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 = (struct 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->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 int
 111 isert_conn_setup_qp(struct isert_conn *isert_conn, struct rdma_cm_id *cma_id)
 112 {
 113         struct isert_device *device = isert_conn->conn_device;
 114         struct ib_qp_init_attr attr;
 115         struct isert_comp *comp;
 116         int ret, i, min = 0;
 117
 118         mutex_lock(&device_list_mutex);
 119         for (i = 0; i < device->comps_used; i++)
 120                 if (device->comps[i].active_qps <
 121                     device->comps[min].active_qps)
 122                         min = i;
 123         comp = &device->comps[min];
 124         comp->active_qps++;
 125         isert_info("conn %p, using comp %p min_index: %d\n",
 126                    isert_conn, comp, min);
 127         mutex_unlock(&device_list_mutex);
 128
 129         memset(&attr, 0, sizeof(struct ib_qp_init_attr));
 130         attr.event_handler = isert_qp_event_callback;
 131         attr.qp_context = isert_conn;
 132         attr.send_cq = comp->cq;
 133         attr.recv_cq = comp->cq;
 134         attr.cap.max_send_wr = ISERT_QP_MAX_REQ_DTOS;
 135         attr.cap.max_recv_wr = ISERT_QP_MAX_RECV_DTOS + 1;
 136         /*
 137          * FIXME: Use devattr.max_sge - 2 for max_send_sge as
 138          * work-around for RDMA_READs with ConnectX-2.
 139          *
 140          * Also, still make sure to have at least two SGEs for
 141          * outgoing control PDU responses.
 142          */
 143         attr.cap.max_send_sge = max(2, device->dev_attr.max_sge - 2);
 144         isert_conn->max_sge = attr.cap.max_send_sge;
 145
 146         attr.cap.max_recv_sge = 1;
 147         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 148         attr.qp_type = IB_QPT_RC;
 149         if (device->pi_capable)
 150                 attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
 151
 152         ret = rdma_create_qp(cma_id, isert_conn->conn_pd, &attr);
 153         if (ret) {
 154                 isert_err("rdma_create_qp failed for cma_id %d\n", ret);
 155                 goto err;
 156         }
 157         isert_conn->conn_qp = cma_id->qp;
 158
 159         return 0;
 160 err:
 161         mutex_lock(&device_list_mutex);
 162         comp->active_qps--;
 163         mutex_unlock(&device_list_mutex);
 164
 165         return ret;
 166 }
 167
 168 static void
 169 isert_cq_event_callback(struct ib_event *e, void *context)
 170 {
 171         isert_dbg("event: %d\n", e->event);
 172 }
 173
 174 static int
 175 isert_alloc_rx_descriptors(struct isert_conn *isert_conn)
 176 {
 177         struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
 178         struct iser_rx_desc *rx_desc;
 179         struct ib_sge *rx_sg;
 180         u64 dma_addr;
 181         int i, j;
 182
 183         isert_conn->conn_rx_descs = kzalloc(ISERT_QP_MAX_RECV_DTOS *
 184                                 sizeof(struct iser_rx_desc), GFP_KERNEL);
 185         if (!isert_conn->conn_rx_descs)
 186                 goto fail;
 187
 188         rx_desc = isert_conn->conn_rx_descs;
 189
 190         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 191                 dma_addr = ib_dma_map_single(ib_dev, (void *)rx_desc,
 192                                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 193                 if (ib_dma_mapping_error(ib_dev, dma_addr))
 194                         goto dma_map_fail;
 195
 196                 rx_desc->dma_addr = dma_addr;
 197
 198                 rx_sg = &rx_desc->rx_sg;
 199                 rx_sg->addr = rx_desc->dma_addr;
 200                 rx_sg->length = ISER_RX_PAYLOAD_SIZE;
 201                 rx_sg->lkey = isert_conn->conn_mr->lkey;
 202         }
 203
 204         isert_conn->conn_rx_desc_head = 0;
 205
 206         return 0;
 207
 208 dma_map_fail:
 209         rx_desc = isert_conn->conn_rx_descs;
 210         for (j = 0; j < i; j++, rx_desc++) {
 211                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 212                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 213         }
 214         kfree(isert_conn->conn_rx_descs);
 215         isert_conn->conn_rx_descs = NULL;
 216 fail:
 217         isert_err("conn %p failed to allocate rx descriptors\n", isert_conn);
 218
 219         return -ENOMEM;
 220 }
 221
 222 static void
 223 isert_free_rx_descriptors(struct isert_conn *isert_conn)
 224 {
 225         struct ib_device *ib_dev = isert_conn->conn_device->ib_device;
 226         struct iser_rx_desc *rx_desc;
 227         int i;
 228
 229         if (!isert_conn->conn_rx_descs)
 230                 return;
 231
 232         rx_desc = isert_conn->conn_rx_descs;
 233         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 234                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 235                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 236         }
 237
 238         kfree(isert_conn->conn_rx_descs);
 239         isert_conn->conn_rx_descs = NULL;
 240 }
 241
 242 static void isert_cq_work(struct work_struct *);
 243 static void isert_cq_callback(struct ib_cq *, void *);
 244
 245 static int
 246 isert_create_device_ib_res(struct isert_device *device)
 247 {
 248         struct ib_device *ib_dev = device->ib_device;
 249         struct ib_device_attr *dev_attr;
 250         int ret = 0, i;
 251         int max_cqe;
 252
 253         dev_attr = &device->dev_attr;
 254         ret = isert_query_device(ib_dev, dev_attr);
 255         if (ret)
 256                 return ret;
 257
 258         max_cqe = min(ISER_MAX_CQ_LEN, dev_attr->max_cqe);
 259
 260         /* asign function handlers */
 261         if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS &&
 262             dev_attr->device_cap_flags & IB_DEVICE_SIGNATURE_HANDOVER) {
 263                 device->use_fastreg = 1;
 264                 device->reg_rdma_mem = isert_reg_rdma;
 265                 device->unreg_rdma_mem = isert_unreg_rdma;
 266         } else {
 267                 device->use_fastreg = 0;
 268                 device->reg_rdma_mem = isert_map_rdma;
 269                 device->unreg_rdma_mem = isert_unmap_cmd;
 270         }
 271
 272         /* Check signature cap */
 273         device->pi_capable = dev_attr->device_cap_flags &
 274                              IB_DEVICE_SIGNATURE_HANDOVER ? true : false;
 275
 276         device->comps_used = min(ISERT_MAX_CQ, min_t(int, num_online_cpus(),
 277                                         device->ib_device->num_comp_vectors));
 278         isert_info("Using %d CQs, %s supports %d vectors support "
 279                    "Fast registration %d pi_capable %d\n",
 280                    device->comps_used, device->ib_device->name,
 281                    device->ib_device->num_comp_vectors, device->use_fastreg,
 282                    device->pi_capable);
 283
 284         device->comps = kcalloc(device->comps_used, sizeof(struct isert_comp),
 285                                 GFP_KERNEL);
 286         if (!device->comps) {
 287                 isert_err("Unable to allocate completion contexts\n");
 288                 return -ENOMEM;
 289         }
 290
 291         for (i = 0; i < device->comps_used; i++) {
 292                 struct isert_comp *comp = &device->comps[i];
 293
 294                 comp->device = device;
 295                 INIT_WORK(&comp->work, isert_cq_work);
 296                 comp->cq = ib_create_cq(device->ib_device,
 297                                         isert_cq_callback,
 298                                         isert_cq_event_callback,
 299                                         (void *)comp,
 300                                         max_cqe, i);
 301                 if (IS_ERR(comp->cq)) {
 302                         ret = PTR_ERR(comp->cq);
 303                         comp->cq = NULL;
 304                         goto out_cq;
 305                 }
 306
 307                 ret = ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP);
 308                 if (ret)
 309                         goto out_cq;
 310         }
 311
 312         return 0;
 313
 314 out_cq:
 315         for (i = 0; i < device->comps_used; i++) {
 316                 struct isert_comp *comp = &device->comps[i];
 317
 318                 if (comp->cq) {
 319                         cancel_work_sync(&comp->work);
 320                         ib_destroy_cq(comp->cq);
 321                 }
 322         }
 323         kfree(device->comps);
 324
 325         return ret;
 326 }
 327
 328 static void
 329 isert_free_device_ib_res(struct isert_device *device)
 330 {
 331         int i;
 332
 333         isert_info("device %p\n", device);
 334
 335         for (i = 0; i < device->comps_used; i++) {
 336                 struct isert_comp *comp = &device->comps[i];
 337
 338                 cancel_work_sync(&comp->work);
 339                 ib_destroy_cq(comp->cq);
 340                 comp->cq = NULL;
 341         }
 342         kfree(device->comps);
 343 }
 344
 345 static void
 346 isert_device_try_release(struct isert_device *device)
 347 {
 348         mutex_lock(&device_list_mutex);
 349         device->refcount--;
 350         isert_info("device %p refcount %d\n", device, device->refcount);
 351         if (!device->refcount) {
 352                 isert_free_device_ib_res(device);
 353                 list_del(&device->dev_node);
 354                 kfree(device);
 355         }
 356         mutex_unlock(&device_list_mutex);
 357 }
 358
 359 static struct isert_device *
 360 isert_device_find_by_ib_dev(struct rdma_cm_id *cma_id)
 361 {
 362         struct isert_device *device;
 363         int ret;
 364
 365         mutex_lock(&device_list_mutex);
 366         list_for_each_entry(device, &device_list, dev_node) {
 367                 if (device->ib_device->node_guid == cma_id->device->node_guid) {
 368                         device->refcount++;
 369                         isert_info("Found iser device %p refcount %d\n",
 370                                    device, device->refcount);
 371                         mutex_unlock(&device_list_mutex);
 372                         return device;
 373                 }
 374         }
 375
 376         device = kzalloc(sizeof(struct isert_device), GFP_KERNEL);
 377         if (!device) {
 378                 mutex_unlock(&device_list_mutex);
 379                 return ERR_PTR(-ENOMEM);
 380         }
 381
 382         INIT_LIST_HEAD(&device->dev_node);
 383
 384         device->ib_device = cma_id->device;
 385         ret = isert_create_device_ib_res(device);
 386         if (ret) {
 387                 kfree(device);
 388                 mutex_unlock(&device_list_mutex);
 389                 return ERR_PTR(ret);
 390         }
 391
 392         device->refcount++;
 393         list_add_tail(&device->dev_node, &device_list);
 394         isert_info("Created a new iser device %p refcount %d\n",
 395                    device, device->refcount);
 396         mutex_unlock(&device_list_mutex);
 397
 398         return device;
 399 }
 400
 401 static void
 402 isert_conn_free_fastreg_pool(struct isert_conn *isert_conn)
 403 {
 404         struct fast_reg_descriptor *fr_desc, *tmp;
 405         int i = 0;
 406
 407         if (list_empty(&isert_conn->conn_fr_pool))
 408                 return;
 409
 410         isert_info("Freeing conn %p fastreg pool", isert_conn);
 411
 412         list_for_each_entry_safe(fr_desc, tmp,
 413                                  &isert_conn->conn_fr_pool, list) {
 414                 list_del(&fr_desc->list);
 415                 ib_free_fast_reg_page_list(fr_desc->data_frpl);
 416                 ib_dereg_mr(fr_desc->data_mr);
 417                 if (fr_desc->pi_ctx) {
 418                         ib_free_fast_reg_page_list(fr_desc->pi_ctx->prot_frpl);
 419                         ib_dereg_mr(fr_desc->pi_ctx->prot_mr);
 420                         ib_destroy_mr(fr_desc->pi_ctx->sig_mr);
 421                         kfree(fr_desc->pi_ctx);
 422                 }
 423                 kfree(fr_desc);
 424                 ++i;
 425         }
 426
 427         if (i < isert_conn->conn_fr_pool_size)
 428                 isert_warn("Pool still has %d regions registered\n",
 429                         isert_conn->conn_fr_pool_size - i);
 430 }
 431
 432 static int
 433 isert_create_pi_ctx(struct fast_reg_descriptor *desc,
 434                     struct ib_device *device,
 435                     struct ib_pd *pd)
 436 {
 437         struct ib_mr_init_attr mr_init_attr;
 438         struct pi_context *pi_ctx;
 439         int ret;
 440
 441         pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
 442         if (!pi_ctx) {
 443                 isert_err("Failed to allocate pi context\n");
 444                 return -ENOMEM;
 445         }
 446
 447         pi_ctx->prot_frpl = ib_alloc_fast_reg_page_list(device,
 448                                             ISCSI_ISER_SG_TABLESIZE);
 449         if (IS_ERR(pi_ctx->prot_frpl)) {
 450                 isert_err("Failed to allocate prot frpl err=%ld\n",
 451                           PTR_ERR(pi_ctx->prot_frpl));
 452                 ret = PTR_ERR(pi_ctx->prot_frpl);
 453                 goto err_pi_ctx;
 454         }
 455
 456         pi_ctx->prot_mr = ib_alloc_fast_reg_mr(pd, ISCSI_ISER_SG_TABLESIZE);
 457         if (IS_ERR(pi_ctx->prot_mr)) {
 458                 isert_err("Failed to allocate prot frmr err=%ld\n",
 459                           PTR_ERR(pi_ctx->prot_mr));
 460                 ret = PTR_ERR(pi_ctx->prot_mr);
 461                 goto err_prot_frpl;
 462         }
 463         desc->ind |= ISERT_PROT_KEY_VALID;
 464
 465         memset(&mr_init_attr, 0, sizeof(mr_init_attr));
 466         mr_init_attr.max_reg_descriptors = 2;
 467         mr_init_attr.flags |= IB_MR_SIGNATURE_EN;
 468         pi_ctx->sig_mr = ib_create_mr(pd, &mr_init_attr);
 469         if (IS_ERR(pi_ctx->sig_mr)) {
 470                 isert_err("Failed to allocate signature enabled mr err=%ld\n",
 471                           PTR_ERR(pi_ctx->sig_mr));
 472                 ret = PTR_ERR(pi_ctx->sig_mr);
 473                 goto err_prot_mr;
 474         }
 475
 476         desc->pi_ctx = pi_ctx;
 477         desc->ind |= ISERT_SIG_KEY_VALID;
 478         desc->ind &= ~ISERT_PROTECTED;
 479
 480         return 0;
 481
 482 err_prot_mr:
 483         ib_dereg_mr(desc->pi_ctx->prot_mr);
 484 err_prot_frpl:
 485         ib_free_fast_reg_page_list(desc->pi_ctx->prot_frpl);
 486 err_pi_ctx:
 487         kfree(desc->pi_ctx);
 488
 489         return ret;
 490 }
 491
 492 static int
 493 isert_create_fr_desc(struct ib_device *ib_device, struct ib_pd *pd,
 494                      struct fast_reg_descriptor *fr_desc)
 495 {
 496         int ret;
 497
 498         fr_desc->data_frpl = ib_alloc_fast_reg_page_list(ib_device,
 499                                                          ISCSI_ISER_SG_TABLESIZE);
 500         if (IS_ERR(fr_desc->data_frpl)) {
 501                 isert_err("Failed to allocate data frpl err=%ld\n",
 502                           PTR_ERR(fr_desc->data_frpl));
 503                 return PTR_ERR(fr_desc->data_frpl);
 504         }
 505
 506         fr_desc->data_mr = ib_alloc_fast_reg_mr(pd, ISCSI_ISER_SG_TABLESIZE);
 507         if (IS_ERR(fr_desc->data_mr)) {
 508                 isert_err("Failed to allocate data frmr err=%ld\n",
 509                           PTR_ERR(fr_desc->data_mr));
 510                 ret = PTR_ERR(fr_desc->data_mr);
 511                 goto err_data_frpl;
 512         }
 513         fr_desc->ind |= ISERT_DATA_KEY_VALID;
 514
 515         isert_dbg("Created fr_desc %p\n", fr_desc);
 516
 517         return 0;
 518
 519 err_data_frpl:
 520         ib_free_fast_reg_page_list(fr_desc->data_frpl);
 521
 522         return ret;
 523 }
 524
 525 static int
 526 isert_conn_create_fastreg_pool(struct isert_conn *isert_conn)
 527 {
 528         struct fast_reg_descriptor *fr_desc;
 529         struct isert_device *device = isert_conn->conn_device;
 530         struct se_session *se_sess = isert_conn->conn->sess->se_sess;
 531         struct se_node_acl *se_nacl = se_sess->se_node_acl;
 532         int i, ret, tag_num;
 533         /*
 534          * Setup the number of FRMRs based upon the number of tags
 535          * available to session in iscsi_target_locate_portal().
 536          */
 537         tag_num = max_t(u32, ISCSIT_MIN_TAGS, se_nacl->queue_depth);
 538         tag_num = (tag_num * 2) + ISCSIT_EXTRA_TAGS;
 539
 540         isert_conn->conn_fr_pool_size = 0;
 541         for (i = 0; i < tag_num; i++) {
 542                 fr_desc = kzalloc(sizeof(*fr_desc), GFP_KERNEL);
 543                 if (!fr_desc) {
 544                         isert_err("Failed to allocate fast_reg descriptor\n");
 545                         ret = -ENOMEM;
 546                         goto err;
 547                 }
 548
 549                 ret = isert_create_fr_desc(device->ib_device,
 550                                            isert_conn->conn_pd, fr_desc);
 551                 if (ret) {
 552                         isert_err("Failed to create fastreg descriptor err=%d\n",
 553                                ret);
 554                         kfree(fr_desc);
 555                         goto err;
 556                 }
 557
 558                 list_add_tail(&fr_desc->list, &isert_conn->conn_fr_pool);
 559                 isert_conn->conn_fr_pool_size++;
 560         }
 561
 562         isert_dbg("Creating conn %p fastreg pool size=%d",
 563                  isert_conn, isert_conn->conn_fr_pool_size);
 564
 565         return 0;
 566
 567 err:
 568         isert_conn_free_fastreg_pool(isert_conn);
 569         return ret;
 570 }
 571
 572 static int
 573 isert_connect_request(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 574 {
 575         struct isert_np *isert_np = cma_id->context;
 576         struct iscsi_np *np = isert_np->np;
 577         struct isert_conn *isert_conn;
 578         struct isert_device *device;
 579         struct ib_device *ib_dev = cma_id->device;
 580         int ret = 0;
 581
 582         spin_lock_bh(&np->np_thread_lock);
 583         if (!np->enabled) {
 584                 spin_unlock_bh(&np->np_thread_lock);
 585                 isert_dbg("iscsi_np is not enabled, reject connect request\n");
 586                 return rdma_reject(cma_id, NULL, 0);
 587         }
 588         spin_unlock_bh(&np->np_thread_lock);
 589
 590         isert_dbg("cma_id: %p, portal: %p\n",
 591                  cma_id, cma_id->context);
 592
 593         isert_conn = kzalloc(sizeof(struct isert_conn), GFP_KERNEL);
 594         if (!isert_conn) {
 595                 isert_err("Unable to allocate isert_conn\n");
 596                 return -ENOMEM;
 597         }
 598         isert_conn->state = ISER_CONN_INIT;
 599         INIT_LIST_HEAD(&isert_conn->conn_accept_node);
 600         init_completion(&isert_conn->conn_login_comp);
 601         init_completion(&isert_conn->login_req_comp);
 602         init_completion(&isert_conn->conn_wait);
 603         kref_init(&isert_conn->conn_kref);
 604         mutex_init(&isert_conn->conn_mutex);
 605         spin_lock_init(&isert_conn->conn_lock);
 606         INIT_LIST_HEAD(&isert_conn->conn_fr_pool);
 607
 608         isert_conn->conn_cm_id = cma_id;
 609
 610         isert_conn->login_buf = kzalloc(ISCSI_DEF_MAX_RECV_SEG_LEN +
 611                                         ISER_RX_LOGIN_SIZE, GFP_KERNEL);
 612         if (!isert_conn->login_buf) {
 613                 isert_err("Unable to allocate isert_conn->login_buf\n");
 614                 ret = -ENOMEM;
 615                 goto out;
 616         }
 617
 618         isert_conn->login_req_buf = isert_conn->login_buf;
 619         isert_conn->login_rsp_buf = isert_conn->login_buf +
 620                                     ISCSI_DEF_MAX_RECV_SEG_LEN;
 621         isert_dbg("Set login_buf: %p login_req_buf: %p login_rsp_buf: %p\n",
 622                  isert_conn->login_buf, isert_conn->login_req_buf,
 623                  isert_conn->login_rsp_buf);
 624
 625         isert_conn->login_req_dma = ib_dma_map_single(ib_dev,
 626                                 (void *)isert_conn->login_req_buf,
 627                                 ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_FROM_DEVICE);
 628
 629         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_req_dma);
 630         if (ret) {
 631                 isert_err("ib_dma_mapping_error failed for login_req_dma: %d\n",
 632                        ret);
 633                 isert_conn->login_req_dma = 0;
 634                 goto out_login_buf;
 635         }
 636
 637         isert_conn->login_rsp_dma = ib_dma_map_single(ib_dev,
 638                                         (void *)isert_conn->login_rsp_buf,
 639                                         ISER_RX_LOGIN_SIZE, DMA_TO_DEVICE);
 640
 641         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_rsp_dma);
 642         if (ret) {
 643                 isert_err("ib_dma_mapping_error failed for login_rsp_dma: %d\n",
 644                        ret);
 645                 isert_conn->login_rsp_dma = 0;
 646                 goto out_req_dma_map;
 647         }
 648
 649         device = isert_device_find_by_ib_dev(cma_id);
 650         if (IS_ERR(device)) {
 651                 ret = PTR_ERR(device);
 652                 goto out_rsp_dma_map;
 653         }
 654
 655         /* Set max inflight RDMA READ requests */
 656         isert_conn->initiator_depth = min_t(u8,
 657                                 event->param.conn.initiator_depth,
 658                                 device->dev_attr.max_qp_init_rd_atom);
 659         isert_dbg("Using initiator_depth: %u\n", isert_conn->initiator_depth);
 660
 661         isert_conn->conn_device = device;
 662         isert_conn->conn_pd = ib_alloc_pd(isert_conn->conn_device->ib_device);
 663         if (IS_ERR(isert_conn->conn_pd)) {
 664                 ret = PTR_ERR(isert_conn->conn_pd);
 665                 isert_err("ib_alloc_pd failed for conn %p: ret=%d\n",
 666                        isert_conn, ret);
 667                 goto out_pd;
 668         }
 669
 670         isert_conn->conn_mr = ib_get_dma_mr(isert_conn->conn_pd,
 671                                            IB_ACCESS_LOCAL_WRITE);
 672         if (IS_ERR(isert_conn->conn_mr)) {
 673                 ret = PTR_ERR(isert_conn->conn_mr);
 674                 isert_err("ib_get_dma_mr failed for conn %p: ret=%d\n",
 675                        isert_conn, ret);
 676                 goto out_mr;
 677         }
 678
 679         ret = isert_conn_setup_qp(isert_conn, cma_id);
 680         if (ret)
 681                 goto out_conn_dev;
 682
 683         ret = isert_rdma_post_recvl(isert_conn);
 684         if (ret)
 685                 goto out_conn_dev;
 686
 687         ret = isert_rdma_accept(isert_conn);
 688         if (ret)
 689                 goto out_conn_dev;
 690
 691         mutex_lock(&isert_np->np_accept_mutex);
 692         list_add_tail(&isert_conn->conn_accept_node, &isert_np->np_accept_list);
 693         mutex_unlock(&isert_np->np_accept_mutex);
 694
 695         isert_info("np %p: Allow accept_np to continue\n", np);
 696         up(&isert_np->np_sem);
 697         return 0;
 698
 699 out_conn_dev:
 700         ib_dereg_mr(isert_conn->conn_mr);
 701 out_mr:
 702         ib_dealloc_pd(isert_conn->conn_pd);
 703 out_pd:
 704         isert_device_try_release(device);
 705 out_rsp_dma_map:
 706         ib_dma_unmap_single(ib_dev, isert_conn->login_rsp_dma,
 707                             ISER_RX_LOGIN_SIZE, DMA_TO_DEVICE);
 708 out_req_dma_map:
 709         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 710                             ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_FROM_DEVICE);
 711 out_login_buf:
 712         kfree(isert_conn->login_buf);
 713 out:
 714         kfree(isert_conn);
 715         rdma_reject(cma_id, NULL, 0);
 716         return ret;
 717 }
 718
 719 static void
 720 isert_connect_release(struct isert_conn *isert_conn)
 721 {
 722         struct isert_device *device = isert_conn->conn_device;
 723         struct ib_device *ib_dev = device->ib_device;
 724
 725         isert_dbg("conn %p\n", isert_conn);
 726
 727         if (device && device->use_fastreg)
 728                 isert_conn_free_fastreg_pool(isert_conn);
 729
 730         isert_free_rx_descriptors(isert_conn);
 731         if (isert_conn->conn_cm_id)
 732                 rdma_destroy_id(isert_conn->conn_cm_id);
 733
 734         if (isert_conn->conn_qp) {
 735                 struct isert_comp *comp = isert_conn->conn_qp->recv_cq->cq_context;
 736
 737                 isert_dbg("dec completion context %p active_qps\n", comp);
 738                 mutex_lock(&device_list_mutex);
 739                 comp->active_qps--;
 740                 mutex_unlock(&device_list_mutex);
 741
 742                 ib_destroy_qp(isert_conn->conn_qp);
 743         }
 744
 745         ib_dereg_mr(isert_conn->conn_mr);
 746         ib_dealloc_pd(isert_conn->conn_pd);
 747
 748         if (isert_conn->login_buf) {
 749                 ib_dma_unmap_single(ib_dev, isert_conn->login_rsp_dma,
 750                                     ISER_RX_LOGIN_SIZE, DMA_TO_DEVICE);
 751                 ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 752                                     ISCSI_DEF_MAX_RECV_SEG_LEN,
 753                                     DMA_FROM_DEVICE);
 754                 kfree(isert_conn->login_buf);
 755         }
 756         kfree(isert_conn);
 757
 758         if (device)
 759                 isert_device_try_release(device);
 760 }
 761
 762 static void
 763 isert_connected_handler(struct rdma_cm_id *cma_id)
 764 {
 765         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 766
 767         isert_info("conn %p\n", isert_conn);
 768
 769         if (!kref_get_unless_zero(&isert_conn->conn_kref)) {
 770                 isert_warn("conn %p connect_release is running\n", isert_conn);
 771                 return;
 772         }
 773
 774         mutex_lock(&isert_conn->conn_mutex);
 775         if (isert_conn->state != ISER_CONN_FULL_FEATURE)
 776                 isert_conn->state = ISER_CONN_UP;
 777         mutex_unlock(&isert_conn->conn_mutex);
 778 }
 779
 780 static void
 781 isert_release_conn_kref(struct kref *kref)
 782 {
 783         struct isert_conn *isert_conn = container_of(kref,
 784                                 struct isert_conn, conn_kref);
 785
 786         isert_info("conn %p final kref %s/%d\n", isert_conn, current->comm,
 787                    current->pid);
 788
 789         isert_connect_release(isert_conn);
 790 }
 791
 792 static void
 793 isert_put_conn(struct isert_conn *isert_conn)
 794 {
 795         kref_put(&isert_conn->conn_kref, isert_release_conn_kref);
 796 }
 797
 798 /**
 799  * isert_conn_terminate() - Initiate connection termination
 800  * @isert_conn: isert connection struct
 801  *
 802  * Notes:
 803  * In case the connection state is FULL_FEATURE, move state
 804  * to TEMINATING and start teardown sequence (rdma_disconnect).
 805  * In case the connection state is UP, complete flush as well.
 806  *
 807  * This routine must be called with conn_mutex held. Thus it is
 808  * safe to call multiple times.
 809  */
 810 static void
 811 isert_conn_terminate(struct isert_conn *isert_conn)
 812 {
 813         int err;
 814
 815         switch (isert_conn->state) {
 816         case ISER_CONN_TERMINATING:
 817                 break;
 818         case ISER_CONN_UP:
 819         case ISER_CONN_FULL_FEATURE: /* FALLTHRU */
 820                 isert_info("Terminating conn %p state %d\n",
 821                            isert_conn, isert_conn->state);
 822                 isert_conn->state = ISER_CONN_TERMINATING;
 823                 err = rdma_disconnect(isert_conn->conn_cm_id);
 824                 if (err)
 825                         isert_warn("Failed rdma_disconnect isert_conn %p\n",
 826                                    isert_conn);
 827                 break;
 828         default:
 829                 isert_warn("conn %p teminating in state %d\n",
 830                            isert_conn, isert_conn->state);
 831         }
 832 }
 833
 834 static int
 835 isert_np_cma_handler(struct isert_np *isert_np,
 836                      enum rdma_cm_event_type event)
 837 {
 838         isert_dbg("isert np %p, handling event %d\n", isert_np, event);
 839
 840         switch (event) {
 841         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 842                 isert_np->np_cm_id = NULL;
 843                 break;
 844         case RDMA_CM_EVENT_ADDR_CHANGE:
 845                 isert_np->np_cm_id = isert_setup_id(isert_np);
 846                 if (IS_ERR(isert_np->np_cm_id)) {
 847                         isert_err("isert np %p setup id failed: %ld\n",
 848                                   isert_np, PTR_ERR(isert_np->np_cm_id));
 849                         isert_np->np_cm_id = NULL;
 850                 }
 851                 break;
 852         default:
 853                 isert_err("isert np %p Unexpected event %d\n",
 854                           isert_np, event);
 855         }
 856
 857         return -1;
 858 }
 859
 860 static int
 861 isert_disconnected_handler(struct rdma_cm_id *cma_id,
 862                            enum rdma_cm_event_type event)
 863 {
 864         struct isert_np *isert_np = cma_id->context;
 865         struct isert_conn *isert_conn;
 866
 867         if (isert_np->np_cm_id == cma_id)
 868                 return isert_np_cma_handler(cma_id->context, event);
 869
 870         isert_conn = cma_id->qp->qp_context;
 871
 872         mutex_lock(&isert_conn->conn_mutex);
 873         isert_conn_terminate(isert_conn);
 874         mutex_unlock(&isert_conn->conn_mutex);
 875
 876         isert_info("conn %p completing conn_wait\n", isert_conn);
 877         complete(&isert_conn->conn_wait);
 878
 879         return 0;
 880 }
 881
 882 static int
 883 isert_connect_error(struct rdma_cm_id *cma_id)
 884 {
 885         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 886
 887         isert_conn->conn_cm_id = NULL;
 888         isert_put_conn(isert_conn);
 889
 890         return -1;
 891 }
 892
 893 static int
 894 isert_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 895 {
 896         int ret = 0;
 897
 898         isert_info("event %d status %d id %p np %p\n", event->event,
 899                    event->status, cma_id, cma_id->context);
 900
 901         switch (event->event) {
 902         case RDMA_CM_EVENT_CONNECT_REQUEST:
 903                 ret = isert_connect_request(cma_id, event);
 904                 if (ret)
 905                         isert_err("failed handle connect request %d\n", ret);
 906                 break;
 907         case RDMA_CM_EVENT_ESTABLISHED:
 908                 isert_connected_handler(cma_id);
 909                 break;
 910         case RDMA_CM_EVENT_ADDR_CHANGE:    /* FALLTHRU */
 911         case RDMA_CM_EVENT_DISCONNECTED:   /* FALLTHRU */
 912         case RDMA_CM_EVENT_DEVICE_REMOVAL: /* FALLTHRU */
 913         case RDMA_CM_EVENT_TIMEWAIT_EXIT:  /* FALLTHRU */
 914                 ret = isert_disconnected_handler(cma_id, event->event);
 915                 break;
 916         case RDMA_CM_EVENT_REJECTED:       /* FALLTHRU */
 917         case RDMA_CM_EVENT_UNREACHABLE:    /* FALLTHRU */
 918         case RDMA_CM_EVENT_CONNECT_ERROR:
 919                 ret = isert_connect_error(cma_id);
 920                 break;
 921         default:
 922                 isert_err("Unhandled RDMA CMA event: %d\n", event->event);
 923                 break;
 924         }
 925
 926         return ret;
 927 }
 928
 929 static int
 930 isert_post_recv(struct isert_conn *isert_conn, u32 count)
 931 {
 932         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 933         int i, ret;
 934         unsigned int rx_head = isert_conn->conn_rx_desc_head;
 935         struct iser_rx_desc *rx_desc;
 936
 937         for (rx_wr = isert_conn->conn_rx_wr, i = 0; i < count; i++, rx_wr++) {
 938                 rx_desc         = &isert_conn->conn_rx_descs[rx_head];
 939                 rx_wr->wr_id    = (uintptr_t)rx_desc;
 940                 rx_wr->sg_list  = &rx_desc->rx_sg;
 941                 rx_wr->num_sge  = 1;
 942                 rx_wr->next     = rx_wr + 1;
 943                 rx_head = (rx_head + 1) & (ISERT_QP_MAX_RECV_DTOS - 1);
 944         }
 945
 946         rx_wr--;
 947         rx_wr->next = NULL; /* mark end of work requests list */
 948
 949         isert_conn->post_recv_buf_count += count;
 950         ret = ib_post_recv(isert_conn->conn_qp, isert_conn->conn_rx_wr,
 951                                 &rx_wr_failed);
 952         if (ret) {
 953                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
 954                 isert_conn->post_recv_buf_count -= count;
 955         } else {
 956                 isert_dbg("Posted %d RX buffers\n", count);
 957                 isert_conn->conn_rx_desc_head = rx_head;
 958         }
 959         return ret;
 960 }
 961
 962 static int
 963 isert_post_send(struct isert_conn *isert_conn, struct iser_tx_desc *tx_desc)
 964 {
 965         struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
 966         struct ib_send_wr send_wr, *send_wr_failed;
 967         int ret;
 968
 969         ib_dma_sync_single_for_device(ib_dev, tx_desc->dma_addr,
 970                                       ISER_HEADERS_LEN, DMA_TO_DEVICE);
 971
 972         send_wr.next    = NULL;
 973         send_wr.wr_id   = (uintptr_t)tx_desc;
 974         send_wr.sg_list = tx_desc->tx_sg;
 975         send_wr.num_sge = tx_desc->num_sge;
 976         send_wr.opcode  = IB_WR_SEND;
 977         send_wr.send_flags = IB_SEND_SIGNALED;
 978
 979         ret = ib_post_send(isert_conn->conn_qp, &send_wr, &send_wr_failed);
 980         if (ret)
 981                 isert_err("ib_post_send() failed, ret: %d\n", ret);
 982
 983         return ret;
 984 }
 985
 986 static void
 987 isert_create_send_desc(struct isert_conn *isert_conn,
 988                        struct isert_cmd *isert_cmd,
 989                        struct iser_tx_desc *tx_desc)
 990 {
 991         struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
 992
 993         ib_dma_sync_single_for_cpu(ib_dev, tx_desc->dma_addr,
 994                                    ISER_HEADERS_LEN, DMA_TO_DEVICE);
 995
 996         memset(&tx_desc->iser_header, 0, sizeof(struct iser_hdr));
 997         tx_desc->iser_header.flags = ISER_VER;
 998
 999         tx_desc->num_sge = 1;
1000         tx_desc->isert_cmd = isert_cmd;
1001
1002         if (tx_desc->tx_sg[0].lkey != isert_conn->conn_mr->lkey) {
1003                 tx_desc->tx_sg[0].lkey = isert_conn->conn_mr->lkey;
1004                 isert_dbg("tx_desc %p lkey mismatch, fixing\n", tx_desc);
1005         }
1006 }
1007
1008 static int
1009 isert_init_tx_hdrs(struct isert_conn *isert_conn,
1010                    struct iser_tx_desc *tx_desc)
1011 {
1012         struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
1013         u64 dma_addr;
1014
1015         dma_addr = ib_dma_map_single(ib_dev, (void *)tx_desc,
1016                         ISER_HEADERS_LEN, DMA_TO_DEVICE);
1017         if (ib_dma_mapping_error(ib_dev, dma_addr)) {
1018                 isert_err("ib_dma_mapping_error() failed\n");
1019                 return -ENOMEM;
1020         }
1021
1022         tx_desc->dma_addr = dma_addr;
1023         tx_desc->tx_sg[0].addr  = tx_desc->dma_addr;
1024         tx_desc->tx_sg[0].length = ISER_HEADERS_LEN;
1025         tx_desc->tx_sg[0].lkey = isert_conn->conn_mr->lkey;
1026
1027         isert_dbg("Setup tx_sg[0].addr: 0x%llx length: %u lkey: 0x%x\n",
1028                   tx_desc->tx_sg[0].addr, tx_desc->tx_sg[0].length,
1029                   tx_desc->tx_sg[0].lkey);
1030
1031         return 0;
1032 }
1033
1034 static void
1035 isert_init_send_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1036                    struct ib_send_wr *send_wr)
1037 {
1038         struct iser_tx_desc *tx_desc = &isert_cmd->tx_desc;
1039
1040         isert_cmd->rdma_wr.iser_ib_op = ISER_IB_SEND;
1041         send_wr->wr_id = (uintptr_t)&isert_cmd->tx_desc;
1042         send_wr->opcode = IB_WR_SEND;
1043         send_wr->sg_list = &tx_desc->tx_sg[0];
1044         send_wr->num_sge = isert_cmd->tx_desc.num_sge;
1045         send_wr->send_flags = IB_SEND_SIGNALED;
1046 }
1047
1048 static int
1049 isert_rdma_post_recvl(struct isert_conn *isert_conn)
1050 {
1051         struct ib_recv_wr rx_wr, *rx_wr_fail;
1052         struct ib_sge sge;
1053         int ret;
1054
1055         memset(&sge, 0, sizeof(struct ib_sge));
1056         sge.addr = isert_conn->login_req_dma;
1057         sge.length = ISER_RX_LOGIN_SIZE;
1058         sge.lkey = isert_conn->conn_mr->lkey;
1059
1060         isert_dbg("Setup sge: addr: %llx length: %d 0x%08x\n",
1061                 sge.addr, sge.length, sge.lkey);
1062
1063         memset(&rx_wr, 0, sizeof(struct ib_recv_wr));
1064         rx_wr.wr_id = (uintptr_t)isert_conn->login_req_buf;
1065         rx_wr.sg_list = &sge;
1066         rx_wr.num_sge = 1;
1067
1068         isert_conn->post_recv_buf_count++;
1069         ret = ib_post_recv(isert_conn->conn_qp, &rx_wr, &rx_wr_fail);
1070         if (ret) {
1071                 isert_err("ib_post_recv() failed: %d\n", ret);
1072                 isert_conn->post_recv_buf_count--;
1073         }
1074
1075         return ret;
1076 }
1077
1078 static int
1079 isert_put_login_tx(struct iscsi_conn *conn, struct iscsi_login *login,
1080                    u32 length)
1081 {
1082         struct isert_conn *isert_conn = conn->context;
1083         struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
1084         struct iser_tx_desc *tx_desc = &isert_conn->conn_login_tx_desc;
1085         int ret;
1086
1087         isert_create_send_desc(isert_conn, NULL, tx_desc);
1088
1089         memcpy(&tx_desc->iscsi_header, &login->rsp[0],
1090                sizeof(struct iscsi_hdr));
1091
1092         isert_init_tx_hdrs(isert_conn, tx_desc);
1093
1094         if (length > 0) {
1095                 struct ib_sge *tx_dsg = &tx_desc->tx_sg[1];
1096
1097                 ib_dma_sync_single_for_cpu(ib_dev, isert_conn->login_rsp_dma,
1098                                            length, DMA_TO_DEVICE);
1099
1100                 memcpy(isert_conn->login_rsp_buf, login->rsp_buf, length);
1101
1102                 ib_dma_sync_single_for_device(ib_dev, isert_conn->login_rsp_dma,
1103                                               length, DMA_TO_DEVICE);
1104
1105                 tx_dsg->addr    = isert_conn->login_rsp_dma;
1106                 tx_dsg->length  = length;
1107                 tx_dsg->lkey    = isert_conn->conn_mr->lkey;
1108                 tx_desc->num_sge = 2;
1109         }
1110         if (!login->login_failed) {
1111                 if (login->login_complete) {
1112                         if (!conn->sess->sess_ops->SessionType &&
1113                             isert_conn->conn_device->use_fastreg) {
1114                                 ret = isert_conn_create_fastreg_pool(isert_conn);
1115                                 if (ret) {
1116                                         isert_err("Conn: %p failed to create"
1117                                                " fastreg pool\n", isert_conn);
1118                                         return ret;
1119                                 }
1120                         }
1121
1122                         ret = isert_alloc_rx_descriptors(isert_conn);
1123                         if (ret)
1124                                 return ret;
1125
1126                         ret = isert_post_recv(isert_conn, ISERT_MIN_POSTED_RX);
1127                         if (ret)
1128                                 return ret;
1129
1130                         /* Now we are in FULL_FEATURE phase */
1131                         mutex_lock(&isert_conn->conn_mutex);
1132                         isert_conn->state = ISER_CONN_FULL_FEATURE;
1133                         mutex_unlock(&isert_conn->conn_mutex);
1134                         goto post_send;
1135                 }
1136
1137                 ret = isert_rdma_post_recvl(isert_conn);
1138                 if (ret)
1139                         return ret;
1140         }
1141 post_send:
1142         ret = isert_post_send(isert_conn, tx_desc);
1143         if (ret)
1144                 return ret;
1145
1146         return 0;
1147 }
1148
1149 static void
1150 isert_rx_login_req(struct isert_conn *isert_conn)
1151 {
1152         struct iser_rx_desc *rx_desc = (void *)isert_conn->login_req_buf;
1153         int rx_buflen = isert_conn->login_req_len;
1154         struct iscsi_conn *conn = isert_conn->conn;
1155         struct iscsi_login *login = conn->conn_login;
1156         int size;
1157
1158         isert_info("conn %p\n", isert_conn);
1159
1160         WARN_ON_ONCE(!login);
1161
1162         if (login->first_request) {
1163                 struct iscsi_login_req *login_req =
1164                         (struct iscsi_login_req *)&rx_desc->iscsi_header;
1165                 /*
1166                  * Setup the initial iscsi_login values from the leading
1167                  * login request PDU.
1168                  */
1169                 login->leading_connection = (!login_req->tsih) ? 1 : 0;
1170                 login->current_stage =
1171                         (login_req->flags & ISCSI_FLAG_LOGIN_CURRENT_STAGE_MASK)
1172                          >> 2;
1173                 login->version_min      = login_req->min_version;
1174                 login->version_max      = login_req->max_version;
1175                 memcpy(login->isid, login_req->isid, 6);
1176                 login->cmd_sn           = be32_to_cpu(login_req->cmdsn);
1177                 login->init_task_tag    = login_req->itt;
1178                 login->initial_exp_statsn = be32_to_cpu(login_req->exp_statsn);
1179                 login->cid              = be16_to_cpu(login_req->cid);
1180                 login->tsih             = be16_to_cpu(login_req->tsih);
1181         }
1182
1183         memcpy(&login->req[0], (void *)&rx_desc->iscsi_header, ISCSI_HDR_LEN);
1184
1185         size = min(rx_buflen, MAX_KEY_VALUE_PAIRS);
1186         isert_dbg("Using login payload size: %d, rx_buflen: %d "
1187                   "MAX_KEY_VALUE_PAIRS: %d\n", size, rx_buflen,
1188                   MAX_KEY_VALUE_PAIRS);
1189         memcpy(login->req_buf, &rx_desc->data[0], size);
1190
1191         if (login->first_request) {
1192                 complete(&isert_conn->conn_login_comp);
1193                 return;
1194         }
1195         schedule_delayed_work(&conn->login_work, 0);
1196 }
1197
1198 static struct iscsi_cmd
1199 *isert_allocate_cmd(struct iscsi_conn *conn)
1200 {
1201         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
1202         struct isert_cmd *isert_cmd;
1203         struct iscsi_cmd *cmd;
1204
1205         cmd = iscsit_allocate_cmd(conn, TASK_INTERRUPTIBLE);
1206         if (!cmd) {
1207                 isert_err("Unable to allocate iscsi_cmd + isert_cmd\n");
1208                 return NULL;
1209         }
1210         isert_cmd = iscsit_priv_cmd(cmd);
1211         isert_cmd->conn = isert_conn;
1212         isert_cmd->iscsi_cmd = cmd;
1213
1214         return cmd;
1215 }
1216
1217 static int
1218 isert_handle_scsi_cmd(struct isert_conn *isert_conn,
1219                       struct isert_cmd *isert_cmd, struct iscsi_cmd *cmd,
1220                       struct iser_rx_desc *rx_desc, unsigned char *buf)
1221 {
1222         struct iscsi_conn *conn = isert_conn->conn;
1223         struct iscsi_scsi_req *hdr = (struct iscsi_scsi_req *)buf;
1224         struct scatterlist *sg;
1225         int imm_data, imm_data_len, unsol_data, sg_nents, rc;
1226         bool dump_payload = false;
1227
1228         rc = iscsit_setup_scsi_cmd(conn, cmd, buf);
1229         if (rc < 0)
1230                 return rc;
1231
1232         imm_data = cmd->immediate_data;
1233         imm_data_len = cmd->first_burst_len;
1234         unsol_data = cmd->unsolicited_data;
1235
1236         rc = iscsit_process_scsi_cmd(conn, cmd, hdr);
1237         if (rc < 0) {
1238                 return 0;
1239         } else if (rc > 0) {
1240                 dump_payload = true;
1241                 goto sequence_cmd;
1242         }
1243
1244         if (!imm_data)
1245                 return 0;
1246
1247         sg = &cmd->se_cmd.t_data_sg[0];
1248         sg_nents = max(1UL, DIV_ROUND_UP(imm_data_len, PAGE_SIZE));
1249
1250         isert_dbg("Copying Immediate SG: %p sg_nents: %u from %p imm_data_len: %d\n",
1251                   sg, sg_nents, &rx_desc->data[0], imm_data_len);
1252
1253         sg_copy_from_buffer(sg, sg_nents, &rx_desc->data[0], imm_data_len);
1254
1255         cmd->write_data_done += imm_data_len;
1256
1257         if (cmd->write_data_done == cmd->se_cmd.data_length) {
1258                 spin_lock_bh(&cmd->istate_lock);
1259                 cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1260                 cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1261                 spin_unlock_bh(&cmd->istate_lock);
1262         }
1263
1264 sequence_cmd:
1265         rc = iscsit_sequence_cmd(conn, cmd, buf, hdr->cmdsn);
1266
1267         if (!rc && dump_payload == false && unsol_data)
1268                 iscsit_set_unsoliticed_dataout(cmd);
1269         else if (dump_payload && imm_data)
1270                 target_put_sess_cmd(conn->sess->se_sess, &cmd->se_cmd);
1271
1272         return 0;
1273 }
1274
1275 static int
1276 isert_handle_iscsi_dataout(struct isert_conn *isert_conn,
1277                            struct iser_rx_desc *rx_desc, unsigned char *buf)
1278 {
1279         struct scatterlist *sg_start;
1280         struct iscsi_conn *conn = isert_conn->conn;
1281         struct iscsi_cmd *cmd = NULL;
1282         struct iscsi_data *hdr = (struct iscsi_data *)buf;
1283         u32 unsol_data_len = ntoh24(hdr->dlength);
1284         int rc, sg_nents, sg_off, page_off;
1285
1286         rc = iscsit_check_dataout_hdr(conn, buf, &cmd);
1287         if (rc < 0)
1288                 return rc;
1289         else if (!cmd)
1290                 return 0;
1291         /*
1292          * FIXME: Unexpected unsolicited_data out
1293          */
1294         if (!cmd->unsolicited_data) {
1295                 isert_err("Received unexpected solicited data payload\n");
1296                 dump_stack();
1297                 return -1;
1298         }
1299
1300         isert_dbg("Unsolicited DataOut unsol_data_len: %u, "
1301                   "write_data_done: %u, data_length: %u\n",
1302                   unsol_data_len,  cmd->write_data_done,
1303                   cmd->se_cmd.data_length);
1304
1305         sg_off = cmd->write_data_done / PAGE_SIZE;
1306         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
1307         sg_nents = max(1UL, DIV_ROUND_UP(unsol_data_len, PAGE_SIZE));
1308         page_off = cmd->write_data_done % PAGE_SIZE;
1309         /*
1310          * FIXME: Non page-aligned unsolicited_data out
1311          */
1312         if (page_off) {
1313                 isert_err("unexpected non-page aligned data payload\n");
1314                 dump_stack();
1315                 return -1;
1316         }
1317         isert_dbg("Copying DataOut: sg_start: %p, sg_off: %u "
1318                   "sg_nents: %u from %p %u\n", sg_start, sg_off,
1319                   sg_nents, &rx_desc->data[0], unsol_data_len);
1320
1321         sg_copy_from_buffer(sg_start, sg_nents, &rx_desc->data[0],
1322                             unsol_data_len);
1323
1324         rc = iscsit_check_dataout_payload(cmd, hdr, false);
1325         if (rc < 0)
1326                 return rc;
1327
1328         return 0;
1329 }
1330
1331 static int
1332 isert_handle_nop_out(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1333                      struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1334                      unsigned char *buf)
1335 {
1336         struct iscsi_conn *conn = isert_conn->conn;
1337         struct iscsi_nopout *hdr = (struct iscsi_nopout *)buf;
1338         int rc;
1339
1340         rc = iscsit_setup_nop_out(conn, cmd, hdr);
1341         if (rc < 0)
1342                 return rc;
1343         /*
1344          * FIXME: Add support for NOPOUT payload using unsolicited RDMA payload
1345          */
1346
1347         return iscsit_process_nop_out(conn, cmd, hdr);
1348 }
1349
1350 static int
1351 isert_handle_text_cmd(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1352                       struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1353                       struct iscsi_text *hdr)
1354 {
1355         struct iscsi_conn *conn = isert_conn->conn;
1356         u32 payload_length = ntoh24(hdr->dlength);
1357         int rc;
1358         unsigned char *text_in = NULL;
1359
1360         rc = iscsit_setup_text_cmd(conn, cmd, hdr);
1361         if (rc < 0)
1362                 return rc;
1363
1364         if (payload_length) {
1365                 text_in = kzalloc(payload_length, GFP_KERNEL);
1366                 if (!text_in) {
1367                         isert_err("Unable to allocate text_in of payload_length: %u\n",
1368                                   payload_length);
1369                         return -ENOMEM;
1370                 }
1371         }
1372         cmd->text_in_ptr = text_in;
1373
1374         memcpy(cmd->text_in_ptr, &rx_desc->data[0], payload_length);
1375
1376         return iscsit_process_text_cmd(conn, cmd, hdr);
1377 }
1378
1379 static int
1380 isert_rx_opcode(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc,
1381                 uint32_t read_stag, uint64_t read_va,
1382                 uint32_t write_stag, uint64_t write_va)
1383 {
1384         struct iscsi_hdr *hdr = &rx_desc->iscsi_header;
1385         struct iscsi_conn *conn = isert_conn->conn;
1386         struct iscsi_session *sess = conn->sess;
1387         struct iscsi_cmd *cmd;
1388         struct isert_cmd *isert_cmd;
1389         int ret = -EINVAL;
1390         u8 opcode = (hdr->opcode & ISCSI_OPCODE_MASK);
1391
1392         if (sess->sess_ops->SessionType &&
1393            (!(opcode & ISCSI_OP_TEXT) || !(opcode & ISCSI_OP_LOGOUT))) {
1394                 isert_err("Got illegal opcode: 0x%02x in SessionType=Discovery,"
1395                           " ignoring\n", opcode);
1396                 return 0;
1397         }
1398
1399         switch (opcode) {
1400         case ISCSI_OP_SCSI_CMD:
1401                 cmd = isert_allocate_cmd(conn);
1402                 if (!cmd)
1403                         break;
1404
1405                 isert_cmd = iscsit_priv_cmd(cmd);
1406                 isert_cmd->read_stag = read_stag;
1407                 isert_cmd->read_va = read_va;
1408                 isert_cmd->write_stag = write_stag;
1409                 isert_cmd->write_va = write_va;
1410
1411                 ret = isert_handle_scsi_cmd(isert_conn, isert_cmd, cmd,
1412                                         rx_desc, (unsigned char *)hdr);
1413                 break;
1414         case ISCSI_OP_NOOP_OUT:
1415                 cmd = isert_allocate_cmd(conn);
1416                 if (!cmd)
1417                         break;
1418
1419                 isert_cmd = iscsit_priv_cmd(cmd);
1420                 ret = isert_handle_nop_out(isert_conn, isert_cmd, cmd,
1421                                            rx_desc, (unsigned char *)hdr);
1422                 break;
1423         case ISCSI_OP_SCSI_DATA_OUT:
1424                 ret = isert_handle_iscsi_dataout(isert_conn, rx_desc,
1425                                                 (unsigned char *)hdr);
1426                 break;
1427         case ISCSI_OP_SCSI_TMFUNC:
1428                 cmd = isert_allocate_cmd(conn);
1429                 if (!cmd)
1430                         break;
1431
1432                 ret = iscsit_handle_task_mgt_cmd(conn, cmd,
1433                                                 (unsigned char *)hdr);
1434                 break;
1435         case ISCSI_OP_LOGOUT:
1436                 cmd = isert_allocate_cmd(conn);
1437                 if (!cmd)
1438                         break;
1439
1440                 ret = iscsit_handle_logout_cmd(conn, cmd, (unsigned char *)hdr);
1441                 break;
1442         case ISCSI_OP_TEXT:
1443                 if (be32_to_cpu(hdr->ttt) != 0xFFFFFFFF) {
1444                         cmd = iscsit_find_cmd_from_itt(conn, hdr->itt);
1445                         if (!cmd)
1446                                 break;
1447                 } else {
1448                         cmd = isert_allocate_cmd(conn);
1449                         if (!cmd)
1450                                 break;
1451                 }
1452
1453                 isert_cmd = iscsit_priv_cmd(cmd);
1454                 ret = isert_handle_text_cmd(isert_conn, isert_cmd, cmd,
1455                                             rx_desc, (struct iscsi_text *)hdr);
1456                 break;
1457         default:
1458                 isert_err("Got unknown iSCSI OpCode: 0x%02x\n", opcode);
1459                 dump_stack();
1460                 break;
1461         }
1462
1463         return ret;
1464 }
1465
1466 static void
1467 isert_rx_do_work(struct iser_rx_desc *rx_desc, struct isert_conn *isert_conn)
1468 {
1469         struct iser_hdr *iser_hdr = &rx_desc->iser_header;
1470         uint64_t read_va = 0, write_va = 0;
1471         uint32_t read_stag = 0, write_stag = 0;
1472         int rc;
1473
1474         switch (iser_hdr->flags & 0xF0) {
1475         case ISCSI_CTRL:
1476                 if (iser_hdr->flags & ISER_RSV) {
1477                         read_stag = be32_to_cpu(iser_hdr->read_stag);
1478                         read_va = be64_to_cpu(iser_hdr->read_va);
1479                         isert_dbg("ISER_RSV: read_stag: 0x%x read_va: 0x%llx\n",
1480                                   read_stag, (unsigned long long)read_va);
1481                 }
1482                 if (iser_hdr->flags & ISER_WSV) {
1483                         write_stag = be32_to_cpu(iser_hdr->write_stag);
1484                         write_va = be64_to_cpu(iser_hdr->write_va);
1485                         isert_dbg("ISER_WSV: write_stag: 0x%x write_va: 0x%llx\n",
1486                                   write_stag, (unsigned long long)write_va);
1487                 }
1488
1489                 isert_dbg("ISER ISCSI_CTRL PDU\n");
1490                 break;
1491         case ISER_HELLO:
1492                 isert_err("iSER Hello message\n");
1493                 break;
1494         default:
1495                 isert_warn("Unknown iSER hdr flags: 0x%02x\n", iser_hdr->flags);
1496                 break;
1497         }
1498
1499         rc = isert_rx_opcode(isert_conn, rx_desc,
1500                              read_stag, read_va, write_stag, write_va);
1501 }
1502
1503 static void
1504 isert_rx_completion(struct iser_rx_desc *desc, struct isert_conn *isert_conn,
1505                     u32 xfer_len)
1506 {
1507         struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
1508         struct iscsi_hdr *hdr;
1509         u64 rx_dma;
1510         int rx_buflen, outstanding;
1511
1512         if ((char *)desc == isert_conn->login_req_buf) {
1513                 rx_dma = isert_conn->login_req_dma;
1514                 rx_buflen = ISER_RX_LOGIN_SIZE;
1515                 isert_dbg("login_buf: Using rx_dma: 0x%llx, rx_buflen: %d\n",
1516                          rx_dma, rx_buflen);
1517         } else {
1518                 rx_dma = desc->dma_addr;
1519                 rx_buflen = ISER_RX_PAYLOAD_SIZE;
1520                 isert_dbg("req_buf: Using rx_dma: 0x%llx, rx_buflen: %d\n",
1521                          rx_dma, rx_buflen);
1522         }
1523
1524         ib_dma_sync_single_for_cpu(ib_dev, rx_dma, rx_buflen, DMA_FROM_DEVICE);
1525
1526         hdr = &desc->iscsi_header;
1527         isert_dbg("iSCSI opcode: 0x%02x, ITT: 0x%08x, flags: 0x%02x dlen: %d\n",
1528                  hdr->opcode, hdr->itt, hdr->flags,
1529                  (int)(xfer_len - ISER_HEADERS_LEN));
1530
1531         if ((char *)desc == isert_conn->login_req_buf) {
1532                 isert_conn->login_req_len = xfer_len - ISER_HEADERS_LEN;
1533                 if (isert_conn->conn) {
1534                         struct iscsi_login *login = isert_conn->conn->conn_login;
1535
1536                         if (login && !login->first_request)
1537                                 isert_rx_login_req(isert_conn);
1538                 }
1539                 mutex_lock(&isert_conn->conn_mutex);
1540                 complete(&isert_conn->login_req_comp);
1541                 mutex_unlock(&isert_conn->conn_mutex);
1542         } else {
1543                 isert_rx_do_work(desc, isert_conn);
1544         }
1545
1546         ib_dma_sync_single_for_device(ib_dev, rx_dma, rx_buflen,
1547                                       DMA_FROM_DEVICE);
1548
1549         isert_conn->post_recv_buf_count--;
1550         isert_dbg("Decremented post_recv_buf_count: %d\n",
1551                   isert_conn->post_recv_buf_count);
1552
1553         if ((char *)desc == isert_conn->login_req_buf)
1554                 return;
1555
1556         outstanding = isert_conn->post_recv_buf_count;
1557         if (outstanding + ISERT_MIN_POSTED_RX <= ISERT_QP_MAX_RECV_DTOS) {
1558                 int err, count = min(ISERT_QP_MAX_RECV_DTOS - outstanding,
1559                                 ISERT_MIN_POSTED_RX);
1560                 err = isert_post_recv(isert_conn, count);
1561                 if (err) {
1562                         isert_err("isert_post_recv() count: %d failed, %d\n",
1563                                count, err);
1564                 }
1565         }
1566 }
1567
1568 static int
1569 isert_map_data_buf(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1570                    struct scatterlist *sg, u32 nents, u32 length, u32 offset,
1571                    enum iser_ib_op_code op, struct isert_data_buf *data)
1572 {
1573         struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
1574
1575         data->dma_dir = op == ISER_IB_RDMA_WRITE ?
1576                               DMA_TO_DEVICE : DMA_FROM_DEVICE;
1577
1578         data->len = length - offset;
1579         data->offset = offset;
1580         data->sg_off = data->offset / PAGE_SIZE;
1581
1582         data->sg = &sg[data->sg_off];
1583         data->nents = min_t(unsigned int, nents - data->sg_off,
1584                                           ISCSI_ISER_SG_TABLESIZE);
1585         data->len = min_t(unsigned int, data->len, ISCSI_ISER_SG_TABLESIZE *
1586                                         PAGE_SIZE);
1587
1588         data->dma_nents = ib_dma_map_sg(ib_dev, data->sg, data->nents,
1589                                         data->dma_dir);
1590         if (unlikely(!data->dma_nents)) {
1591                 isert_err("Cmd: unable to dma map SGs %p\n", sg);
1592                 return -EINVAL;
1593         }
1594
1595         isert_dbg("Mapped cmd: %p count: %u sg: %p sg_nents: %u rdma_len %d\n",
1596                   isert_cmd, data->dma_nents, data->sg, data->nents, data->len);
1597
1598         return 0;
1599 }
1600
1601 static void
1602 isert_unmap_data_buf(struct isert_conn *isert_conn, struct isert_data_buf *data)
1603 {
1604         struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
1605
1606         ib_dma_unmap_sg(ib_dev, data->sg, data->nents, data->dma_dir);
1607         memset(data, 0, sizeof(*data));
1608 }
1609
1610
1611
1612 static void
1613 isert_unmap_cmd(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn)
1614 {
1615         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1616
1617         isert_dbg("Cmd %p\n", isert_cmd);
1618
1619         if (wr->data.sg) {
1620                 isert_dbg("Cmd %p unmap_sg op\n", isert_cmd);
1621                 isert_unmap_data_buf(isert_conn, &wr->data);
1622         }
1623
1624         if (wr->send_wr) {
1625                 isert_dbg("Cmd %p free send_wr\n", isert_cmd);
1626                 kfree(wr->send_wr);
1627                 wr->send_wr = NULL;
1628         }
1629
1630         if (wr->ib_sge) {
1631                 isert_dbg("Cmd %p free ib_sge\n", isert_cmd);
1632                 kfree(wr->ib_sge);
1633                 wr->ib_sge = NULL;
1634         }
1635 }
1636
1637 static void
1638 isert_unreg_rdma(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn)
1639 {
1640         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1641         LIST_HEAD(unmap_list);
1642
1643         isert_dbg("Cmd %p\n", isert_cmd);
1644
1645         if (wr->fr_desc) {
1646                 isert_dbg("Cmd %p free fr_desc %p\n", isert_cmd, wr->fr_desc);
1647                 if (wr->fr_desc->ind & ISERT_PROTECTED) {
1648                         isert_unmap_data_buf(isert_conn, &wr->prot);
1649                         wr->fr_desc->ind &= ~ISERT_PROTECTED;
1650                 }
1651                 spin_lock_bh(&isert_conn->conn_lock);
1652                 list_add_tail(&wr->fr_desc->list, &isert_conn->conn_fr_pool);
1653                 spin_unlock_bh(&isert_conn->conn_lock);
1654                 wr->fr_desc = NULL;
1655         }
1656
1657         if (wr->data.sg) {
1658                 isert_dbg("Cmd %p unmap_sg op\n", isert_cmd);
1659                 isert_unmap_data_buf(isert_conn, &wr->data);
1660         }
1661
1662         wr->ib_sge = NULL;
1663         wr->send_wr = NULL;
1664 }
1665
1666 static void
1667 isert_put_cmd(struct isert_cmd *isert_cmd, bool comp_err)
1668 {
1669         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1670         struct isert_conn *isert_conn = isert_cmd->conn;
1671         struct iscsi_conn *conn = isert_conn->conn;
1672         struct isert_device *device = isert_conn->conn_device;
1673         struct iscsi_text_rsp *hdr;
1674
1675         isert_dbg("Cmd %p\n", isert_cmd);
1676
1677         switch (cmd->iscsi_opcode) {
1678         case ISCSI_OP_SCSI_CMD:
1679                 spin_lock_bh(&conn->cmd_lock);
1680                 if (!list_empty(&cmd->i_conn_node))
1681                         list_del_init(&cmd->i_conn_node);
1682                 spin_unlock_bh(&conn->cmd_lock);
1683
1684                 if (cmd->data_direction == DMA_TO_DEVICE) {
1685                         iscsit_stop_dataout_timer(cmd);
1686                         /*
1687                          * Check for special case during comp_err where
1688                          * WRITE_PENDING has been handed off from core,
1689                          * but requires an extra target_put_sess_cmd()
1690                          * before transport_generic_free_cmd() below.
1691                          */
1692                         if (comp_err &&
1693                             cmd->se_cmd.t_state == TRANSPORT_WRITE_PENDING) {
1694                                 struct se_cmd *se_cmd = &cmd->se_cmd;
1695
1696                                 target_put_sess_cmd(se_cmd->se_sess, se_cmd);
1697                         }
1698                 }
1699
1700                 device->unreg_rdma_mem(isert_cmd, isert_conn);
1701                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1702                 break;
1703         case ISCSI_OP_SCSI_TMFUNC:
1704                 spin_lock_bh(&conn->cmd_lock);
1705                 if (!list_empty(&cmd->i_conn_node))
1706                         list_del_init(&cmd->i_conn_node);
1707                 spin_unlock_bh(&conn->cmd_lock);
1708
1709                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1710                 break;
1711         case ISCSI_OP_REJECT:
1712         case ISCSI_OP_NOOP_OUT:
1713         case ISCSI_OP_TEXT:
1714                 hdr = (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
1715                 /* If the continue bit is on, keep the command alive */
1716                 if (hdr->flags & ISCSI_FLAG_TEXT_CONTINUE)
1717                         break;
1718
1719                 spin_lock_bh(&conn->cmd_lock);
1720                 if (!list_empty(&cmd->i_conn_node))
1721                         list_del_init(&cmd->i_conn_node);
1722                 spin_unlock_bh(&conn->cmd_lock);
1723
1724                 /*
1725                  * Handle special case for REJECT when iscsi_add_reject*() has
1726                  * overwritten the original iscsi_opcode assignment, and the
1727                  * associated cmd->se_cmd needs to be released.
1728                  */
1729                 if (cmd->se_cmd.se_tfo != NULL) {
1730                         isert_dbg("Calling transport_generic_free_cmd for 0x%02x\n",
1731                                  cmd->iscsi_opcode);
1732                         transport_generic_free_cmd(&cmd->se_cmd, 0);
1733                         break;
1734                 }
1735                 /*
1736                  * Fall-through
1737                  */
1738         default:
1739                 iscsit_release_cmd(cmd);
1740                 break;
1741         }
1742 }
1743
1744 static void
1745 isert_unmap_tx_desc(struct iser_tx_desc *tx_desc, struct ib_device *ib_dev)
1746 {
1747         if (tx_desc->dma_addr != 0) {
1748                 isert_dbg("unmap single for tx_desc->dma_addr\n");
1749                 ib_dma_unmap_single(ib_dev, tx_desc->dma_addr,
1750                                     ISER_HEADERS_LEN, DMA_TO_DEVICE);
1751                 tx_desc->dma_addr = 0;
1752         }
1753 }
1754
1755 static void
1756 isert_completion_put(struct iser_tx_desc *tx_desc, struct isert_cmd *isert_cmd,
1757                      struct ib_device *ib_dev, bool comp_err)
1758 {
1759         if (isert_cmd->pdu_buf_dma != 0) {
1760                 isert_dbg("unmap single for isert_cmd->pdu_buf_dma\n");
1761                 ib_dma_unmap_single(ib_dev, isert_cmd->pdu_buf_dma,
1762                                     isert_cmd->pdu_buf_len, DMA_TO_DEVICE);
1763                 isert_cmd->pdu_buf_dma = 0;
1764         }
1765
1766         isert_unmap_tx_desc(tx_desc, ib_dev);
1767         isert_put_cmd(isert_cmd, comp_err);
1768 }
1769
1770 static int
1771 isert_check_pi_status(struct se_cmd *se_cmd, struct ib_mr *sig_mr)
1772 {
1773         struct ib_mr_status mr_status;
1774         int ret;
1775
1776         ret = ib_check_mr_status(sig_mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
1777         if (ret) {
1778                 isert_err("ib_check_mr_status failed, ret %d\n", ret);
1779                 goto fail_mr_status;
1780         }
1781
1782         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1783                 u64 sec_offset_err;
1784                 u32 block_size = se_cmd->se_dev->dev_attrib.block_size + 8;
1785
1786                 switch (mr_status.sig_err.err_type) {
1787                 case IB_SIG_BAD_GUARD:
1788                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_GUARD_CHECK_FAILED;
1789                         break;
1790                 case IB_SIG_BAD_REFTAG:
1791                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_REF_TAG_CHECK_FAILED;
1792                         break;
1793                 case IB_SIG_BAD_APPTAG:
1794                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_APP_TAG_CHECK_FAILED;
1795                         break;
1796                 }
1797                 sec_offset_err = mr_status.sig_err.sig_err_offset;
1798                 do_div(sec_offset_err, block_size);
1799                 se_cmd->bad_sector = sec_offset_err + se_cmd->t_task_lba;
1800
1801                 isert_err("PI error found type %d at sector 0x%llx "
1802                           "expected 0x%x vs actual 0x%x\n",
1803                           mr_status.sig_err.err_type,
1804                           (unsigned long long)se_cmd->bad_sector,
1805                           mr_status.sig_err.expected,
1806                           mr_status.sig_err.actual);
1807                 ret = 1;
1808         }
1809
1810 fail_mr_status:
1811         return ret;
1812 }
1813
1814 static void
1815 isert_completion_rdma_write(struct iser_tx_desc *tx_desc,
1816                             struct isert_cmd *isert_cmd)
1817 {
1818         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1819         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1820         struct se_cmd *se_cmd = &cmd->se_cmd;
1821         struct isert_conn *isert_conn = isert_cmd->conn;
1822         struct isert_device *device = isert_conn->conn_device;
1823         int ret = 0;
1824
1825         if (wr->fr_desc && wr->fr_desc->ind & ISERT_PROTECTED) {
1826                 ret = isert_check_pi_status(se_cmd,
1827                                             wr->fr_desc->pi_ctx->sig_mr);
1828                 wr->fr_desc->ind &= ~ISERT_PROTECTED;
1829         }
1830
1831         device->unreg_rdma_mem(isert_cmd, isert_conn);
1832         wr->send_wr_num = 0;
1833         if (ret)
1834                 transport_send_check_condition_and_sense(se_cmd,
1835                                                          se_cmd->pi_err, 0);
1836         else
1837                 isert_put_response(isert_conn->conn, cmd);
1838 }
1839
1840 static void
1841 isert_completion_rdma_read(struct iser_tx_desc *tx_desc,
1842                            struct isert_cmd *isert_cmd)
1843 {
1844         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1845         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1846         struct se_cmd *se_cmd = &cmd->se_cmd;
1847         struct isert_conn *isert_conn = isert_cmd->conn;
1848         struct isert_device *device = isert_conn->conn_device;
1849         int ret = 0;
1850
1851         if (wr->fr_desc && wr->fr_desc->ind & ISERT_PROTECTED) {
1852                 ret = isert_check_pi_status(se_cmd,
1853                                             wr->fr_desc->pi_ctx->sig_mr);
1854                 wr->fr_desc->ind &= ~ISERT_PROTECTED;
1855         }
1856
1857         iscsit_stop_dataout_timer(cmd);
1858         device->unreg_rdma_mem(isert_cmd, isert_conn);
1859         cmd->write_data_done = wr->data.len;
1860         wr->send_wr_num = 0;
1861
1862         isert_dbg("Cmd: %p RDMA_READ comp calling execute_cmd\n", isert_cmd);
1863         spin_lock_bh(&cmd->istate_lock);
1864         cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1865         cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1866         spin_unlock_bh(&cmd->istate_lock);
1867
1868         if (ret) {
1869                 target_put_sess_cmd(se_cmd->se_sess, se_cmd);
1870                 transport_send_check_condition_and_sense(se_cmd,
1871                                                          se_cmd->pi_err, 0);
1872         } else {
1873                 target_execute_cmd(se_cmd);
1874         }
1875 }
1876
1877 static void
1878 isert_do_control_comp(struct work_struct *work)
1879 {
1880         struct isert_cmd *isert_cmd = container_of(work,
1881                         struct isert_cmd, comp_work);
1882         struct isert_conn *isert_conn = isert_cmd->conn;
1883         struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
1884         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1885
1886         isert_dbg("Cmd %p i_state %d\n", isert_cmd, cmd->i_state);
1887
1888         switch (cmd->i_state) {
1889         case ISTATE_SEND_TASKMGTRSP:
1890                 iscsit_tmr_post_handler(cmd, cmd->conn);
1891         case ISTATE_SEND_REJECT:   /* FALLTHRU */
1892         case ISTATE_SEND_TEXTRSP:  /* FALLTHRU */
1893                 cmd->i_state = ISTATE_SENT_STATUS;
1894                 isert_completion_put(&isert_cmd->tx_desc, isert_cmd,
1895                                      ib_dev, false);
1896                 break;
1897         case ISTATE_SEND_LOGOUTRSP:
1898                 iscsit_logout_post_handler(cmd, cmd->conn);
1899                 break;
1900         default:
1901                 isert_err("Unknown i_state %d\n", cmd->i_state);
1902                 dump_stack();
1903                 break;
1904         }
1905 }
1906
1907 static void
1908 isert_response_completion(struct iser_tx_desc *tx_desc,
1909                           struct isert_cmd *isert_cmd,
1910                           struct isert_conn *isert_conn,
1911                           struct ib_device *ib_dev)
1912 {
1913         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1914
1915         if (cmd->i_state == ISTATE_SEND_TASKMGTRSP ||
1916             cmd->i_state == ISTATE_SEND_LOGOUTRSP ||
1917             cmd->i_state == ISTATE_SEND_REJECT ||
1918             cmd->i_state == ISTATE_SEND_TEXTRSP) {
1919                 isert_unmap_tx_desc(tx_desc, ib_dev);
1920
1921                 INIT_WORK(&isert_cmd->comp_work, isert_do_control_comp);
1922                 queue_work(isert_comp_wq, &isert_cmd->comp_work);
1923                 return;
1924         }
1925
1926         cmd->i_state = ISTATE_SENT_STATUS;
1927         isert_completion_put(tx_desc, isert_cmd, ib_dev, false);
1928 }
1929
1930 static void
1931 isert_send_completion(struct iser_tx_desc *tx_desc,
1932                       struct isert_conn *isert_conn)
1933 {
1934         struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
1935         struct isert_cmd *isert_cmd = tx_desc->isert_cmd;
1936         struct isert_rdma_wr *wr;
1937
1938         if (!isert_cmd) {
1939                 isert_unmap_tx_desc(tx_desc, ib_dev);
1940                 return;
1941         }
1942         wr = &isert_cmd->rdma_wr;
1943
1944         isert_dbg("Cmd %p iser_ib_op %d\n", isert_cmd, wr->iser_ib_op);
1945
1946         switch (wr->iser_ib_op) {
1947         case ISER_IB_RECV:
1948                 isert_err("Got ISER_IB_RECV\n");
1949                 dump_stack();
1950                 break;
1951         case ISER_IB_SEND:
1952                 isert_response_completion(tx_desc, isert_cmd,
1953                                           isert_conn, ib_dev);
1954                 break;
1955         case ISER_IB_RDMA_WRITE:
1956                 isert_completion_rdma_write(tx_desc, isert_cmd);
1957                 break;
1958         case ISER_IB_RDMA_READ:
1959                 isert_completion_rdma_read(tx_desc, isert_cmd);
1960                 break;
1961         default:
1962                 isert_err("Unknown wr->iser_ib_op: 0x%x\n", wr->iser_ib_op);
1963                 dump_stack();
1964                 break;
1965         }
1966 }
1967
1968 /**
1969  * is_isert_tx_desc() - Indicate if the completion wr_id
1970  *     is a TX descriptor or not.
1971  * @isert_conn: iser connection
1972  * @wr_id: completion WR identifier
1973  *
1974  * Since we cannot rely on wc opcode in FLUSH errors
1975  * we must work around it by checking if the wr_id address
1976  * falls in the iser connection rx_descs buffer. If so
1977  * it is an RX descriptor, otherwize it is a TX.
1978  */
1979 static inline bool
1980 is_isert_tx_desc(struct isert_conn *isert_conn, void *wr_id)
1981 {
1982         void *start = isert_conn->conn_rx_descs;
1983         int len = ISERT_QP_MAX_RECV_DTOS * sizeof(*isert_conn->conn_rx_descs);
1984
1985         if (wr_id >= start && wr_id < start + len)
1986                 return false;
1987
1988         return true;
1989 }
1990
1991 static void
1992 isert_cq_comp_err(struct isert_conn *isert_conn, struct ib_wc *wc)
1993 {
1994         if (wc->wr_id == ISER_BEACON_WRID) {
1995                 isert_info("conn %p completing conn_wait_comp_err\n",
1996                            isert_conn);
1997                 complete(&isert_conn->conn_wait_comp_err);
1998         } else if (is_isert_tx_desc(isert_conn, (void *)(uintptr_t)wc->wr_id)) {
1999                 struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
2000                 struct isert_cmd *isert_cmd;
2001                 struct iser_tx_desc *desc;
2002
2003                 desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
2004                 isert_cmd = desc->isert_cmd;
2005                 if (!isert_cmd)
2006                         isert_unmap_tx_desc(desc, ib_dev);
2007                 else
2008                         isert_completion_put(desc, isert_cmd, ib_dev, true);
2009         } else {
2010                 isert_conn->post_recv_buf_count--;
2011                 if (!isert_conn->post_recv_buf_count)
2012                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
2013         }
2014 }
2015
2016 static void
2017 isert_handle_wc(struct ib_wc *wc)
2018 {
2019         struct isert_conn *isert_conn;
2020         struct iser_tx_desc *tx_desc;
2021         struct iser_rx_desc *rx_desc;
2022
2023         isert_conn = wc->qp->qp_context;
2024         if (likely(wc->status == IB_WC_SUCCESS)) {
2025                 if (wc->opcode == IB_WC_RECV) {
2026                         rx_desc = (struct iser_rx_desc *)(uintptr_t)wc->wr_id;
2027                         isert_rx_completion(rx_desc, isert_conn, wc->byte_len);
2028                 } else {
2029                         tx_desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
2030                         isert_send_completion(tx_desc, isert_conn);
2031                 }
2032         } else {
2033                 if (wc->status != IB_WC_WR_FLUSH_ERR)
2034                         isert_err("wr id %llx status %d vend_err %x\n",
2035                                   wc->wr_id, wc->status, wc->vendor_err);
2036                 else
2037                         isert_dbg("flush error: wr id %llx\n", wc->wr_id);
2038
2039                 if (wc->wr_id != ISER_FASTREG_LI_WRID)
2040                         isert_cq_comp_err(isert_conn, wc);
2041         }
2042 }
2043
2044 static void
2045 isert_cq_work(struct work_struct *work)
2046 {
2047         enum { isert_poll_budget = 65536 };
2048         struct isert_comp *comp = container_of(work, struct isert_comp,
2049                                                work);
2050         struct ib_wc *const wcs = comp->wcs;
2051         int i, n, completed = 0;
2052
2053         while ((n = ib_poll_cq(comp->cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
2054                 for (i = 0; i < n; i++)
2055                         isert_handle_wc(&wcs[i]);
2056
2057                 completed += n;
2058                 if (completed >= isert_poll_budget)
2059                         break;
2060         }
2061
2062         ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP);
2063 }
2064
2065 static void
2066 isert_cq_callback(struct ib_cq *cq, void *context)
2067 {
2068         struct isert_comp *comp = context;
2069
2070         queue_work(isert_comp_wq, &comp->work);
2071 }
2072
2073 static int
2074 isert_post_response(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd)
2075 {
2076         struct ib_send_wr *wr_failed;
2077         int ret;
2078
2079         ret = ib_post_send(isert_conn->conn_qp, &isert_cmd->tx_desc.send_wr,
2080                            &wr_failed);
2081         if (ret) {
2082                 isert_err("ib_post_send failed with %d\n", ret);
2083                 return ret;
2084         }
2085         return ret;
2086 }
2087
2088 static int
2089 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2090 {
2091         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2092         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2093         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2094         struct iscsi_scsi_rsp *hdr = (struct iscsi_scsi_rsp *)
2095                                 &isert_cmd->tx_desc.iscsi_header;
2096
2097         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2098         iscsit_build_rsp_pdu(cmd, conn, true, hdr);
2099         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2100         /*
2101          * Attach SENSE DATA payload to iSCSI Response PDU
2102          */
2103         if (cmd->se_cmd.sense_buffer &&
2104             ((cmd->se_cmd.se_cmd_flags & SCF_TRANSPORT_TASK_SENSE) ||
2105             (cmd->se_cmd.se_cmd_flags & SCF_EMULATED_TASK_SENSE))) {
2106                 struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
2107                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2108                 u32 padding, pdu_len;
2109
2110                 put_unaligned_be16(cmd->se_cmd.scsi_sense_length,
2111                                    cmd->sense_buffer);
2112                 cmd->se_cmd.scsi_sense_length += sizeof(__be16);
2113
2114                 padding = -(cmd->se_cmd.scsi_sense_length) & 3;
2115                 hton24(hdr->dlength, (u32)cmd->se_cmd.scsi_sense_length);
2116                 pdu_len = cmd->se_cmd.scsi_sense_length + padding;
2117
2118                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2119                                 (void *)cmd->sense_buffer, pdu_len,
2120                                 DMA_TO_DEVICE);
2121
2122                 isert_cmd->pdu_buf_len = pdu_len;
2123                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2124                 tx_dsg->length  = pdu_len;
2125                 tx_dsg->lkey    = isert_conn->conn_mr->lkey;
2126                 isert_cmd->tx_desc.num_sge = 2;
2127         }
2128
2129         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2130
2131         isert_dbg("Posting SCSI Response\n");
2132
2133         return isert_post_response(isert_conn, isert_cmd);
2134 }
2135
2136 static void
2137 isert_aborted_task(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2138 {
2139         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2140         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2141         struct isert_device *device = isert_conn->conn_device;
2142
2143         spin_lock_bh(&conn->cmd_lock);
2144         if (!list_empty(&cmd->i_conn_node))
2145                 list_del_init(&cmd->i_conn_node);
2146         spin_unlock_bh(&conn->cmd_lock);
2147
2148         if (cmd->data_direction == DMA_TO_DEVICE)
2149                 iscsit_stop_dataout_timer(cmd);
2150
2151         device->unreg_rdma_mem(isert_cmd, isert_conn);
2152 }
2153
2154 static enum target_prot_op
2155 isert_get_sup_prot_ops(struct iscsi_conn *conn)
2156 {
2157         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2158         struct isert_device *device = isert_conn->conn_device;
2159
2160         if (conn->tpg->tpg_attrib.t10_pi) {
2161                 if (device->pi_capable) {
2162                         isert_info("conn %p PI offload enabled\n", isert_conn);
2163                         isert_conn->pi_support = true;
2164                         return TARGET_PROT_ALL;
2165                 }
2166         }
2167
2168         isert_info("conn %p PI offload disabled\n", isert_conn);
2169         isert_conn->pi_support = false;
2170
2171         return TARGET_PROT_NORMAL;
2172 }
2173
2174 static int
2175 isert_put_nopin(struct iscsi_cmd *cmd, struct iscsi_conn *conn,
2176                 bool nopout_response)
2177 {
2178         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2179         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2180         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2181
2182         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2183         iscsit_build_nopin_rsp(cmd, conn, (struct iscsi_nopin *)
2184                                &isert_cmd->tx_desc.iscsi_header,
2185                                nopout_response);
2186         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2187         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2188
2189         isert_dbg("conn %p Posting NOPIN Response\n", isert_conn);
2190
2191         return isert_post_response(isert_conn, isert_cmd);
2192 }
2193
2194 static int
2195 isert_put_logout_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2196 {
2197         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2198         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2199         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2200
2201         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2202         iscsit_build_logout_rsp(cmd, conn, (struct iscsi_logout_rsp *)
2203                                 &isert_cmd->tx_desc.iscsi_header);
2204         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2205         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2206
2207         isert_dbg("conn %p Posting Logout Response\n", isert_conn);
2208
2209         return isert_post_response(isert_conn, isert_cmd);
2210 }
2211
2212 static int
2213 isert_put_tm_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2214 {
2215         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2216         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2217         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2218
2219         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2220         iscsit_build_task_mgt_rsp(cmd, conn, (struct iscsi_tm_rsp *)
2221                                   &isert_cmd->tx_desc.iscsi_header);
2222         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2223         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2224
2225         isert_dbg("conn %p Posting Task Management Response\n", isert_conn);
2226
2227         return isert_post_response(isert_conn, isert_cmd);
2228 }
2229
2230 static int
2231 isert_put_reject(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2232 {
2233         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2234         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2235         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2236         struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
2237         struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2238         struct iscsi_reject *hdr =
2239                 (struct iscsi_reject *)&isert_cmd->tx_desc.iscsi_header;
2240
2241         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2242         iscsit_build_reject(cmd, conn, hdr);
2243         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2244
2245         hton24(hdr->dlength, ISCSI_HDR_LEN);
2246         isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2247                         (void *)cmd->buf_ptr, ISCSI_HDR_LEN,
2248                         DMA_TO_DEVICE);
2249         isert_cmd->pdu_buf_len = ISCSI_HDR_LEN;
2250         tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2251         tx_dsg->length  = ISCSI_HDR_LEN;
2252         tx_dsg->lkey    = isert_conn->conn_mr->lkey;
2253         isert_cmd->tx_desc.num_sge = 2;
2254
2255         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2256
2257         isert_dbg("conn %p Posting Reject\n", isert_conn);
2258
2259         return isert_post_response(isert_conn, isert_cmd);
2260 }
2261
2262 static int
2263 isert_put_text_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2264 {
2265         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2266         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2267         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2268         struct iscsi_text_rsp *hdr =
2269                 (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
2270         u32 txt_rsp_len;
2271         int rc;
2272
2273         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2274         rc = iscsit_build_text_rsp(cmd, conn, hdr, ISCSI_INFINIBAND);
2275         if (rc < 0)
2276                 return rc;
2277
2278         txt_rsp_len = rc;
2279         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2280
2281         if (txt_rsp_len) {
2282                 struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
2283                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2284                 void *txt_rsp_buf = cmd->buf_ptr;
2285
2286                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2287                                 txt_rsp_buf, txt_rsp_len, DMA_TO_DEVICE);
2288
2289                 isert_cmd->pdu_buf_len = txt_rsp_len;
2290                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2291                 tx_dsg->length  = txt_rsp_len;
2292                 tx_dsg->lkey    = isert_conn->conn_mr->lkey;
2293                 isert_cmd->tx_desc.num_sge = 2;
2294         }
2295         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2296
2297         isert_dbg("conn %p Text Response\n", isert_conn);
2298
2299         return isert_post_response(isert_conn, isert_cmd);
2300 }
2301
2302 static int
2303 isert_build_rdma_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
2304                     struct ib_sge *ib_sge, struct ib_send_wr *send_wr,
2305                     u32 data_left, u32 offset)
2306 {
2307         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
2308         struct scatterlist *sg_start, *tmp_sg;
2309         struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
2310         u32 sg_off, page_off;
2311         int i = 0, sg_nents;
2312
2313         sg_off = offset / PAGE_SIZE;
2314         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
2315         sg_nents = min(cmd->se_cmd.t_data_nents - sg_off, isert_conn->max_sge);
2316         page_off = offset % PAGE_SIZE;
2317
2318         send_wr->sg_list = ib_sge;
2319         send_wr->num_sge = sg_nents;
2320         send_wr->wr_id = (uintptr_t)&isert_cmd->tx_desc;
2321         /*
2322          * Perform mapping of TCM scatterlist memory ib_sge dma_addr.
2323          */
2324         for_each_sg(sg_start, tmp_sg, sg_nents, i) {
2325                 isert_dbg("RDMA from SGL dma_addr: 0x%llx dma_len: %u, "
2326                           "page_off: %u\n",
2327                           (unsigned long long)tmp_sg->dma_address,
2328                           tmp_sg->length, page_off);
2329
2330                 ib_sge->addr = ib_sg_dma_address(ib_dev, tmp_sg) + page_off;
2331                 ib_sge->length = min_t(u32, data_left,
2332                                 ib_sg_dma_len(ib_dev, tmp_sg) - page_off);
2333                 ib_sge->lkey = isert_conn->conn_mr->lkey;
2334
2335                 isert_dbg("RDMA ib_sge: addr: 0x%llx  length: %u lkey: %x\n",
2336                           ib_sge->addr, ib_sge->length, ib_sge->lkey);
2337                 page_off = 0;
2338                 data_left -= ib_sge->length;
2339                 ib_sge++;
2340                 isert_dbg("Incrementing ib_sge pointer to %p\n", ib_sge);
2341         }
2342
2343         isert_dbg("Set outgoing sg_list: %p num_sg: %u from TCM SGLs\n",
2344                   send_wr->sg_list, send_wr->num_sge);
2345
2346         return sg_nents;
2347 }
2348
2349 static int
2350 isert_map_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2351                struct isert_rdma_wr *wr)
2352 {
2353         struct se_cmd *se_cmd = &cmd->se_cmd;
2354         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2355         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2356         struct isert_data_buf *data = &wr->data;
2357         struct ib_send_wr *send_wr;
2358         struct ib_sge *ib_sge;
2359         u32 offset, data_len, data_left, rdma_write_max, va_offset = 0;
2360         int ret = 0, i, ib_sge_cnt;
2361
2362         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2363
2364         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2365         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2366                                  se_cmd->t_data_nents, se_cmd->data_length,
2367                                  offset, wr->iser_ib_op, &wr->data);
2368         if (ret)
2369                 return ret;
2370
2371         data_left = data->len;
2372         offset = data->offset;
2373
2374         ib_sge = kzalloc(sizeof(struct ib_sge) * data->nents, GFP_KERNEL);
2375         if (!ib_sge) {
2376                 isert_warn("Unable to allocate ib_sge\n");
2377                 ret = -ENOMEM;
2378                 goto unmap_cmd;
2379         }
2380         wr->ib_sge = ib_sge;
2381
2382         wr->send_wr_num = DIV_ROUND_UP(data->nents, isert_conn->max_sge);
2383         wr->send_wr = kzalloc(sizeof(struct ib_send_wr) * wr->send_wr_num,
2384                                 GFP_KERNEL);
2385         if (!wr->send_wr) {
2386                 isert_dbg("Unable to allocate wr->send_wr\n");
2387                 ret = -ENOMEM;
2388                 goto unmap_cmd;
2389         }
2390
2391         wr->isert_cmd = isert_cmd;
2392         rdma_write_max = isert_conn->max_sge * PAGE_SIZE;
2393
2394         for (i = 0; i < wr->send_wr_num; i++) {
2395                 send_wr = &isert_cmd->rdma_wr.send_wr[i];
2396                 data_len = min(data_left, rdma_write_max);
2397
2398                 send_wr->send_flags = 0;
2399                 if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2400                         send_wr->opcode = IB_WR_RDMA_WRITE;
2401                         send_wr->wr.rdma.remote_addr = isert_cmd->read_va + offset;
2402                         send_wr->wr.rdma.rkey = isert_cmd->read_stag;
2403                         if (i + 1 == wr->send_wr_num)
2404                                 send_wr->next = &isert_cmd->tx_desc.send_wr;
2405                         else
2406                                 send_wr->next = &wr->send_wr[i + 1];
2407                 } else {
2408                         send_wr->opcode = IB_WR_RDMA_READ;
2409                         send_wr->wr.rdma.remote_addr = isert_cmd->write_va + va_offset;
2410                         send_wr->wr.rdma.rkey = isert_cmd->write_stag;
2411                         if (i + 1 == wr->send_wr_num)
2412                                 send_wr->send_flags = IB_SEND_SIGNALED;
2413                         else
2414                                 send_wr->next = &wr->send_wr[i + 1];
2415                 }
2416
2417                 ib_sge_cnt = isert_build_rdma_wr(isert_conn, isert_cmd, ib_sge,
2418                                         send_wr, data_len, offset);
2419                 ib_sge += ib_sge_cnt;
2420
2421                 offset += data_len;
2422                 va_offset += data_len;
2423                 data_left -= data_len;
2424         }
2425
2426         return 0;
2427 unmap_cmd:
2428         isert_unmap_data_buf(isert_conn, data);
2429
2430         return ret;
2431 }
2432
2433 static int
2434 isert_map_fr_pagelist(struct ib_device *ib_dev,
2435                       struct scatterlist *sg_start, int sg_nents, u64 *fr_pl)
2436 {
2437         u64 start_addr, end_addr, page, chunk_start = 0;
2438         struct scatterlist *tmp_sg;
2439         int i = 0, new_chunk, last_ent, n_pages;
2440
2441         n_pages = 0;
2442         new_chunk = 1;
2443         last_ent = sg_nents - 1;
2444         for_each_sg(sg_start, tmp_sg, sg_nents, i) {
2445                 start_addr = ib_sg_dma_address(ib_dev, tmp_sg);
2446                 if (new_chunk)
2447                         chunk_start = start_addr;
2448                 end_addr = start_addr + ib_sg_dma_len(ib_dev, tmp_sg);
2449
2450                 isert_dbg("SGL[%d] dma_addr: 0x%llx len: %u\n",
2451                           i, (unsigned long long)tmp_sg->dma_address,
2452                           tmp_sg->length);
2453
2454                 if ((end_addr & ~PAGE_MASK) && i < last_ent) {
2455                         new_chunk = 0;
2456                         continue;
2457                 }
2458                 new_chunk = 1;
2459
2460                 page = chunk_start & PAGE_MASK;
2461                 do {
2462                         fr_pl[n_pages++] = page;
2463                         isert_dbg("Mapped page_list[%d] page_addr: 0x%llx\n",
2464                                   n_pages - 1, page);
2465                         page += PAGE_SIZE;
2466                 } while (page < end_addr);
2467         }
2468
2469         return n_pages;
2470 }
2471
2472 static inline void
2473 isert_inv_rkey(struct ib_send_wr *inv_wr, struct ib_mr *mr)
2474 {
2475         u32 rkey;
2476
2477         memset(inv_wr, 0, sizeof(*inv_wr));
2478         inv_wr->wr_id = ISER_FASTREG_LI_WRID;
2479         inv_wr->opcode = IB_WR_LOCAL_INV;
2480         inv_wr->ex.invalidate_rkey = mr->rkey;
2481
2482         /* Bump the key */
2483         rkey = ib_inc_rkey(mr->rkey);
2484         ib_update_fast_reg_key(mr, rkey);
2485 }
2486
2487 static int
2488 isert_fast_reg_mr(struct isert_conn *isert_conn,
2489                   struct fast_reg_descriptor *fr_desc,
2490                   struct isert_data_buf *mem,
2491                   enum isert_indicator ind,
2492                   struct ib_sge *sge)
2493 {
2494         struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
2495         struct ib_mr *mr;
2496         struct ib_fast_reg_page_list *frpl;
2497         struct ib_send_wr fr_wr, inv_wr;
2498         struct ib_send_wr *bad_wr, *wr = NULL;
2499         int ret, pagelist_len;
2500         u32 page_off;
2501
2502         if (mem->dma_nents == 1) {
2503                 sge->lkey = isert_conn->conn_mr->lkey;
2504                 sge->addr = ib_sg_dma_address(ib_dev, &mem->sg[0]);
2505                 sge->length = ib_sg_dma_len(ib_dev, &mem->sg[0]);
2506                 isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2507                          sge->addr, sge->length, sge->lkey);
2508                 return 0;
2509         }
2510
2511         if (ind == ISERT_DATA_KEY_VALID) {
2512                 /* Registering data buffer */
2513                 mr = fr_desc->data_mr;
2514                 frpl = fr_desc->data_frpl;
2515         } else {
2516                 /* Registering protection buffer */
2517                 mr = fr_desc->pi_ctx->prot_mr;
2518                 frpl = fr_desc->pi_ctx->prot_frpl;
2519         }
2520
2521         page_off = mem->offset % PAGE_SIZE;
2522
2523         isert_dbg("Use fr_desc %p sg_nents %d offset %u\n",
2524                   fr_desc, mem->nents, mem->offset);
2525
2526         pagelist_len = isert_map_fr_pagelist(ib_dev, mem->sg, mem->nents,
2527                                              &frpl->page_list[0]);
2528
2529         if (!(fr_desc->ind & ind)) {
2530                 isert_inv_rkey(&inv_wr, mr);
2531                 wr = &inv_wr;
2532         }
2533
2534         /* Prepare FASTREG WR */
2535         memset(&fr_wr, 0, sizeof(fr_wr));
2536         fr_wr.wr_id = ISER_FASTREG_LI_WRID;
2537         fr_wr.opcode = IB_WR_FAST_REG_MR;
2538         fr_wr.wr.fast_reg.iova_start = frpl->page_list[0] + page_off;
2539         fr_wr.wr.fast_reg.page_list = frpl;
2540         fr_wr.wr.fast_reg.page_list_len = pagelist_len;
2541         fr_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
2542         fr_wr.wr.fast_reg.length = mem->len;
2543         fr_wr.wr.fast_reg.rkey = mr->rkey;
2544         fr_wr.wr.fast_reg.access_flags = IB_ACCESS_LOCAL_WRITE;
2545
2546         if (!wr)
2547                 wr = &fr_wr;
2548         else
2549                 wr->next = &fr_wr;
2550
2551         ret = ib_post_send(isert_conn->conn_qp, wr, &bad_wr);
2552         if (ret) {
2553                 isert_err("fast registration failed, ret:%d\n", ret);
2554                 return ret;
2555         }
2556         fr_desc->ind &= ~ind;
2557
2558         sge->lkey = mr->lkey;
2559         sge->addr = frpl->page_list[0] + page_off;
2560         sge->length = mem->len;
2561
2562         isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2563                   sge->addr, sge->length, sge->lkey);
2564
2565         return ret;
2566 }
2567
2568 static inline void
2569 isert_set_dif_domain(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs,
2570                      struct ib_sig_domain *domain)
2571 {
2572         domain->sig_type = IB_SIG_TYPE_T10_DIF;
2573         domain->sig.dif.bg_type = IB_T10DIF_CRC;
2574         domain->sig.dif.pi_interval = se_cmd->se_dev->dev_attrib.block_size;
2575         domain->sig.dif.ref_tag = se_cmd->reftag_seed;
2576         /*
2577          * At the moment we hard code those, but if in the future
2578          * the target core would like to use it, we will take it
2579          * from se_cmd.
2580          */
2581         domain->sig.dif.apptag_check_mask = 0xffff;
2582         domain->sig.dif.app_escape = true;
2583         domain->sig.dif.ref_escape = true;
2584         if (se_cmd->prot_type == TARGET_DIF_TYPE1_PROT ||
2585             se_cmd->prot_type == TARGET_DIF_TYPE2_PROT)
2586                 domain->sig.dif.ref_remap = true;
2587 };
2588
2589 static int
2590 isert_set_sig_attrs(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs)
2591 {
2592         switch (se_cmd->prot_op) {
2593         case TARGET_PROT_DIN_INSERT:
2594         case TARGET_PROT_DOUT_STRIP:
2595                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
2596                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2597                 break;
2598         case TARGET_PROT_DOUT_INSERT:
2599         case TARGET_PROT_DIN_STRIP:
2600                 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
2601                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2602                 break;
2603         case TARGET_PROT_DIN_PASS:
2604         case TARGET_PROT_DOUT_PASS:
2605                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2606                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2607                 break;
2608         default:
2609                 isert_err("Unsupported PI operation %d\n", se_cmd->prot_op);
2610                 return -EINVAL;
2611         }
2612
2613         return 0;
2614 }
2615
2616 static inline u8
2617 isert_set_prot_checks(u8 prot_checks)
2618 {
2619         return (prot_checks & TARGET_DIF_CHECK_GUARD  ? 0xc0 : 0) |
2620                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x30 : 0) |
2621                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x0f : 0);
2622 }
2623
2624 static int
2625 isert_reg_sig_mr(struct isert_conn *isert_conn,
2626                  struct se_cmd *se_cmd,
2627                  struct isert_rdma_wr *rdma_wr,
2628                  struct fast_reg_descriptor *fr_desc)
2629 {
2630         struct ib_send_wr sig_wr, inv_wr;
2631         struct ib_send_wr *bad_wr, *wr = NULL;
2632         struct pi_context *pi_ctx = fr_desc->pi_ctx;
2633         struct ib_sig_attrs sig_attrs;
2634         int ret;
2635
2636         memset(&sig_attrs, 0, sizeof(sig_attrs));
2637         ret = isert_set_sig_attrs(se_cmd, &sig_attrs);
2638         if (ret)
2639                 goto err;
2640
2641         sig_attrs.check_mask = isert_set_prot_checks(se_cmd->prot_checks);
2642
2643         if (!(fr_desc->ind & ISERT_SIG_KEY_VALID)) {
2644                 isert_inv_rkey(&inv_wr, pi_ctx->sig_mr);
2645                 wr = &inv_wr;
2646         }
2647
2648         memset(&sig_wr, 0, sizeof(sig_wr));
2649         sig_wr.opcode = IB_WR_REG_SIG_MR;
2650         sig_wr.wr_id = ISER_FASTREG_LI_WRID;
2651         sig_wr.sg_list = &rdma_wr->ib_sg[DATA];
2652         sig_wr.num_sge = 1;
2653         sig_wr.wr.sig_handover.access_flags = IB_ACCESS_LOCAL_WRITE;
2654         sig_wr.wr.sig_handover.sig_attrs = &sig_attrs;
2655         sig_wr.wr.sig_handover.sig_mr = pi_ctx->sig_mr;
2656         if (se_cmd->t_prot_sg)
2657                 sig_wr.wr.sig_handover.prot = &rdma_wr->ib_sg[PROT];
2658
2659         if (!wr)
2660                 wr = &sig_wr;
2661         else
2662                 wr->next = &sig_wr;
2663
2664         ret = ib_post_send(isert_conn->conn_qp, wr, &bad_wr);
2665         if (ret) {
2666                 isert_err("fast registration failed, ret:%d\n", ret);
2667                 goto err;
2668         }
2669         fr_desc->ind &= ~ISERT_SIG_KEY_VALID;
2670
2671         rdma_wr->ib_sg[SIG].lkey = pi_ctx->sig_mr->lkey;
2672         rdma_wr->ib_sg[SIG].addr = 0;
2673         rdma_wr->ib_sg[SIG].length = se_cmd->data_length;
2674         if (se_cmd->prot_op != TARGET_PROT_DIN_STRIP &&
2675             se_cmd->prot_op != TARGET_PROT_DOUT_INSERT)
2676                 /*
2677                  * We have protection guards on the wire
2678                  * so we need to set a larget transfer
2679                  */
2680                 rdma_wr->ib_sg[SIG].length += se_cmd->prot_length;
2681
2682         isert_dbg("sig_sge: addr: 0x%llx  length: %u lkey: %x\n",
2683                   rdma_wr->ib_sg[SIG].addr, rdma_wr->ib_sg[SIG].length,
2684                   rdma_wr->ib_sg[SIG].lkey);
2685 err:
2686         return ret;
2687 }
2688
2689 static int
2690 isert_handle_prot_cmd(struct isert_conn *isert_conn,
2691                       struct isert_cmd *isert_cmd,
2692                       struct isert_rdma_wr *wr)
2693 {
2694         struct isert_device *device = isert_conn->conn_device;
2695         struct se_cmd *se_cmd = &isert_cmd->iscsi_cmd->se_cmd;
2696         int ret;
2697
2698         if (!wr->fr_desc->pi_ctx) {
2699                 ret = isert_create_pi_ctx(wr->fr_desc,
2700                                           device->ib_device,
2701                                           isert_conn->conn_pd);
2702                 if (ret) {
2703                         isert_err("conn %p failed to allocate pi_ctx\n",
2704                                   isert_conn);
2705                         return ret;
2706                 }
2707         }
2708
2709         if (se_cmd->t_prot_sg) {
2710                 ret = isert_map_data_buf(isert_conn, isert_cmd,
2711                                          se_cmd->t_prot_sg,
2712                                          se_cmd->t_prot_nents,
2713                                          se_cmd->prot_length,
2714                                          0, wr->iser_ib_op, &wr->prot);
2715                 if (ret) {
2716                         isert_err("conn %p failed to map protection buffer\n",
2717                                   isert_conn);
2718                         return ret;
2719                 }
2720
2721                 memset(&wr->ib_sg[PROT], 0, sizeof(wr->ib_sg[PROT]));
2722                 ret = isert_fast_reg_mr(isert_conn, wr->fr_desc, &wr->prot,
2723                                         ISERT_PROT_KEY_VALID, &wr->ib_sg[PROT]);
2724                 if (ret) {
2725                         isert_err("conn %p failed to fast reg mr\n",
2726                                   isert_conn);
2727                         goto unmap_prot_cmd;
2728                 }
2729         }
2730
2731         ret = isert_reg_sig_mr(isert_conn, se_cmd, wr, wr->fr_desc);
2732         if (ret) {
2733                 isert_err("conn %p failed to fast reg mr\n",
2734                           isert_conn);
2735                 goto unmap_prot_cmd;
2736         }
2737         wr->fr_desc->ind |= ISERT_PROTECTED;
2738
2739         return 0;
2740
2741 unmap_prot_cmd:
2742         if (se_cmd->t_prot_sg)
2743                 isert_unmap_data_buf(isert_conn, &wr->prot);
2744
2745         return ret;
2746 }
2747
2748 static int
2749 isert_reg_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2750                struct isert_rdma_wr *wr)
2751 {
2752         struct se_cmd *se_cmd = &cmd->se_cmd;
2753         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2754         struct isert_conn *isert_conn = conn->context;
2755         struct fast_reg_descriptor *fr_desc = NULL;
2756         struct ib_send_wr *send_wr;
2757         struct ib_sge *ib_sg;
2758         u32 offset;
2759         int ret = 0;
2760         unsigned long flags;
2761
2762         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2763
2764         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2765         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2766                                  se_cmd->t_data_nents, se_cmd->data_length,
2767                                  offset, wr->iser_ib_op, &wr->data);
2768         if (ret)
2769                 return ret;
2770
2771         if (wr->data.dma_nents != 1 || isert_prot_cmd(isert_conn, se_cmd)) {
2772                 spin_lock_irqsave(&isert_conn->conn_lock, flags);
2773                 fr_desc = list_first_entry(&isert_conn->conn_fr_pool,
2774                                            struct fast_reg_descriptor, list);
2775                 list_del(&fr_desc->list);
2776                 spin_unlock_irqrestore(&isert_conn->conn_lock, flags);
2777                 wr->fr_desc = fr_desc;
2778         }
2779
2780         ret = isert_fast_reg_mr(isert_conn, fr_desc, &wr->data,
2781                                 ISERT_DATA_KEY_VALID, &wr->ib_sg[DATA]);
2782         if (ret)
2783                 goto unmap_cmd;
2784
2785         if (isert_prot_cmd(isert_conn, se_cmd)) {
2786                 ret = isert_handle_prot_cmd(isert_conn, isert_cmd, wr);
2787                 if (ret)
2788                         goto unmap_cmd;
2789
2790                 ib_sg = &wr->ib_sg[SIG];
2791         } else {
2792                 ib_sg = &wr->ib_sg[DATA];
2793         }
2794
2795         memcpy(&wr->s_ib_sge, ib_sg, sizeof(*ib_sg));
2796         wr->ib_sge = &wr->s_ib_sge;
2797         wr->send_wr_num = 1;
2798         memset(&wr->s_send_wr, 0, sizeof(*send_wr));
2799         wr->send_wr = &wr->s_send_wr;
2800         wr->isert_cmd = isert_cmd;
2801
2802         send_wr = &isert_cmd->rdma_wr.s_send_wr;
2803         send_wr->sg_list = &wr->s_ib_sge;
2804         send_wr->num_sge = 1;
2805         send_wr->wr_id = (uintptr_t)&isert_cmd->tx_desc;
2806         if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2807                 send_wr->opcode = IB_WR_RDMA_WRITE;
2808                 send_wr->wr.rdma.remote_addr = isert_cmd->read_va;
2809                 send_wr->wr.rdma.rkey = isert_cmd->read_stag;
2810                 send_wr->send_flags = !isert_prot_cmd(isert_conn, se_cmd) ?
2811                                       0 : IB_SEND_SIGNALED;
2812         } else {
2813                 send_wr->opcode = IB_WR_RDMA_READ;
2814                 send_wr->wr.rdma.remote_addr = isert_cmd->write_va;
2815                 send_wr->wr.rdma.rkey = isert_cmd->write_stag;
2816                 send_wr->send_flags = IB_SEND_SIGNALED;
2817         }
2818
2819         return 0;
2820
2821 unmap_cmd:
2822         if (fr_desc) {
2823                 spin_lock_irqsave(&isert_conn->conn_lock, flags);
2824                 list_add_tail(&fr_desc->list, &isert_conn->conn_fr_pool);
2825                 spin_unlock_irqrestore(&isert_conn->conn_lock, flags);
2826         }
2827         isert_unmap_data_buf(isert_conn, &wr->data);
2828
2829         return ret;
2830 }
2831
2832 static int
2833 isert_put_datain(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2834 {
2835         struct se_cmd *se_cmd = &cmd->se_cmd;
2836         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2837         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2838         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2839         struct isert_device *device = isert_conn->conn_device;
2840         struct ib_send_wr *wr_failed;
2841         int rc;
2842
2843         isert_dbg("Cmd: %p RDMA_WRITE data_length: %u\n",
2844                  isert_cmd, se_cmd->data_length);
2845
2846         wr->iser_ib_op = ISER_IB_RDMA_WRITE;
2847         rc = device->reg_rdma_mem(conn, cmd, wr);
2848         if (rc) {
2849                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2850                 return rc;
2851         }
2852
2853         if (!isert_prot_cmd(isert_conn, se_cmd)) {
2854                 /*
2855                  * Build isert_conn->tx_desc for iSCSI response PDU and attach
2856                  */
2857                 isert_create_send_desc(isert_conn, isert_cmd,
2858                                        &isert_cmd->tx_desc);
2859                 iscsit_build_rsp_pdu(cmd, conn, true, (struct iscsi_scsi_rsp *)
2860                                      &isert_cmd->tx_desc.iscsi_header);
2861                 isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2862                 isert_init_send_wr(isert_conn, isert_cmd,
2863                                    &isert_cmd->tx_desc.send_wr);
2864                 isert_cmd->rdma_wr.s_send_wr.next = &isert_cmd->tx_desc.send_wr;
2865                 wr->send_wr_num += 1;
2866         }
2867
2868         rc = ib_post_send(isert_conn->conn_qp, wr->send_wr, &wr_failed);
2869         if (rc)
2870                 isert_warn("ib_post_send() failed for IB_WR_RDMA_WRITE\n");
2871
2872         if (!isert_prot_cmd(isert_conn, se_cmd))
2873                 isert_dbg("Cmd: %p posted RDMA_WRITE + Response for iSER Data "
2874                          "READ\n", isert_cmd);
2875         else
2876                 isert_dbg("Cmd: %p posted RDMA_WRITE for iSER Data READ\n",
2877                          isert_cmd);
2878
2879         return 1;
2880 }
2881
2882 static int
2883 isert_get_dataout(struct iscsi_conn *conn, struct iscsi_cmd *cmd, bool recovery)
2884 {
2885         struct se_cmd *se_cmd = &cmd->se_cmd;
2886         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2887         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2888         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2889         struct isert_device *device = isert_conn->conn_device;
2890         struct ib_send_wr *wr_failed;
2891         int rc;
2892
2893         isert_dbg("Cmd: %p RDMA_READ data_length: %u write_data_done: %u\n",
2894                  isert_cmd, se_cmd->data_length, cmd->write_data_done);
2895         wr->iser_ib_op = ISER_IB_RDMA_READ;
2896         rc = device->reg_rdma_mem(conn, cmd, wr);
2897         if (rc) {
2898                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2899                 return rc;
2900         }
2901
2902         rc = ib_post_send(isert_conn->conn_qp, wr->send_wr, &wr_failed);
2903         if (rc)
2904                 isert_warn("ib_post_send() failed for IB_WR_RDMA_READ\n");
2905
2906         isert_dbg("Cmd: %p posted RDMA_READ memory for ISER Data WRITE\n",
2907                  isert_cmd);
2908
2909         return 0;
2910 }
2911
2912 static int
2913 isert_immediate_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2914 {
2915         int ret;
2916
2917         switch (state) {
2918         case ISTATE_SEND_NOPIN_WANT_RESPONSE:
2919                 ret = isert_put_nopin(cmd, conn, false);
2920                 break;
2921         default:
2922                 isert_err("Unknown immediate state: 0x%02x\n", state);
2923                 ret = -EINVAL;
2924                 break;
2925         }
2926
2927         return ret;
2928 }
2929
2930 static int
2931 isert_response_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2932 {
2933         struct isert_conn *isert_conn = conn->context;
2934         int ret;
2935
2936         switch (state) {
2937         case ISTATE_SEND_LOGOUTRSP:
2938                 ret = isert_put_logout_rsp(cmd, conn);
2939                 if (!ret)
2940                         isert_conn->logout_posted = true;
2941                 break;
2942         case ISTATE_SEND_NOPIN:
2943                 ret = isert_put_nopin(cmd, conn, true);
2944                 break;
2945         case ISTATE_SEND_TASKMGTRSP:
2946                 ret = isert_put_tm_rsp(cmd, conn);
2947                 break;
2948         case ISTATE_SEND_REJECT:
2949                 ret = isert_put_reject(cmd, conn);
2950                 break;
2951         case ISTATE_SEND_TEXTRSP:
2952                 ret = isert_put_text_rsp(cmd, conn);
2953                 break;
2954         case ISTATE_SEND_STATUS:
2955                 /*
2956                  * Special case for sending non GOOD SCSI status from TX thread
2957                  * context during pre se_cmd excecution failure.
2958                  */
2959                 ret = isert_put_response(conn, cmd);
2960                 break;
2961         default:
2962                 isert_err("Unknown response state: 0x%02x\n", state);
2963                 ret = -EINVAL;
2964                 break;
2965         }
2966
2967         return ret;
2968 }
2969
2970 struct rdma_cm_id *
2971 isert_setup_id(struct isert_np *isert_np)
2972 {
2973         struct iscsi_np *np = isert_np->np;
2974         struct rdma_cm_id *id;
2975         struct sockaddr *sa;
2976         int ret;
2977
2978         sa = (struct sockaddr *)&np->np_sockaddr;
2979         isert_dbg("ksockaddr: %p, sa: %p\n", &np->np_sockaddr, sa);
2980
2981         id = rdma_create_id(isert_cma_handler, isert_np,
2982                             RDMA_PS_TCP, IB_QPT_RC);
2983         if (IS_ERR(id)) {
2984                 isert_err("rdma_create_id() failed: %ld\n", PTR_ERR(id));
2985                 ret = PTR_ERR(id);
2986                 goto out;
2987         }
2988         isert_dbg("id %p context %p\n", id, id->context);
2989
2990         ret = rdma_bind_addr(id, sa);
2991         if (ret) {
2992                 isert_err("rdma_bind_addr() failed: %d\n", ret);
2993                 goto out_id;
2994         }
2995
2996         ret = rdma_listen(id, ISERT_RDMA_LISTEN_BACKLOG);
2997         if (ret) {
2998                 isert_err("rdma_listen() failed: %d\n", ret);
2999                 goto out_id;
3000         }
3001
3002         return id;
3003 out_id:
3004         rdma_destroy_id(id);
3005 out:
3006         return ERR_PTR(ret);
3007 }
3008
3009 static int
3010 isert_setup_np(struct iscsi_np *np,
3011                struct __kernel_sockaddr_storage *ksockaddr)
3012 {
3013         struct isert_np *isert_np;
3014         struct rdma_cm_id *isert_lid;
3015         int ret;
3016
3017         isert_np = kzalloc(sizeof(struct isert_np), GFP_KERNEL);
3018         if (!isert_np) {
3019                 isert_err("Unable to allocate struct isert_np\n");
3020                 return -ENOMEM;
3021         }
3022         sema_init(&isert_np->np_sem, 0);
3023         mutex_init(&isert_np->np_accept_mutex);
3024         INIT_LIST_HEAD(&isert_np->np_accept_list);
3025         init_completion(&isert_np->np_login_comp);
3026         isert_np->np = np;
3027
3028         /*
3029          * Setup the np->np_sockaddr from the passed sockaddr setup
3030          * in iscsi_target_configfs.c code..
3031          */
3032         memcpy(&np->np_sockaddr, ksockaddr,
3033                sizeof(struct __kernel_sockaddr_storage));
3034
3035         isert_lid = isert_setup_id(isert_np);
3036         if (IS_ERR(isert_lid)) {
3037                 ret = PTR_ERR(isert_lid);
3038                 goto out;
3039         }
3040
3041         isert_np->np_cm_id = isert_lid;
3042         np->np_context = isert_np;
3043
3044         return 0;
3045
3046 out:
3047         kfree(isert_np);
3048
3049         return ret;
3050 }
3051
3052 static int
3053 isert_rdma_accept(struct isert_conn *isert_conn)
3054 {
3055         struct rdma_cm_id *cm_id = isert_conn->conn_cm_id;
3056         struct rdma_conn_param cp;
3057         int ret;
3058
3059         memset(&cp, 0, sizeof(struct rdma_conn_param));
3060         cp.initiator_depth = isert_conn->initiator_depth;
3061         cp.retry_count = 7;
3062         cp.rnr_retry_count = 7;
3063
3064         ret = rdma_accept(cm_id, &cp);
3065         if (ret) {
3066                 isert_err("rdma_accept() failed with: %d\n", ret);
3067                 return ret;
3068         }
3069
3070         return 0;
3071 }
3072
3073 static int
3074 isert_get_login_rx(struct iscsi_conn *conn, struct iscsi_login *login)
3075 {
3076         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
3077         int ret;
3078
3079         isert_info("before login_req comp conn: %p\n", isert_conn);
3080         ret = wait_for_completion_interruptible(&isert_conn->login_req_comp);
3081         if (ret) {
3082                 isert_err("isert_conn %p interrupted before got login req\n",
3083                           isert_conn);
3084                 return ret;
3085         }
3086         reinit_completion(&isert_conn->login_req_comp);
3087
3088         /*
3089          * For login requests after the first PDU, isert_rx_login_req() will
3090          * kick schedule_delayed_work(&conn->login_work) as the packet is
3091          * received, which turns this callback from iscsi_target_do_login_rx()
3092          * into a NOP.
3093          */
3094         if (!login->first_request)
3095                 return 0;
3096
3097         isert_rx_login_req(isert_conn);
3098
3099         isert_info("before conn_login_comp conn: %p\n", conn);
3100         ret = wait_for_completion_interruptible(&isert_conn->conn_login_comp);
3101         if (ret)
3102                 return ret;
3103
3104         isert_info("processing login->req: %p\n", login->req);
3105
3106         return 0;
3107 }
3108
3109 static void
3110 isert_set_conn_info(struct iscsi_np *np, struct iscsi_conn *conn,
3111                     struct isert_conn *isert_conn)
3112 {
3113         struct rdma_cm_id *cm_id = isert_conn->conn_cm_id;
3114         struct rdma_route *cm_route = &cm_id->route;
3115         struct sockaddr_in *sock_in;
3116         struct sockaddr_in6 *sock_in6;
3117
3118         conn->login_family = np->np_sockaddr.ss_family;
3119
3120         if (np->np_sockaddr.ss_family == AF_INET6) {
3121                 sock_in6 = (struct sockaddr_in6 *)&cm_route->addr.dst_addr;
3122                 snprintf(conn->login_ip, sizeof(conn->login_ip), "%pI6c",
3123                          &sock_in6->sin6_addr.in6_u);
3124                 conn->login_port = ntohs(sock_in6->sin6_port);
3125
3126                 sock_in6 = (struct sockaddr_in6 *)&cm_route->addr.src_addr;
3127                 snprintf(conn->local_ip, sizeof(conn->local_ip), "%pI6c",
3128                          &sock_in6->sin6_addr.in6_u);
3129                 conn->local_port = ntohs(sock_in6->sin6_port);
3130         } else {
3131                 sock_in = (struct sockaddr_in *)&cm_route->addr.dst_addr;
3132                 sprintf(conn->login_ip, "%pI4",
3133                         &sock_in->sin_addr.s_addr);
3134                 conn->login_port = ntohs(sock_in->sin_port);
3135
3136                 sock_in = (struct sockaddr_in *)&cm_route->addr.src_addr;
3137                 sprintf(conn->local_ip, "%pI4",
3138                         &sock_in->sin_addr.s_addr);
3139                 conn->local_port = ntohs(sock_in->sin_port);
3140         }
3141 }
3142
3143 static int
3144 isert_accept_np(struct iscsi_np *np, struct iscsi_conn *conn)
3145 {
3146         struct isert_np *isert_np = (struct isert_np *)np->np_context;
3147         struct isert_conn *isert_conn;
3148         int max_accept = 0, ret;
3149
3150 accept_wait:
3151         ret = down_interruptible(&isert_np->np_sem);
3152         if (ret || max_accept > 5)
3153                 return -ENODEV;
3154
3155         spin_lock_bh(&np->np_thread_lock);
3156         if (np->np_thread_state >= ISCSI_NP_THREAD_RESET) {
3157                 spin_unlock_bh(&np->np_thread_lock);
3158                 isert_dbg("np_thread_state %d\n",
3159                          np->np_thread_state);
3160                 /**
3161                  * No point in stalling here when np_thread
3162                  * is in state RESET/SHUTDOWN/EXIT - bail
3163                  **/
3164                 return -ENODEV;
3165         }
3166         spin_unlock_bh(&np->np_thread_lock);
3167
3168         mutex_lock(&isert_np->np_accept_mutex);
3169         if (list_empty(&isert_np->np_accept_list)) {
3170                 mutex_unlock(&isert_np->np_accept_mutex);
3171                 max_accept++;
3172                 goto accept_wait;
3173         }
3174         isert_conn = list_first_entry(&isert_np->np_accept_list,
3175                         struct isert_conn, conn_accept_node);
3176         list_del_init(&isert_conn->conn_accept_node);
3177         mutex_unlock(&isert_np->np_accept_mutex);
3178
3179         conn->context = isert_conn;
3180         isert_conn->conn = conn;
3181         max_accept = 0;
3182
3183         isert_set_conn_info(np, conn, isert_conn);
3184
3185         isert_dbg("Processing isert_conn: %p\n", isert_conn);
3186
3187         return 0;
3188 }
3189
3190 static void
3191 isert_free_np(struct iscsi_np *np)
3192 {
3193         struct isert_np *isert_np = (struct isert_np *)np->np_context;
3194         struct isert_conn *isert_conn, *n;
3195
3196         if (isert_np->np_cm_id)
3197                 rdma_destroy_id(isert_np->np_cm_id);
3198
3199         /*
3200          * FIXME: At this point we don't have a good way to insure
3201          * that at this point we don't have hanging connections that
3202          * completed RDMA establishment but didn't start iscsi login
3203          * process. So work-around this by cleaning up what ever piled
3204          * up in np_accept_list.
3205          */
3206         mutex_lock(&isert_np->np_accept_mutex);
3207         if (!list_empty(&isert_np->np_accept_list)) {
3208                 isert_info("Still have isert connections, cleaning up...\n");
3209                 list_for_each_entry_safe(isert_conn, n,
3210                                          &isert_np->np_accept_list,
3211                                          conn_accept_node) {
3212                         isert_info("cleaning isert_conn %p state (%d)\n",
3213                                    isert_conn, isert_conn->state);
3214                         isert_connect_release(isert_conn);
3215                 }
3216         }
3217         mutex_unlock(&isert_np->np_accept_mutex);
3218
3219         np->np_context = NULL;
3220         kfree(isert_np);
3221 }
3222
3223 static void isert_release_work(struct work_struct *work)
3224 {
3225         struct isert_conn *isert_conn = container_of(work,
3226                                                      struct isert_conn,
3227                                                      release_work);
3228
3229         isert_info("Starting release conn %p\n", isert_conn);
3230
3231         wait_for_completion(&isert_conn->conn_wait);
3232
3233         mutex_lock(&isert_conn->conn_mutex);
3234         isert_conn->state = ISER_CONN_DOWN;
3235         mutex_unlock(&isert_conn->conn_mutex);
3236
3237         isert_info("Destroying conn %p\n", isert_conn);
3238         isert_put_conn(isert_conn);
3239 }
3240
3241 static void
3242 isert_wait4logout(struct isert_conn *isert_conn)
3243 {
3244         struct iscsi_conn *conn = isert_conn->conn;
3245
3246         isert_info("conn %p\n", isert_conn);
3247
3248         if (isert_conn->logout_posted) {
3249                 isert_info("conn %p wait for conn_logout_comp\n", isert_conn);
3250                 wait_for_completion_timeout(&conn->conn_logout_comp,
3251                                             SECONDS_FOR_LOGOUT_COMP * HZ);
3252         }
3253 }
3254
3255 static void
3256 isert_wait4cmds(struct iscsi_conn *conn)
3257 {
3258         isert_info("iscsi_conn %p\n", conn);
3259
3260         if (conn->sess) {
3261                 target_sess_cmd_list_set_waiting(conn->sess->se_sess);
3262                 target_wait_for_sess_cmds(conn->sess->se_sess);
3263         }
3264 }
3265
3266 static void
3267 isert_wait4flush(struct isert_conn *isert_conn)
3268 {
3269         struct ib_recv_wr *bad_wr;
3270
3271         isert_info("conn %p\n", isert_conn);
3272
3273         init_completion(&isert_conn->conn_wait_comp_err);
3274         isert_conn->beacon.wr_id = ISER_BEACON_WRID;
3275         /* post an indication that all flush errors were consumed */
3276         if (ib_post_recv(isert_conn->conn_qp, &isert_conn->beacon, &bad_wr)) {
3277                 isert_err("conn %p failed to post beacon", isert_conn);
3278                 return;
3279         }
3280
3281         wait_for_completion(&isert_conn->conn_wait_comp_err);
3282 }
3283
3284 static void isert_wait_conn(struct iscsi_conn *conn)
3285 {
3286         struct isert_conn *isert_conn = conn->context;
3287
3288         isert_info("Starting conn %p\n", isert_conn);
3289
3290         mutex_lock(&isert_conn->conn_mutex);
3291         /*
3292          * Only wait for conn_wait_comp_err if the isert_conn made it
3293          * into full feature phase..
3294          */
3295         if (isert_conn->state == ISER_CONN_INIT) {
3296                 mutex_unlock(&isert_conn->conn_mutex);
3297                 return;
3298         }
3299         isert_conn_terminate(isert_conn);
3300         mutex_unlock(&isert_conn->conn_mutex);
3301
3302         isert_wait4cmds(conn);
3303         isert_wait4flush(isert_conn);
3304         isert_wait4logout(isert_conn);
3305
3306         INIT_WORK(&isert_conn->release_work, isert_release_work);
3307         queue_work(isert_release_wq, &isert_conn->release_work);
3308 }
3309
3310 static void isert_free_conn(struct iscsi_conn *conn)
3311 {
3312         struct isert_conn *isert_conn = conn->context;
3313
3314         isert_put_conn(isert_conn);
3315 }
3316
3317 static struct iscsit_transport iser_target_transport = {
3318         .name                   = "IB/iSER",
3319         .transport_type         = ISCSI_INFINIBAND,
3320         .priv_size              = sizeof(struct isert_cmd),
3321         .owner                  = THIS_MODULE,
3322         .iscsit_setup_np        = isert_setup_np,
3323         .iscsit_accept_np       = isert_accept_np,
3324         .iscsit_free_np         = isert_free_np,
3325         .iscsit_wait_conn       = isert_wait_conn,
3326         .iscsit_free_conn       = isert_free_conn,
3327         .iscsit_get_login_rx    = isert_get_login_rx,
3328         .iscsit_put_login_tx    = isert_put_login_tx,
3329         .iscsit_immediate_queue = isert_immediate_queue,
3330         .iscsit_response_queue  = isert_response_queue,
3331         .iscsit_get_dataout     = isert_get_dataout,
3332         .iscsit_queue_data_in   = isert_put_datain,
3333         .iscsit_queue_status    = isert_put_response,
3334         .iscsit_aborted_task    = isert_aborted_task,
3335         .iscsit_get_sup_prot_ops = isert_get_sup_prot_ops,
3336 };
3337
3338 static int __init isert_init(void)
3339 {
3340         int ret;
3341
3342         isert_comp_wq = alloc_workqueue("isert_comp_wq",
3343                                         WQ_UNBOUND | WQ_HIGHPRI, 0);
3344         if (!isert_comp_wq) {
3345                 isert_err("Unable to allocate isert_comp_wq\n");
3346                 ret = -ENOMEM;
3347                 return -ENOMEM;
3348         }
3349
3350         isert_release_wq = alloc_workqueue("isert_release_wq", WQ_UNBOUND,
3351                                         WQ_UNBOUND_MAX_ACTIVE);
3352         if (!isert_release_wq) {
3353                 isert_err("Unable to allocate isert_release_wq\n");
3354                 ret = -ENOMEM;
3355                 goto destroy_comp_wq;
3356         }
3357
3358         iscsit_register_transport(&iser_target_transport);
3359         isert_info("iSER_TARGET[0] - Loaded iser_target_transport\n");
3360
3361         return 0;
3362
3363 destroy_comp_wq:
3364         destroy_workqueue(isert_comp_wq);
3365
3366         return ret;
3367 }
3368
3369 static void __exit isert_exit(void)
3370 {
3371         flush_scheduled_work();
3372         destroy_workqueue(isert_release_wq);
3373         destroy_workqueue(isert_comp_wq);
3374         iscsit_unregister_transport(&iser_target_transport);
3375         isert_info("iSER_TARGET[0] - Released iser_target_transport\n");
3376 }
3377
3378 MODULE_DESCRIPTION("iSER-Target for mainline target infrastructure");
3379 MODULE_VERSION("0.1");
3380 MODULE_AUTHOR("nab@Linux-iSCSI.org");
3381 MODULE_LICENSE("GPL");
3382
3383 module_init(isert_init);
3384 module_exit(isert_exit);