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