desc->num_sge;
1038         send_wr.opcode  = IB_WR_SEND;
1039         send_wr.send_flags = IB_SEND_SIGNALED;
1040
1041         ret = ib_post_send(isert_conn->qp, &send_wr, &send_wr_failed);
1042         if (ret)
1043                 isert_err("ib_post_send() failed, ret: %d\n", ret);
1044
1045         return ret;
1046 }
1047
1048 static void
1049 isert_create_send_desc(struct isert_conn *isert_conn,
1050                        struct isert_cmd *isert_cmd,
1051                        struct iser_tx_desc *tx_desc)
1052 {
1053         struct isert_device *device = isert_conn->device;
1054         struct ib_device *ib_dev = device->ib_device;
1055
1056         ib_dma_sync_single_for_cpu(ib_dev, tx_desc->dma_addr,
1057                                    ISER_HEADERS_LEN, DMA_TO_DEVICE);
1058
1059         memset(&tx_desc->iser_header, 0, sizeof(struct iser_hdr));
1060         tx_desc->iser_header.flags = ISER_VER;
1061
1062         tx_desc->num_sge = 1;
1063         tx_desc->isert_cmd = isert_cmd;
1064
1065         if (tx_desc->tx_sg[0].lkey != device->mr->lkey) {
1066                 tx_desc->tx_sg[0].lkey = device->mr->lkey;
1067                 isert_dbg("tx_desc %p lkey mismatch, fixing\n", tx_desc);
1068         }
1069 }
1070
1071 static int
1072 isert_init_tx_hdrs(struct isert_conn *isert_conn,
1073                    struct iser_tx_desc *tx_desc)
1074 {
1075         struct isert_device *device = isert_conn->device;
1076         struct ib_device *ib_dev = device->ib_device;
1077         u64 dma_addr;
1078
1079         dma_addr = ib_dma_map_single(ib_dev, (void *)tx_desc,
1080                         ISER_HEADERS_LEN, DMA_TO_DEVICE);
1081         if (ib_dma_mapping_error(ib_dev, dma_addr)) {
1082                 isert_err("ib_dma_mapping_error() failed\n");
1083                 return -ENOMEM;
1084         }
1085
1086         tx_desc->dma_addr = dma_addr;
1087         tx_desc->tx_sg[0].addr  = tx_desc->dma_addr;
1088         tx_desc->tx_sg[0].length = ISER_HEADERS_LEN;
1089         tx_desc->tx_sg[0].lkey = device->mr->lkey;
1090
1091         isert_dbg("Setup tx_sg[0].addr: 0x%llx length: %u lkey: 0x%x\n",
1092                   tx_desc->tx_sg[0].addr, tx_desc->tx_sg[0].length,
1093                   tx_desc->tx_sg[0].lkey);
1094
1095         return 0;
1096 }
1097
1098 static void
1099 isert_init_send_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1100                    struct ib_send_wr *send_wr)
1101 {
1102         struct iser_tx_desc *tx_desc = &isert_cmd->tx_desc;
1103
1104         isert_cmd->rdma_wr.iser_ib_op = ISER_IB_SEND;
1105         send_wr->wr_id = (uintptr_t)&isert_cmd->tx_desc;
1106         send_wr->opcode = IB_WR_SEND;
1107         send_wr->sg_list = &tx_desc->tx_sg[0];
1108         send_wr->num_sge = isert_cmd->tx_desc.num_sge;
1109         send_wr->send_flags = IB_SEND_SIGNALED;
1110 }
1111
1112 static int
1113 isert_rdma_post_recvl(struct isert_conn *isert_conn)
1114 {
1115         struct ib_recv_wr rx_wr, *rx_wr_fail;
1116         struct ib_sge sge;
1117         int ret;
1118
1119         memset(&sge, 0, sizeof(struct ib_sge));
1120         sge.addr = isert_conn->login_req_dma;
1121         sge.length = ISER_RX_LOGIN_SIZE;
1122         sge.lkey = isert_conn->device->mr->lkey;
1123
1124         isert_dbg("Setup sge: addr: %llx length: %d 0x%08x\n",
1125                 sge.addr, sge.length, sge.lkey);
1126
1127         memset(&rx_wr, 0, sizeof(struct ib_recv_wr));
1128         rx_wr.wr_id = (uintptr_t)isert_conn->login_req_buf;
1129         rx_wr.sg_list = &sge;
1130         rx_wr.num_sge = 1;
1131
1132         isert_conn->post_recv_buf_count++;
1133         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_fail);
1134         if (ret) {
1135                 isert_err("ib_post_recv() failed: %d\n", ret);
1136                 isert_conn->post_recv_buf_count--;
1137         }
1138
1139         return ret;
1140 }
1141
1142 static int
1143 isert_put_login_tx(struct iscsi_conn *conn, struct iscsi_login *login,
1144                    u32 length)
1145 {
1146         struct isert_conn *isert_conn = conn->context;
1147         struct isert_device *device = isert_conn->device;
1148         struct ib_device *ib_dev = device->ib_device;
1149         struct iser_tx_desc *tx_desc = &isert_conn->login_tx_desc;
1150         int ret;
1151
1152         isert_create_send_desc(isert_conn, NULL, tx_desc);
1153
1154         memcpy(&tx_desc->iscsi_header, &login->rsp[0],
1155                sizeof(struct iscsi_hdr));
1156
1157         isert_init_tx_hdrs(isert_conn, tx_desc);
1158
1159         if (length > 0) {
1160                 struct ib_sge *tx_dsg = &tx_desc->tx_sg[1];
1161
1162                 ib_dma_sync_single_for_cpu(ib_dev, isert_conn->login_rsp_dma,
1163                                            length, DMA_TO_DEVICE);
1164
1165                 memcpy(isert_conn->login_rsp_buf, login->rsp_buf, length);
1166
1167                 ib_dma_sync_single_for_device(ib_dev, isert_conn->login_rsp_dma,
1168                                               length, DMA_TO_DEVICE);
1169
1170                 tx_dsg->addr    = isert_conn->login_rsp_dma;
1171                 tx_dsg->length  = length;
1172                 tx_dsg->lkey    = isert_conn->device->mr->lkey;
1173                 tx_desc->num_sge = 2;
1174         }
1175         if (!login->login_failed) {
1176                 if (login->login_complete) {
1177                         if (!conn->sess->sess_ops->SessionType &&
1178                             isert_conn->device->use_fastreg) {
1179                                 ret = isert_conn_create_fastreg_pool(isert_conn);
1180                                 if (ret) {
1181                                         isert_err("Conn: %p failed to create"
1182                                                " fastreg pool\n", isert_conn);
1183                                         return ret;
1184                                 }
1185                         }
1186
1187                         ret = isert_alloc_rx_descriptors(isert_conn);
1188                         if (ret)
1189                                 return ret;
1190
1191                         ret = isert_post_recv(isert_conn, ISERT_MIN_POSTED_RX);
1192                         if (ret)
1193                                 return ret;
1194
1195                         /* Now we are in FULL_FEATURE phase */
1196                         mutex_lock(&isert_conn->mutex);
1197                         isert_conn->state = ISER_CONN_FULL_FEATURE;
1198                         mutex_unlock(&isert_conn->mutex);
1199                         goto post_send;
1200                 }
1201
1202                 ret = isert_rdma_post_recvl(isert_conn);
1203                 if (ret)
1204                         return ret;
1205         }
1206 post_send:
1207         ret = isert_post_send(isert_conn, tx_desc);
1208         if (ret)
1209                 return ret;
1210
1211         return 0;
1212 }
1213
1214 static void
1215 isert_rx_login_req(struct isert_conn *isert_conn)
1216 {
1217         struct iser_rx_desc *rx_desc = (void *)isert_conn->login_req_buf;
1218         int rx_buflen = isert_conn->login_req_len;
1219         struct iscsi_conn *conn = isert_conn->conn;
1220         struct iscsi_login *login = conn->conn_login;
1221         int size;
1222
1223         isert_info("conn %p\n", isert_conn);
1224
1225         WARN_ON_ONCE(!login);
1226
1227         if (login->first_request) {
1228                 struct iscsi_login_req *login_req =
1229                         (struct iscsi_login_req *)&rx_desc->iscsi_header;
1230                 /*
1231                  * Setup the initial iscsi_login values from the leading
1232                  * login request PDU.
1233                  */
1234                 login->leading_connection = (!login_req->tsih) ? 1 : 0;
1235                 login->current_stage =
1236                         (login_req->flags & ISCSI_FLAG_LOGIN_CURRENT_STAGE_MASK)
1237                          >> 2;
1238                 login->version_min      = login_req->min_version;
1239                 login->version_max      = login_req->max_version;
1240                 memcpy(login->isid, login_req->isid, 6);
1241                 login->cmd_sn           = be32_to_cpu(login_req->cmdsn);
1242                 login->init_task_tag    = login_req->itt;
1243                 login->initial_exp_statsn = be32_to_cpu(login_req->exp_statsn);
1244                 login->cid              = be16_to_cpu(login_req->cid);
1245                 login->tsih             = be16_to_cpu(login_req->tsih);
1246         }
1247
1248         memcpy(&login->req[0], (void *)&rx_desc->iscsi_header, ISCSI_HDR_LEN);
1249
1250         size = min(rx_buflen, MAX_KEY_VALUE_PAIRS);
1251         isert_dbg("Using login payload size: %d, rx_buflen: %d "
1252                   "MAX_KEY_VALUE_PAIRS: %d\n", size, rx_buflen,
1253                   MAX_KEY_VALUE_PAIRS);
1254         memcpy(login->req_buf, &rx_desc->data[0], size);
1255
1256         if (login->first_request) {
1257                 complete(&isert_conn->login_comp);
1258                 return;
1259         }
1260         schedule_delayed_work(&conn->login_work, 0);
1261 }
1262
1263 static struct iscsi_cmd
1264 *isert_allocate_cmd(struct iscsi_conn *conn)
1265 {
1266         struct isert_conn *isert_conn = conn->context;
1267         struct isert_cmd *isert_cmd;
1268         struct iscsi_cmd *cmd;
1269
1270         cmd = iscsit_allocate_cmd(conn, TASK_INTERRUPTIBLE);
1271         if (!cmd) {
1272                 isert_err("Unable to allocate iscsi_cmd + isert_cmd\n");
1273                 return NULL;
1274         }
1275         isert_cmd = iscsit_priv_cmd(cmd);
1276         isert_cmd->conn = isert_conn;
1277         isert_cmd->iscsi_cmd = cmd;
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         struct scatterlist *sg;
1290         int imm_data, imm_data_len, unsol_data, sg_nents, rc;
1291         bool dump_payload = false;
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
1301         rc = iscsit_process_scsi_cmd(conn, cmd, hdr);
1302         if (rc < 0) {
1303                 return 0;
1304         } else if (rc > 0) {
1305                 dump_payload = true;
1306                 goto sequence_cmd;
1307         }
1308
1309         if (!imm_data)
1310                 return 0;
1311
1312         sg = &cmd->se_cmd.t_data_sg[0];
1313         sg_nents = max(1UL, DIV_ROUND_UP(imm_data_len, PAGE_SIZE));
1314
1315         isert_dbg("Copying Immediate SG: %p sg_nents: %u from %p imm_data_len: %d\n",
1316                   sg, sg_nents, &rx_desc->data[0], imm_data_len);
1317
1318         sg_copy_from_buffer(sg, sg_nents, &rx_desc->data[0], imm_data_len);
1319
1320         cmd->write_data_done += imm_data_len;
1321
1322         if (cmd->write_data_done == cmd->se_cmd.data_length) {
1323                 spin_lock_bh(&cmd->istate_lock);
1324                 cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1325                 cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1326                 spin_unlock_bh(&cmd->istate_lock);
1327         }
1328
1329 sequence_cmd:
1330         rc = iscsit_sequence_cmd(conn, cmd, buf, hdr->cmdsn);
1331
1332         if (!rc && dump_payload == false && unsol_data)
1333                 iscsit_set_unsoliticed_dataout(cmd);
1334         else if (dump_payload && imm_data)
1335                 target_put_sess_cmd(conn->sess->se_sess, &cmd->se_cmd);
1336
1337         return 0;
1338 }
1339
1340 static int
1341 isert_handle_iscsi_dataout(struct isert_conn *isert_conn,
1342                            struct iser_rx_desc *rx_desc, unsigned char *buf)
1343 {
1344         struct scatterlist *sg_start;
1345         struct iscsi_conn *conn = isert_conn->conn;
1346         struct iscsi_cmd *cmd = NULL;
1347         struct iscsi_data *hdr = (struct iscsi_data *)buf;
1348         u32 unsol_data_len = ntoh24(hdr->dlength);
1349         int rc, sg_nents, sg_off, page_off;
1350
1351         rc = iscsit_check_dataout_hdr(conn, buf, &cmd);
1352         if (rc < 0)
1353                 return rc;
1354         else if (!cmd)
1355                 return 0;
1356         /*
1357          * FIXME: Unexpected unsolicited_data out
1358          */
1359         if (!cmd->unsolicited_data) {
1360                 isert_err("Received unexpected solicited data payload\n");
1361                 dump_stack();
1362                 return -1;
1363         }
1364
1365         isert_dbg("Unsolicited DataOut unsol_data_len: %u, "
1366                   "write_data_done: %u, data_length: %u\n",
1367                   unsol_data_len,  cmd->write_data_done,
1368                   cmd->se_cmd.data_length);
1369
1370         sg_off = cmd->write_data_done / PAGE_SIZE;
1371         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
1372         sg_nents = max(1UL, DIV_ROUND_UP(unsol_data_len, PAGE_SIZE));
1373         page_off = cmd->write_data_done % PAGE_SIZE;
1374         /*
1375          * FIXME: Non page-aligned unsolicited_data out
1376          */
1377         if (page_off) {
1378                 isert_err("unexpected non-page aligned data payload\n");
1379                 dump_stack();
1380                 return -1;
1381         }
1382         isert_dbg("Copying DataOut: sg_start: %p, sg_off: %u "
1383                   "sg_nents: %u from %p %u\n", sg_start, sg_off,
1384                   sg_nents, &rx_desc->data[0], unsol_data_len);
1385
1386         sg_copy_from_buffer(sg_start, sg_nents, &rx_desc->data[0],
1387                             unsol_data_len);
1388
1389         rc = iscsit_check_dataout_payload(cmd, hdr, false);
1390         if (rc < 0)
1391                 return rc;
1392
1393         return 0;
1394 }
1395
1396 static int
1397 isert_handle_nop_out(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1398                      struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1399                      unsigned char *buf)
1400 {
1401         struct iscsi_conn *conn = isert_conn->conn;
1402         struct iscsi_nopout *hdr = (struct iscsi_nopout *)buf;
1403         int rc;
1404
1405         rc = iscsit_setup_nop_out(conn, cmd, hdr);
1406         if (rc < 0)
1407                 return rc;
1408         /*
1409          * FIXME: Add support for NOPOUT payload using unsolicited RDMA payload
1410          */
1411
1412         return iscsit_process_nop_out(conn, cmd, hdr);
1413 }
1414
1415 static int
1416 isert_handle_text_cmd(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1417                       struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1418                       struct iscsi_text *hdr)
1419 {
1420         struct iscsi_conn *conn = isert_conn->conn;
1421         u32 payload_length = ntoh24(hdr->dlength);
1422         int rc;
1423         unsigned char *text_in = NULL;
1424
1425         rc = iscsit_setup_text_cmd(conn, cmd, hdr);
1426         if (rc < 0)
1427                 return rc;
1428
1429         if (payload_length) {
1430                 text_in = kzalloc(payload_length, GFP_KERNEL);
1431                 if (!text_in) {
1432                         isert_err("Unable to allocate text_in of payload_length: %u\n",
1433                                   payload_length);
1434                         return -ENOMEM;
1435                 }
1436         }
1437         cmd->text_in_ptr = text_in;
1438
1439         memcpy(cmd->text_in_ptr, &rx_desc->data[0], payload_length);
1440
1441         return iscsit_process_text_cmd(conn, cmd, hdr);
1442 }
1443
1444 static int
1445 isert_rx_opcode(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc,
1446                 uint32_t read_stag, uint64_t read_va,
1447                 uint32_t write_stag, uint64_t write_va)
1448 {
1449         struct iscsi_hdr *hdr = &rx_desc->iscsi_header;
1450         struct iscsi_conn *conn = isert_conn->conn;
1451         struct iscsi_cmd *cmd;
1452         struct isert_cmd *isert_cmd;
1453         int ret = -EINVAL;
1454         u8 opcode = (hdr->opcode & ISCSI_OPCODE_MASK);
1455
1456         if (conn->sess->sess_ops->SessionType &&
1457            (!(opcode & ISCSI_OP_TEXT) || !(opcode & ISCSI_OP_LOGOUT))) {
1458                 isert_err("Got illegal opcode: 0x%02x in SessionType=Discovery,"
1459                           " ignoring\n", opcode);
1460                 return 0;
1461         }
1462
1463         switch (opcode) {
1464         case ISCSI_OP_SCSI_CMD:
1465                 cmd = isert_allocate_cmd(conn);
1466                 if (!cmd)
1467                         break;
1468
1469                 isert_cmd = iscsit_priv_cmd(cmd);
1470                 isert_cmd->read_stag = read_stag;
1471                 isert_cmd->read_va = read_va;
1472                 isert_cmd->write_stag = write_stag;
1473                 isert_cmd->write_va = write_va;
1474
1475                 ret = isert_handle_scsi_cmd(isert_conn, isert_cmd, cmd,
1476                                         rx_desc, (unsigned char *)hdr);
1477                 break;
1478         case ISCSI_OP_NOOP_OUT:
1479                 cmd = isert_allocate_cmd(conn);
1480                 if (!cmd)
1481                         break;
1482
1483                 isert_cmd = iscsit_priv_cmd(cmd);
1484                 ret = isert_handle_nop_out(isert_conn, isert_cmd, cmd,
1485                                            rx_desc, (unsigned char *)hdr);
1486                 break;
1487         case ISCSI_OP_SCSI_DATA_OUT:
1488                 ret = isert_handle_iscsi_dataout(isert_conn, rx_desc,
1489                                                 (unsigned char *)hdr);
1490                 break;
1491         case ISCSI_OP_SCSI_TMFUNC:
1492                 cmd = isert_allocate_cmd(conn);
1493                 if (!cmd)
1494                         break;
1495
1496                 ret = iscsit_handle_task_mgt_cmd(conn, cmd,
1497                                                 (unsigned char *)hdr);
1498                 break;
1499         case ISCSI_OP_LOGOUT:
1500                 cmd = isert_allocate_cmd(conn);
1501                 if (!cmd)
1502                         break;
1503
1504                 ret = iscsit_handle_logout_cmd(conn, cmd, (unsigned char *)hdr);
1505                 break;
1506         case ISCSI_OP_TEXT:
1507                 if (be32_to_cpu(hdr->ttt) != 0xFFFFFFFF) {
1508                         cmd = iscsit_find_cmd_from_itt(conn, hdr->itt);
1509                         if (!cmd)
1510                                 break;
1511                 } else {
1512                         cmd = isert_allocate_cmd(conn);
1513                         if (!cmd)
1514                                 break;
1515                 }
1516
1517                 isert_cmd = iscsit_priv_cmd(cmd);
1518                 ret = isert_handle_text_cmd(isert_conn, isert_cmd, cmd,
1519                                             rx_desc, (struct iscsi_text *)hdr);
1520                 break;
1521         default:
1522                 isert_err("Got unknown iSCSI OpCode: 0x%02x\n", opcode);
1523                 dump_stack();
1524                 break;
1525         }
1526
1527         return ret;
1528 }
1529
1530 static void
1531 isert_rx_do_work(struct iser_rx_desc *rx_desc, struct isert_conn *isert_conn)
1532 {
1533         struct iser_hdr *iser_hdr = &rx_desc->iser_header;
1534         uint64_t read_va = 0, write_va = 0;
1535         uint32_t read_stag = 0, write_stag = 0;
1536         int rc;
1537
1538         switch (iser_hdr->flags & 0xF0) {
1539         case ISCSI_CTRL:
1540                 if (iser_hdr->flags & ISER_RSV) {
1541                         read_stag = be32_to_cpu(iser_hdr->read_stag);
1542                         read_va = be64_to_cpu(iser_hdr->read_va);
1543                         isert_dbg("ISER_RSV: read_stag: 0x%x read_va: 0x%llx\n",
1544                                   read_stag, (unsigned long long)read_va);
1545                 }
1546                 if (iser_hdr->flags & ISER_WSV) {
1547                         write_stag = be32_to_cpu(iser_hdr->write_stag);
1548                         write_va = be64_to_cpu(iser_hdr->write_va);
1549                         isert_dbg("ISER_WSV: write_stag: 0x%x write_va: 0x%llx\n",
1550                                   write_stag, (unsigned long long)write_va);
1551                 }
1552
1553                 isert_dbg("ISER ISCSI_CTRL PDU\n");
1554                 break;
1555         case ISER_HELLO:
1556                 isert_err("iSER Hello message\n");
1557                 break;
1558         default:
1559                 isert_warn("Unknown iSER hdr flags: 0x%02x\n", iser_hdr->flags);
1560                 break;
1561         }
1562
1563         rc = isert_rx_opcode(isert_conn, rx_desc,
1564                              read_stag, read_va, write_stag, write_va);
1565 }
1566
1567 static void
1568 isert_rcv_completion(struct iser_rx_desc *desc,
1569                      struct isert_conn *isert_conn,
1570                      u32 xfer_len)
1571 {
1572         struct ib_device *ib_dev = isert_conn->cm_id->device;
1573         struct iscsi_hdr *hdr;
1574         u64 rx_dma;
1575         int rx_buflen, outstanding;
1576
1577         if ((char *)desc == isert_conn->login_req_buf) {
1578                 rx_dma = isert_conn->login_req_dma;
1579                 rx_buflen = ISER_RX_LOGIN_SIZE;
1580                 isert_dbg("login_buf: Using rx_dma: 0x%llx, rx_buflen: %d\n",
1581                          rx_dma, rx_buflen);
1582         } else {
1583                 rx_dma = desc->dma_addr;
1584                 rx_buflen = ISER_RX_PAYLOAD_SIZE;
1585                 isert_dbg("req_buf: Using rx_dma: 0x%llx, rx_buflen: %d\n",
1586                          rx_dma, rx_buflen);
1587         }
1588
1589         ib_dma_sync_single_for_cpu(ib_dev, rx_dma, rx_buflen, DMA_FROM_DEVICE);
1590
1591         hdr = &desc->iscsi_header;
1592         isert_dbg("iSCSI opcode: 0x%02x, ITT: 0x%08x, flags: 0x%02x dlen: %d\n",
1593                  hdr->opcode, hdr->itt, hdr->flags,
1594                  (int)(xfer_len - ISER_HEADERS_LEN));
1595
1596         if ((char *)desc == isert_conn->login_req_buf) {
1597                 isert_conn->login_req_len = xfer_len - ISER_HEADERS_LEN;
1598                 if (isert_conn->conn) {
1599                         struct iscsi_login *login = isert_conn->conn->conn_login;
1600
1601                         if (login && !login->first_request)
1602                                 isert_rx_login_req(isert_conn);
1603                 }
1604                 mutex_lock(&isert_conn->mutex);
1605                 complete(&isert_conn->login_req_comp);
1606                 mutex_unlock(&isert_conn->mutex);
1607         } else {
1608                 isert_rx_do_work(desc, isert_conn);
1609         }
1610
1611         ib_dma_sync_single_for_device(ib_dev, rx_dma, rx_buflen,
1612                                       DMA_FROM_DEVICE);
1613
1614         isert_conn->post_recv_buf_count--;
1615         isert_dbg("Decremented post_recv_buf_count: %d\n",
1616                   isert_conn->post_recv_buf_count);
1617
1618         if ((char *)desc == isert_conn->login_req_buf)
1619                 return;
1620
1621         outstanding = isert_conn->post_recv_buf_count;
1622         if (outstanding + ISERT_MIN_POSTED_RX <= ISERT_QP_MAX_RECV_DTOS) {
1623                 int err, count = min(ISERT_QP_MAX_RECV_DTOS - outstanding,
1624                                 ISERT_MIN_POSTED_RX);
1625                 err = isert_post_recv(isert_conn, count);
1626                 if (err) {
1627                         isert_err("isert_post_recv() count: %d failed, %d\n",
1628                                count, err);
1629                 }
1630         }
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->send_wr) {
1690                 isert_dbg("Cmd %p free send_wr\n", isert_cmd);
1691                 kfree(wr->send_wr);
1692                 wr->send_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->send_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->se_sess, 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->send_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->send_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->se_sess, 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("wr id %llx status %d vend_err %x\n",
2095                                   wc->wr_id, wc->status, wc->vendor_err);
2096                 else
2097                         isert_dbg("flush error: wr id %llx\n", wc->wr_id);
2098
2099                 if (wc->wr_id != ISER_FASTREG_LI_WRID)
2100                         isert_cq_comp_err(isert_conn, wc);
2101         }
2102 }
2103
2104 static void
2105 isert_cq_work(struct work_struct *work)
2106 {
2107         enum { isert_poll_budget = 65536 };
2108         struct isert_comp *comp = container_of(work, struct isert_comp,
2109                                                work);
2110         struct ib_wc *const wcs = comp->wcs;
2111         int i, n, completed = 0;
2112
2113         while ((n = ib_poll_cq(comp->cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
2114                 for (i = 0; i < n; i++)
2115                         isert_handle_wc(&wcs[i]);
2116
2117                 completed += n;
2118                 if (completed >= isert_poll_budget)
2119                         break;
2120         }
2121
2122         ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP);
2123 }
2124
2125 static void
2126 isert_cq_callback(struct ib_cq *cq, void *context)
2127 {
2128         struct isert_comp *comp = context;
2129
2130         queue_work(isert_comp_wq, &comp->work);
2131 }
2132
2133 static int
2134 isert_post_response(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd)
2135 {
2136         struct ib_send_wr *wr_failed;
2137         int ret;
2138
2139         ret = ib_post_send(isert_conn->qp, &isert_cmd->tx_desc.send_wr,
2140                            &wr_failed);
2141         if (ret) {
2142                 isert_err("ib_post_send failed with %d\n", ret);
2143                 return ret;
2144         }
2145         return ret;
2146 }
2147
2148 static int
2149 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2150 {
2151         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2152         struct isert_conn *isert_conn = conn->context;
2153         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2154         struct iscsi_scsi_rsp *hdr = (struct iscsi_scsi_rsp *)
2155                                 &isert_cmd->tx_desc.iscsi_header;
2156
2157         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2158         iscsit_build_rsp_pdu(cmd, conn, true, hdr);
2159         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2160         /*
2161          * Attach SENSE DATA payload to iSCSI Response PDU
2162          */
2163         if (cmd->se_cmd.sense_buffer &&
2164             ((cmd->se_cmd.se_cmd_flags & SCF_TRANSPORT_TASK_SENSE) ||
2165             (cmd->se_cmd.se_cmd_flags & SCF_EMULATED_TASK_SENSE))) {
2166                 struct isert_device *device = isert_conn->device;
2167                 struct ib_device *ib_dev = device->ib_device;
2168                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2169                 u32 padding, pdu_len;
2170
2171                 put_unaligned_be16(cmd->se_cmd.scsi_sense_length,
2172                                    cmd->sense_buffer);
2173                 cmd->se_cmd.scsi_sense_length += sizeof(__be16);
2174
2175                 padding = -(cmd->se_cmd.scsi_sense_length) & 3;
2176                 hton24(hdr->dlength, (u32)cmd->se_cmd.scsi_sense_length);
2177                 pdu_len = cmd->se_cmd.scsi_sense_length + padding;
2178
2179                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2180                                 (void *)cmd->sense_buffer, pdu_len,
2181                                 DMA_TO_DEVICE);
2182
2183                 isert_cmd->pdu_buf_len = pdu_len;
2184                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2185                 tx_dsg->length  = pdu_len;
2186                 tx_dsg->lkey    = device->mr->lkey;
2187                 isert_cmd->tx_desc.num_sge = 2;
2188         }
2189
2190         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2191
2192         isert_dbg("Posting SCSI Response\n");
2193
2194         return isert_post_response(isert_conn, isert_cmd);
2195 }
2196
2197 static void
2198 isert_aborted_task(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2199 {
2200         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2201         struct isert_conn *isert_conn = conn->context;
2202         struct isert_device *device = isert_conn->device;
2203
2204         spin_lock_bh(&conn->cmd_lock);
2205         if (!list_empty(&cmd->i_conn_node))
2206                 list_del_init(&cmd->i_conn_node);
2207         spin_unlock_bh(&conn->cmd_lock);
2208
2209         if (cmd->data_direction == DMA_TO_DEVICE)
2210                 iscsit_stop_dataout_timer(cmd);
2211
2212         device->unreg_rdma_mem(isert_cmd, isert_conn);
2213 }
2214
2215 static enum target_prot_op
2216 isert_get_sup_prot_ops(struct iscsi_conn *conn)
2217 {
2218         struct isert_conn *isert_conn = conn->context;
2219         struct isert_device *device = isert_conn->device;
2220
2221         if (conn->tpg->tpg_attrib.t10_pi) {
2222                 if (device->pi_capable) {
2223                         isert_info("conn %p PI offload enabled\n", isert_conn);
2224                         isert_conn->pi_support = true;
2225                         return TARGET_PROT_ALL;
2226                 }
2227         }
2228
2229         isert_info("conn %p PI offload disabled\n", isert_conn);
2230         isert_conn->pi_support = false;
2231
2232         return TARGET_PROT_NORMAL;
2233 }
2234
2235 static int
2236 isert_put_nopin(struct iscsi_cmd *cmd, struct iscsi_conn *conn,
2237                 bool nopout_response)
2238 {
2239         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2240         struct isert_conn *isert_conn = conn->context;
2241         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2242
2243         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2244         iscsit_build_nopin_rsp(cmd, conn, (struct iscsi_nopin *)
2245                                &isert_cmd->tx_desc.iscsi_header,
2246                                nopout_response);
2247         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2248         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2249
2250         isert_dbg("conn %p Posting NOPIN Response\n", isert_conn);
2251
2252         return isert_post_response(isert_conn, isert_cmd);
2253 }
2254
2255 static int
2256 isert_put_logout_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2257 {
2258         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2259         struct isert_conn *isert_conn = conn->context;
2260         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2261
2262         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2263         iscsit_build_logout_rsp(cmd, conn, (struct iscsi_logout_rsp *)
2264                                 &isert_cmd->tx_desc.iscsi_header);
2265         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2266         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2267
2268         isert_dbg("conn %p Posting Logout Response\n", isert_conn);
2269
2270         return isert_post_response(isert_conn, isert_cmd);
2271 }
2272
2273 static int
2274 isert_put_tm_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2275 {
2276         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2277         struct isert_conn *isert_conn = conn->context;
2278         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2279
2280         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2281         iscsit_build_task_mgt_rsp(cmd, conn, (struct iscsi_tm_rsp *)
2282                                   &isert_cmd->tx_desc.iscsi_header);
2283         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2284         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2285
2286         isert_dbg("conn %p Posting Task Management Response\n", isert_conn);
2287
2288         return isert_post_response(isert_conn, isert_cmd);
2289 }
2290
2291 static int
2292 isert_put_reject(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2293 {
2294         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2295         struct isert_conn *isert_conn = conn->context;
2296         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2297         struct isert_device *device = isert_conn->device;
2298         struct ib_device *ib_dev = device->ib_device;
2299         struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2300         struct iscsi_reject *hdr =
2301                 (struct iscsi_reject *)&isert_cmd->tx_desc.iscsi_header;
2302
2303         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2304         iscsit_build_reject(cmd, conn, hdr);
2305         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2306
2307         hton24(hdr->dlength, ISCSI_HDR_LEN);
2308         isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2309                         (void *)cmd->buf_ptr, ISCSI_HDR_LEN,
2310                         DMA_TO_DEVICE);
2311         isert_cmd->pdu_buf_len = ISCSI_HDR_LEN;
2312         tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2313         tx_dsg->length  = ISCSI_HDR_LEN;
2314         tx_dsg->lkey    = device->mr->lkey;
2315         isert_cmd->tx_desc.num_sge = 2;
2316
2317         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2318
2319         isert_dbg("conn %p Posting Reject\n", isert_conn);
2320
2321         return isert_post_response(isert_conn, isert_cmd);
2322 }
2323
2324 static int
2325 isert_put_text_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2326 {
2327         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2328         struct isert_conn *isert_conn = conn->context;
2329         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2330         struct iscsi_text_rsp *hdr =
2331                 (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
2332         u32 txt_rsp_len;
2333         int rc;
2334
2335         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2336         rc = iscsit_build_text_rsp(cmd, conn, hdr, ISCSI_INFINIBAND);
2337         if (rc < 0)
2338                 return rc;
2339
2340         txt_rsp_len = rc;
2341         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2342
2343         if (txt_rsp_len) {
2344                 struct isert_device *device = isert_conn->device;
2345                 struct ib_device *ib_dev = device->ib_device;
2346                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2347                 void *txt_rsp_buf = cmd->buf_ptr;
2348
2349                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2350                                 txt_rsp_buf, txt_rsp_len, DMA_TO_DEVICE);
2351
2352                 isert_cmd->pdu_buf_len = txt_rsp_len;
2353                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2354                 tx_dsg->length  = txt_rsp_len;
2355                 tx_dsg->lkey    = device->mr->lkey;
2356                 isert_cmd->tx_desc.num_sge = 2;
2357         }
2358         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2359
2360         isert_dbg("conn %p Text Response\n", isert_conn);
2361
2362         return isert_post_response(isert_conn, isert_cmd);
2363 }
2364
2365 static int
2366 isert_build_rdma_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
2367                     struct ib_sge *ib_sge, struct ib_send_wr *send_wr,
2368                     u32 data_left, u32 offset)
2369 {
2370         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
2371         struct scatterlist *sg_start, *tmp_sg;
2372         struct isert_device *device = isert_conn->device;
2373         struct ib_device *ib_dev = device->ib_device;
2374         u32 sg_off, page_off;
2375         int i = 0, sg_nents;
2376
2377         sg_off = offset / PAGE_SIZE;
2378         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
2379         sg_nents = min(cmd->se_cmd.t_data_nents - sg_off, isert_conn->max_sge);
2380         page_off = offset % PAGE_SIZE;
2381
2382         send_wr->sg_list = ib_sge;
2383         send_wr->wr_id = (uintptr_t)&isert_cmd->tx_desc;
2384         /*
2385          * Perform mapping of TCM scatterlist memory ib_sge dma_addr.
2386          */
2387         for_each_sg(sg_start, tmp_sg, sg_nents, i) {
2388                 isert_dbg("RDMA from SGL dma_addr: 0x%llx dma_len: %u, "
2389                           "page_off: %u\n",
2390                           (unsigned long long)tmp_sg->dma_address,
2391                           tmp_sg->length, page_off);
2392
2393                 ib_sge->addr = ib_sg_dma_address(ib_dev, tmp_sg) + page_off;
2394                 ib_sge->length = min_t(u32, data_left,
2395                                 ib_sg_dma_len(ib_dev, tmp_sg) - page_off);
2396                 ib_sge->lkey = device->mr->lkey;
2397
2398                 isert_dbg("RDMA ib_sge: addr: 0x%llx  length: %u lkey: %x\n",
2399                           ib_sge->addr, ib_sge->length, ib_sge->lkey);
2400                 page_off = 0;
2401                 data_left -= ib_sge->length;
2402                 if (!data_left)
2403                         break;
2404                 ib_sge++;
2405                 isert_dbg("Incrementing ib_sge pointer to %p\n", ib_sge);
2406         }
2407
2408         send_wr->num_sge = ++i;
2409         isert_dbg("Set outgoing sg_list: %p num_sg: %u from TCM SGLs\n",
2410                   send_wr->sg_list, send_wr->num_sge);
2411
2412         return send_wr->num_sge;
2413 }
2414
2415 static int
2416 isert_map_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2417                struct isert_rdma_wr *wr)
2418 {
2419         struct se_cmd *se_cmd = &cmd->se_cmd;
2420         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2421         struct isert_conn *isert_conn = conn->context;
2422         struct isert_data_buf *data = &wr->data;
2423         struct ib_send_wr *send_wr;
2424         struct ib_sge *ib_sge;
2425         u32 offset, data_len, data_left, rdma_write_max, va_offset = 0;
2426         int ret = 0, i, ib_sge_cnt;
2427
2428         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2429
2430         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2431         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2432                                  se_cmd->t_data_nents, se_cmd->data_length,
2433                                  offset, wr->iser_ib_op, &wr->data);
2434         if (ret)
2435                 return ret;
2436
2437         data_left = data->len;
2438         offset = data->offset;
2439
2440         ib_sge = kzalloc(sizeof(struct ib_sge) * data->nents, GFP_KERNEL);
2441         if (!ib_sge) {
2442                 isert_warn("Unable to allocate ib_sge\n");
2443                 ret = -ENOMEM;
2444                 goto unmap_cmd;
2445         }
2446         wr->ib_sge = ib_sge;
2447
2448         wr->send_wr_num = DIV_ROUND_UP(data->nents, isert_conn->max_sge);
2449         wr->send_wr = kzalloc(sizeof(struct ib_send_wr) * wr->send_wr_num,
2450                                 GFP_KERNEL);
2451         if (!wr->send_wr) {
2452                 isert_dbg("Unable to allocate wr->send_wr\n");
2453                 ret = -ENOMEM;
2454                 goto unmap_cmd;
2455         }
2456
2457         wr->isert_cmd = isert_cmd;
2458         rdma_write_max = isert_conn->max_sge * PAGE_SIZE;
2459
2460         for (i = 0; i < wr->send_wr_num; i++) {
2461                 send_wr = &isert_cmd->rdma_wr.send_wr[i];
2462                 data_len = min(data_left, rdma_write_max);
2463
2464                 send_wr->send_flags = 0;
2465                 if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2466                         send_wr->opcode = IB_WR_RDMA_WRITE;
2467                         send_wr->wr.rdma.remote_addr = isert_cmd->read_va + offset;
2468                         send_wr->wr.rdma.rkey = isert_cmd->read_stag;
2469                         if (i + 1 == wr->send_wr_num)
2470                                 send_wr->next = &isert_cmd->tx_desc.send_wr;
2471                         else
2472                                 send_wr->next = &wr->send_wr[i + 1];
2473                 } else {
2474                         send_wr->opcode = IB_WR_RDMA_READ;
2475                         send_wr->wr.rdma.remote_addr = isert_cmd->write_va + va_offset;
2476                         send_wr->wr.rdma.rkey = isert_cmd->write_stag;
2477                         if (i + 1 == wr->send_wr_num)
2478                                 send_wr->send_flags = IB_SEND_SIGNALED;
2479                         else
2480                                 send_wr->next = &wr->send_wr[i + 1];
2481                 }
2482
2483                 ib_sge_cnt = isert_build_rdma_wr(isert_conn, isert_cmd, ib_sge,
2484                                         send_wr, data_len, offset);
2485                 ib_sge += ib_sge_cnt;
2486
2487                 offset += data_len;
2488                 va_offset += data_len;
2489                 data_left -= data_len;
2490         }
2491
2492         return 0;
2493 unmap_cmd:
2494         isert_unmap_data_buf(isert_conn, data);
2495
2496         return ret;
2497 }
2498
2499 static int
2500 isert_map_fr_pagelist(struct ib_device *ib_dev,
2501                       struct scatterlist *sg_start, int sg_nents, u64 *fr_pl)
2502 {
2503         u64 start_addr, end_addr, page, chunk_start = 0;
2504         struct scatterlist *tmp_sg;
2505         int i = 0, new_chunk, last_ent, n_pages;
2506
2507         n_pages = 0;
2508         new_chunk = 1;
2509         last_ent = sg_nents - 1;
2510         for_each_sg(sg_start, tmp_sg, sg_nents, i) {
2511                 start_addr = ib_sg_dma_address(ib_dev, tmp_sg);
2512                 if (new_chunk)
2513                         chunk_start = start_addr;
2514                 end_addr = start_addr + ib_sg_dma_len(ib_dev, tmp_sg);
2515
2516                 isert_dbg("SGL[%d] dma_addr: 0x%llx len: %u\n",
2517                           i, (unsigned long long)tmp_sg->dma_address,
2518                           tmp_sg->length);
2519
2520                 if ((end_addr & ~PAGE_MASK) && i < last_ent) {
2521                         new_chunk = 0;
2522                         continue;
2523                 }
2524                 new_chunk = 1;
2525
2526                 page = chunk_start & PAGE_MASK;
2527                 do {
2528                         fr_pl[n_pages++] = page;
2529                         isert_dbg("Mapped page_list[%d] page_addr: 0x%llx\n",
2530                                   n_pages - 1, page);
2531                         page += PAGE_SIZE;
2532                 } while (page < end_addr);
2533         }
2534
2535         return n_pages;
2536 }
2537
2538 static inline void
2539 isert_inv_rkey(struct ib_send_wr *inv_wr, struct ib_mr *mr)
2540 {
2541         u32 rkey;
2542
2543         memset(inv_wr, 0, sizeof(*inv_wr));
2544         inv_wr->wr_id = ISER_FASTREG_LI_WRID;
2545         inv_wr->opcode = IB_WR_LOCAL_INV;
2546         inv_wr->ex.invalidate_rkey = mr->rkey;
2547
2548         /* Bump the key */
2549         rkey = ib_inc_rkey(mr->rkey);
2550         ib_update_fast_reg_key(mr, rkey);
2551 }
2552
2553 static int
2554 isert_fast_reg_mr(struct isert_conn *isert_conn,
2555                   struct fast_reg_descriptor *fr_desc,
2556                   struct isert_data_buf *mem,
2557                   enum isert_indicator ind,
2558                   struct ib_sge *sge)
2559 {
2560         struct isert_device *device = isert_conn->device;
2561         struct ib_device *ib_dev = device->ib_device;
2562         struct ib_mr *mr;
2563         struct ib_fast_reg_page_list *frpl;
2564         struct ib_send_wr fr_wr, inv_wr;
2565         struct ib_send_wr *bad_wr, *wr = NULL;
2566         int ret, pagelist_len;
2567         u32 page_off;
2568
2569         if (mem->dma_nents == 1) {
2570                 sge->lkey = device->mr->lkey;
2571                 sge->addr = ib_sg_dma_address(ib_dev, &mem->sg[0]);
2572                 sge->length = ib_sg_dma_len(ib_dev, &mem->sg[0]);
2573                 isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2574                          sge->addr, sge->length, sge->lkey);
2575                 return 0;
2576         }
2577
2578         if (ind == ISERT_DATA_KEY_VALID) {
2579                 /* Registering data buffer */
2580                 mr = fr_desc->data_mr;
2581                 frpl = fr_desc->data_frpl;
2582         } else {
2583                 /* Registering protection buffer */
2584                 mr = fr_desc->pi_ctx->prot_mr;
2585                 frpl = fr_desc->pi_ctx->prot_frpl;
2586         }
2587
2588         page_off = mem->offset % PAGE_SIZE;
2589
2590         isert_dbg("Use fr_desc %p sg_nents %d offset %u\n",
2591                   fr_desc, mem->nents, mem->offset);
2592
2593         pagelist_len = isert_map_fr_pagelist(ib_dev, mem->sg, mem->nents,
2594                                              &frpl->page_list[0]);
2595
2596         if (!(fr_desc->ind & ind)) {
2597                 isert_inv_rkey(&inv_wr, mr);
2598                 wr = &inv_wr;
2599         }
2600
2601         /* Prepare FASTREG WR */
2602         memset(&fr_wr, 0, sizeof(fr_wr));
2603         fr_wr.wr_id = ISER_FASTREG_LI_WRID;
2604         fr_wr.opcode = IB_WR_FAST_REG_MR;
2605         fr_wr.wr.fast_reg.iova_start = frpl->page_list[0] + page_off;
2606         fr_wr.wr.fast_reg.page_list = frpl;
2607         fr_wr.wr.fast_reg.page_list_len = pagelist_len;
2608         fr_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
2609         fr_wr.wr.fast_reg.length = mem->len;
2610         fr_wr.wr.fast_reg.rkey = mr->rkey;
2611         fr_wr.wr.fast_reg.access_flags = IB_ACCESS_LOCAL_WRITE;
2612
2613         if (!wr)
2614                 wr = &fr_wr;
2615         else
2616                 wr->next = &fr_wr;
2617
2618         ret = ib_post_send(isert_conn->qp, wr, &bad_wr);
2619         if (ret) {
2620                 isert_err("fast registration failed, ret:%d\n", ret);
2621                 return ret;
2622         }
2623         fr_desc->ind &= ~ind;
2624
2625         sge->lkey = mr->lkey;
2626         sge->addr = frpl->page_list[0] + page_off;
2627         sge->length = mem->len;
2628
2629         isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2630                   sge->addr, sge->length, sge->lkey);
2631
2632         return ret;
2633 }
2634
2635 static inline void
2636 isert_set_dif_domain(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs,
2637                      struct ib_sig_domain *domain)
2638 {
2639         domain->sig_type = IB_SIG_TYPE_T10_DIF;
2640         domain->sig.dif.bg_type = IB_T10DIF_CRC;
2641         domain->sig.dif.pi_interval = se_cmd->se_dev->dev_attrib.block_size;
2642         domain->sig.dif.ref_tag = se_cmd->reftag_seed;
2643         /*
2644          * At the moment we hard code those, but if in the future
2645          * the target core would like to use it, we will take it
2646          * from se_cmd.
2647          */
2648         domain->sig.dif.apptag_check_mask = 0xffff;
2649         domain->sig.dif.app_escape = true;
2650         domain->sig.dif.ref_escape = true;
2651         if (se_cmd->prot_type == TARGET_DIF_TYPE1_PROT ||
2652             se_cmd->prot_type == TARGET_DIF_TYPE2_PROT)
2653                 domain->sig.dif.ref_remap = true;
2654 };
2655
2656 static int
2657 isert_set_sig_attrs(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs)
2658 {
2659         switch (se_cmd->prot_op) {
2660         case TARGET_PROT_DIN_INSERT:
2661         case TARGET_PROT_DOUT_STRIP:
2662                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
2663                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2664                 break;
2665         case TARGET_PROT_DOUT_INSERT:
2666         case TARGET_PROT_DIN_STRIP:
2667                 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
2668                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2669                 break;
2670         case TARGET_PROT_DIN_PASS:
2671         case TARGET_PROT_DOUT_PASS:
2672                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2673                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2674                 break;
2675         default:
2676                 isert_err("Unsupported PI operation %d\n", se_cmd->prot_op);
2677                 return -EINVAL;
2678         }
2679
2680         return 0;
2681 }
2682
2683 static inline u8
2684 isert_set_prot_checks(u8 prot_checks)
2685 {
2686         return (prot_checks & TARGET_DIF_CHECK_GUARD  ? 0xc0 : 0) |
2687                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x30 : 0) |
2688                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x0f : 0);
2689 }
2690
2691 static int
2692 isert_reg_sig_mr(struct isert_conn *isert_conn,
2693                  struct se_cmd *se_cmd,
2694                  struct isert_rdma_wr *rdma_wr,
2695                  struct fast_reg_descriptor *fr_desc)
2696 {
2697         struct ib_send_wr sig_wr, inv_wr;
2698         struct ib_send_wr *bad_wr, *wr = NULL;
2699         struct pi_context *pi_ctx = fr_desc->pi_ctx;
2700         struct ib_sig_attrs sig_attrs;
2701         int ret;
2702
2703         memset(&sig_attrs, 0, sizeof(sig_attrs));
2704         ret = isert_set_sig_attrs(se_cmd, &sig_attrs);
2705         if (ret)
2706                 goto err;
2707
2708         sig_attrs.check_mask = isert_set_prot_checks(se_cmd->prot_checks);
2709
2710         if (!(fr_desc->ind & ISERT_SIG_KEY_VALID)) {
2711                 isert_inv_rkey(&inv_wr, pi_ctx->sig_mr);
2712                 wr = &inv_wr;
2713         }
2714
2715         memset(&sig_wr, 0, sizeof(sig_wr));
2716         sig_wr.opcode = IB_WR_REG_SIG_MR;
2717         sig_wr.wr_id = ISER_FASTREG_LI_WRID;
2718         sig_wr.sg_list = &rdma_wr->ib_sg[DATA];
2719         sig_wr.num_sge = 1;
2720         sig_wr.wr.sig_handover.access_flags = IB_ACCESS_LOCAL_WRITE;
2721         sig_wr.wr.sig_handover.sig_attrs = &sig_attrs;
2722         sig_wr.wr.sig_handover.sig_mr = pi_ctx->sig_mr;
2723         if (se_cmd->t_prot_sg)
2724                 sig_wr.wr.sig_handover.prot = &rdma_wr->ib_sg[PROT];
2725
2726         if (!wr)
2727                 wr = &sig_wr;
2728         else
2729                 wr->next = &sig_wr;
2730
2731         ret = ib_post_send(isert_conn->qp, wr, &bad_wr);
2732         if (ret) {
2733                 isert_err("fast registration failed, ret:%d\n", ret);
2734                 goto err;
2735         }
2736         fr_desc->ind &= ~ISERT_SIG_KEY_VALID;
2737
2738         rdma_wr->ib_sg[SIG].lkey = pi_ctx->sig_mr->lkey;
2739         rdma_wr->ib_sg[SIG].addr = 0;
2740         rdma_wr->ib_sg[SIG].length = se_cmd->data_length;
2741         if (se_cmd->prot_op != TARGET_PROT_DIN_STRIP &&
2742             se_cmd->prot_op != TARGET_PROT_DOUT_INSERT)
2743                 /*
2744                  * We have protection guards on the wire
2745                  * so we need to set a larget transfer
2746                  */
2747                 rdma_wr->ib_sg[SIG].length += se_cmd->prot_length;
2748
2749         isert_dbg("sig_sge: addr: 0x%llx  length: %u lkey: %x\n",
2750                   rdma_wr->ib_sg[SIG].addr, rdma_wr->ib_sg[SIG].length,
2751                   rdma_wr->ib_sg[SIG].lkey);
2752 err:
2753         return ret;
2754 }
2755
2756 static int
2757 isert_handle_prot_cmd(struct isert_conn *isert_conn,
2758                       struct isert_cmd *isert_cmd,
2759                       struct isert_rdma_wr *wr)
2760 {
2761         struct isert_device *device = isert_conn->device;
2762         struct se_cmd *se_cmd = &isert_cmd->iscsi_cmd->se_cmd;
2763         int ret;
2764
2765         if (!wr->fr_desc->pi_ctx) {
2766                 ret = isert_create_pi_ctx(wr->fr_desc,
2767                                           device->ib_device,
2768                                           device->pd);
2769                 if (ret) {
2770                         isert_err("conn %p failed to allocate pi_ctx\n",
2771                                   isert_conn);
2772                         return ret;
2773                 }
2774         }
2775
2776         if (se_cmd->t_prot_sg) {
2777                 ret = isert_map_data_buf(isert_conn, isert_cmd,
2778                                          se_cmd->t_prot_sg,
2779                                          se_cmd->t_prot_nents,
2780                                          se_cmd->prot_length,
2781                                          0, wr->iser_ib_op, &wr->prot);
2782                 if (ret) {
2783                         isert_err("conn %p failed to map protection buffer\n",
2784                                   isert_conn);
2785                         return ret;
2786                 }
2787
2788                 memset(&wr->ib_sg[PROT], 0, sizeof(wr->ib_sg[PROT]));
2789                 ret = isert_fast_reg_mr(isert_conn, wr->fr_desc, &wr->prot,
2790                                         ISERT_PROT_KEY_VALID, &wr->ib_sg[PROT]);
2791                 if (ret) {
2792                         isert_err("conn %p failed to fast reg mr\n",
2793                                   isert_conn);
2794                         goto unmap_prot_cmd;
2795                 }
2796         }
2797
2798         ret = isert_reg_sig_mr(isert_conn, se_cmd, wr, wr->fr_desc);
2799         if (ret) {
2800                 isert_err("conn %p failed to fast reg mr\n",
2801                           isert_conn);
2802                 goto unmap_prot_cmd;
2803         }
2804         wr->fr_desc->ind |= ISERT_PROTECTED;
2805
2806         return 0;
2807
2808 unmap_prot_cmd:
2809         if (se_cmd->t_prot_sg)
2810                 isert_unmap_data_buf(isert_conn, &wr->prot);
2811
2812         return ret;
2813 }
2814
2815 static int
2816 isert_reg_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2817                struct isert_rdma_wr *wr)
2818 {
2819         struct se_cmd *se_cmd = &cmd->se_cmd;
2820         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2821         struct isert_conn *isert_conn = conn->context;
2822         struct fast_reg_descriptor *fr_desc = NULL;
2823         struct ib_send_wr *send_wr;
2824         struct ib_sge *ib_sg;
2825         u32 offset;
2826         int ret = 0;
2827         unsigned long flags;
2828
2829         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2830
2831         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2832         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2833                                  se_cmd->t_data_nents, se_cmd->data_length,
2834                                  offset, wr->iser_ib_op, &wr->data);
2835         if (ret)
2836                 return ret;
2837
2838         if (wr->data.dma_nents != 1 || isert_prot_cmd(isert_conn, se_cmd)) {
2839                 spin_lock_irqsave(&isert_conn->pool_lock, flags);
2840                 fr_desc = list_first_entry(&isert_conn->fr_pool,
2841                                            struct fast_reg_descriptor, list);
2842                 list_del(&fr_desc->list);
2843                 spin_unlock_irqrestore(&isert_conn->pool_lock, flags);
2844                 wr->fr_desc = fr_desc;
2845         }
2846
2847         ret = isert_fast_reg_mr(isert_conn, fr_desc, &wr->data,
2848                                 ISERT_DATA_KEY_VALID, &wr->ib_sg[DATA]);
2849         if (ret)
2850                 goto unmap_cmd;
2851
2852         if (isert_prot_cmd(isert_conn, se_cmd)) {
2853                 ret = isert_handle_prot_cmd(isert_conn, isert_cmd, wr);
2854                 if (ret)
2855                         goto unmap_cmd;
2856
2857                 ib_sg = &wr->ib_sg[SIG];
2858         } else {
2859                 ib_sg = &wr->ib_sg[DATA];
2860         }
2861
2862         memcpy(&wr->s_ib_sge, ib_sg, sizeof(*ib_sg));
2863         wr->ib_sge = &wr->s_ib_sge;
2864         wr->send_wr_num = 1;
2865         memset(&wr->s_send_wr, 0, sizeof(*send_wr));
2866         wr->send_wr = &wr->s_send_wr;
2867         wr->isert_cmd = isert_cmd;
2868
2869         send_wr = &isert_cmd->rdma_wr.s_send_wr;
2870         send_wr->sg_list = &wr->s_ib_sge;
2871         send_wr->num_sge = 1;
2872         send_wr->wr_id = (uintptr_t)&isert_cmd->tx_desc;
2873         if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2874                 send_wr->opcode = IB_WR_RDMA_WRITE;
2875                 send_wr->wr.rdma.remote_addr = isert_cmd->read_va;
2876                 send_wr->wr.rdma.rkey = isert_cmd->read_stag;
2877                 send_wr->send_flags = !isert_prot_cmd(isert_conn, se_cmd) ?
2878                                       0 : IB_SEND_SIGNALED;
2879         } else {
2880                 send_wr->opcode = IB_WR_RDMA_READ;
2881                 send_wr->wr.rdma.remote_addr = isert_cmd->write_va;
2882                 send_wr->wr.rdma.rkey = isert_cmd->write_stag;
2883                 send_wr->send_flags = IB_SEND_SIGNALED;
2884         }
2885
2886         return 0;
2887
2888 unmap_cmd:
2889         if (fr_desc) {
2890                 spin_lock_irqsave(&isert_conn->pool_lock, flags);
2891                 list_add_tail(&fr_desc->list, &isert_conn->fr_pool);
2892                 spin_unlock_irqrestore(&isert_conn->pool_lock, flags);
2893         }
2894         isert_unmap_data_buf(isert_conn, &wr->data);
2895
2896         return ret;
2897 }
2898
2899 static int
2900 isert_put_datain(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2901 {
2902         struct se_cmd *se_cmd = &cmd->se_cmd;
2903         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2904         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2905         struct isert_conn *isert_conn = conn->context;
2906         struct isert_device *device = isert_conn->device;
2907         struct ib_send_wr *wr_failed;
2908         int rc;
2909
2910         isert_dbg("Cmd: %p RDMA_WRITE data_length: %u\n",
2911                  isert_cmd, se_cmd->data_length);
2912
2913         wr->iser_ib_op = ISER_IB_RDMA_WRITE;
2914         rc = device->reg_rdma_mem(conn, cmd, wr);
2915         if (rc) {
2916                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2917                 return rc;
2918         }
2919
2920         if (!isert_prot_cmd(isert_conn, se_cmd)) {
2921                 /*
2922                  * Build isert_conn->tx_desc for iSCSI response PDU and attach
2923                  */
2924                 isert_create_send_desc(isert_conn, isert_cmd,
2925                                        &isert_cmd->tx_desc);
2926                 iscsit_build_rsp_pdu(cmd, conn, true, (struct iscsi_scsi_rsp *)
2927                                      &isert_cmd->tx_desc.iscsi_header);
2928                 isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2929                 isert_init_send_wr(isert_conn, isert_cmd,
2930                                    &isert_cmd->tx_desc.send_wr);
2931                 isert_cmd->rdma_wr.s_send_wr.next = &isert_cmd->tx_desc.send_wr;
2932                 wr->send_wr_num += 1;
2933         }
2934
2935         rc = ib_post_send(isert_conn->qp, wr->send_wr, &wr_failed);
2936         if (rc)
2937                 isert_warn("ib_post_send() failed for IB_WR_RDMA_WRITE\n");
2938
2939         if (!isert_prot_cmd(isert_conn, se_cmd))
2940                 isert_dbg("Cmd: %p posted RDMA_WRITE + Response for iSER Data "
2941                          "READ\n", isert_cmd);
2942         else
2943                 isert_dbg("Cmd: %p posted RDMA_WRITE for iSER Data READ\n",
2944                          isert_cmd);
2945
2946         return 1;
2947 }
2948
2949 static int
2950 isert_get_dataout(struct iscsi_conn *conn, struct iscsi_cmd *cmd, bool recovery)
2951 {
2952         struct se_cmd *se_cmd = &cmd->se_cmd;
2953         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2954         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2955         struct isert_conn *isert_conn = conn->context;
2956         struct isert_device *device = isert_conn->device;
2957         struct ib_send_wr *wr_failed;
2958         int rc;
2959
2960         isert_dbg("Cmd: %p RDMA_READ data_length: %u write_data_done: %u\n",
2961                  isert_cmd, se_cmd->data_length, cmd->write_data_done);
2962         wr->iser_ib_op = ISER_IB_RDMA_READ;
2963         rc = device->reg_rdma_mem(conn, cmd, wr);
2964         if (rc) {
2965                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2966                 return rc;
2967         }
2968
2969         rc = ib_post_send(isert_conn->qp, wr->send_wr, &wr_failed);
2970         if (rc)
2971                 isert_warn("ib_post_send() failed for IB_WR_RDMA_READ\n");
2972
2973         isert_dbg("Cmd: %p posted RDMA_READ memory for ISER Data WRITE\n",
2974                  isert_cmd);
2975
2976         return 0;
2977 }
2978
2979 static int
2980 isert_immediate_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2981 {
2982         int ret;
2983
2984         switch (state) {
2985         case ISTATE_SEND_NOPIN_WANT_RESPONSE:
2986                 ret = isert_put_nopin(cmd, conn, false);
2987                 break;
2988         default:
2989                 isert_err("Unknown immediate state: 0x%02x\n", state);
2990                 ret = -EINVAL;
2991                 break;
2992         }
2993
2994         return ret;
2995 }
2996
2997 static int
2998 isert_response_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2999 {
3000         struct isert_conn *isert_conn = conn->context;
3001         int ret;
3002
3003         switch (state) {
3004         case ISTATE_SEND_LOGOUTRSP:
3005                 ret = isert_put_logout_rsp(cmd, conn);
3006                 if (!ret)
3007                         isert_conn->logout_posted = true;
3008                 break;
3009         case ISTATE_SEND_NOPIN:
3010                 ret = isert_put_nopin(cmd, conn, true);
3011                 break;
3012         case ISTATE_SEND_TASKMGTRSP:
3013                 ret = isert_put_tm_rsp(cmd, conn);
3014                 break;
3015         case ISTATE_SEND_REJECT:
3016                 ret = isert_put_reject(cmd, conn);
3017                 break;
3018         case ISTATE_SEND_TEXTRSP:
3019                 ret = isert_put_text_rsp(cmd, conn);
3020                 break;
3021         case ISTATE_SEND_STATUS:
3022                 /*
3023                  * Special case for sending non GOOD SCSI status from TX thread
3024                  * context during pre se_cmd excecution failure.
3025                  */
3026                 ret = isert_put_response(conn, cmd);
3027                 break;
3028         default:
3029                 isert_err("Unknown response state: 0x%02x\n", state);
3030                 ret = -EINVAL;
3031                 break;
3032         }
3033
3034         return ret;
3035 }
3036
3037 struct rdma_cm_id *
3038 isert_setup_id(struct isert_np *isert_np)
3039 {
3040         struct iscsi_np *np = isert_np->np;
3041         struct rdma_cm_id *id;
3042         struct sockaddr *sa;
3043         int ret;
3044
3045         sa = (struct sockaddr *)&np->np_sockaddr;
3046         isert_dbg("ksockaddr: %p, sa: %p\n", &np->np_sockaddr, sa);
3047
3048         id = rdma_create_id(isert_cma_handler, isert_np,
3049                             RDMA_PS_TCP, IB_QPT_RC);
3050         if (IS_ERR(id)) {
3051                 isert_err("rdma_create_id() failed: %ld\n", PTR_ERR(id));
3052                 ret = PTR_ERR(id);
3053                 goto out;
3054         }
3055         isert_dbg("id %p context %p\n", id, id->context);
3056
3057         ret = rdma_bind_addr(id, sa);
3058         if (ret) {
3059                 isert_err("rdma_bind_addr() failed: %d\n", ret);
3060                 goto out_id;
3061         }
3062
3063         ret = rdma_listen(id, 0);
3064         if (ret) {
3065                 isert_err("rdma_listen() failed: %d\n", ret);
3066                 goto out_id;
3067         }
3068
3069         return id;
3070 out_id:
3071         rdma_destroy_id(id);
3072 out:
3073         return ERR_PTR(ret);
3074 }
3075
3076 static int
3077 isert_setup_np(struct iscsi_np *np,
3078                struct __kernel_sockaddr_storage *ksockaddr)
3079 {
3080         struct isert_np *isert_np;
3081         struct rdma_cm_id *isert_lid;
3082         int ret;
3083
3084         isert_np = kzalloc(sizeof(struct isert_np), GFP_KERNEL);
3085         if (!isert_np) {
3086                 isert_err("Unable to allocate struct isert_np\n");
3087                 return -ENOMEM;
3088         }
3089         sema_init(&isert_np->np_sem, 0);
3090         mutex_init(&isert_np->np_accept_mutex);
3091         INIT_LIST_HEAD(&isert_np->np_accept_list);
3092         init_completion(&isert_np->np_login_comp);
3093         isert_np->np = np;
3094
3095         /*
3096          * Setup the np->np_sockaddr from the passed sockaddr setup
3097          * in iscsi_target_configfs.c code..
3098          */
3099         memcpy(&np->np_sockaddr, ksockaddr,
3100                sizeof(struct __kernel_sockaddr_storage));
3101
3102         isert_lid = isert_setup_id(isert_np);
3103         if (IS_ERR(isert_lid)) {
3104                 ret = PTR_ERR(isert_lid);
3105                 goto out;
3106         }
3107
3108         isert_np->np_cm_id = isert_lid;
3109         np->np_context = isert_np;
3110
3111         return 0;
3112
3113 out:
3114         kfree(isert_np);
3115
3116         return ret;
3117 }
3118
3119 static int
3120 isert_rdma_accept(struct isert_conn *isert_conn)
3121 {
3122         struct rdma_cm_id *cm_id = isert_conn->cm_id;
3123         struct rdma_conn_param cp;
3124         int ret;
3125
3126         memset(&cp, 0, sizeof(struct rdma_conn_param));
3127         cp.initiator_depth = isert_conn->initiator_depth;
3128         cp.retry_count = 7;
3129         cp.rnr_retry_count = 7;
3130
3131         ret = rdma_accept(cm_id, &cp);
3132         if (ret) {
3133                 isert_err("rdma_accept() failed with: %d\n", ret);
3134                 return ret;
3135         }
3136
3137         return 0;
3138 }
3139
3140 static int
3141 isert_get_login_rx(struct iscsi_conn *conn, struct iscsi_login *login)
3142 {
3143         struct isert_conn *isert_conn = conn->context;
3144         int ret;
3145
3146         isert_info("before login_req comp conn: %p\n", isert_conn);
3147         ret = wait_for_completion_interruptible(&isert_conn->login_req_comp);
3148         if (ret) {
3149                 isert_err("isert_conn %p interrupted before got login req\n",
3150                           isert_conn);
3151                 return ret;
3152         }
3153         reinit_completion(&isert_conn->login_req_comp);
3154
3155         /*
3156          * For login requests after the first PDU, isert_rx_login_req() will
3157          * kick schedule_delayed_work(&conn->login_work) as the packet is
3158          * received, which turns this callback from iscsi_target_do_login_rx()
3159          * into a NOP.
3160          */
3161         if (!login->first_request)
3162                 return 0;
3163
3164         isert_rx_login_req(isert_conn);
3165
3166         isert_info("before login_comp conn: %p\n", conn);
3167         ret = wait_for_completion_interruptible(&isert_conn->login_comp);
3168         if (ret)
3169                 return ret;
3170
3171         isert_info("processing login->req: %p\n", login->req);
3172
3173         return 0;
3174 }
3175
3176 static void
3177 isert_set_conn_info(struct iscsi_np *np, struct iscsi_conn *conn,
3178                     struct isert_conn *isert_conn)
3179 {
3180         struct rdma_cm_id *cm_id = isert_conn->cm_id;
3181         struct rdma_route *cm_route = &cm_id->route;
3182         struct sockaddr_in *sock_in;
3183         struct sockaddr_in6 *sock_in6;
3184
3185         conn->login_family = np->np_sockaddr.ss_family;
3186
3187         if (np->np_sockaddr.ss_family == AF_INET6) {
3188                 sock_in6 = (struct sockaddr_in6 *)&cm_route->addr.dst_addr;
3189                 snprintf(conn->login_ip, sizeof(conn->login_ip), "%pI6c",
3190                          &sock_in6->sin6_addr.in6_u);
3191                 conn->login_port = ntohs(sock_in6->sin6_port);
3192
3193                 sock_in6 = (struct sockaddr_in6 *)&cm_route->addr.src_addr;
3194                 snprintf(conn->local_ip, sizeof(conn->local_ip), "%pI6c",
3195                          &sock_in6->sin6_addr.in6_u);
3196                 conn->local_port = ntohs(sock_in6->sin6_port);
3197         } else {
3198                 sock_in = (struct sockaddr_in *)&cm_route->addr.dst_addr;
3199                 sprintf(conn->login_ip, "%pI4",
3200                         &sock_in->sin_addr.s_addr);
3201                 conn->login_port = ntohs(sock_in->sin_port);
3202
3203                 sock_in = (struct sockaddr_in *)&cm_route->addr.src_addr;
3204                 sprintf(conn->local_ip, "%pI4",
3205                         &sock_in->sin_addr.s_addr);
3206                 conn->local_port = ntohs(sock_in->sin_port);
3207         }
3208 }
3209
3210 static int
3211 isert_accept_np(struct iscsi_np *np, struct iscsi_conn *conn)
3212 {
3213         struct isert_np *isert_np = np->np_context;
3214         struct isert_conn *isert_conn;
3215         int ret;
3216
3217 accept_wait:
3218         ret = down_interruptible(&isert_np->np_sem);
3219         if (ret)
3220                 return -ENODEV;
3221
3222         spin_lock_bh(&np->np_thread_lock);
3223         if (np->np_thread_state >= ISCSI_NP_THREAD_RESET) {
3224                 spin_unlock_bh(&np->np_thread_lock);
3225                 isert_dbg("np_thread_state %d\n",
3226                          np->np_thread_state);
3227                 /**
3228                  * No point in stalling here when np_thread
3229                  * is in state RESET/SHUTDOWN/EXIT - bail
3230                  **/
3231                 return -ENODEV;
3232         }
3233         spin_unlock_bh(&np->np_thread_lock);
3234
3235         mutex_lock(&isert_np->np_accept_mutex);
3236         if (list_empty(&isert_np->np_accept_list)) {
3237                 mutex_unlock(&isert_np->np_accept_mutex);
3238                 goto accept_wait;
3239         }
3240         isert_conn = list_first_entry(&isert_np->np_accept_list,
3241                         struct isert_conn, accept_node);
3242         list_del_init(&isert_conn->accept_node);
3243         mutex_unlock(&isert_np->np_accept_mutex);
3244
3245         conn->context = isert_conn;
3246         isert_conn->conn = conn;
3247
3248         isert_set_conn_info(np, conn, isert_conn);
3249
3250         isert_dbg("Processing isert_conn: %p\n", isert_conn);
3251
3252         return 0;
3253 }
3254
3255 static void
3256 isert_free_np(struct iscsi_np *np)
3257 {
3258         struct isert_np *isert_np = np->np_context;
3259         struct isert_conn *isert_conn, *n;
3260
3261         if (isert_np->np_cm_id)
3262                 rdma_destroy_id(isert_np->np_cm_id);
3263
3264         /*
3265          * FIXME: At this point we don't have a good way to insure
3266          * that at this point we don't have hanging connections that
3267          * completed RDMA establishment but didn't start iscsi login
3268          * process. So work-around this by cleaning up what ever piled
3269          * up in np_accept_list.
3270          */
3271         mutex_lock(&isert_np->np_accept_mutex);
3272         if (!list_empty(&isert_np->np_accept_list)) {
3273                 isert_info("Still have isert connections, cleaning up...\n");
3274                 list_for_each_entry_safe(isert_conn, n,
3275                                          &isert_np->np_accept_list,
3276                                          accept_node) {
3277                         isert_info("cleaning isert_conn %p state (%d)\n",
3278                                    isert_conn, isert_conn->state);
3279                         isert_connect_release(isert_conn);
3280                 }
3281         }
3282         mutex_unlock(&isert_np->np_accept_mutex);
3283
3284         np->np_context = NULL;
3285         kfree(isert_np);
3286 }
3287
3288 static void isert_release_work(struct work_struct *work)
3289 {
3290         struct isert_conn *isert_conn = container_of(work,
3291                                                      struct isert_conn,
3292                                                      release_work);
3293
3294         isert_info("Starting release conn %p\n", isert_conn);
3295
3296         wait_for_completion(&isert_conn->wait);
3297
3298         mutex_lock(&isert_conn->mutex);
3299         isert_conn->state = ISER_CONN_DOWN;
3300         mutex_unlock(&isert_conn->mutex);
3301
3302         isert_info("Destroying conn %p\n", isert_conn);
3303         isert_put_conn(isert_conn);
3304 }
3305
3306 static void
3307 isert_wait4logout(struct isert_conn *isert_conn)
3308 {
3309         struct iscsi_conn *conn = isert_conn->conn;
3310
3311         isert_info("conn %p\n", isert_conn);
3312
3313         if (isert_conn->logout_posted) {
3314                 isert_info("conn %p wait for conn_logout_comp\n", isert_conn);
3315                 wait_for_completion_timeout(&conn->conn_logout_comp,
3316                                             SECONDS_FOR_LOGOUT_COMP * HZ);
3317         }
3318 }
3319
3320 static void
3321 isert_wait4cmds(struct iscsi_conn *conn)
3322 {
3323         isert_info("iscsi_conn %p\n", conn);
3324
3325         if (conn->sess) {
3326                 target_sess_cmd_list_set_waiting(conn->sess->se_sess);
3327                 target_wait_for_sess_cmds(conn->sess->se_sess);
3328         }
3329 }
3330
3331 static void
3332 isert_wait4flush(struct isert_conn *isert_conn)
3333 {
3334         struct ib_recv_wr *bad_wr;
3335
3336         isert_info("conn %p\n", isert_conn);
3337
3338         init_completion(&isert_conn->wait_comp_err);
3339         isert_conn->beacon.wr_id = ISER_BEACON_WRID;
3340         /* post an indication that all flush errors were consumed */
3341         if (ib_post_recv(isert_conn->qp, &isert_conn->beacon, &bad_wr)) {
3342                 isert_err("conn %p failed to post beacon", isert_conn);
3343                 return;
3344         }
3345
3346         wait_for_completion(&isert_conn->wait_comp_err);
3347 }
3348
3349 static void isert_wait_conn(struct iscsi_conn *conn)
3350 {
3351         struct isert_conn *isert_conn = conn->context;
3352
3353         isert_info("Starting conn %p\n", isert_conn);
3354
3355         mutex_lock(&isert_conn->mutex);
3356         /*
3357          * Only wait for wait_comp_err if the isert_conn made it
3358          * into full feature phase..
3359          */
3360         if (isert_conn->state == ISER_CONN_INIT) {
3361                 mutex_unlock(&isert_conn->mutex);
3362                 return;
3363         }
3364         isert_conn_terminate(isert_conn);
3365         mutex_unlock(&isert_conn->mutex);
3366
3367         isert_wait4cmds(conn);
3368         isert_wait4flush(isert_conn);
3369         isert_wait4logout(isert_conn);
3370
3371         INIT_WORK(&isert_conn->release_work, isert_release_work);
3372         queue_work(isert_release_wq, &isert_conn->release_work);
3373 }
3374
3375 static void isert_free_conn(struct iscsi_conn *conn)
3376 {
3377         struct isert_conn *isert_conn = conn->context;
3378
3379         isert_put_conn(isert_conn);
3380 }
3381
3382 static struct iscsit_transport iser_target_transport = {
3383         .name                   = "IB/iSER",
3384         .transport_type         = ISCSI_INFINIBAND,
3385         .priv_size              = sizeof(struct isert_cmd),
3386         .owner                  = THIS_MODULE,
3387         .iscsit_setup_np        = isert_setup_np,
3388         .iscsit_accept_np       = isert_accept_np,
3389         .iscsit_free_np         = isert_free_np,
3390         .iscsit_wait_conn       = isert_wait_conn,
3391         .iscsit_free_conn       = isert_free_conn,
3392         .iscsit_get_login_rx    = isert_get_login_rx,
3393         .iscsit_put_login_tx    = isert_put_login_tx,
3394         .iscsit_immediate_queue = isert_immediate_queue,
3395         .iscsit_response_queue  = isert_response_queue,
3396         .iscsit_get_dataout     = isert_get_dataout,
3397         .iscsit_queue_data_in   = isert_put_datain,
3398         .iscsit_queue_status    = isert_put_response,
3399         .iscsit_aborted_task    = isert_aborted_task,
3400         .iscsit_get_sup_prot_ops = isert_get_sup_prot_ops,
3401 };
3402
3403 static int __init isert_init(void)
3404 {
3405         int ret;
3406
3407         isert_comp_wq = alloc_workqueue("isert_comp_wq",
3408                                         WQ_UNBOUND | WQ_HIGHPRI, 0);
3409         if (!isert_comp_wq) {
3410                 isert_err("Unable to allocate isert_comp_wq\n");
3411                 ret = -ENOMEM;
3412                 return -ENOMEM;
3413         }
3414
3415         isert_release_wq = alloc_workqueue("isert_release_wq", WQ_UNBOUND,
3416                                         WQ_UNBOUND_MAX_ACTIVE);
3417         if (!isert_release_wq) {
3418                 isert_err("Unable to allocate isert_release_wq\n");
3419                 ret = -ENOMEM;
3420                 goto destroy_comp_wq;
3421         }
3422
3423         iscsit_register_transport(&iser_target_transport);
3424         isert_info("iSER_TARGET[0] - Loaded iser_target_transport\n");
3425
3426         return 0;
3427
3428 destroy_comp_wq:
3429         destroy_workqueue(isert_comp_wq);
3430
3431         return ret;
3432 }
3433
3434 static void __exit isert_exit(void)
3435 {
3436         flush_scheduled_work();
3437         destroy_workqueue(isert_release_wq);
3438         destroy_workqueue(isert_comp_wq);
3439         iscsit_unregister_transport(&iser_target_transport);
3440         isert_info("iSER_TARGET[0] - Released iser_target_transport\n");
3441 }
3442
3443 MODULE_DESCRIPTION("iSER-Target for mainline target infrastructure");
3444 MODULE_VERSION("1.0");
3445 MODULE_AUTHOR("nab@Linux-iSCSI.org");
3446 MODULE_LICENSE("GPL");
3447
3448 module_init(isert_init);
3449 module_exit(isert_exit);