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, device->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 isert_device *device = isert_conn->conn_device;
 178         struct ib_device *ib_dev = device->ib_device;
 179         struct iser_rx_desc *rx_desc;
 180         struct ib_sge *rx_sg;
 181         u64 dma_addr;
 182         int i, j;
 183
 184         isert_conn->conn_rx_descs = kzalloc(ISERT_QP_MAX_RECV_DTOS *
 185                                 sizeof(struct iser_rx_desc), GFP_KERNEL);
 186         if (!isert_conn->conn_rx_descs)
 187                 goto fail;
 188
 189         rx_desc = isert_conn->conn_rx_descs;
 190
 191         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 192                 dma_addr = ib_dma_map_single(ib_dev, (void *)rx_desc,
 193                                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 194                 if (ib_dma_mapping_error(ib_dev, dma_addr))
 195                         goto dma_map_fail;
 196
 197                 rx_desc->dma_addr = dma_addr;
 198
 199                 rx_sg = &rx_desc->rx_sg;
 200                 rx_sg->addr = rx_desc->dma_addr;
 201                 rx_sg->length = ISER_RX_PAYLOAD_SIZE;
 202                 rx_sg->lkey = device->mr->lkey;
 203         }
 204
 205         isert_conn->conn_rx_desc_head = 0;
 206
 207         return 0;
 208
 209 dma_map_fail:
 210         rx_desc = isert_conn->conn_rx_descs;
 211         for (j = 0; j < i; j++, rx_desc++) {
 212                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 213                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 214         }
 215         kfree(isert_conn->conn_rx_descs);
 216         isert_conn->conn_rx_descs = NULL;
 217 fail:
 218         isert_err("conn %p failed to allocate rx descriptors\n", isert_conn);
 219
 220         return -ENOMEM;
 221 }
 222
 223 static void
 224 isert_free_rx_descriptors(struct isert_conn *isert_conn)
 225 {
 226         struct ib_device *ib_dev = isert_conn->conn_device->ib_device;
 227         struct iser_rx_desc *rx_desc;
 228         int i;
 229
 230         if (!isert_conn->conn_rx_descs)
 231                 return;
 232
 233         rx_desc = isert_conn->conn_rx_descs;
 234         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 235                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 236                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 237         }
 238
 239         kfree(isert_conn->conn_rx_descs);
 240         isert_conn->conn_rx_descs = NULL;
 241 }
 242
 243 static void isert_cq_work(struct work_struct *);
 244 static void isert_cq_callback(struct ib_cq *, void *);
 245
 246 static int
 247 isert_create_device_ib_res(struct isert_device *device)
 248 {
 249         struct ib_device *ib_dev = device->ib_device;
 250         struct ib_device_attr *dev_attr;
 251         int ret = 0, i;
 252         int max_cqe;
 253
 254         dev_attr = &device->dev_attr;
 255         ret = isert_query_device(ib_dev, dev_attr);
 256         if (ret)
 257                 return ret;
 258
 259         max_cqe = min(ISER_MAX_CQ_LEN, dev_attr->max_cqe);
 260
 261         /* asign function handlers */
 262         if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS &&
 263             dev_attr->device_cap_flags & IB_DEVICE_SIGNATURE_HANDOVER) {
 264                 device->use_fastreg = 1;
 265                 device->reg_rdma_mem = isert_reg_rdma;
 266                 device->unreg_rdma_mem = isert_unreg_rdma;
 267         } else {
 268                 device->use_fastreg = 0;
 269                 device->reg_rdma_mem = isert_map_rdma;
 270                 device->unreg_rdma_mem = isert_unmap_cmd;
 271         }
 272
 273         /* Check signature cap */
 274         device->pi_capable = dev_attr->device_cap_flags &
 275                              IB_DEVICE_SIGNATURE_HANDOVER ? true : false;
 276
 277         device->comps_used = min(ISERT_MAX_CQ, min_t(int, num_online_cpus(),
 278                                         device->ib_device->num_comp_vectors));
 279         isert_info("Using %d CQs, %s supports %d vectors support "
 280                    "Fast registration %d pi_capable %d\n",
 281                    device->comps_used, device->ib_device->name,
 282                    device->ib_device->num_comp_vectors, device->use_fastreg,
 283                    device->pi_capable);
 284
 285         device->comps = kcalloc(device->comps_used, sizeof(struct isert_comp),
 286                                 GFP_KERNEL);
 287         if (!device->comps) {
 288                 isert_err("Unable to allocate completion contexts\n");
 289                 return -ENOMEM;
 290         }
 291
 292         for (i = 0; i < device->comps_used; i++) {
 293                 struct isert_comp *comp = &device->comps[i];
 294
 295                 comp->device = device;
 296                 INIT_WORK(&comp->work, isert_cq_work);
 297                 comp->cq = ib_create_cq(device->ib_device,
 298                                         isert_cq_callback,
 299                                         isert_cq_event_callback,
 300                                         (void *)comp,
 301                                         max_cqe, i);
 302                 if (IS_ERR(comp->cq)) {
 303                         ret = PTR_ERR(comp->cq);
 304                         comp->cq = NULL;
 305                         goto out_cq;
 306                 }
 307
 308                 ret = ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP);
 309                 if (ret)
 310                         goto out_cq;
 311         }
 312
 313         device->pd = ib_alloc_pd(device->ib_device);
 314         if (IS_ERR(device->pd)) {
 315                 ret = PTR_ERR(device->pd);
 316                 isert_err("failed to allocate pd, device %p, ret=%d\n",
 317                           device, ret);
 318                 goto out_cq;
 319         }
 320
 321         device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE);
 322         if (IS_ERR(device->mr)) {
 323                 ret = PTR_ERR(device->mr);
 324                 isert_err("failed to create dma mr, device %p, ret=%d\n",
 325                           device, ret);
 326                 goto out_mr;
 327         }
 328
 329
 330         return 0;
 331
 332 out_mr:
 333         ib_dealloc_pd(device->pd);
 334 out_cq:
 335         for (i = 0; i < device->comps_used; i++) {
 336                 struct isert_comp *comp = &device->comps[i];
 337
 338                 if (comp->cq) {
 339                         cancel_work_sync(&comp->work);
 340                         ib_destroy_cq(comp->cq);
 341                 }
 342         }
 343         kfree(device->comps);
 344
 345         return ret;
 346 }
 347
 348 static void
 349 isert_free_device_ib_res(struct isert_device *device)
 350 {
 351         int i;
 352
 353         isert_info("device %p\n", device);
 354
 355         ib_dereg_mr(device->mr);
 356         ib_dealloc_pd(device->pd);
 357         for (i = 0; i < device->comps_used; i++) {
 358                 struct isert_comp *comp = &device->comps[i];
 359
 360                 cancel_work_sync(&comp->work);
 361                 ib_destroy_cq(comp->cq);
 362                 comp->cq = NULL;
 363         }
 364         kfree(device->comps);
 365 }
 366
 367 static void
 368 isert_device_try_release(struct isert_device *device)
 369 {
 370         mutex_lock(&device_list_mutex);
 371         device->refcount--;
 372         isert_info("device %p refcount %d\n", device, device->refcount);
 373         if (!device->refcount) {
 374                 isert_free_device_ib_res(device);
 375                 list_del(&device->dev_node);
 376                 kfree(device);
 377         }
 378         mutex_unlock(&device_list_mutex);
 379 }
 380
 381 static struct isert_device *
 382 isert_device_find_by_ib_dev(struct rdma_cm_id *cma_id)
 383 {
 384         struct isert_device *device;
 385         int ret;
 386
 387         mutex_lock(&device_list_mutex);
 388         list_for_each_entry(device, &device_list, dev_node) {
 389                 if (device->ib_device->node_guid == cma_id->device->node_guid) {
 390                         device->refcount++;
 391                         isert_info("Found iser device %p refcount %d\n",
 392                                    device, device->refcount);
 393                         mutex_unlock(&device_list_mutex);
 394                         return device;
 395                 }
 396         }
 397
 398         device = kzalloc(sizeof(struct isert_device), GFP_KERNEL);
 399         if (!device) {
 400                 mutex_unlock(&device_list_mutex);
 401                 return ERR_PTR(-ENOMEM);
 402         }
 403
 404         INIT_LIST_HEAD(&device->dev_node);
 405
 406         device->ib_device = cma_id->device;
 407         ret = isert_create_device_ib_res(device);
 408         if (ret) {
 409                 kfree(device);
 410                 mutex_unlock(&device_list_mutex);
 411                 return ERR_PTR(ret);
 412         }
 413
 414         device->refcount++;
 415         list_add_tail(&device->dev_node, &device_list);
 416         isert_info("Created a new iser device %p refcount %d\n",
 417                    device, device->refcount);
 418         mutex_unlock(&device_list_mutex);
 419
 420         return device;
 421 }
 422
 423 static void
 424 isert_conn_free_fastreg_pool(struct isert_conn *isert_conn)
 425 {
 426         struct fast_reg_descriptor *fr_desc, *tmp;
 427         int i = 0;
 428
 429         if (list_empty(&isert_conn->conn_fr_pool))
 430                 return;
 431
 432         isert_info("Freeing conn %p fastreg pool", isert_conn);
 433
 434         list_for_each_entry_safe(fr_desc, tmp,
 435                                  &isert_conn->conn_fr_pool, list) {
 436                 list_del(&fr_desc->list);
 437                 ib_free_fast_reg_page_list(fr_desc->data_frpl);
 438                 ib_dereg_mr(fr_desc->data_mr);
 439                 if (fr_desc->pi_ctx) {
 440                         ib_free_fast_reg_page_list(fr_desc->pi_ctx->prot_frpl);
 441                         ib_dereg_mr(fr_desc->pi_ctx->prot_mr);
 442                         ib_destroy_mr(fr_desc->pi_ctx->sig_mr);
 443                         kfree(fr_desc->pi_ctx);
 444                 }
 445                 kfree(fr_desc);
 446                 ++i;
 447         }
 448
 449         if (i < isert_conn->conn_fr_pool_size)
 450                 isert_warn("Pool still has %d regions registered\n",
 451                         isert_conn->conn_fr_pool_size - i);
 452 }
 453
 454 static int
 455 isert_create_pi_ctx(struct fast_reg_descriptor *desc,
 456                     struct ib_device *device,
 457                     struct ib_pd *pd)
 458 {
 459         struct ib_mr_init_attr mr_init_attr;
 460         struct pi_context *pi_ctx;
 461         int ret;
 462
 463         pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
 464         if (!pi_ctx) {
 465                 isert_err("Failed to allocate pi context\n");
 466                 return -ENOMEM;
 467         }
 468
 469         pi_ctx->prot_frpl = ib_alloc_fast_reg_page_list(device,
 470                                             ISCSI_ISER_SG_TABLESIZE);
 471         if (IS_ERR(pi_ctx->prot_frpl)) {
 472                 isert_err("Failed to allocate prot frpl err=%ld\n",
 473                           PTR_ERR(pi_ctx->prot_frpl));
 474                 ret = PTR_ERR(pi_ctx->prot_frpl);
 475                 goto err_pi_ctx;
 476         }
 477
 478         pi_ctx->prot_mr = ib_alloc_fast_reg_mr(pd, ISCSI_ISER_SG_TABLESIZE);
 479         if (IS_ERR(pi_ctx->prot_mr)) {
 480                 isert_err("Failed to allocate prot frmr err=%ld\n",
 481                           PTR_ERR(pi_ctx->prot_mr));
 482                 ret = PTR_ERR(pi_ctx->prot_mr);
 483                 goto err_prot_frpl;
 484         }
 485         desc->ind |= ISERT_PROT_KEY_VALID;
 486
 487         memset(&mr_init_attr, 0, sizeof(mr_init_attr));
 488         mr_init_attr.max_reg_descriptors = 2;
 489         mr_init_attr.flags |= IB_MR_SIGNATURE_EN;
 490         pi_ctx->sig_mr = ib_create_mr(pd, &mr_init_attr);
 491         if (IS_ERR(pi_ctx->sig_mr)) {
 492                 isert_err("Failed to allocate signature enabled mr err=%ld\n",
 493                           PTR_ERR(pi_ctx->sig_mr));
 494                 ret = PTR_ERR(pi_ctx->sig_mr);
 495                 goto err_prot_mr;
 496         }
 497
 498         desc->pi_ctx = pi_ctx;
 499         desc->ind |= ISERT_SIG_KEY_VALID;
 500         desc->ind &= ~ISERT_PROTECTED;
 501
 502         return 0;
 503
 504 err_prot_mr:
 505         ib_dereg_mr(desc->pi_ctx->prot_mr);
 506 err_prot_frpl:
 507         ib_free_fast_reg_page_list(desc->pi_ctx->prot_frpl);
 508 err_pi_ctx:
 509         kfree(desc->pi_ctx);
 510
 511         return ret;
 512 }
 513
 514 static int
 515 isert_create_fr_desc(struct ib_device *ib_device, struct ib_pd *pd,
 516                      struct fast_reg_descriptor *fr_desc)
 517 {
 518         int ret;
 519
 520         fr_desc->data_frpl = ib_alloc_fast_reg_page_list(ib_device,
 521                                                          ISCSI_ISER_SG_TABLESIZE);
 522         if (IS_ERR(fr_desc->data_frpl)) {
 523                 isert_err("Failed to allocate data frpl err=%ld\n",
 524                           PTR_ERR(fr_desc->data_frpl));
 525                 return PTR_ERR(fr_desc->data_frpl);
 526         }
 527
 528         fr_desc->data_mr = ib_alloc_fast_reg_mr(pd, ISCSI_ISER_SG_TABLESIZE);
 529         if (IS_ERR(fr_desc->data_mr)) {
 530                 isert_err("Failed to allocate data frmr err=%ld\n",
 531                           PTR_ERR(fr_desc->data_mr));
 532                 ret = PTR_ERR(fr_desc->data_mr);
 533                 goto err_data_frpl;
 534         }
 535         fr_desc->ind |= ISERT_DATA_KEY_VALID;
 536
 537         isert_dbg("Created fr_desc %p\n", fr_desc);
 538
 539         return 0;
 540
 541 err_data_frpl:
 542         ib_free_fast_reg_page_list(fr_desc->data_frpl);
 543
 544         return ret;
 545 }
 546
 547 static int
 548 isert_conn_create_fastreg_pool(struct isert_conn *isert_conn)
 549 {
 550         struct fast_reg_descriptor *fr_desc;
 551         struct isert_device *device = isert_conn->conn_device;
 552         struct se_session *se_sess = isert_conn->conn->sess->se_sess;
 553         struct se_node_acl *se_nacl = se_sess->se_node_acl;
 554         int i, ret, tag_num;
 555         /*
 556          * Setup the number of FRMRs based upon the number of tags
 557          * available to session in iscsi_target_locate_portal().
 558          */
 559         tag_num = max_t(u32, ISCSIT_MIN_TAGS, se_nacl->queue_depth);
 560         tag_num = (tag_num * 2) + ISCSIT_EXTRA_TAGS;
 561
 562         isert_conn->conn_fr_pool_size = 0;
 563         for (i = 0; i < tag_num; i++) {
 564                 fr_desc = kzalloc(sizeof(*fr_desc), GFP_KERNEL);
 565                 if (!fr_desc) {
 566                         isert_err("Failed to allocate fast_reg descriptor\n");
 567                         ret = -ENOMEM;
 568                         goto err;
 569                 }
 570
 571                 ret = isert_create_fr_desc(device->ib_device,
 572                                            device->pd, fr_desc);
 573                 if (ret) {
 574                         isert_err("Failed to create fastreg descriptor err=%d\n",
 575                                ret);
 576                         kfree(fr_desc);
 577                         goto err;
 578                 }
 579
 580                 list_add_tail(&fr_desc->list, &isert_conn->conn_fr_pool);
 581                 isert_conn->conn_fr_pool_size++;
 582         }
 583
 584         isert_dbg("Creating conn %p fastreg pool size=%d",
 585                  isert_conn, isert_conn->conn_fr_pool_size);
 586
 587         return 0;
 588
 589 err:
 590         isert_conn_free_fastreg_pool(isert_conn);
 591         return ret;
 592 }
 593
 594 static int
 595 isert_connect_request(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 596 {
 597         struct isert_np *isert_np = cma_id->context;
 598         struct iscsi_np *np = isert_np->np;
 599         struct isert_conn *isert_conn;
 600         struct isert_device *device;
 601         struct ib_device *ib_dev = cma_id->device;
 602         int ret = 0;
 603
 604         spin_lock_bh(&np->np_thread_lock);
 605         if (!np->enabled) {
 606                 spin_unlock_bh(&np->np_thread_lock);
 607                 isert_dbg("iscsi_np is not enabled, reject connect request\n");
 608                 return rdma_reject(cma_id, NULL, 0);
 609         }
 610         spin_unlock_bh(&np->np_thread_lock);
 611
 612         isert_dbg("cma_id: %p, portal: %p\n",
 613                  cma_id, cma_id->context);
 614
 615         isert_conn = kzalloc(sizeof(struct isert_conn), GFP_KERNEL);
 616         if (!isert_conn) {
 617                 isert_err("Unable to allocate isert_conn\n");
 618                 return -ENOMEM;
 619         }
 620         isert_conn->state = ISER_CONN_INIT;
 621         INIT_LIST_HEAD(&isert_conn->conn_accept_node);
 622         init_completion(&isert_conn->conn_login_comp);
 623         init_completion(&isert_conn->login_req_comp);
 624         init_completion(&isert_conn->conn_wait);
 625         kref_init(&isert_conn->conn_kref);
 626         mutex_init(&isert_conn->conn_mutex);
 627         spin_lock_init(&isert_conn->conn_lock);
 628         INIT_LIST_HEAD(&isert_conn->conn_fr_pool);
 629
 630         isert_conn->conn_cm_id = cma_id;
 631
 632         isert_conn->login_buf = kzalloc(ISCSI_DEF_MAX_RECV_SEG_LEN +
 633                                         ISER_RX_LOGIN_SIZE, GFP_KERNEL);
 634         if (!isert_conn->login_buf) {
 635                 isert_err("Unable to allocate isert_conn->login_buf\n");
 636                 ret = -ENOMEM;
 637                 goto out;
 638         }
 639
 640         isert_conn->login_req_buf = isert_conn->login_buf;
 641         isert_conn->login_rsp_buf = isert_conn->login_buf +
 642                                     ISCSI_DEF_MAX_RECV_SEG_LEN;
 643         isert_dbg("Set login_buf: %p login_req_buf: %p login_rsp_buf: %p\n",
 644                  isert_conn->login_buf, isert_conn->login_req_buf,
 645                  isert_conn->login_rsp_buf);
 646
 647         isert_conn->login_req_dma = ib_dma_map_single(ib_dev,
 648                                 (void *)isert_conn->login_req_buf,
 649                                 ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_FROM_DEVICE);
 650
 651         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_req_dma);
 652         if (ret) {
 653                 isert_err("ib_dma_mapping_error failed for login_req_dma: %d\n",
 654                        ret);
 655                 isert_conn->login_req_dma = 0;
 656                 goto out_login_buf;
 657         }
 658
 659         isert_conn->login_rsp_dma = ib_dma_map_single(ib_dev,
 660                                         (void *)isert_conn->login_rsp_buf,
 661                                         ISER_RX_LOGIN_SIZE, DMA_TO_DEVICE);
 662
 663         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_rsp_dma);
 664         if (ret) {
 665                 isert_err("ib_dma_mapping_error failed for login_rsp_dma: %d\n",
 666                        ret);
 667                 isert_conn->login_rsp_dma = 0;
 668                 goto out_req_dma_map;
 669         }
 670
 671         device = isert_device_find_by_ib_dev(cma_id);
 672         if (IS_ERR(device)) {
 673                 ret = PTR_ERR(device);
 674                 goto out_rsp_dma_map;
 675         }
 676
 677         /* Set max inflight RDMA READ requests */
 678         isert_conn->initiator_depth = min_t(u8,
 679                                 event->param.conn.initiator_depth,
 680                                 device->dev_attr.max_qp_init_rd_atom);
 681         isert_dbg("Using initiator_depth: %u\n", isert_conn->initiator_depth);
 682
 683         isert_conn->conn_device = device;
 684
 685         ret = isert_conn_setup_qp(isert_conn, cma_id);
 686         if (ret)
 687                 goto out_conn_dev;
 688
 689         ret = isert_rdma_post_recvl(isert_conn);
 690         if (ret)
 691                 goto out_conn_dev;
 692
 693         ret = isert_rdma_accept(isert_conn);
 694         if (ret)
 695                 goto out_conn_dev;
 696
 697         mutex_lock(&isert_np->np_accept_mutex);
 698         list_add_tail(&isert_conn->conn_accept_node, &isert_np->np_accept_list);
 699         mutex_unlock(&isert_np->np_accept_mutex);
 700
 701         isert_info("np %p: Allow accept_np to continue\n", np);
 702         up(&isert_np->np_sem);
 703         return 0;
 704
 705 out_conn_dev:
 706         isert_device_try_release(device);
 707 out_rsp_dma_map:
 708         ib_dma_unmap_single(ib_dev, isert_conn->login_rsp_dma,
 709                             ISER_RX_LOGIN_SIZE, DMA_TO_DEVICE);
 710 out_req_dma_map:
 711         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 712                             ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_FROM_DEVICE);
 713 out_login_buf:
 714         kfree(isert_conn->login_buf);
 715 out:
 716         kfree(isert_conn);
 717         rdma_reject(cma_id, NULL, 0);
 718         return ret;
 719 }
 720
 721 static void
 722 isert_connect_release(struct isert_conn *isert_conn)
 723 {
 724         struct isert_device *device = isert_conn->conn_device;
 725         struct ib_device *ib_dev = device->ib_device;
 726
 727         isert_dbg("conn %p\n", isert_conn);
 728
 729         if (device && device->use_fastreg)
 730                 isert_conn_free_fastreg_pool(isert_conn);
 731
 732         isert_free_rx_descriptors(isert_conn);
 733         if (isert_conn->conn_cm_id)
 734                 rdma_destroy_id(isert_conn->conn_cm_id);
 735
 736         if (isert_conn->conn_qp) {
 737                 struct isert_comp *comp = isert_conn->conn_qp->recv_cq->cq_context;
 738
 739                 isert_dbg("dec completion context %p active_qps\n", comp);
 740                 mutex_lock(&device_list_mutex);
 741                 comp->active_qps--;
 742                 mutex_unlock(&device_list_mutex);
 743
 744                 ib_destroy_qp(isert_conn->conn_qp);
 745         }
 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 int
 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_conn->conn_cm_id = NULL;
 887         isert_put_conn(isert_conn);
 888
 889         return -1;
 890 }
 891
 892 static int
 893 isert_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 894 {
 895         int ret = 0;
 896
 897         isert_info("event %d status %d id %p np %p\n", event->event,
 898                    event->status, cma_id, cma_id->context);
 899
 900         switch (event->event) {
 901         case RDMA_CM_EVENT_CONNECT_REQUEST:
 902                 ret = isert_connect_request(cma_id, event);
 903                 if (ret)
 904                         isert_err("failed handle connect request %d\n", ret);
 905                 break;
 906         case RDMA_CM_EVENT_ESTABLISHED:
 907                 isert_connected_handler(cma_id);
 908                 break;
 909         case RDMA_CM_EVENT_ADDR_CHANGE:    /* FALLTHRU */
 910         case RDMA_CM_EVENT_DISCONNECTED:   /* FALLTHRU */
 911         case RDMA_CM_EVENT_DEVICE_REMOVAL: /* FALLTHRU */
 912         case RDMA_CM_EVENT_TIMEWAIT_EXIT:  /* FALLTHRU */
 913                 ret = isert_disconnected_handler(cma_id, event->event);
 914                 break;
 915         case RDMA_CM_EVENT_REJECTED:       /* FALLTHRU */
 916         case RDMA_CM_EVENT_UNREACHABLE:    /* FALLTHRU */
 917         case RDMA_CM_EVENT_CONNECT_ERROR:
 918                 ret = isert_connect_error(cma_id);
 919                 break;
 920         default:
 921                 isert_err("Unhandled RDMA CMA event: %d\n", event->event);
 922                 break;
 923         }
 924
 925         return ret;
 926 }
 927
 928 static int
 929 isert_post_recv(struct isert_conn *isert_conn, u32 count)
 930 {
 931         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 932         int i, ret;
 933         unsigned int rx_head = isert_conn->conn_rx_desc_head;
 934         struct iser_rx_desc *rx_desc;
 935
 936         for (rx_wr = isert_conn->conn_rx_wr, i = 0; i < count; i++, rx_wr++) {
 937                 rx_desc         = &isert_conn->conn_rx_descs[rx_head];
 938                 rx_wr->wr_id    = (uintptr_t)rx_desc;
 939                 rx_wr->sg_list  = &rx_desc->rx_sg;
 940                 rx_wr->num_sge  = 1;
 941                 rx_wr->next     = rx_wr + 1;
 942                 rx_head = (rx_head + 1) & (ISERT_QP_MAX_RECV_DTOS - 1);
 943         }
 944
 945         rx_wr--;
 946         rx_wr->next = NULL; /* mark end of work requests list */
 947
 948         isert_conn->post_recv_buf_count += count;
 949         ret = ib_post_recv(isert_conn->conn_qp, isert_conn->conn_rx_wr,
 950                                 &rx_wr_failed);
 951         if (ret) {
 952                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
 953                 isert_conn->post_recv_buf_count -= count;
 954         } else {
 955                 isert_dbg("Posted %d RX buffers\n", count);
 956                 isert_conn->conn_rx_desc_head = rx_head;
 957         }
 958         return ret;
 959 }
 960
 961 static int
 962 isert_post_send(struct isert_conn *isert_conn, struct iser_tx_desc *tx_desc)
 963 {
 964         struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
 965         struct ib_send_wr send_wr, *send_wr_failed;
 966         int ret;
 967
 968         ib_dma_sync_single_for_device(ib_dev, tx_desc->dma_addr,
 969                                       ISER_HEADERS_LEN, DMA_TO_DEVICE);
 970
 971         send_wr.next    = NULL;
 972         send_wr.wr_id   = (uintptr_t)tx_desc;
 973         send_wr.sg_list = tx_desc->tx_sg;
 974         send_wr.num_sge = tx_desc->num_sge;
 975         send_wr.opcode  = IB_WR_SEND;
 976         send_wr.send_flags = IB_SEND_SIGNALED;
 977
 978         ret = ib_post_send(isert_conn->conn_qp, &send_wr, &send_wr_failed);
 979         if (ret)
 980                 isert_err("ib_post_send() failed, ret: %d\n", ret);
 981
 982         return ret;
 983 }
 984
 985 static void
 986 isert_create_send_desc(struct isert_conn *isert_conn,
 987                        struct isert_cmd *isert_cmd,
 988                        struct iser_tx_desc *tx_desc)
 989 {
 990         struct isert_device *device = isert_conn->conn_device;
 991         struct ib_device *ib_dev = device->ib_device;
 992
 993         ib_dma_sync_single_for_cpu(ib_dev, tx_desc->dma_addr,
 994                                    ISER_HEADERS_LEN, DMA_TO_DEVICE);
 995
 996         memset(&tx_desc->iser_header, 0, sizeof(struct iser_hdr));
 997         tx_desc->iser_header.flags = ISER_VER;
 998
 999         tx_desc->num_sge = 1;
1000         tx_desc->isert_cmd = isert_cmd;
1001
1002         if (tx_desc->tx_sg[0].lkey != device->mr->lkey) {
1003                 tx_desc->tx_sg[0].lkey = device->mr->lkey;
1004                 isert_dbg("tx_desc %p lkey mismatch, fixing\n", tx_desc);
1005         }
1006 }
1007
1008 static int
1009 isert_init_tx_hdrs(struct isert_conn *isert_conn,
1010                    struct iser_tx_desc *tx_desc)
1011 {
1012         struct isert_device *device = isert_conn->conn_device;
1013         struct ib_device *ib_dev = device->ib_device;
1014         u64 dma_addr;
1015
1016         dma_addr = ib_dma_map_single(ib_dev, (void *)tx_desc,
1017                         ISER_HEADERS_LEN, DMA_TO_DEVICE);
1018         if (ib_dma_mapping_error(ib_dev, dma_addr)) {
1019                 isert_err("ib_dma_mapping_error() failed\n");
1020                 return -ENOMEM;
1021         }
1022
1023         tx_desc->dma_addr = dma_addr;
1024         tx_desc->tx_sg[0].addr  = tx_desc->dma_addr;
1025         tx_desc->tx_sg[0].length = ISER_HEADERS_LEN;
1026         tx_desc->tx_sg[0].lkey = device->mr->lkey;
1027
1028         isert_dbg("Setup tx_sg[0].addr: 0x%llx length: %u lkey: 0x%x\n",
1029                   tx_desc->tx_sg[0].addr, tx_desc->tx_sg[0].length,
1030                   tx_desc->tx_sg[0].lkey);
1031
1032         return 0;
1033 }
1034
1035 static void
1036 isert_init_send_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1037                    struct ib_send_wr *send_wr)
1038 {
1039         struct iser_tx_desc *tx_desc = &isert_cmd->tx_desc;
1040
1041         isert_cmd->rdma_wr.iser_ib_op = ISER_IB_SEND;
1042         send_wr->wr_id = (uintptr_t)&isert_cmd->tx_desc;
1043         send_wr->opcode = IB_WR_SEND;
1044         send_wr->sg_list = &tx_desc->tx_sg[0];
1045         send_wr->num_sge = isert_cmd->tx_desc.num_sge;
1046         send_wr->send_flags = IB_SEND_SIGNALED;
1047 }
1048
1049 static int
1050 isert_rdma_post_recvl(struct isert_conn *isert_conn)
1051 {
1052         struct ib_recv_wr rx_wr, *rx_wr_fail;
1053         struct ib_sge sge;
1054         int ret;
1055
1056         memset(&sge, 0, sizeof(struct ib_sge));
1057         sge.addr = isert_conn->login_req_dma;
1058         sge.length = ISER_RX_LOGIN_SIZE;
1059         sge.lkey = isert_conn->conn_device->mr->lkey;
1060
1061         isert_dbg("Setup sge: addr: %llx length: %d 0x%08x\n",
1062                 sge.addr, sge.length, sge.lkey);
1063
1064         memset(&rx_wr, 0, sizeof(struct ib_recv_wr));
1065         rx_wr.wr_id = (uintptr_t)isert_conn->login_req_buf;
1066         rx_wr.sg_list = &sge;
1067         rx_wr.num_sge = 1;
1068
1069         isert_conn->post_recv_buf_count++;
1070         ret = ib_post_recv(isert_conn->conn_qp, &rx_wr, &rx_wr_fail);
1071         if (ret) {
1072                 isert_err("ib_post_recv() failed: %d\n", ret);
1073                 isert_conn->post_recv_buf_count--;
1074         }
1075
1076         return ret;
1077 }
1078
1079 static int
1080 isert_put_login_tx(struct iscsi_conn *conn, struct iscsi_login *login,
1081                    u32 length)
1082 {
1083         struct isert_conn *isert_conn = conn->context;
1084         struct isert_device *device = isert_conn->conn_device;
1085         struct ib_device *ib_dev = device->ib_device;
1086         struct iser_tx_desc *tx_desc = &isert_conn->conn_login_tx_desc;
1087         int ret;
1088
1089         isert_create_send_desc(isert_conn, NULL, tx_desc);
1090
1091         memcpy(&tx_desc->iscsi_header, &login->rsp[0],
1092                sizeof(struct iscsi_hdr));
1093
1094         isert_init_tx_hdrs(isert_conn, tx_desc);
1095
1096         if (length > 0) {
1097                 struct ib_sge *tx_dsg = &tx_desc->tx_sg[1];
1098
1099                 ib_dma_sync_single_for_cpu(ib_dev, isert_conn->login_rsp_dma,
1100                                            length, DMA_TO_DEVICE);
1101
1102                 memcpy(isert_conn->login_rsp_buf, login->rsp_buf, length);
1103
1104                 ib_dma_sync_single_for_device(ib_dev, isert_conn->login_rsp_dma,
1105                                               length, DMA_TO_DEVICE);
1106
1107                 tx_dsg->addr    = isert_conn->login_rsp_dma;
1108                 tx_dsg->length  = length;
1109                 tx_dsg->lkey    = isert_conn->conn_device->mr->lkey;
1110                 tx_desc->num_sge = 2;
1111         }
1112         if (!login->login_failed) {
1113                 if (login->login_complete) {
1114                         if (!conn->sess->sess_ops->SessionType &&
1115                             isert_conn->conn_device->use_fastreg) {
1116                                 ret = isert_conn_create_fastreg_pool(isert_conn);
1117                                 if (ret) {
1118                                         isert_err("Conn: %p failed to create"
1119                                                " fastreg pool\n", isert_conn);
1120                                         return ret;
1121                                 }
1122                         }
1123
1124                         ret = isert_alloc_rx_descriptors(isert_conn);
1125                         if (ret)
1126                                 return ret;
1127
1128                         ret = isert_post_recv(isert_conn, ISERT_MIN_POSTED_RX);
1129                         if (ret)
1130                                 return ret;
1131
1132                         /* Now we are in FULL_FEATURE phase */
1133                         mutex_lock(&isert_conn->conn_mutex);
1134                         isert_conn->state = ISER_CONN_FULL_FEATURE;
1135                         mutex_unlock(&isert_conn->conn_mutex);
1136                         goto post_send;
1137                 }
1138
1139                 ret = isert_rdma_post_recvl(isert_conn);
1140                 if (ret)
1141                         return ret;
1142         }
1143 post_send:
1144         ret = isert_post_send(isert_conn, tx_desc);
1145         if (ret)
1146                 return ret;
1147
1148         return 0;
1149 }
1150
1151 static void
1152 isert_rx_login_req(struct isert_conn *isert_conn)
1153 {
1154         struct iser_rx_desc *rx_desc = (void *)isert_conn->login_req_buf;
1155         int rx_buflen = isert_conn->login_req_len;
1156         struct iscsi_conn *conn = isert_conn->conn;
1157         struct iscsi_login *login = conn->conn_login;
1158         int size;
1159
1160         isert_info("conn %p\n", isert_conn);
1161
1162         WARN_ON_ONCE(!login);
1163
1164         if (login->first_request) {
1165                 struct iscsi_login_req *login_req =
1166                         (struct iscsi_login_req *)&rx_desc->iscsi_header;
1167                 /*
1168                  * Setup the initial iscsi_login values from the leading
1169                  * login request PDU.
1170                  */
1171                 login->leading_connection = (!login_req->tsih) ? 1 : 0;
1172                 login->current_stage =
1173                         (login_req->flags & ISCSI_FLAG_LOGIN_CURRENT_STAGE_MASK)
1174                          >> 2;
1175                 login->version_min      = login_req->min_version;
1176                 login->version_max      = login_req->max_version;
1177                 memcpy(login->isid, login_req->isid, 6);
1178                 login->cmd_sn           = be32_to_cpu(login_req->cmdsn);
1179                 login->init_task_tag    = login_req->itt;
1180                 login->initial_exp_statsn = be32_to_cpu(login_req->exp_statsn);
1181                 login->cid              = be16_to_cpu(login_req->cid);
1182                 login->tsih             = be16_to_cpu(login_req->tsih);
1183         }
1184
1185         memcpy(&login->req[0], (void *)&rx_desc->iscsi_header, ISCSI_HDR_LEN);
1186
1187         size = min(rx_buflen, MAX_KEY_VALUE_PAIRS);
1188         isert_dbg("Using login payload size: %d, rx_buflen: %d "
1189                   "MAX_KEY_VALUE_PAIRS: %d\n", size, rx_buflen,
1190                   MAX_KEY_VALUE_PAIRS);
1191         memcpy(login->req_buf, &rx_desc->data[0], size);
1192
1193         if (login->first_request) {
1194                 complete(&isert_conn->conn_login_comp);
1195                 return;
1196         }
1197         schedule_delayed_work(&conn->login_work, 0);
1198 }
1199
1200 static struct iscsi_cmd
1201 *isert_allocate_cmd(struct iscsi_conn *conn)
1202 {
1203         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
1204         struct isert_cmd *isert_cmd;
1205         struct iscsi_cmd *cmd;
1206
1207         cmd = iscsit_allocate_cmd(conn, TASK_INTERRUPTIBLE);
1208         if (!cmd) {
1209                 isert_err("Unable to allocate iscsi_cmd + isert_cmd\n");
1210                 return NULL;
1211         }
1212         isert_cmd = iscsit_priv_cmd(cmd);
1213         isert_cmd->conn = isert_conn;
1214         isert_cmd->iscsi_cmd = cmd;
1215
1216         return cmd;
1217 }
1218
1219 static int
1220 isert_handle_scsi_cmd(struct isert_conn *isert_conn,
1221                       struct isert_cmd *isert_cmd, struct iscsi_cmd *cmd,
1222                       struct iser_rx_desc *rx_desc, unsigned char *buf)
1223 {
1224         struct iscsi_conn *conn = isert_conn->conn;
1225         struct iscsi_scsi_req *hdr = (struct iscsi_scsi_req *)buf;
1226         struct scatterlist *sg;
1227         int imm_data, imm_data_len, unsol_data, sg_nents, rc;
1228         bool dump_payload = false;
1229
1230         rc = iscsit_setup_scsi_cmd(conn, cmd, buf);
1231         if (rc < 0)
1232                 return rc;
1233
1234         imm_data = cmd->immediate_data;
1235         imm_data_len = cmd->first_burst_len;
1236         unsol_data = cmd->unsolicited_data;
1237
1238         rc = iscsit_process_scsi_cmd(conn, cmd, hdr);
1239         if (rc < 0) {
1240                 return 0;
1241         } else if (rc > 0) {
1242                 dump_payload = true;
1243                 goto sequence_cmd;
1244         }
1245
1246         if (!imm_data)
1247                 return 0;
1248
1249         sg = &cmd->se_cmd.t_data_sg[0];
1250         sg_nents = max(1UL, DIV_ROUND_UP(imm_data_len, PAGE_SIZE));
1251
1252         isert_dbg("Copying Immediate SG: %p sg_nents: %u from %p imm_data_len: %d\n",
1253                   sg, sg_nents, &rx_desc->data[0], imm_data_len);
1254
1255         sg_copy_from_buffer(sg, sg_nents, &rx_desc->data[0], imm_data_len);
1256
1257         cmd->write_data_done += imm_data_len;
1258
1259         if (cmd->write_data_done == cmd->se_cmd.data_length) {
1260                 spin_lock_bh(&cmd->istate_lock);
1261                 cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1262                 cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1263                 spin_unlock_bh(&cmd->istate_lock);
1264         }
1265
1266 sequence_cmd:
1267         rc = iscsit_sequence_cmd(conn, cmd, buf, hdr->cmdsn);
1268
1269         if (!rc && dump_payload == false && unsol_data)
1270                 iscsit_set_unsoliticed_dataout(cmd);
1271         else if (dump_payload && imm_data)
1272                 target_put_sess_cmd(conn->sess->se_sess, &cmd->se_cmd);
1273
1274         return 0;
1275 }
1276
1277 static int
1278 isert_handle_iscsi_dataout(struct isert_conn *isert_conn,
1279                            struct iser_rx_desc *rx_desc, unsigned char *buf)
1280 {
1281         struct scatterlist *sg_start;
1282         struct iscsi_conn *conn = isert_conn->conn;
1283         struct iscsi_cmd *cmd = NULL;
1284         struct iscsi_data *hdr = (struct iscsi_data *)buf;
1285         u32 unsol_data_len = ntoh24(hdr->dlength);
1286         int rc, sg_nents, sg_off, page_off;
1287
1288         rc = iscsit_check_dataout_hdr(conn, buf, &cmd);
1289         if (rc < 0)
1290                 return rc;
1291         else if (!cmd)
1292                 return 0;
1293         /*
1294          * FIXME: Unexpected unsolicited_data out
1295          */
1296         if (!cmd->unsolicited_data) {
1297                 isert_err("Received unexpected solicited data payload\n");
1298                 dump_stack();
1299                 return -1;
1300         }
1301
1302         isert_dbg("Unsolicited DataOut unsol_data_len: %u, "
1303                   "write_data_done: %u, data_length: %u\n",
1304                   unsol_data_len,  cmd->write_data_done,
1305                   cmd->se_cmd.data_length);
1306
1307         sg_off = cmd->write_data_done / PAGE_SIZE;
1308         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
1309         sg_nents = max(1UL, DIV_ROUND_UP(unsol_data_len, PAGE_SIZE));
1310         page_off = cmd->write_data_done % PAGE_SIZE;
1311         /*
1312          * FIXME: Non page-aligned unsolicited_data out
1313          */
1314         if (page_off) {
1315                 isert_err("unexpected non-page aligned data payload\n");
1316                 dump_stack();
1317                 return -1;
1318         }
1319         isert_dbg("Copying DataOut: sg_start: %p, sg_off: %u "
1320                   "sg_nents: %u from %p %u\n", sg_start, sg_off,
1321                   sg_nents, &rx_desc->data[0], unsol_data_len);
1322
1323         sg_copy_from_buffer(sg_start, sg_nents, &rx_desc->data[0],
1324                             unsol_data_len);
1325
1326         rc = iscsit_check_dataout_payload(cmd, hdr, false);
1327         if (rc < 0)
1328                 return rc;
1329
1330         return 0;
1331 }
1332
1333 static int
1334 isert_handle_nop_out(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1335                      struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1336                      unsigned char *buf)
1337 {
1338         struct iscsi_conn *conn = isert_conn->conn;
1339         struct iscsi_nopout *hdr = (struct iscsi_nopout *)buf;
1340         int rc;
1341
1342         rc = iscsit_setup_nop_out(conn, cmd, hdr);
1343         if (rc < 0)
1344                 return rc;
1345         /*
1346          * FIXME: Add support for NOPOUT payload using unsolicited RDMA payload
1347          */
1348
1349         return iscsit_process_nop_out(conn, cmd, hdr);
1350 }
1351
1352 static int
1353 isert_handle_text_cmd(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1354                       struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1355                       struct iscsi_text *hdr)
1356 {
1357         struct iscsi_conn *conn = isert_conn->conn;
1358         u32 payload_length = ntoh24(hdr->dlength);
1359         int rc;
1360         unsigned char *text_in = NULL;
1361
1362         rc = iscsit_setup_text_cmd(conn, cmd, hdr);
1363         if (rc < 0)
1364                 return rc;
1365
1366         if (payload_length) {
1367                 text_in = kzalloc(payload_length, GFP_KERNEL);
1368                 if (!text_in) {
1369                         isert_err("Unable to allocate text_in of payload_length: %u\n",
1370                                   payload_length);
1371                         return -ENOMEM;
1372                 }
1373         }
1374         cmd->text_in_ptr = text_in;
1375
1376         memcpy(cmd->text_in_ptr, &rx_desc->data[0], payload_length);
1377
1378         return iscsit_process_text_cmd(conn, cmd, hdr);
1379 }
1380
1381 static int
1382 isert_rx_opcode(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc,
1383                 uint32_t read_stag, uint64_t read_va,
1384                 uint32_t write_stag, uint64_t write_va)
1385 {
1386         struct iscsi_hdr *hdr = &rx_desc->iscsi_header;
1387         struct iscsi_conn *conn = isert_conn->conn;
1388         struct iscsi_session *sess = conn->sess;
1389         struct iscsi_cmd *cmd;
1390         struct isert_cmd *isert_cmd;
1391         int ret = -EINVAL;
1392         u8 opcode = (hdr->opcode & ISCSI_OPCODE_MASK);
1393
1394         if (sess->sess_ops->SessionType &&
1395            (!(opcode & ISCSI_OP_TEXT) || !(opcode & ISCSI_OP_LOGOUT))) {
1396                 isert_err("Got illegal opcode: 0x%02x in SessionType=Discovery,"
1397                           " ignoring\n", opcode);
1398                 return 0;
1399         }
1400
1401         switch (opcode) {
1402         case ISCSI_OP_SCSI_CMD:
1403                 cmd = isert_allocate_cmd(conn);
1404                 if (!cmd)
1405                         break;
1406
1407                 isert_cmd = iscsit_priv_cmd(cmd);
1408                 isert_cmd->read_stag = read_stag;
1409                 isert_cmd->read_va = read_va;
1410                 isert_cmd->write_stag = write_stag;
1411                 isert_cmd->write_va = write_va;
1412
1413                 ret = isert_handle_scsi_cmd(isert_conn, isert_cmd, cmd,
1414                                         rx_desc, (unsigned char *)hdr);
1415                 break;
1416         case ISCSI_OP_NOOP_OUT:
1417                 cmd = isert_allocate_cmd(conn);
1418                 if (!cmd)
1419                         break;
1420
1421                 isert_cmd = iscsit_priv_cmd(cmd);
1422                 ret = isert_handle_nop_out(isert_conn, isert_cmd, cmd,
1423                                            rx_desc, (unsigned char *)hdr);
1424                 break;
1425         case ISCSI_OP_SCSI_DATA_OUT:
1426                 ret = isert_handle_iscsi_dataout(isert_conn, rx_desc,
1427                                                 (unsigned char *)hdr);
1428                 break;
1429         case ISCSI_OP_SCSI_TMFUNC:
1430                 cmd = isert_allocate_cmd(conn);
1431                 if (!cmd)
1432                         break;
1433
1434                 ret = iscsit_handle_task_mgt_cmd(conn, cmd,
1435                                                 (unsigned char *)hdr);
1436                 break;
1437         case ISCSI_OP_LOGOUT:
1438                 cmd = isert_allocate_cmd(conn);
1439                 if (!cmd)
1440                         break;
1441
1442                 ret = iscsit_handle_logout_cmd(conn, cmd, (unsigned char *)hdr);
1443                 break;
1444         case ISCSI_OP_TEXT:
1445                 if (be32_to_cpu(hdr->ttt) != 0xFFFFFFFF) {
1446                         cmd = iscsit_find_cmd_from_itt(conn, hdr->itt);
1447                         if (!cmd)
1448                                 break;
1449                 } else {
1450                         cmd = isert_allocate_cmd(conn);
1451                         if (!cmd)
1452                                 break;
1453                 }
1454
1455                 isert_cmd = iscsit_priv_cmd(cmd);
1456                 ret = isert_handle_text_cmd(isert_conn, isert_cmd, cmd,
1457                                             rx_desc, (struct iscsi_text *)hdr);
1458                 break;
1459         default:
1460                 isert_err("Got unknown iSCSI OpCode: 0x%02x\n", opcode);
1461                 dump_stack();
1462                 break;
1463         }
1464
1465         return ret;
1466 }
1467
1468 static void
1469 isert_rx_do_work(struct iser_rx_desc *rx_desc, struct isert_conn *isert_conn)
1470 {
1471         struct iser_hdr *iser_hdr = &rx_desc->iser_header;
1472         uint64_t read_va = 0, write_va = 0;
1473         uint32_t read_stag = 0, write_stag = 0;
1474         int rc;
1475
1476         switch (iser_hdr->flags & 0xF0) {
1477         case ISCSI_CTRL:
1478                 if (iser_hdr->flags & ISER_RSV) {
1479                         read_stag = be32_to_cpu(iser_hdr->read_stag);
1480                         read_va = be64_to_cpu(iser_hdr->read_va);
1481                         isert_dbg("ISER_RSV: read_stag: 0x%x read_va: 0x%llx\n",
1482                                   read_stag, (unsigned long long)read_va);
1483                 }
1484                 if (iser_hdr->flags & ISER_WSV) {
1485                         write_stag = be32_to_cpu(iser_hdr->write_stag);
1486                         write_va = be64_to_cpu(iser_hdr->write_va);
1487                         isert_dbg("ISER_WSV: write_stag: 0x%x write_va: 0x%llx\n",
1488                                   write_stag, (unsigned long long)write_va);
1489                 }
1490
1491                 isert_dbg("ISER ISCSI_CTRL PDU\n");
1492                 break;
1493         case ISER_HELLO:
1494                 isert_err("iSER Hello message\n");
1495                 break;
1496         default:
1497                 isert_warn("Unknown iSER hdr flags: 0x%02x\n", iser_hdr->flags);
1498                 break;
1499         }
1500
1501         rc = isert_rx_opcode(isert_conn, rx_desc,
1502                              read_stag, read_va, write_stag, write_va);
1503 }
1504
1505 static void
1506 isert_rx_completion(struct iser_rx_desc *desc, struct isert_conn *isert_conn,
1507                     u32 xfer_len)
1508 {
1509         struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
1510         struct iscsi_hdr *hdr;
1511         u64 rx_dma;
1512         int rx_buflen, outstanding;
1513
1514         if ((char *)desc == isert_conn->login_req_buf) {
1515                 rx_dma = isert_conn->login_req_dma;
1516                 rx_buflen = ISER_RX_LOGIN_SIZE;
1517                 isert_dbg("login_buf: Using rx_dma: 0x%llx, rx_buflen: %d\n",
1518                          rx_dma, rx_buflen);
1519         } else {
1520                 rx_dma = desc->dma_addr;
1521                 rx_buflen = ISER_RX_PAYLOAD_SIZE;
1522                 isert_dbg("req_buf: Using rx_dma: 0x%llx, rx_buflen: %d\n",
1523                          rx_dma, rx_buflen);
1524         }
1525
1526         ib_dma_sync_single_for_cpu(ib_dev, rx_dma, rx_buflen, DMA_FROM_DEVICE);
1527
1528         hdr = &desc->iscsi_header;
1529         isert_dbg("iSCSI opcode: 0x%02x, ITT: 0x%08x, flags: 0x%02x dlen: %d\n",
1530                  hdr->opcode, hdr->itt, hdr->flags,
1531                  (int)(xfer_len - ISER_HEADERS_LEN));
1532
1533         if ((char *)desc == isert_conn->login_req_buf) {
1534                 isert_conn->login_req_len = xfer_len - ISER_HEADERS_LEN;
1535                 if (isert_conn->conn) {
1536                         struct iscsi_login *login = isert_conn->conn->conn_login;
1537
1538                         if (login && !login->first_request)
1539                                 isert_rx_login_req(isert_conn);
1540                 }
1541                 mutex_lock(&isert_conn->conn_mutex);
1542                 complete(&isert_conn->login_req_comp);
1543                 mutex_unlock(&isert_conn->conn_mutex);
1544         } else {
1545                 isert_rx_do_work(desc, isert_conn);
1546         }
1547
1548         ib_dma_sync_single_for_device(ib_dev, rx_dma, rx_buflen,
1549                                       DMA_FROM_DEVICE);
1550
1551         isert_conn->post_recv_buf_count--;
1552         isert_dbg("Decremented post_recv_buf_count: %d\n",
1553                   isert_conn->post_recv_buf_count);
1554
1555         if ((char *)desc == isert_conn->login_req_buf)
1556                 return;
1557
1558         outstanding = isert_conn->post_recv_buf_count;
1559         if (outstanding + ISERT_MIN_POSTED_RX <= ISERT_QP_MAX_RECV_DTOS) {
1560                 int err, count = min(ISERT_QP_MAX_RECV_DTOS - outstanding,
1561                                 ISERT_MIN_POSTED_RX);
1562                 err = isert_post_recv(isert_conn, count);
1563                 if (err) {
1564                         isert_err("isert_post_recv() count: %d failed, %d\n",
1565                                count, err);
1566                 }
1567         }
1568 }
1569
1570 static int
1571 isert_map_data_buf(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1572                    struct scatterlist *sg, u32 nents, u32 length, u32 offset,
1573                    enum iser_ib_op_code op, struct isert_data_buf *data)
1574 {
1575         struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
1576
1577         data->dma_dir = op == ISER_IB_RDMA_WRITE ?
1578                               DMA_TO_DEVICE : DMA_FROM_DEVICE;
1579
1580         data->len = length - offset;
1581         data->offset = offset;
1582         data->sg_off = data->offset / PAGE_SIZE;
1583
1584         data->sg = &sg[data->sg_off];
1585         data->nents = min_t(unsigned int, nents - data->sg_off,
1586                                           ISCSI_ISER_SG_TABLESIZE);
1587         data->len = min_t(unsigned int, data->len, ISCSI_ISER_SG_TABLESIZE *
1588                                         PAGE_SIZE);
1589
1590         data->dma_nents = ib_dma_map_sg(ib_dev, data->sg, data->nents,
1591                                         data->dma_dir);
1592         if (unlikely(!data->dma_nents)) {
1593                 isert_err("Cmd: unable to dma map SGs %p\n", sg);
1594                 return -EINVAL;
1595         }
1596
1597         isert_dbg("Mapped cmd: %p count: %u sg: %p sg_nents: %u rdma_len %d\n",
1598                   isert_cmd, data->dma_nents, data->sg, data->nents, data->len);
1599
1600         return 0;
1601 }
1602
1603 static void
1604 isert_unmap_data_buf(struct isert_conn *isert_conn, struct isert_data_buf *data)
1605 {
1606         struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
1607
1608         ib_dma_unmap_sg(ib_dev, data->sg, data->nents, data->dma_dir);
1609         memset(data, 0, sizeof(*data));
1610 }
1611
1612
1613
1614 static void
1615 isert_unmap_cmd(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn)
1616 {
1617         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1618
1619         isert_dbg("Cmd %p\n", isert_cmd);
1620
1621         if (wr->data.sg) {
1622                 isert_dbg("Cmd %p unmap_sg op\n", isert_cmd);
1623                 isert_unmap_data_buf(isert_conn, &wr->data);
1624         }
1625
1626         if (wr->send_wr) {
1627                 isert_dbg("Cmd %p free send_wr\n", isert_cmd);
1628                 kfree(wr->send_wr);
1629                 wr->send_wr = NULL;
1630         }
1631
1632         if (wr->ib_sge) {
1633                 isert_dbg("Cmd %p free ib_sge\n", isert_cmd);
1634                 kfree(wr->ib_sge);
1635                 wr->ib_sge = NULL;
1636         }
1637 }
1638
1639 static void
1640 isert_unreg_rdma(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn)
1641 {
1642         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1643         LIST_HEAD(unmap_list);
1644
1645         isert_dbg("Cmd %p\n", isert_cmd);
1646
1647         if (wr->fr_desc) {
1648                 isert_dbg("Cmd %p free fr_desc %p\n", isert_cmd, wr->fr_desc);
1649                 if (wr->fr_desc->ind & ISERT_PROTECTED) {
1650                         isert_unmap_data_buf(isert_conn, &wr->prot);
1651                         wr->fr_desc->ind &= ~ISERT_PROTECTED;
1652                 }
1653                 spin_lock_bh(&isert_conn->conn_lock);
1654                 list_add_tail(&wr->fr_desc->list, &isert_conn->conn_fr_pool);
1655                 spin_unlock_bh(&isert_conn->conn_lock);
1656                 wr->fr_desc = NULL;
1657         }
1658
1659         if (wr->data.sg) {
1660                 isert_dbg("Cmd %p unmap_sg op\n", isert_cmd);
1661                 isert_unmap_data_buf(isert_conn, &wr->data);
1662         }
1663
1664         wr->ib_sge = NULL;
1665         wr->send_wr = NULL;
1666 }
1667
1668 static void
1669 isert_put_cmd(struct isert_cmd *isert_cmd, bool comp_err)
1670 {
1671         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1672         struct isert_conn *isert_conn = isert_cmd->conn;
1673         struct iscsi_conn *conn = isert_conn->conn;
1674         struct isert_device *device = isert_conn->conn_device;
1675         struct iscsi_text_rsp *hdr;
1676
1677         isert_dbg("Cmd %p\n", isert_cmd);
1678
1679         switch (cmd->iscsi_opcode) {
1680         case ISCSI_OP_SCSI_CMD:
1681                 spin_lock_bh(&conn->cmd_lock);
1682                 if (!list_empty(&cmd->i_conn_node))
1683                         list_del_init(&cmd->i_conn_node);
1684                 spin_unlock_bh(&conn->cmd_lock);
1685
1686                 if (cmd->data_direction == DMA_TO_DEVICE) {
1687                         iscsit_stop_dataout_timer(cmd);
1688                         /*
1689                          * Check for special case during comp_err where
1690                          * WRITE_PENDING has been handed off from core,
1691                          * but requires an extra target_put_sess_cmd()
1692                          * before transport_generic_free_cmd() below.
1693                          */
1694                         if (comp_err &&
1695                             cmd->se_cmd.t_state == TRANSPORT_WRITE_PENDING) {
1696                                 struct se_cmd *se_cmd = &cmd->se_cmd;
1697
1698                                 target_put_sess_cmd(se_cmd->se_sess, se_cmd);
1699                         }
1700                 }
1701
1702                 device->unreg_rdma_mem(isert_cmd, isert_conn);
1703                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1704                 break;
1705         case ISCSI_OP_SCSI_TMFUNC:
1706                 spin_lock_bh(&conn->cmd_lock);
1707                 if (!list_empty(&cmd->i_conn_node))
1708                         list_del_init(&cmd->i_conn_node);
1709                 spin_unlock_bh(&conn->cmd_lock);
1710
1711                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1712                 break;
1713         case ISCSI_OP_REJECT:
1714         case ISCSI_OP_NOOP_OUT:
1715         case ISCSI_OP_TEXT:
1716                 hdr = (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
1717                 /* If the continue bit is on, keep the command alive */
1718                 if (hdr->flags & ISCSI_FLAG_TEXT_CONTINUE)
1719                         break;
1720
1721                 spin_lock_bh(&conn->cmd_lock);
1722                 if (!list_empty(&cmd->i_conn_node))
1723                         list_del_init(&cmd->i_conn_node);
1724                 spin_unlock_bh(&conn->cmd_lock);
1725
1726                 /*
1727                  * Handle special case for REJECT when iscsi_add_reject*() has
1728                  * overwritten the original iscsi_opcode assignment, and the
1729                  * associated cmd->se_cmd needs to be released.
1730                  */
1731                 if (cmd->se_cmd.se_tfo != NULL) {
1732                         isert_dbg("Calling transport_generic_free_cmd for 0x%02x\n",
1733                                  cmd->iscsi_opcode);
1734                         transport_generic_free_cmd(&cmd->se_cmd, 0);
1735                         break;
1736                 }
1737                 /*
1738                  * Fall-through
1739                  */
1740         default:
1741                 iscsit_release_cmd(cmd);
1742                 break;
1743         }
1744 }
1745
1746 static void
1747 isert_unmap_tx_desc(struct iser_tx_desc *tx_desc, struct ib_device *ib_dev)
1748 {
1749         if (tx_desc->dma_addr != 0) {
1750                 isert_dbg("unmap single for tx_desc->dma_addr\n");
1751                 ib_dma_unmap_single(ib_dev, tx_desc->dma_addr,
1752                                     ISER_HEADERS_LEN, DMA_TO_DEVICE);
1753                 tx_desc->dma_addr = 0;
1754         }
1755 }
1756
1757 static void
1758 isert_completion_put(struct iser_tx_desc *tx_desc, struct isert_cmd *isert_cmd,
1759                      struct ib_device *ib_dev, bool comp_err)
1760 {
1761         if (isert_cmd->pdu_buf_dma != 0) {
1762                 isert_dbg("unmap single for isert_cmd->pdu_buf_dma\n");
1763                 ib_dma_unmap_single(ib_dev, isert_cmd->pdu_buf_dma,
1764                                     isert_cmd->pdu_buf_len, DMA_TO_DEVICE);
1765                 isert_cmd->pdu_buf_dma = 0;
1766         }
1767
1768         isert_unmap_tx_desc(tx_desc, ib_dev);
1769         isert_put_cmd(isert_cmd, comp_err);
1770 }
1771
1772 static int
1773 isert_check_pi_status(struct se_cmd *se_cmd, struct ib_mr *sig_mr)
1774 {
1775         struct ib_mr_status mr_status;
1776         int ret;
1777
1778         ret = ib_check_mr_status(sig_mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
1779         if (ret) {
1780                 isert_err("ib_check_mr_status failed, ret %d\n", ret);
1781                 goto fail_mr_status;
1782         }
1783
1784         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1785                 u64 sec_offset_err;
1786                 u32 block_size = se_cmd->se_dev->dev_attrib.block_size + 8;
1787
1788                 switch (mr_status.sig_err.err_type) {
1789                 case IB_SIG_BAD_GUARD:
1790                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_GUARD_CHECK_FAILED;
1791                         break;
1792                 case IB_SIG_BAD_REFTAG:
1793                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_REF_TAG_CHECK_FAILED;
1794                         break;
1795                 case IB_SIG_BAD_APPTAG:
1796                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_APP_TAG_CHECK_FAILED;
1797                         break;
1798                 }
1799                 sec_offset_err = mr_status.sig_err.sig_err_offset;
1800                 do_div(sec_offset_err, block_size);
1801                 se_cmd->bad_sector = sec_offset_err + se_cmd->t_task_lba;
1802
1803                 isert_err("PI error found type %d at sector 0x%llx "
1804                           "expected 0x%x vs actual 0x%x\n",
1805                           mr_status.sig_err.err_type,
1806                           (unsigned long long)se_cmd->bad_sector,
1807                           mr_status.sig_err.expected,
1808                           mr_status.sig_err.actual);
1809                 ret = 1;
1810         }
1811
1812 fail_mr_status:
1813         return ret;
1814 }
1815
1816 static void
1817 isert_completion_rdma_write(struct iser_tx_desc *tx_desc,
1818                             struct isert_cmd *isert_cmd)
1819 {
1820         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1821         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1822         struct se_cmd *se_cmd = &cmd->se_cmd;
1823         struct isert_conn *isert_conn = isert_cmd->conn;
1824         struct isert_device *device = isert_conn->conn_device;
1825         int ret = 0;
1826
1827         if (wr->fr_desc && wr->fr_desc->ind & ISERT_PROTECTED) {
1828                 ret = isert_check_pi_status(se_cmd,
1829                                             wr->fr_desc->pi_ctx->sig_mr);
1830                 wr->fr_desc->ind &= ~ISERT_PROTECTED;
1831         }
1832
1833         device->unreg_rdma_mem(isert_cmd, isert_conn);
1834         wr->send_wr_num = 0;
1835         if (ret)
1836                 transport_send_check_condition_and_sense(se_cmd,
1837                                                          se_cmd->pi_err, 0);
1838         else
1839                 isert_put_response(isert_conn->conn, cmd);
1840 }
1841
1842 static void
1843 isert_completion_rdma_read(struct iser_tx_desc *tx_desc,
1844                            struct isert_cmd *isert_cmd)
1845 {
1846         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1847         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1848         struct se_cmd *se_cmd = &cmd->se_cmd;
1849         struct isert_conn *isert_conn = isert_cmd->conn;
1850         struct isert_device *device = isert_conn->conn_device;
1851         int ret = 0;
1852
1853         if (wr->fr_desc && wr->fr_desc->ind & ISERT_PROTECTED) {
1854                 ret = isert_check_pi_status(se_cmd,
1855                                             wr->fr_desc->pi_ctx->sig_mr);
1856                 wr->fr_desc->ind &= ~ISERT_PROTECTED;
1857         }
1858
1859         iscsit_stop_dataout_timer(cmd);
1860         device->unreg_rdma_mem(isert_cmd, isert_conn);
1861         cmd->write_data_done = wr->data.len;
1862         wr->send_wr_num = 0;
1863
1864         isert_dbg("Cmd: %p RDMA_READ comp calling execute_cmd\n", isert_cmd);
1865         spin_lock_bh(&cmd->istate_lock);
1866         cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1867         cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1868         spin_unlock_bh(&cmd->istate_lock);
1869
1870         if (ret) {
1871                 target_put_sess_cmd(se_cmd->se_sess, se_cmd);
1872                 transport_send_check_condition_and_sense(se_cmd,
1873                                                          se_cmd->pi_err, 0);
1874         } else {
1875                 target_execute_cmd(se_cmd);
1876         }
1877 }
1878
1879 static void
1880 isert_do_control_comp(struct work_struct *work)
1881 {
1882         struct isert_cmd *isert_cmd = container_of(work,
1883                         struct isert_cmd, comp_work);
1884         struct isert_conn *isert_conn = isert_cmd->conn;
1885         struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
1886         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1887
1888         isert_dbg("Cmd %p i_state %d\n", isert_cmd, cmd->i_state);
1889
1890         switch (cmd->i_state) {
1891         case ISTATE_SEND_TASKMGTRSP:
1892                 iscsit_tmr_post_handler(cmd, cmd->conn);
1893         case ISTATE_SEND_REJECT:   /* FALLTHRU */
1894         case ISTATE_SEND_TEXTRSP:  /* FALLTHRU */
1895                 cmd->i_state = ISTATE_SENT_STATUS;
1896                 isert_completion_put(&isert_cmd->tx_desc, isert_cmd,
1897                                      ib_dev, false);
1898                 break;
1899         case ISTATE_SEND_LOGOUTRSP:
1900                 iscsit_logout_post_handler(cmd, cmd->conn);
1901                 break;
1902         default:
1903                 isert_err("Unknown i_state %d\n", cmd->i_state);
1904                 dump_stack();
1905                 break;
1906         }
1907 }
1908
1909 static void
1910 isert_response_completion(struct iser_tx_desc *tx_desc,
1911                           struct isert_cmd *isert_cmd,
1912                           struct isert_conn *isert_conn,
1913                           struct ib_device *ib_dev)
1914 {
1915         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1916
1917         if (cmd->i_state == ISTATE_SEND_TASKMGTRSP ||
1918             cmd->i_state == ISTATE_SEND_LOGOUTRSP ||
1919             cmd->i_state == ISTATE_SEND_REJECT ||
1920             cmd->i_state == ISTATE_SEND_TEXTRSP) {
1921                 isert_unmap_tx_desc(tx_desc, ib_dev);
1922
1923                 INIT_WORK(&isert_cmd->comp_work, isert_do_control_comp);
1924                 queue_work(isert_comp_wq, &isert_cmd->comp_work);
1925                 return;
1926         }
1927
1928         cmd->i_state = ISTATE_SENT_STATUS;
1929         isert_completion_put(tx_desc, isert_cmd, ib_dev, false);
1930 }
1931
1932 static void
1933 isert_send_completion(struct iser_tx_desc *tx_desc,
1934                       struct isert_conn *isert_conn)
1935 {
1936         struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
1937         struct isert_cmd *isert_cmd = tx_desc->isert_cmd;
1938         struct isert_rdma_wr *wr;
1939
1940         if (!isert_cmd) {
1941                 isert_unmap_tx_desc(tx_desc, ib_dev);
1942                 return;
1943         }
1944         wr = &isert_cmd->rdma_wr;
1945
1946         isert_dbg("Cmd %p iser_ib_op %d\n", isert_cmd, wr->iser_ib_op);
1947
1948         switch (wr->iser_ib_op) {
1949         case ISER_IB_SEND:
1950                 isert_response_completion(tx_desc, isert_cmd,
1951                                           isert_conn, ib_dev);
1952                 break;
1953         case ISER_IB_RDMA_WRITE:
1954                 isert_completion_rdma_write(tx_desc, isert_cmd);
1955                 break;
1956         case ISER_IB_RDMA_READ:
1957                 isert_completion_rdma_read(tx_desc, isert_cmd);
1958                 break;
1959         default:
1960                 isert_err("Unknown wr->iser_ib_op: 0x%x\n", wr->iser_ib_op);
1961                 dump_stack();
1962                 break;
1963         }
1964 }
1965
1966 /**
1967  * is_isert_tx_desc() - Indicate if the completion wr_id
1968  *     is a TX descriptor or not.
1969  * @isert_conn: iser connection
1970  * @wr_id: completion WR identifier
1971  *
1972  * Since we cannot rely on wc opcode in FLUSH errors
1973  * we must work around it by checking if the wr_id address
1974  * falls in the iser connection rx_descs buffer. If so
1975  * it is an RX descriptor, otherwize it is a TX.
1976  */
1977 static inline bool
1978 is_isert_tx_desc(struct isert_conn *isert_conn, void *wr_id)
1979 {
1980         void *start = isert_conn->conn_rx_descs;
1981         int len = ISERT_QP_MAX_RECV_DTOS * sizeof(*isert_conn->conn_rx_descs);
1982
1983         if (wr_id >= start && wr_id < start + len)
1984                 return false;
1985
1986         return true;
1987 }
1988
1989 static void
1990 isert_cq_comp_err(struct isert_conn *isert_conn, struct ib_wc *wc)
1991 {
1992         if (wc->wr_id == ISER_BEACON_WRID) {
1993                 isert_info("conn %p completing conn_wait_comp_err\n",
1994                            isert_conn);
1995                 complete(&isert_conn->conn_wait_comp_err);
1996         } else if (is_isert_tx_desc(isert_conn, (void *)(uintptr_t)wc->wr_id)) {
1997                 struct ib_device *ib_dev = isert_conn->conn_cm_id->device;
1998                 struct isert_cmd *isert_cmd;
1999                 struct iser_tx_desc *desc;
2000
2001                 desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
2002                 isert_cmd = desc->isert_cmd;
2003                 if (!isert_cmd)
2004                         isert_unmap_tx_desc(desc, ib_dev);
2005                 else
2006                         isert_completion_put(desc, isert_cmd, ib_dev, true);
2007         } else {
2008                 isert_conn->post_recv_buf_count--;
2009                 if (!isert_conn->post_recv_buf_count)
2010                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
2011         }
2012 }
2013
2014 static void
2015 isert_handle_wc(struct ib_wc *wc)
2016 {
2017         struct isert_conn *isert_conn;
2018         struct iser_tx_desc *tx_desc;
2019         struct iser_rx_desc *rx_desc;
2020
2021         isert_conn = wc->qp->qp_context;
2022         if (likely(wc->status == IB_WC_SUCCESS)) {
2023                 if (wc->opcode == IB_WC_RECV) {
2024                         rx_desc = (struct iser_rx_desc *)(uintptr_t)wc->wr_id;
2025                         isert_rx_completion(rx_desc, isert_conn, wc->byte_len);
2026                 } else {
2027                         tx_desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
2028                         isert_send_completion(tx_desc, isert_conn);
2029                 }
2030         } else {
2031                 if (wc->status != IB_WC_WR_FLUSH_ERR)
2032                         isert_err("wr id %llx status %d vend_err %x\n",
2033                                   wc->wr_id, wc->status, wc->vendor_err);
2034                 else
2035                         isert_dbg("flush error: wr id %llx\n", wc->wr_id);
2036
2037                 if (wc->wr_id != ISER_FASTREG_LI_WRID)
2038                         isert_cq_comp_err(isert_conn, wc);
2039         }
2040 }
2041
2042 static void
2043 isert_cq_work(struct work_struct *work)
2044 {
2045         enum { isert_poll_budget = 65536 };
2046         struct isert_comp *comp = container_of(work, struct isert_comp,
2047                                                work);
2048         struct ib_wc *const wcs = comp->wcs;
2049         int i, n, completed = 0;
2050
2051         while ((n = ib_poll_cq(comp->cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
2052                 for (i = 0; i < n; i++)
2053                         isert_handle_wc(&wcs[i]);
2054
2055                 completed += n;
2056                 if (completed >= isert_poll_budget)
2057                         break;
2058         }
2059
2060         ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP);
2061 }
2062
2063 static void
2064 isert_cq_callback(struct ib_cq *cq, void *context)
2065 {
2066         struct isert_comp *comp = context;
2067
2068         queue_work(isert_comp_wq, &comp->work);
2069 }
2070
2071 static int
2072 isert_post_response(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd)
2073 {
2074         struct ib_send_wr *wr_failed;
2075         int ret;
2076
2077         ret = ib_post_send(isert_conn->conn_qp, &isert_cmd->tx_desc.send_wr,
2078                            &wr_failed);
2079         if (ret) {
2080                 isert_err("ib_post_send failed with %d\n", ret);
2081                 return ret;
2082         }
2083         return ret;
2084 }
2085
2086 static int
2087 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2088 {
2089         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2090         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2091         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2092         struct iscsi_scsi_rsp *hdr = (struct iscsi_scsi_rsp *)
2093                                 &isert_cmd->tx_desc.iscsi_header;
2094
2095         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2096         iscsit_build_rsp_pdu(cmd, conn, true, hdr);
2097         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2098         /*
2099          * Attach SENSE DATA payload to iSCSI Response PDU
2100          */
2101         if (cmd->se_cmd.sense_buffer &&
2102             ((cmd->se_cmd.se_cmd_flags & SCF_TRANSPORT_TASK_SENSE) ||
2103             (cmd->se_cmd.se_cmd_flags & SCF_EMULATED_TASK_SENSE))) {
2104                 struct isert_device *device = isert_conn->conn_device;
2105                 struct ib_device *ib_dev = device->ib_device;
2106                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2107                 u32 padding, pdu_len;
2108
2109                 put_unaligned_be16(cmd->se_cmd.scsi_sense_length,
2110                                    cmd->sense_buffer);
2111                 cmd->se_cmd.scsi_sense_length += sizeof(__be16);
2112
2113                 padding = -(cmd->se_cmd.scsi_sense_length) & 3;
2114                 hton24(hdr->dlength, (u32)cmd->se_cmd.scsi_sense_length);
2115                 pdu_len = cmd->se_cmd.scsi_sense_length + padding;
2116
2117                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2118                                 (void *)cmd->sense_buffer, pdu_len,
2119                                 DMA_TO_DEVICE);
2120
2121                 isert_cmd->pdu_buf_len = pdu_len;
2122                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2123                 tx_dsg->length  = pdu_len;
2124                 tx_dsg->lkey    = device->mr->lkey;
2125                 isert_cmd->tx_desc.num_sge = 2;
2126         }
2127
2128         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2129
2130         isert_dbg("Posting SCSI Response\n");
2131
2132         return isert_post_response(isert_conn, isert_cmd);
2133 }
2134
2135 static void
2136 isert_aborted_task(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2137 {
2138         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2139         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2140         struct isert_device *device = isert_conn->conn_device;
2141
2142         spin_lock_bh(&conn->cmd_lock);
2143         if (!list_empty(&cmd->i_conn_node))
2144                 list_del_init(&cmd->i_conn_node);
2145         spin_unlock_bh(&conn->cmd_lock);
2146
2147         if (cmd->data_direction == DMA_TO_DEVICE)
2148                 iscsit_stop_dataout_timer(cmd);
2149
2150         device->unreg_rdma_mem(isert_cmd, isert_conn);
2151 }
2152
2153 static enum target_prot_op
2154 isert_get_sup_prot_ops(struct iscsi_conn *conn)
2155 {
2156         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2157         struct isert_device *device = isert_conn->conn_device;
2158
2159         if (conn->tpg->tpg_attrib.t10_pi) {
2160                 if (device->pi_capable) {
2161                         isert_info("conn %p PI offload enabled\n", isert_conn);
2162                         isert_conn->pi_support = true;
2163                         return TARGET_PROT_ALL;
2164                 }
2165         }
2166
2167         isert_info("conn %p PI offload disabled\n", isert_conn);
2168         isert_conn->pi_support = false;
2169
2170         return TARGET_PROT_NORMAL;
2171 }
2172
2173 static int
2174 isert_put_nopin(struct iscsi_cmd *cmd, struct iscsi_conn *conn,
2175                 bool nopout_response)
2176 {
2177         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2178         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2179         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2180
2181         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2182         iscsit_build_nopin_rsp(cmd, conn, (struct iscsi_nopin *)
2183                                &isert_cmd->tx_desc.iscsi_header,
2184                                nopout_response);
2185         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2186         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2187
2188         isert_dbg("conn %p Posting NOPIN Response\n", isert_conn);
2189
2190         return isert_post_response(isert_conn, isert_cmd);
2191 }
2192
2193 static int
2194 isert_put_logout_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2195 {
2196         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2197         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2198         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2199
2200         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2201         iscsit_build_logout_rsp(cmd, conn, (struct iscsi_logout_rsp *)
2202                                 &isert_cmd->tx_desc.iscsi_header);
2203         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2204         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2205
2206         isert_dbg("conn %p Posting Logout Response\n", isert_conn);
2207
2208         return isert_post_response(isert_conn, isert_cmd);
2209 }
2210
2211 static int
2212 isert_put_tm_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2213 {
2214         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2215         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2216         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2217
2218         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2219         iscsit_build_task_mgt_rsp(cmd, conn, (struct iscsi_tm_rsp *)
2220                                   &isert_cmd->tx_desc.iscsi_header);
2221         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2222         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2223
2224         isert_dbg("conn %p Posting Task Management Response\n", isert_conn);
2225
2226         return isert_post_response(isert_conn, isert_cmd);
2227 }
2228
2229 static int
2230 isert_put_reject(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2231 {
2232         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2233         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2234         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2235         struct isert_device *device = isert_conn->conn_device;
2236         struct ib_device *ib_dev = device->ib_device;
2237         struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2238         struct iscsi_reject *hdr =
2239                 (struct iscsi_reject *)&isert_cmd->tx_desc.iscsi_header;
2240
2241         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2242         iscsit_build_reject(cmd, conn, hdr);
2243         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2244
2245         hton24(hdr->dlength, ISCSI_HDR_LEN);
2246         isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2247                         (void *)cmd->buf_ptr, ISCSI_HDR_LEN,
2248                         DMA_TO_DEVICE);
2249         isert_cmd->pdu_buf_len = ISCSI_HDR_LEN;
2250         tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2251         tx_dsg->length  = ISCSI_HDR_LEN;
2252         tx_dsg->lkey    = device->mr->lkey;
2253         isert_cmd->tx_desc.num_sge = 2;
2254
2255         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2256
2257         isert_dbg("conn %p Posting Reject\n", isert_conn);
2258
2259         return isert_post_response(isert_conn, isert_cmd);
2260 }
2261
2262 static int
2263 isert_put_text_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2264 {
2265         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2266         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2267         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2268         struct iscsi_text_rsp *hdr =
2269                 (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
2270         u32 txt_rsp_len;
2271         int rc;
2272
2273         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2274         rc = iscsit_build_text_rsp(cmd, conn, hdr, ISCSI_INFINIBAND);
2275         if (rc < 0)
2276                 return rc;
2277
2278         txt_rsp_len = rc;
2279         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2280
2281         if (txt_rsp_len) {
2282                 struct isert_device *device = isert_conn->conn_device;
2283                 struct ib_device *ib_dev = device->ib_device;
2284                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2285                 void *txt_rsp_buf = cmd->buf_ptr;
2286
2287                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2288                                 txt_rsp_buf, txt_rsp_len, DMA_TO_DEVICE);
2289
2290                 isert_cmd->pdu_buf_len = txt_rsp_len;
2291                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2292                 tx_dsg->length  = txt_rsp_len;
2293                 tx_dsg->lkey    = device->mr->lkey;
2294                 isert_cmd->tx_desc.num_sge = 2;
2295         }
2296         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2297
2298         isert_dbg("conn %p Text Response\n", isert_conn);
2299
2300         return isert_post_response(isert_conn, isert_cmd);
2301 }
2302
2303 static int
2304 isert_build_rdma_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
2305                     struct ib_sge *ib_sge, struct ib_send_wr *send_wr,
2306                     u32 data_left, u32 offset)
2307 {
2308         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
2309         struct scatterlist *sg_start, *tmp_sg;
2310         struct isert_device *device = isert_conn->conn_device;
2311         struct ib_device *ib_dev = device->ib_device;
2312         u32 sg_off, page_off;
2313         int i = 0, sg_nents;
2314
2315         sg_off = offset / PAGE_SIZE;
2316         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
2317         sg_nents = min(cmd->se_cmd.t_data_nents - sg_off, isert_conn->max_sge);
2318         page_off = offset % PAGE_SIZE;
2319
2320         send_wr->sg_list = ib_sge;
2321         send_wr->num_sge = sg_nents;
2322         send_wr->wr_id = (uintptr_t)&isert_cmd->tx_desc;
2323         /*
2324          * Perform mapping of TCM scatterlist memory ib_sge dma_addr.
2325          */
2326         for_each_sg(sg_start, tmp_sg, sg_nents, i) {
2327                 isert_dbg("RDMA from SGL dma_addr: 0x%llx dma_len: %u, "
2328                           "page_off: %u\n",
2329                           (unsigned long long)tmp_sg->dma_address,
2330                           tmp_sg->length, page_off);
2331
2332                 ib_sge->addr = ib_sg_dma_address(ib_dev, tmp_sg) + page_off;
2333                 ib_sge->length = min_t(u32, data_left,
2334                                 ib_sg_dma_len(ib_dev, tmp_sg) - page_off);
2335                 ib_sge->lkey = device->mr->lkey;
2336
2337                 isert_dbg("RDMA ib_sge: addr: 0x%llx  length: %u lkey: %x\n",
2338                           ib_sge->addr, ib_sge->length, ib_sge->lkey);
2339                 page_off = 0;
2340                 data_left -= ib_sge->length;
2341                 ib_sge++;
2342                 isert_dbg("Incrementing ib_sge pointer to %p\n", ib_sge);
2343         }
2344
2345         isert_dbg("Set outgoing sg_list: %p num_sg: %u from TCM SGLs\n",
2346                   send_wr->sg_list, send_wr->num_sge);
2347
2348         return sg_nents;
2349 }
2350
2351 static int
2352 isert_map_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2353                struct isert_rdma_wr *wr)
2354 {
2355         struct se_cmd *se_cmd = &cmd->se_cmd;
2356         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2357         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2358         struct isert_data_buf *data = &wr->data;
2359         struct ib_send_wr *send_wr;
2360         struct ib_sge *ib_sge;
2361         u32 offset, data_len, data_left, rdma_write_max, va_offset = 0;
2362         int ret = 0, i, ib_sge_cnt;
2363
2364         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2365
2366         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2367         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2368                                  se_cmd->t_data_nents, se_cmd->data_length,
2369                                  offset, wr->iser_ib_op, &wr->data);
2370         if (ret)
2371                 return ret;
2372
2373         data_left = data->len;
2374         offset = data->offset;
2375
2376         ib_sge = kzalloc(sizeof(struct ib_sge) * data->nents, GFP_KERNEL);
2377         if (!ib_sge) {
2378                 isert_warn("Unable to allocate ib_sge\n");
2379                 ret = -ENOMEM;
2380                 goto unmap_cmd;
2381         }
2382         wr->ib_sge = ib_sge;
2383
2384         wr->send_wr_num = DIV_ROUND_UP(data->nents, isert_conn->max_sge);
2385         wr->send_wr = kzalloc(sizeof(struct ib_send_wr) * wr->send_wr_num,
2386                                 GFP_KERNEL);
2387         if (!wr->send_wr) {
2388                 isert_dbg("Unable to allocate wr->send_wr\n");
2389                 ret = -ENOMEM;
2390                 goto unmap_cmd;
2391         }
2392
2393         wr->isert_cmd = isert_cmd;
2394         rdma_write_max = isert_conn->max_sge * PAGE_SIZE;
2395
2396         for (i = 0; i < wr->send_wr_num; i++) {
2397                 send_wr = &isert_cmd->rdma_wr.send_wr[i];
2398                 data_len = min(data_left, rdma_write_max);
2399
2400                 send_wr->send_flags = 0;
2401                 if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2402                         send_wr->opcode = IB_WR_RDMA_WRITE;
2403                         send_wr->wr.rdma.remote_addr = isert_cmd->read_va + offset;
2404                         send_wr->wr.rdma.rkey = isert_cmd->read_stag;
2405                         if (i + 1 == wr->send_wr_num)
2406                                 send_wr->next = &isert_cmd->tx_desc.send_wr;
2407                         else
2408                                 send_wr->next = &wr->send_wr[i + 1];
2409                 } else {
2410                         send_wr->opcode = IB_WR_RDMA_READ;
2411                         send_wr->wr.rdma.remote_addr = isert_cmd->write_va + va_offset;
2412                         send_wr->wr.rdma.rkey = isert_cmd->write_stag;
2413                         if (i + 1 == wr->send_wr_num)
2414                                 send_wr->send_flags = IB_SEND_SIGNALED;
2415                         else
2416                                 send_wr->next = &wr->send_wr[i + 1];
2417                 }
2418
2419                 ib_sge_cnt = isert_build_rdma_wr(isert_conn, isert_cmd, ib_sge,
2420                                         send_wr, data_len, offset);
2421                 ib_sge += ib_sge_cnt;
2422
2423                 offset += data_len;
2424                 va_offset += data_len;
2425                 data_left -= data_len;
2426         }
2427
2428         return 0;
2429 unmap_cmd:
2430         isert_unmap_data_buf(isert_conn, data);
2431
2432         return ret;
2433 }
2434
2435 static int
2436 isert_map_fr_pagelist(struct ib_device *ib_dev,
2437                       struct scatterlist *sg_start, int sg_nents, u64 *fr_pl)
2438 {
2439         u64 start_addr, end_addr, page, chunk_start = 0;
2440         struct scatterlist *tmp_sg;
2441         int i = 0, new_chunk, last_ent, n_pages;
2442
2443         n_pages = 0;
2444         new_chunk = 1;
2445         last_ent = sg_nents - 1;
2446         for_each_sg(sg_start, tmp_sg, sg_nents, i) {
2447                 start_addr = ib_sg_dma_address(ib_dev, tmp_sg);
2448                 if (new_chunk)
2449                         chunk_start = start_addr;
2450                 end_addr = start_addr + ib_sg_dma_len(ib_dev, tmp_sg);
2451
2452                 isert_dbg("SGL[%d] dma_addr: 0x%llx len: %u\n",
2453                           i, (unsigned long long)tmp_sg->dma_address,
2454                           tmp_sg->length);
2455
2456                 if ((end_addr & ~PAGE_MASK) && i < last_ent) {
2457                         new_chunk = 0;
2458                         continue;
2459                 }
2460                 new_chunk = 1;
2461
2462                 page = chunk_start & PAGE_MASK;
2463                 do {
2464                         fr_pl[n_pages++] = page;
2465                         isert_dbg("Mapped page_list[%d] page_addr: 0x%llx\n",
2466                                   n_pages - 1, page);
2467                         page += PAGE_SIZE;
2468                 } while (page < end_addr);
2469         }
2470
2471         return n_pages;
2472 }
2473
2474 static inline void
2475 isert_inv_rkey(struct ib_send_wr *inv_wr, struct ib_mr *mr)
2476 {
2477         u32 rkey;
2478
2479         memset(inv_wr, 0, sizeof(*inv_wr));
2480         inv_wr->wr_id = ISER_FASTREG_LI_WRID;
2481         inv_wr->opcode = IB_WR_LOCAL_INV;
2482         inv_wr->ex.invalidate_rkey = mr->rkey;
2483
2484         /* Bump the key */
2485         rkey = ib_inc_rkey(mr->rkey);
2486         ib_update_fast_reg_key(mr, rkey);
2487 }
2488
2489 static int
2490 isert_fast_reg_mr(struct isert_conn *isert_conn,
2491                   struct fast_reg_descriptor *fr_desc,
2492                   struct isert_data_buf *mem,
2493                   enum isert_indicator ind,
2494                   struct ib_sge *sge)
2495 {
2496         struct isert_device *device = isert_conn->conn_device;
2497         struct ib_device *ib_dev = device->ib_device;
2498         struct ib_mr *mr;
2499         struct ib_fast_reg_page_list *frpl;
2500         struct ib_send_wr fr_wr, inv_wr;
2501         struct ib_send_wr *bad_wr, *wr = NULL;
2502         int ret, pagelist_len;
2503         u32 page_off;
2504
2505         if (mem->dma_nents == 1) {
2506                 sge->lkey = device->mr->lkey;
2507                 sge->addr = ib_sg_dma_address(ib_dev, &mem->sg[0]);
2508                 sge->length = ib_sg_dma_len(ib_dev, &mem->sg[0]);
2509                 isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2510                          sge->addr, sge->length, sge->lkey);
2511                 return 0;
2512         }
2513
2514         if (ind == ISERT_DATA_KEY_VALID) {
2515                 /* Registering data buffer */
2516                 mr = fr_desc->data_mr;
2517                 frpl = fr_desc->data_frpl;
2518         } else {
2519                 /* Registering protection buffer */
2520                 mr = fr_desc->pi_ctx->prot_mr;
2521                 frpl = fr_desc->pi_ctx->prot_frpl;
2522         }
2523
2524         page_off = mem->offset % PAGE_SIZE;
2525
2526         isert_dbg("Use fr_desc %p sg_nents %d offset %u\n",
2527                   fr_desc, mem->nents, mem->offset);
2528
2529         pagelist_len = isert_map_fr_pagelist(ib_dev, mem->sg, mem->nents,
2530                                              &frpl->page_list[0]);
2531
2532         if (!(fr_desc->ind & ind)) {
2533                 isert_inv_rkey(&inv_wr, mr);
2534                 wr = &inv_wr;
2535         }
2536
2537         /* Prepare FASTREG WR */
2538         memset(&fr_wr, 0, sizeof(fr_wr));
2539         fr_wr.wr_id = ISER_FASTREG_LI_WRID;
2540         fr_wr.opcode = IB_WR_FAST_REG_MR;
2541         fr_wr.wr.fast_reg.iova_start = frpl->page_list[0] + page_off;
2542         fr_wr.wr.fast_reg.page_list = frpl;
2543         fr_wr.wr.fast_reg.page_list_len = pagelist_len;
2544         fr_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
2545         fr_wr.wr.fast_reg.length = mem->len;
2546         fr_wr.wr.fast_reg.rkey = mr->rkey;
2547         fr_wr.wr.fast_reg.access_flags = IB_ACCESS_LOCAL_WRITE;
2548
2549         if (!wr)
2550                 wr = &fr_wr;
2551         else
2552                 wr->next = &fr_wr;
2553
2554         ret = ib_post_send(isert_conn->conn_qp, wr, &bad_wr);
2555         if (ret) {
2556                 isert_err("fast registration failed, ret:%d\n", ret);
2557                 return ret;
2558         }
2559         fr_desc->ind &= ~ind;
2560
2561         sge->lkey = mr->lkey;
2562         sge->addr = frpl->page_list[0] + page_off;
2563         sge->length = mem->len;
2564
2565         isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2566                   sge->addr, sge->length, sge->lkey);
2567
2568         return ret;
2569 }
2570
2571 static inline void
2572 isert_set_dif_domain(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs,
2573                      struct ib_sig_domain *domain)
2574 {
2575         domain->sig_type = IB_SIG_TYPE_T10_DIF;
2576         domain->sig.dif.bg_type = IB_T10DIF_CRC;
2577         domain->sig.dif.pi_interval = se_cmd->se_dev->dev_attrib.block_size;
2578         domain->sig.dif.ref_tag = se_cmd->reftag_seed;
2579         /*
2580          * At the moment we hard code those, but if in the future
2581          * the target core would like to use it, we will take it
2582          * from se_cmd.
2583          */
2584         domain->sig.dif.apptag_check_mask = 0xffff;
2585         domain->sig.dif.app_escape = true;
2586         domain->sig.dif.ref_escape = true;
2587         if (se_cmd->prot_type == TARGET_DIF_TYPE1_PROT ||
2588             se_cmd->prot_type == TARGET_DIF_TYPE2_PROT)
2589                 domain->sig.dif.ref_remap = true;
2590 };
2591
2592 static int
2593 isert_set_sig_attrs(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs)
2594 {
2595         switch (se_cmd->prot_op) {
2596         case TARGET_PROT_DIN_INSERT:
2597         case TARGET_PROT_DOUT_STRIP:
2598                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
2599                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2600                 break;
2601         case TARGET_PROT_DOUT_INSERT:
2602         case TARGET_PROT_DIN_STRIP:
2603                 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
2604                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2605                 break;
2606         case TARGET_PROT_DIN_PASS:
2607         case TARGET_PROT_DOUT_PASS:
2608                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2609                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2610                 break;
2611         default:
2612                 isert_err("Unsupported PI operation %d\n", se_cmd->prot_op);
2613                 return -EINVAL;
2614         }
2615
2616         return 0;
2617 }
2618
2619 static inline u8
2620 isert_set_prot_checks(u8 prot_checks)
2621 {
2622         return (prot_checks & TARGET_DIF_CHECK_GUARD  ? 0xc0 : 0) |
2623                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x30 : 0) |
2624                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x0f : 0);
2625 }
2626
2627 static int
2628 isert_reg_sig_mr(struct isert_conn *isert_conn,
2629                  struct se_cmd *se_cmd,
2630                  struct isert_rdma_wr *rdma_wr,
2631                  struct fast_reg_descriptor *fr_desc)
2632 {
2633         struct ib_send_wr sig_wr, inv_wr;
2634         struct ib_send_wr *bad_wr, *wr = NULL;
2635         struct pi_context *pi_ctx = fr_desc->pi_ctx;
2636         struct ib_sig_attrs sig_attrs;
2637         int ret;
2638
2639         memset(&sig_attrs, 0, sizeof(sig_attrs));
2640         ret = isert_set_sig_attrs(se_cmd, &sig_attrs);
2641         if (ret)
2642                 goto err;
2643
2644         sig_attrs.check_mask = isert_set_prot_checks(se_cmd->prot_checks);
2645
2646         if (!(fr_desc->ind & ISERT_SIG_KEY_VALID)) {
2647                 isert_inv_rkey(&inv_wr, pi_ctx->sig_mr);
2648                 wr = &inv_wr;
2649         }
2650
2651         memset(&sig_wr, 0, sizeof(sig_wr));
2652         sig_wr.opcode = IB_WR_REG_SIG_MR;
2653         sig_wr.wr_id = ISER_FASTREG_LI_WRID;
2654         sig_wr.sg_list = &rdma_wr->ib_sg[DATA];
2655         sig_wr.num_sge = 1;
2656         sig_wr.wr.sig_handover.access_flags = IB_ACCESS_LOCAL_WRITE;
2657         sig_wr.wr.sig_handover.sig_attrs = &sig_attrs;
2658         sig_wr.wr.sig_handover.sig_mr = pi_ctx->sig_mr;
2659         if (se_cmd->t_prot_sg)
2660                 sig_wr.wr.sig_handover.prot = &rdma_wr->ib_sg[PROT];
2661
2662         if (!wr)
2663                 wr = &sig_wr;
2664         else
2665                 wr->next = &sig_wr;
2666
2667         ret = ib_post_send(isert_conn->conn_qp, wr, &bad_wr);
2668         if (ret) {
2669                 isert_err("fast registration failed, ret:%d\n", ret);
2670                 goto err;
2671         }
2672         fr_desc->ind &= ~ISERT_SIG_KEY_VALID;
2673
2674         rdma_wr->ib_sg[SIG].lkey = pi_ctx->sig_mr->lkey;
2675         rdma_wr->ib_sg[SIG].addr = 0;
2676         rdma_wr->ib_sg[SIG].length = se_cmd->data_length;
2677         if (se_cmd->prot_op != TARGET_PROT_DIN_STRIP &&
2678             se_cmd->prot_op != TARGET_PROT_DOUT_INSERT)
2679                 /*
2680                  * We have protection guards on the wire
2681                  * so we need to set a larget transfer
2682                  */
2683                 rdma_wr->ib_sg[SIG].length += se_cmd->prot_length;
2684
2685         isert_dbg("sig_sge: addr: 0x%llx  length: %u lkey: %x\n",
2686                   rdma_wr->ib_sg[SIG].addr, rdma_wr->ib_sg[SIG].length,
2687                   rdma_wr->ib_sg[SIG].lkey);
2688 err:
2689         return ret;
2690 }
2691
2692 static int
2693 isert_handle_prot_cmd(struct isert_conn *isert_conn,
2694                       struct isert_cmd *isert_cmd,
2695                       struct isert_rdma_wr *wr)
2696 {
2697         struct isert_device *device = isert_conn->conn_device;
2698         struct se_cmd *se_cmd = &isert_cmd->iscsi_cmd->se_cmd;
2699         int ret;
2700
2701         if (!wr->fr_desc->pi_ctx) {
2702                 ret = isert_create_pi_ctx(wr->fr_desc,
2703                                           device->ib_device,
2704                                           device->pd);
2705                 if (ret) {
2706                         isert_err("conn %p failed to allocate pi_ctx\n",
2707                                   isert_conn);
2708                         return ret;
2709                 }
2710         }
2711
2712         if (se_cmd->t_prot_sg) {
2713                 ret = isert_map_data_buf(isert_conn, isert_cmd,
2714                                          se_cmd->t_prot_sg,
2715                                          se_cmd->t_prot_nents,
2716                                          se_cmd->prot_length,
2717                                          0, wr->iser_ib_op, &wr->prot);
2718                 if (ret) {
2719                         isert_err("conn %p failed to map protection buffer\n",
2720                                   isert_conn);
2721                         return ret;
2722                 }
2723
2724                 memset(&wr->ib_sg[PROT], 0, sizeof(wr->ib_sg[PROT]));
2725                 ret = isert_fast_reg_mr(isert_conn, wr->fr_desc, &wr->prot,
2726                                         ISERT_PROT_KEY_VALID, &wr->ib_sg[PROT]);
2727                 if (ret) {
2728                         isert_err("conn %p failed to fast reg mr\n",
2729                                   isert_conn);
2730                         goto unmap_prot_cmd;
2731                 }
2732         }
2733
2734         ret = isert_reg_sig_mr(isert_conn, se_cmd, wr, wr->fr_desc);
2735         if (ret) {
2736                 isert_err("conn %p failed to fast reg mr\n",
2737                           isert_conn);
2738                 goto unmap_prot_cmd;
2739         }
2740         wr->fr_desc->ind |= ISERT_PROTECTED;
2741
2742         return 0;
2743
2744 unmap_prot_cmd:
2745         if (se_cmd->t_prot_sg)
2746                 isert_unmap_data_buf(isert_conn, &wr->prot);
2747
2748         return ret;
2749 }
2750
2751 static int
2752 isert_reg_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2753                struct isert_rdma_wr *wr)
2754 {
2755         struct se_cmd *se_cmd = &cmd->se_cmd;
2756         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2757         struct isert_conn *isert_conn = conn->context;
2758         struct fast_reg_descriptor *fr_desc = NULL;
2759         struct ib_send_wr *send_wr;
2760         struct ib_sge *ib_sg;
2761         u32 offset;
2762         int ret = 0;
2763         unsigned long flags;
2764
2765         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2766
2767         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2768         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2769                                  se_cmd->t_data_nents, se_cmd->data_length,
2770                                  offset, wr->iser_ib_op, &wr->data);
2771         if (ret)
2772                 return ret;
2773
2774         if (wr->data.dma_nents != 1 || isert_prot_cmd(isert_conn, se_cmd)) {
2775                 spin_lock_irqsave(&isert_conn->conn_lock, flags);
2776                 fr_desc = list_first_entry(&isert_conn->conn_fr_pool,
2777                                            struct fast_reg_descriptor, list);
2778                 list_del(&fr_desc->list);
2779                 spin_unlock_irqrestore(&isert_conn->conn_lock, flags);
2780                 wr->fr_desc = fr_desc;
2781         }
2782
2783         ret = isert_fast_reg_mr(isert_conn, fr_desc, &wr->data,
2784                                 ISERT_DATA_KEY_VALID, &wr->ib_sg[DATA]);
2785         if (ret)
2786                 goto unmap_cmd;
2787
2788         if (isert_prot_cmd(isert_conn, se_cmd)) {
2789                 ret = isert_handle_prot_cmd(isert_conn, isert_cmd, wr);
2790                 if (ret)
2791                         goto unmap_cmd;
2792
2793                 ib_sg = &wr->ib_sg[SIG];
2794         } else {
2795                 ib_sg = &wr->ib_sg[DATA];
2796         }
2797
2798         memcpy(&wr->s_ib_sge, ib_sg, sizeof(*ib_sg));
2799         wr->ib_sge = &wr->s_ib_sge;
2800         wr->send_wr_num = 1;
2801         memset(&wr->s_send_wr, 0, sizeof(*send_wr));
2802         wr->send_wr = &wr->s_send_wr;
2803         wr->isert_cmd = isert_cmd;
2804
2805         send_wr = &isert_cmd->rdma_wr.s_send_wr;
2806         send_wr->sg_list = &wr->s_ib_sge;
2807         send_wr->num_sge = 1;
2808         send_wr->wr_id = (uintptr_t)&isert_cmd->tx_desc;
2809         if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2810                 send_wr->opcode = IB_WR_RDMA_WRITE;
2811                 send_wr->wr.rdma.remote_addr = isert_cmd->read_va;
2812                 send_wr->wr.rdma.rkey = isert_cmd->read_stag;
2813                 send_wr->send_flags = !isert_prot_cmd(isert_conn, se_cmd) ?
2814                                       0 : IB_SEND_SIGNALED;
2815         } else {
2816                 send_wr->opcode = IB_WR_RDMA_READ;
2817                 send_wr->wr.rdma.remote_addr = isert_cmd->write_va;
2818                 send_wr->wr.rdma.rkey = isert_cmd->write_stag;
2819                 send_wr->send_flags = IB_SEND_SIGNALED;
2820         }
2821
2822         return 0;
2823
2824 unmap_cmd:
2825         if (fr_desc) {
2826                 spin_lock_irqsave(&isert_conn->conn_lock, flags);
2827                 list_add_tail(&fr_desc->list, &isert_conn->conn_fr_pool);
2828                 spin_unlock_irqrestore(&isert_conn->conn_lock, flags);
2829         }
2830         isert_unmap_data_buf(isert_conn, &wr->data);
2831
2832         return ret;
2833 }
2834
2835 static int
2836 isert_put_datain(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2837 {
2838         struct se_cmd *se_cmd = &cmd->se_cmd;
2839         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2840         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2841         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2842         struct isert_device *device = isert_conn->conn_device;
2843         struct ib_send_wr *wr_failed;
2844         int rc;
2845
2846         isert_dbg("Cmd: %p RDMA_WRITE data_length: %u\n",
2847                  isert_cmd, se_cmd->data_length);
2848
2849         wr->iser_ib_op = ISER_IB_RDMA_WRITE;
2850         rc = device->reg_rdma_mem(conn, cmd, wr);
2851         if (rc) {
2852                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2853                 return rc;
2854         }
2855
2856         if (!isert_prot_cmd(isert_conn, se_cmd)) {
2857                 /*
2858                  * Build isert_conn->tx_desc for iSCSI response PDU and attach
2859                  */
2860                 isert_create_send_desc(isert_conn, isert_cmd,
2861                                        &isert_cmd->tx_desc);
2862                 iscsit_build_rsp_pdu(cmd, conn, true, (struct iscsi_scsi_rsp *)
2863                                      &isert_cmd->tx_desc.iscsi_header);
2864                 isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2865                 isert_init_send_wr(isert_conn, isert_cmd,
2866                                    &isert_cmd->tx_desc.send_wr);
2867                 isert_cmd->rdma_wr.s_send_wr.next = &isert_cmd->tx_desc.send_wr;
2868                 wr->send_wr_num += 1;
2869         }
2870
2871         rc = ib_post_send(isert_conn->conn_qp, wr->send_wr, &wr_failed);
2872         if (rc)
2873                 isert_warn("ib_post_send() failed for IB_WR_RDMA_WRITE\n");
2874
2875         if (!isert_prot_cmd(isert_conn, se_cmd))
2876                 isert_dbg("Cmd: %p posted RDMA_WRITE + Response for iSER Data "
2877                          "READ\n", isert_cmd);
2878         else
2879                 isert_dbg("Cmd: %p posted RDMA_WRITE for iSER Data READ\n",
2880                          isert_cmd);
2881
2882         return 1;
2883 }
2884
2885 static int
2886 isert_get_dataout(struct iscsi_conn *conn, struct iscsi_cmd *cmd, bool recovery)
2887 {
2888         struct se_cmd *se_cmd = &cmd->se_cmd;
2889         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2890         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2891         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
2892         struct isert_device *device = isert_conn->conn_device;
2893         struct ib_send_wr *wr_failed;
2894         int rc;
2895
2896         isert_dbg("Cmd: %p RDMA_READ data_length: %u write_data_done: %u\n",
2897                  isert_cmd, se_cmd->data_length, cmd->write_data_done);
2898         wr->iser_ib_op = ISER_IB_RDMA_READ;
2899         rc = device->reg_rdma_mem(conn, cmd, wr);
2900         if (rc) {
2901                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2902                 return rc;
2903         }
2904
2905         rc = ib_post_send(isert_conn->conn_qp, wr->send_wr, &wr_failed);
2906         if (rc)
2907                 isert_warn("ib_post_send() failed for IB_WR_RDMA_READ\n");
2908
2909         isert_dbg("Cmd: %p posted RDMA_READ memory for ISER Data WRITE\n",
2910                  isert_cmd);
2911
2912         return 0;
2913 }
2914
2915 static int
2916 isert_immediate_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2917 {
2918         int ret;
2919
2920         switch (state) {
2921         case ISTATE_SEND_NOPIN_WANT_RESPONSE:
2922                 ret = isert_put_nopin(cmd, conn, false);
2923                 break;
2924         default:
2925                 isert_err("Unknown immediate state: 0x%02x\n", state);
2926                 ret = -EINVAL;
2927                 break;
2928         }
2929
2930         return ret;
2931 }
2932
2933 static int
2934 isert_response_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2935 {
2936         struct isert_conn *isert_conn = conn->context;
2937         int ret;
2938
2939         switch (state) {
2940         case ISTATE_SEND_LOGOUTRSP:
2941                 ret = isert_put_logout_rsp(cmd, conn);
2942                 if (!ret)
2943                         isert_conn->logout_posted = true;
2944                 break;
2945         case ISTATE_SEND_NOPIN:
2946                 ret = isert_put_nopin(cmd, conn, true);
2947                 break;
2948         case ISTATE_SEND_TASKMGTRSP:
2949                 ret = isert_put_tm_rsp(cmd, conn);
2950                 break;
2951         case ISTATE_SEND_REJECT:
2952                 ret = isert_put_reject(cmd, conn);
2953                 break;
2954         case ISTATE_SEND_TEXTRSP:
2955                 ret = isert_put_text_rsp(cmd, conn);
2956                 break;
2957         case ISTATE_SEND_STATUS:
2958                 /*
2959                  * Special case for sending non GOOD SCSI status from TX thread
2960                  * context during pre se_cmd excecution failure.
2961                  */
2962                 ret = isert_put_response(conn, cmd);
2963                 break;
2964         default:
2965                 isert_err("Unknown response state: 0x%02x\n", state);
2966                 ret = -EINVAL;
2967                 break;
2968         }
2969
2970         return ret;
2971 }
2972
2973 struct rdma_cm_id *
2974 isert_setup_id(struct isert_np *isert_np)
2975 {
2976         struct iscsi_np *np = isert_np->np;
2977         struct rdma_cm_id *id;
2978         struct sockaddr *sa;
2979         int ret;
2980
2981         sa = (struct sockaddr *)&np->np_sockaddr;
2982         isert_dbg("ksockaddr: %p, sa: %p\n", &np->np_sockaddr, sa);
2983
2984         id = rdma_create_id(isert_cma_handler, isert_np,
2985                             RDMA_PS_TCP, IB_QPT_RC);
2986         if (IS_ERR(id)) {
2987                 isert_err("rdma_create_id() failed: %ld\n", PTR_ERR(id));
2988                 ret = PTR_ERR(id);
2989                 goto out;
2990         }
2991         isert_dbg("id %p context %p\n", id, id->context);
2992
2993         ret = rdma_bind_addr(id, sa);
2994         if (ret) {
2995                 isert_err("rdma_bind_addr() failed: %d\n", ret);
2996                 goto out_id;
2997         }
2998
2999         ret = rdma_listen(id, ISERT_RDMA_LISTEN_BACKLOG);
3000         if (ret) {
3001                 isert_err("rdma_listen() failed: %d\n", ret);
3002                 goto out_id;
3003         }
3004
3005         return id;
3006 out_id:
3007         rdma_destroy_id(id);
3008 out:
3009         return ERR_PTR(ret);
3010 }
3011
3012 static int
3013 isert_setup_np(struct iscsi_np *np,
3014                struct __kernel_sockaddr_storage *ksockaddr)
3015 {
3016         struct isert_np *isert_np;
3017         struct rdma_cm_id *isert_lid;
3018         int ret;
3019
3020         isert_np = kzalloc(sizeof(struct isert_np), GFP_KERNEL);
3021         if (!isert_np) {
3022                 isert_err("Unable to allocate struct isert_np\n");
3023                 return -ENOMEM;
3024         }
3025         sema_init(&isert_np->np_sem, 0);
3026         mutex_init(&isert_np->np_accept_mutex);
3027         INIT_LIST_HEAD(&isert_np->np_accept_list);
3028         init_completion(&isert_np->np_login_comp);
3029         isert_np->np = np;
3030
3031         /*
3032          * Setup the np->np_sockaddr from the passed sockaddr setup
3033          * in iscsi_target_configfs.c code..
3034          */
3035         memcpy(&np->np_sockaddr, ksockaddr,
3036                sizeof(struct __kernel_sockaddr_storage));
3037
3038         isert_lid = isert_setup_id(isert_np);
3039         if (IS_ERR(isert_lid)) {
3040                 ret = PTR_ERR(isert_lid);
3041                 goto out;
3042         }
3043
3044         isert_np->np_cm_id = isert_lid;
3045         np->np_context = isert_np;
3046
3047         return 0;
3048
3049 out:
3050         kfree(isert_np);
3051
3052         return ret;
3053 }
3054
3055 static int
3056 isert_rdma_accept(struct isert_conn *isert_conn)
3057 {
3058         struct rdma_cm_id *cm_id = isert_conn->conn_cm_id;
3059         struct rdma_conn_param cp;
3060         int ret;
3061
3062         memset(&cp, 0, sizeof(struct rdma_conn_param));
3063         cp.initiator_depth = isert_conn->initiator_depth;
3064         cp.retry_count = 7;
3065         cp.rnr_retry_count = 7;
3066
3067         ret = rdma_accept(cm_id, &cp);
3068         if (ret) {
3069                 isert_err("rdma_accept() failed with: %d\n", ret);
3070                 return ret;
3071         }
3072
3073         return 0;
3074 }
3075
3076 static int
3077 isert_get_login_rx(struct iscsi_conn *conn, struct iscsi_login *login)
3078 {
3079         struct isert_conn *isert_conn = (struct isert_conn *)conn->context;
3080         int ret;
3081
3082         isert_info("before login_req comp conn: %p\n", isert_conn);
3083         ret = wait_for_completion_interruptible(&isert_conn->login_req_comp);
3084         if (ret) {
3085                 isert_err("isert_conn %p interrupted before got login req\n",
3086                           isert_conn);
3087                 return ret;
3088         }
3089         reinit_completion(&isert_conn->login_req_comp);
3090
3091         /*
3092          * For login requests after the first PDU, isert_rx_login_req() will
3093          * kick schedule_delayed_work(&conn->login_work) as the packet is
3094          * received, which turns this callback from iscsi_target_do_login_rx()
3095          * into a NOP.
3096          */
3097         if (!login->first_request)
3098                 return 0;
3099
3100         isert_rx_login_req(isert_conn);
3101
3102         isert_info("before conn_login_comp conn: %p\n", conn);
3103         ret = wait_for_completion_interruptible(&isert_conn->conn_login_comp);
3104         if (ret)
3105                 return ret;
3106
3107         isert_info("processing login->req: %p\n", login->req);
3108
3109         return 0;
3110 }
3111
3112 static void
3113 isert_set_conn_info(struct iscsi_np *np, struct iscsi_conn *conn,
3114                     struct isert_conn *isert_conn)
3115 {
3116         struct rdma_cm_id *cm_id = isert_conn->conn_cm_id;
3117         struct rdma_route *cm_route = &cm_id->route;
3118         struct sockaddr_in *sock_in;
3119         struct sockaddr_in6 *sock_in6;
3120
3121         conn->login_family = np->np_sockaddr.ss_family;
3122
3123         if (np->np_sockaddr.ss_family == AF_INET6) {
3124                 sock_in6 = (struct sockaddr_in6 *)&cm_route->addr.dst_addr;
3125                 snprintf(conn->login_ip, sizeof(conn->login_ip), "%pI6c",
3126                          &sock_in6->sin6_addr.in6_u);
3127                 conn->login_port = ntohs(sock_in6->sin6_port);
3128
3129                 sock_in6 = (struct sockaddr_in6 *)&cm_route->addr.src_addr;
3130                 snprintf(conn->local_ip, sizeof(conn->local_ip), "%pI6c",
3131                          &sock_in6->sin6_addr.in6_u);
3132                 conn->local_port = ntohs(sock_in6->sin6_port);
3133         } else {
3134                 sock_in = (struct sockaddr_in *)&cm_route->addr.dst_addr;
3135                 sprintf(conn->login_ip, "%pI4",
3136                         &sock_in->sin_addr.s_addr);
3137                 conn->login_port = ntohs(sock_in->sin_port);
3138
3139                 sock_in = (struct sockaddr_in *)&cm_route->addr.src_addr;
3140                 sprintf(conn->local_ip, "%pI4",
3141                         &sock_in->sin_addr.s_addr);
3142                 conn->local_port = ntohs(sock_in->sin_port);
3143         }
3144 }
3145
3146 static int
3147 isert_accept_np(struct iscsi_np *np, struct iscsi_conn *conn)
3148 {
3149         struct isert_np *isert_np = (struct isert_np *)np->np_context;
3150         struct isert_conn *isert_conn;
3151         int max_accept = 0, ret;
3152
3153 accept_wait:
3154         ret = down_interruptible(&isert_np->np_sem);
3155         if (ret || max_accept > 5)
3156                 return -ENODEV;
3157
3158         spin_lock_bh(&np->np_thread_lock);
3159         if (np->np_thread_state >= ISCSI_NP_THREAD_RESET) {
3160                 spin_unlock_bh(&np->np_thread_lock);
3161                 isert_dbg("np_thread_state %d\n",
3162                          np->np_thread_state);
3163                 /**
3164                  * No point in stalling here when np_thread
3165                  * is in state RESET/SHUTDOWN/EXIT - bail
3166                  **/
3167                 return -ENODEV;
3168         }
3169         spin_unlock_bh(&np->np_thread_lock);
3170
3171         mutex_lock(&isert_np->np_accept_mutex);
3172         if (list_empty(&isert_np->np_accept_list)) {
3173                 mutex_unlock(&isert_np->np_accept_mutex);
3174                 max_accept++;
3175                 goto accept_wait;
3176         }
3177         isert_conn = list_first_entry(&isert_np->np_accept_list,
3178                         struct isert_conn, conn_accept_node);
3179         list_del_init(&isert_conn->conn_accept_node);
3180         mutex_unlock(&isert_np->np_accept_mutex);
3181
3182         conn->context = isert_conn;
3183         isert_conn->conn = conn;
3184         max_accept = 0;
3185
3186         isert_set_conn_info(np, conn, isert_conn);
3187
3188         isert_dbg("Processing isert_conn: %p\n", isert_conn);
3189
3190         return 0;
3191 }
3192
3193 static void
3194 isert_free_np(struct iscsi_np *np)
3195 {
3196         struct isert_np *isert_np = (struct isert_np *)np->np_context;
3197         struct isert_conn *isert_conn, *n;
3198
3199         if (isert_np->np_cm_id)
3200                 rdma_destroy_id(isert_np->np_cm_id);
3201
3202         /*
3203          * FIXME: At this point we don't have a good way to insure
3204          * that at this point we don't have hanging connections that
3205          * completed RDMA establishment but didn't start iscsi login
3206          * process. So work-around this by cleaning up what ever piled
3207          * up in np_accept_list.
3208          */
3209         mutex_lock(&isert_np->np_accept_mutex);
3210         if (!list_empty(&isert_np->np_accept_list)) {
3211                 isert_info("Still have isert connections, cleaning up...\n");
3212                 list_for_each_entry_safe(isert_conn, n,
3213                                          &isert_np->np_accept_list,
3214                                          conn_accept_node) {
3215                         isert_info("cleaning isert_conn %p state (%d)\n",
3216                                    isert_conn, isert_conn->state);
3217                         isert_connect_release(isert_conn);
3218                 }
3219         }
3220         mutex_unlock(&isert_np->np_accept_mutex);
3221
3222         np->np_context = NULL;
3223         kfree(isert_np);
3224 }
3225
3226 static void isert_release_work(struct work_struct *work)
3227 {
3228         struct isert_conn *isert_conn = container_of(work,
3229                                                      struct isert_conn,
3230                                                      release_work);
3231
3232         isert_info("Starting release conn %p\n", isert_conn);
3233
3234         wait_for_completion(&isert_conn->conn_wait);
3235
3236         mutex_lock(&isert_conn->conn_mutex);
3237         isert_conn->state = ISER_CONN_DOWN;
3238         mutex_unlock(&isert_conn->conn_mutex);
3239
3240         isert_info("Destroying conn %p\n", isert_conn);
3241         isert_put_conn(isert_conn);
3242 }
3243
3244 static void
3245 isert_wait4logout(struct isert_conn *isert_conn)
3246 {
3247         struct iscsi_conn *conn = isert_conn->conn;
3248
3249         isert_info("conn %p\n", isert_conn);
3250
3251         if (isert_conn->logout_posted) {
3252                 isert_info("conn %p wait for conn_logout_comp\n", isert_conn);
3253                 wait_for_completion_timeout(&conn->conn_logout_comp,
3254                                             SECONDS_FOR_LOGOUT_COMP * HZ);
3255         }
3256 }
3257
3258 static void
3259 isert_wait4cmds(struct iscsi_conn *conn)
3260 {
3261         isert_info("iscsi_conn %p\n", conn);
3262
3263         if (conn->sess) {
3264                 target_sess_cmd_list_set_waiting(conn->sess->se_sess);
3265                 target_wait_for_sess_cmds(conn->sess->se_sess);
3266         }
3267 }
3268
3269 static void
3270 isert_wait4flush(struct isert_conn *isert_conn)
3271 {
3272         struct ib_recv_wr *bad_wr;
3273
3274         isert_info("conn %p\n", isert_conn);
3275
3276         init_completion(&isert_conn->conn_wait_comp_err);
3277         isert_conn->beacon.wr_id = ISER_BEACON_WRID;
3278         /* post an indication that all flush errors were consumed */
3279         if (ib_post_recv(isert_conn->conn_qp, &isert_conn->beacon, &bad_wr)) {
3280                 isert_err("conn %p failed to post beacon", isert_conn);
3281                 return;
3282         }
3283
3284         wait_for_completion(&isert_conn->conn_wait_comp_err);
3285 }
3286
3287 static void isert_wait_conn(struct iscsi_conn *conn)
3288 {
3289         struct isert_conn *isert_conn = conn->context;
3290
3291         isert_info("Starting conn %p\n", isert_conn);
3292
3293         mutex_lock(&isert_conn->conn_mutex);
3294         /*
3295          * Only wait for conn_wait_comp_err if the isert_conn made it
3296          * into full feature phase..
3297          */
3298         if (isert_conn->state == ISER_CONN_INIT) {
3299                 mutex_unlock(&isert_conn->conn_mutex);
3300                 return;
3301         }
3302         isert_conn_terminate(isert_conn);
3303         mutex_unlock(&isert_conn->conn_mutex);
3304
3305         isert_wait4cmds(conn);
3306         isert_wait4flush(isert_conn);
3307         isert_wait4logout(isert_conn);
3308
3309         INIT_WORK(&isert_conn->release_work, isert_release_work);
3310         queue_work(isert_release_wq, &isert_conn->release_work);
3311 }
3312
3313 static void isert_free_conn(struct iscsi_conn *conn)
3314 {
3315         struct isert_conn *isert_conn = conn->context;
3316
3317         isert_put_conn(isert_conn);
3318 }
3319
3320 static struct iscsit_transport iser_target_transport = {
3321         .name                   = "IB/iSER",
3322         .transport_type         = ISCSI_INFINIBAND,
3323         .priv_size              = sizeof(struct isert_cmd),
3324         .owner                  = THIS_MODULE,
3325         .iscsit_setup_np        = isert_setup_np,
3326         .iscsit_accept_np       = isert_accept_np,
3327         .iscsit_free_np         = isert_free_np,
3328         .iscsit_wait_conn       = isert_wait_conn,
3329         .iscsit_free_conn       = isert_free_conn,
3330         .iscsit_get_login_rx    = isert_get_login_rx,
3331         .iscsit_put_login_tx    = isert_put_login_tx,
3332         .iscsit_immediate_queue = isert_immediate_queue,
3333         .iscsit_response_queue  = isert_response_queue,
3334         .iscsit_get_dataout     = isert_get_dataout,
3335         .iscsit_queue_data_in   = isert_put_datain,
3336         .iscsit_queue_status    = isert_put_response,
3337         .iscsit_aborted_task    = isert_aborted_task,
3338         .iscsit_get_sup_prot_ops = isert_get_sup_prot_ops,
3339 };
3340
3341 static int __init isert_init(void)
3342 {
3343         int ret;
3344
3345         isert_comp_wq = alloc_workqueue("isert_comp_wq",
3346                                         WQ_UNBOUND | WQ_HIGHPRI, 0);
3347         if (!isert_comp_wq) {
3348                 isert_err("Unable to allocate isert_comp_wq\n");
3349                 ret = -ENOMEM;
3350                 return -ENOMEM;
3351         }
3352
3353         isert_release_wq = alloc_workqueue("isert_release_wq", WQ_UNBOUND,
3354                                         WQ_UNBOUND_MAX_ACTIVE);
3355         if (!isert_release_wq) {
3356                 isert_err("Unable to allocate isert_release_wq\n");
3357                 ret = -ENOMEM;
3358                 goto destroy_comp_wq;
3359         }
3360
3361         iscsit_register_transport(&iser_target_transport);
3362         isert_info("iSER_TARGET[0] - Loaded iser_target_transport\n");
3363
3364         return 0;
3365
3366 destroy_comp_wq:
3367         destroy_workqueue(isert_comp_wq);
3368
3369         return ret;
3370 }
3371
3372 static void __exit isert_exit(void)
3373 {
3374         flush_scheduled_work();
3375         destroy_workqueue(isert_release_wq);
3376         destroy_workqueue(isert_comp_wq);
3377         iscsit_unregister_transport(&iser_target_transport);
3378         isert_info("iSER_TARGET[0] - Released iser_target_transport\n");
3379 }
3380
3381 MODULE_DESCRIPTION("iSER-Target for mainline target infrastructure");
3382 MODULE_VERSION("0.1");
3383 MODULE_AUTHOR("nab@Linux-iSCSI.org");
3384 MODULE_LICENSE("GPL");
3385
3386 module_init(isert_init);
3387 module_exit(isert_exit);