drivers/infiniband/ulp/isert/ib_isert.c

   1 /*******************************************************************************
   2  * This file contains iSCSI extentions for RDMA (iSER) Verbs
   3  *
   4  * (c) Copyright 2013 Datera, Inc.
   5  *
   6  * Nicholas A. Bellinger <nab@linux-iscsi.org>
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License as published by
  10  * the Free Software Foundation; either version 2 of the License, or
  11  * (at your option) any later version.
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  ****************************************************************************/
  18
  19 #include <linux/string.h>
  20 #include <linux/module.h>
  21 #include <linux/scatterlist.h>
  22 #include <linux/socket.h>
  23 #include <linux/in.h>
  24 #include <linux/in6.h>
  25 #include <rdma/ib_verbs.h>
  26 #include <rdma/rdma_cm.h>
  27 #include <target/target_core_base.h>
  28 #include <target/target_core_fabric.h>
  29 #include <target/iscsi/iscsi_transport.h>
  30 #include <linux/semaphore.h>
  31
  32 #include "isert_proto.h"
  33 #include "ib_isert.h"
  34
  35 #define ISERT_MAX_CONN          8
  36 #define ISER_MAX_RX_CQ_LEN      (ISERT_QP_MAX_RECV_DTOS * ISERT_MAX_CONN)
  37 #define ISER_MAX_TX_CQ_LEN      (ISERT_QP_MAX_REQ_DTOS  * ISERT_MAX_CONN)
  38 #define ISER_MAX_CQ_LEN         (ISER_MAX_RX_CQ_LEN + ISER_MAX_TX_CQ_LEN + \
  39                                  ISERT_MAX_CONN)
  40
  41 static int isert_debug_level;
  42 module_param_named(debug_level, isert_debug_level, int, 0644);
  43 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0 (default:0)");
  44
  45 static DEFINE_MUTEX(device_list_mutex);
  46 static LIST_HEAD(device_list);
  47 static struct workqueue_struct *isert_comp_wq;
  48 static struct workqueue_struct *isert_release_wq;
  49
  50 static void
  51 isert_unmap_cmd(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn);
  52 static int
  53 isert_map_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
  54                struct isert_rdma_wr *wr);
  55 static void
  56 isert_unreg_rdma(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn);
  57 static int
  58 isert_reg_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
  59                struct isert_rdma_wr *wr);
  60 static int
  61 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd);
  62 static int
  63 isert_rdma_post_recvl(struct isert_conn *isert_conn);
  64 static int
  65 isert_rdma_accept(struct isert_conn *isert_conn);
  66 struct rdma_cm_id *isert_setup_id(struct isert_np *isert_np);
  67
  68 static void isert_release_work(struct work_struct *work);
  69
  70 static inline bool
  71 isert_prot_cmd(struct isert_conn *conn, struct se_cmd *cmd)
  72 {
  73         return (conn->pi_support &&
  74                 cmd->prot_op != TARGET_PROT_NORMAL);
  75 }
  76
  77
  78 static void
  79 isert_qp_event_callback(struct ib_event *e, void *context)
  80 {
  81         struct isert_conn *isert_conn = context;
  82
  83         isert_err("%s (%d): conn %p\n",
  84                   ib_event_msg(e->event), e->event, isert_conn);
  85
  86         switch (e->event) {
  87         case IB_EVENT_COMM_EST:
  88                 rdma_notify(isert_conn->cm_id, IB_EVENT_COMM_EST);
  89                 break;
  90         case IB_EVENT_QP_LAST_WQE_REACHED:
  91                 isert_warn("Reached TX IB_EVENT_QP_LAST_WQE_REACHED\n");
  92                 break;
  93         default:
  94                 break;
  95         }
  96 }
  97
  98 static int
  99 isert_query_device(struct ib_device *ib_dev, struct ib_device_attr *devattr)
 100 {
 101         int ret;
 102
 103         ret = ib_query_device(ib_dev, devattr);
 104         if (ret) {
 105                 isert_err("ib_query_device() failed: %d\n", ret);
 106                 return ret;
 107         }
 108         isert_dbg("devattr->max_sge: %d\n", devattr->max_sge);
 109         isert_dbg("devattr->max_sge_rd: %d\n", devattr->max_sge_rd);
 110
 111         return 0;
 112 }
 113
 114 static struct isert_comp *
 115 isert_comp_get(struct isert_conn *isert_conn)
 116 {
 117         struct isert_device *device = isert_conn->device;
 118         struct isert_comp *comp;
 119         int i, min = 0;
 120
 121         mutex_lock(&device_list_mutex);
 122         for (i = 0; i < device->comps_used; i++)
 123                 if (device->comps[i].active_qps <
 124                     device->comps[min].active_qps)
 125                         min = i;
 126         comp = &device->comps[min];
 127         comp->active_qps++;
 128         mutex_unlock(&device_list_mutex);
 129
 130         isert_info("conn %p, using comp %p min_index: %d\n",
 131                    isert_conn, comp, min);
 132
 133         return comp;
 134 }
 135
 136 static void
 137 isert_comp_put(struct isert_comp *comp)
 138 {
 139         mutex_lock(&device_list_mutex);
 140         comp->active_qps--;
 141         mutex_unlock(&device_list_mutex);
 142 }
 143
 144 static struct ib_qp *
 145 isert_create_qp(struct isert_conn *isert_conn,
 146                 struct isert_comp *comp,
 147                 struct rdma_cm_id *cma_id)
 148 {
 149         struct isert_device *device = isert_conn->device;
 150         struct ib_qp_init_attr attr;
 151         int ret;
 152
 153         memset(&attr, 0, sizeof(struct ib_qp_init_attr));
 154         attr.event_handler = isert_qp_event_callback;
 155         attr.qp_context = isert_conn;
 156         attr.send_cq = comp->cq;
 157         attr.recv_cq = comp->cq;
 158         attr.cap.max_send_wr = ISERT_QP_MAX_REQ_DTOS;
 159         attr.cap.max_recv_wr = ISERT_QP_MAX_RECV_DTOS + 1;
 160         /*
 161          * FIXME: Use devattr.max_sge - 2 for max_send_sge as
 162          * work-around for RDMA_READs with ConnectX-2.
 163          *
 164          * Also, still make sure to have at least two SGEs for
 165          * outgoing control PDU responses.
 166          */
 167         attr.cap.max_send_sge = max(2, device->dev_attr.max_sge - 2);
 168         isert_conn->max_sge = attr.cap.max_send_sge;
 169
 170         attr.cap.max_recv_sge = 1;
 171         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 172         attr.qp_type = IB_QPT_RC;
 173         if (device->pi_capable)
 174                 attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
 175
 176         ret = rdma_create_qp(cma_id, device->pd, &attr);
 177         if (ret) {
 178                 isert_err("rdma_create_qp failed for cma_id %d\n", ret);
 179                 return ERR_PTR(ret);
 180         }
 181
 182         return cma_id->qp;
 183 }
 184
 185 static int
 186 isert_conn_setup_qp(struct isert_conn *isert_conn, struct rdma_cm_id *cma_id)
 187 {
 188         struct isert_comp *comp;
 189         int ret;
 190
 191         comp = isert_comp_get(isert_conn);
 192         isert_conn->qp = isert_create_qp(isert_conn, comp, cma_id);
 193         if (IS_ERR(isert_conn->qp)) {
 194                 ret = PTR_ERR(isert_conn->qp);
 195                 goto err;
 196         }
 197
 198         return 0;
 199 err:
 200         isert_comp_put(comp);
 201         return ret;
 202 }
 203
 204 static void
 205 isert_cq_event_callback(struct ib_event *e, void *context)
 206 {
 207         isert_dbg("event: %d\n", e->event);
 208 }
 209
 210 static int
 211 isert_alloc_rx_descriptors(struct isert_conn *isert_conn)
 212 {
 213         struct isert_device *device = isert_conn->device;
 214         struct ib_device *ib_dev = device->ib_device;
 215         struct iser_rx_desc *rx_desc;
 216         struct ib_sge *rx_sg;
 217         u64 dma_addr;
 218         int i, j;
 219
 220         isert_conn->rx_descs = kzalloc(ISERT_QP_MAX_RECV_DTOS *
 221                                 sizeof(struct iser_rx_desc), GFP_KERNEL);
 222         if (!isert_conn->rx_descs)
 223                 goto fail;
 224
 225         rx_desc = isert_conn->rx_descs;
 226
 227         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 228                 dma_addr = ib_dma_map_single(ib_dev, (void *)rx_desc,
 229                                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 230                 if (ib_dma_mapping_error(ib_dev, dma_addr))
 231                         goto dma_map_fail;
 232
 233                 rx_desc->dma_addr = dma_addr;
 234
 235                 rx_sg = &rx_desc->rx_sg;
 236                 rx_sg->addr = rx_desc->dma_addr;
 237                 rx_sg->length = ISER_RX_PAYLOAD_SIZE;
 238                 rx_sg->lkey = device->mr->lkey;
 239         }
 240
 241         isert_conn->rx_desc_head = 0;
 242
 243         return 0;
 244
 245 dma_map_fail:
 246         rx_desc = isert_conn->rx_descs;
 247         for (j = 0; j < i; j++, rx_desc++) {
 248                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 249                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 250         }
 251         kfree(isert_conn->rx_descs);
 252         isert_conn->rx_descs = NULL;
 253 fail:
 254         isert_err("conn %p failed to allocate rx descriptors\n", isert_conn);
 255
 256         return -ENOMEM;
 257 }
 258
 259 static void
 260 isert_free_rx_descriptors(struct isert_conn *isert_conn)
 261 {
 262         struct ib_device *ib_dev = isert_conn->device->ib_device;
 263         struct iser_rx_desc *rx_desc;
 264         int i;
 265
 266         if (!isert_conn->rx_descs)
 267                 return;
 268
 269         rx_desc = isert_conn->rx_descs;
 270         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 271                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 272                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 273         }
 274
 275         kfree(isert_conn->rx_descs);
 276         isert_conn->rx_descs = NULL;
 277 }
 278
 279 static void isert_cq_work(struct work_struct *);
 280 static void isert_cq_callback(struct ib_cq *, void *);
 281
 282 static void
 283 isert_free_comps(struct isert_device *device)
 284 {
 285         int i;
 286
 287         for (i = 0; i < device->comps_used; i++) {
 288                 struct isert_comp *comp = &device->comps[i];
 289
 290                 if (comp->cq) {
 291                         cancel_work_sync(&comp->work);
 292                         ib_destroy_cq(comp->cq);
 293                 }
 294         }
 295         kfree(device->comps);
 296 }
 297
 298 static int
 299 isert_alloc_comps(struct isert_device *device,
 300                   struct ib_device_attr *attr)
 301 {
 302         int i, max_cqe, ret = 0;
 303
 304         device->comps_used = min(ISERT_MAX_CQ, min_t(int, num_online_cpus(),
 305                                  device->ib_device->num_comp_vectors));
 306
 307         isert_info("Using %d CQs, %s supports %d vectors support "
 308                    "Fast registration %d pi_capable %d\n",
 309                    device->comps_used, device->ib_device->name,
 310                    device->ib_device->num_comp_vectors, device->use_fastreg,
 311                    device->pi_capable);
 312
 313         device->comps = kcalloc(device->comps_used, sizeof(struct isert_comp),
 314                                 GFP_KERNEL);
 315         if (!device->comps) {
 316                 isert_err("Unable to allocate completion contexts\n");
 317                 return -ENOMEM;
 318         }
 319
 320         max_cqe = min(ISER_MAX_CQ_LEN, attr->max_cqe);
 321
 322         for (i = 0; i < device->comps_used; i++) {
 323                 struct ib_cq_init_attr cq_attr = {};
 324                 struct isert_comp *comp = &device->comps[i];
 325
 326                 comp->device = device;
 327                 INIT_WORK(&comp->work, isert_cq_work);
 328                 cq_attr.cqe = max_cqe;
 329                 cq_attr.comp_vector = i;
 330                 comp->cq = ib_create_cq(device->ib_device,
 331                                         isert_cq_callback,
 332                                         isert_cq_event_callback,
 333                                         (void *)comp,
 334                                         &cq_attr);
 335                 if (IS_ERR(comp->cq)) {
 336                         isert_err("Unable to allocate cq\n");
 337                         ret = PTR_ERR(comp->cq);
 338                         comp->cq = NULL;
 339                         goto out_cq;
 340                 }
 341
 342                 ret = ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP);
 343                 if (ret)
 344                         goto out_cq;
 345         }
 346
 347         return 0;
 348 out_cq:
 349         isert_free_comps(device);
 350         return ret;
 351 }
 352
 353 static int
 354 isert_create_device_ib_res(struct isert_device *device)
 355 {
 356         struct ib_device_attr *dev_attr;
 357         int ret;
 358
 359         dev_attr = &device->dev_attr;
 360         ret = isert_query_device(device->ib_device, dev_attr);
 361         if (ret)
 362                 return ret;
 363
 364         /* asign function handlers */
 365         if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS &&
 366             dev_attr->device_cap_flags & IB_DEVICE_SIGNATURE_HANDOVER) {
 367                 device->use_fastreg = 1;
 368                 device->reg_rdma_mem = isert_reg_rdma;
 369                 device->unreg_rdma_mem = isert_unreg_rdma;
 370         } else {
 371                 device->use_fastreg = 0;
 372                 device->reg_rdma_mem = isert_map_rdma;
 373                 device->unreg_rdma_mem = isert_unmap_cmd;
 374         }
 375
 376         ret = isert_alloc_comps(device, dev_attr);
 377         if (ret)
 378                 return ret;
 379
 380         device->pd = ib_alloc_pd(device->ib_device);
 381         if (IS_ERR(device->pd)) {
 382                 ret = PTR_ERR(device->pd);
 383                 isert_err("failed to allocate pd, device %p, ret=%d\n",
 384                           device, ret);
 385                 goto out_cq;
 386         }
 387
 388         device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE);
 389         if (IS_ERR(device->mr)) {
 390                 ret = PTR_ERR(device->mr);
 391                 isert_err("failed to create dma mr, device %p, ret=%d\n",
 392                           device, ret);
 393                 goto out_mr;
 394         }
 395
 396         /* Check signature cap */
 397         device->pi_capable = dev_attr->device_cap_flags &
 398                              IB_DEVICE_SIGNATURE_HANDOVER ? true : false;
 399
 400         return 0;
 401
 402 out_mr:
 403         ib_dealloc_pd(device->pd);
 404 out_cq:
 405         isert_free_comps(device);
 406         return ret;
 407 }
 408
 409 static void
 410 isert_free_device_ib_res(struct isert_device *device)
 411 {
 412         isert_info("device %p\n", device);
 413
 414         ib_dereg_mr(device->mr);
 415         ib_dealloc_pd(device->pd);
 416         isert_free_comps(device);
 417 }
 418
 419 static void
 420 isert_device_put(struct isert_device *device)
 421 {
 422         mutex_lock(&device_list_mutex);
 423         device->refcount--;
 424         isert_info("device %p refcount %d\n", device, device->refcount);
 425         if (!device->refcount) {
 426                 isert_free_device_ib_res(device);
 427                 list_del(&device->dev_node);
 428                 kfree(device);
 429         }
 430         mutex_unlock(&device_list_mutex);
 431 }
 432
 433 static struct isert_device *
 434 isert_device_get(struct rdma_cm_id *cma_id)
 435 {
 436         struct isert_device *device;
 437         int ret;
 438
 439         mutex_lock(&device_list_mutex);
 440         list_for_each_entry(device, &device_list, dev_node) {
 441                 if (device->ib_device->node_guid == cma_id->device->node_guid) {
 442                         device->refcount++;
 443                         isert_info("Found iser device %p refcount %d\n",
 444                                    device, device->refcount);
 445                         mutex_unlock(&device_list_mutex);
 446                         return device;
 447                 }
 448         }
 449
 450         device = kzalloc(sizeof(struct isert_device), GFP_KERNEL);
 451         if (!device) {
 452                 mutex_unlock(&device_list_mutex);
 453                 return ERR_PTR(-ENOMEM);
 454         }
 455
 456         INIT_LIST_HEAD(&device->dev_node);
 457
 458         device->ib_device = cma_id->device;
 459         ret = isert_create_device_ib_res(device);
 460         if (ret) {
 461                 kfree(device);
 462                 mutex_unlock(&device_list_mutex);
 463                 return ERR_PTR(ret);
 464         }
 465
 466         device->refcount++;
 467         list_add_tail(&device->dev_node, &device_list);
 468         isert_info("Created a new iser device %p refcount %d\n",
 469                    device, device->refcount);
 470         mutex_unlock(&device_list_mutex);
 471
 472         return device;
 473 }
 474
 475 static void
 476 isert_conn_free_fastreg_pool(struct isert_conn *isert_conn)
 477 {
 478         struct fast_reg_descriptor *fr_desc, *tmp;
 479         int i = 0;
 480
 481         if (list_empty(&isert_conn->fr_pool))
 482                 return;
 483
 484         isert_info("Freeing conn %p fastreg pool", isert_conn);
 485
 486         list_for_each_entry_safe(fr_desc, tmp,
 487                                  &isert_conn->fr_pool, list) {
 488                 list_del(&fr_desc->list);
 489                 ib_free_fast_reg_page_list(fr_desc->data_frpl);
 490                 ib_dereg_mr(fr_desc->data_mr);
 491                 if (fr_desc->pi_ctx) {
 492                         ib_free_fast_reg_page_list(fr_desc->pi_ctx->prot_frpl);
 493                         ib_dereg_mr(fr_desc->pi_ctx->prot_mr);
 494                         ib_destroy_mr(fr_desc->pi_ctx->sig_mr);
 495                         kfree(fr_desc->pi_ctx);
 496                 }
 497                 kfree(fr_desc);
 498                 ++i;
 499         }
 500
 501         if (i < isert_conn->fr_pool_size)
 502                 isert_warn("Pool still has %d regions registered\n",
 503                         isert_conn->fr_pool_size - i);
 504 }
 505
 506 static int
 507 isert_create_pi_ctx(struct fast_reg_descriptor *desc,
 508                     struct ib_device *device,
 509                     struct ib_pd *pd)
 510 {
 511         struct ib_mr_init_attr mr_init_attr;
 512         struct pi_context *pi_ctx;
 513         int ret;
 514
 515         pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
 516         if (!pi_ctx) {
 517                 isert_err("Failed to allocate pi context\n");
 518                 return -ENOMEM;
 519         }
 520
 521         pi_ctx->prot_frpl = ib_alloc_fast_reg_page_list(device,
 522                                             ISCSI_ISER_SG_TABLESIZE);
 523         if (IS_ERR(pi_ctx->prot_frpl)) {
 524                 isert_err("Failed to allocate prot frpl err=%ld\n",
 525                           PTR_ERR(pi_ctx->prot_frpl));
 526                 ret = PTR_ERR(pi_ctx->prot_frpl);
 527                 goto err_pi_ctx;
 528         }
 529
 530         pi_ctx->prot_mr = ib_alloc_fast_reg_mr(pd, ISCSI_ISER_SG_TABLESIZE);
 531         if (IS_ERR(pi_ctx->prot_mr)) {
 532                 isert_err("Failed to allocate prot frmr err=%ld\n",
 533                           PTR_ERR(pi_ctx->prot_mr));
 534                 ret = PTR_ERR(pi_ctx->prot_mr);
 535                 goto err_prot_frpl;
 536         }
 537         desc->ind |= ISERT_PROT_KEY_VALID;
 538
 539         memset(&mr_init_attr, 0, sizeof(mr_init_attr));
 540         mr_init_attr.max_reg_descriptors = 2;
 541         mr_init_attr.flags |= IB_MR_SIGNATURE_EN;
 542         pi_ctx->sig_mr = ib_create_mr(pd, &mr_init_attr);
 543         if (IS_ERR(pi_ctx->sig_mr)) {
 544                 isert_err("Failed to allocate signature enabled mr err=%ld\n",
 545                           PTR_ERR(pi_ctx->sig_mr));
 546                 ret = PTR_ERR(pi_ctx->sig_mr);
 547                 goto err_prot_mr;
 548         }
 549
 550         desc->pi_ctx = pi_ctx;
 551         desc->ind |= ISERT_SIG_KEY_VALID;
 552         desc->ind &= ~ISERT_PROTECTED;
 553
 554         return 0;
 555
 556 err_prot_mr:
 557         ib_dereg_mr(pi_ctx->prot_mr);
 558 err_prot_frpl:
 559         ib_free_fast_reg_page_list(pi_ctx->prot_frpl);
 560 err_pi_ctx:
 561         kfree(pi_ctx);
 562
 563         return ret;
 564 }
 565
 566 static int
 567 isert_create_fr_desc(struct ib_device *ib_device, struct ib_pd *pd,
 568                      struct fast_reg_descriptor *fr_desc)
 569 {
 570         int ret;
 571
 572         fr_desc->data_frpl = ib_alloc_fast_reg_page_list(ib_device,
 573                                                          ISCSI_ISER_SG_TABLESIZE);
 574         if (IS_ERR(fr_desc->data_frpl)) {
 575                 isert_err("Failed to allocate data frpl err=%ld\n",
 576                           PTR_ERR(fr_desc->data_frpl));
 577                 return PTR_ERR(fr_desc->data_frpl);
 578         }
 579
 580         fr_desc->data_mr = ib_alloc_fast_reg_mr(pd, ISCSI_ISER_SG_TABLESIZE);
 581         if (IS_ERR(fr_desc->data_mr)) {
 582                 isert_err("Failed to allocate data frmr err=%ld\n",
 583                           PTR_ERR(fr_desc->data_mr));
 584                 ret = PTR_ERR(fr_desc->data_mr);
 585                 goto err_data_frpl;
 586         }
 587         fr_desc->ind |= ISERT_DATA_KEY_VALID;
 588
 589         isert_dbg("Created fr_desc %p\n", fr_desc);
 590
 591         return 0;
 592
 593 err_data_frpl:
 594         ib_free_fast_reg_page_list(fr_desc->data_frpl);
 595
 596         return ret;
 597 }
 598
 599 static int
 600 isert_conn_create_fastreg_pool(struct isert_conn *isert_conn)
 601 {
 602         struct fast_reg_descriptor *fr_desc;
 603         struct isert_device *device = isert_conn->device;
 604         struct se_session *se_sess = isert_conn->conn->sess->se_sess;
 605         struct se_node_acl *se_nacl = se_sess->se_node_acl;
 606         int i, ret, tag_num;
 607         /*
 608          * Setup the number of FRMRs based upon the number of tags
 609          * available to session in iscsi_target_locate_portal().
 610          */
 611         tag_num = max_t(u32, ISCSIT_MIN_TAGS, se_nacl->queue_depth);
 612         tag_num = (tag_num * 2) + ISCSIT_EXTRA_TAGS;
 613
 614         isert_conn->fr_pool_size = 0;
 615         for (i = 0; i < tag_num; i++) {
 616                 fr_desc = kzalloc(sizeof(*fr_desc), GFP_KERNEL);
 617                 if (!fr_desc) {
 618                         isert_err("Failed to allocate fast_reg descriptor\n");
 619                         ret = -ENOMEM;
 620                         goto err;
 621                 }
 622
 623                 ret = isert_create_fr_desc(device->ib_device,
 624                                            device->pd, fr_desc);
 625                 if (ret) {
 626                         isert_err("Failed to create fastreg descriptor err=%d\n",
 627                                ret);
 628                         kfree(fr_desc);
 629                         goto err;
 630                 }
 631
 632                 list_add_tail(&fr_desc->list, &isert_conn->fr_pool);
 633                 isert_conn->fr_pool_size++;
 634         }
 635
 636         isert_dbg("Creating conn %p fastreg pool size=%d",
 637                  isert_conn, isert_conn->fr_pool_size);
 638
 639         return 0;
 640
 641 err:
 642         isert_conn_free_fastreg_pool(isert_conn);
 643         return ret;
 644 }
 645
 646 static void
 647 isert_init_conn(struct isert_conn *isert_conn)
 648 {
 649         isert_conn->state = ISER_CONN_INIT;
 650         INIT_LIST_HEAD(&isert_conn->accept_node);
 651         init_completion(&isert_conn->login_comp);
 652         init_completion(&isert_conn->login_req_comp);
 653         init_completion(&isert_conn->wait);
 654         kref_init(&isert_conn->kref);
 655         mutex_init(&isert_conn->mutex);
 656         spin_lock_init(&isert_conn->pool_lock);
 657         INIT_LIST_HEAD(&isert_conn->fr_pool);
 658         INIT_WORK(&isert_conn->release_work, isert_release_work);
 659 }
 660
 661 static void
 662 isert_free_login_buf(struct isert_conn *isert_conn)
 663 {
 664         struct ib_device *ib_dev = isert_conn->device->ib_device;
 665
 666         ib_dma_unmap_single(ib_dev, isert_conn->login_rsp_dma,
 667                             ISER_RX_LOGIN_SIZE, DMA_TO_DEVICE);
 668         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 669                             ISCSI_DEF_MAX_RECV_SEG_LEN,
 670                             DMA_FROM_DEVICE);
 671         kfree(isert_conn->login_buf);
 672 }
 673
 674 static int
 675 isert_alloc_login_buf(struct isert_conn *isert_conn,
 676                       struct ib_device *ib_dev)
 677 {
 678         int ret;
 679
 680         isert_conn->login_buf = kzalloc(ISCSI_DEF_MAX_RECV_SEG_LEN +
 681                                         ISER_RX_LOGIN_SIZE, GFP_KERNEL);
 682         if (!isert_conn->login_buf) {
 683                 isert_err("Unable to allocate isert_conn->login_buf\n");
 684                 return -ENOMEM;
 685         }
 686
 687         isert_conn->login_req_buf = isert_conn->login_buf;
 688         isert_conn->login_rsp_buf = isert_conn->login_buf +
 689                                     ISCSI_DEF_MAX_RECV_SEG_LEN;
 690
 691         isert_dbg("Set login_buf: %p login_req_buf: %p login_rsp_buf: %p\n",
 692                  isert_conn->login_buf, isert_conn->login_req_buf,
 693                  isert_conn->login_rsp_buf);
 694
 695         isert_conn->login_req_dma = ib_dma_map_single(ib_dev,
 696                                 (void *)isert_conn->login_req_buf,
 697                                 ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_FROM_DEVICE);
 698
 699         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_req_dma);
 700         if (ret) {
 701                 isert_err("login_req_dma mapping error: %d\n", ret);
 702                 isert_conn->login_req_dma = 0;
 703                 goto out_login_buf;
 704         }
 705
 706         isert_conn->login_rsp_dma = ib_dma_map_single(ib_dev,
 707                                         (void *)isert_conn->login_rsp_buf,
 708                                         ISER_RX_LOGIN_SIZE, DMA_TO_DEVICE);
 709
 710         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_rsp_dma);
 711         if (ret) {
 712                 isert_err("login_rsp_dma mapping error: %d\n", ret);
 713                 isert_conn->login_rsp_dma = 0;
 714                 goto out_req_dma_map;
 715         }
 716
 717         return 0;
 718
 719 out_req_dma_map:
 720         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 721                             ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_FROM_DEVICE);
 722 out_login_buf:
 723         kfree(isert_conn->login_buf);
 724         return ret;
 725 }
 726
 727 static int
 728 isert_connect_request(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 729 {
 730         struct isert_np *isert_np = cma_id->context;
 731         struct iscsi_np *np = isert_np->np;
 732         struct isert_conn *isert_conn;
 733         struct isert_device *device;
 734         int ret = 0;
 735
 736         spin_lock_bh(&np->np_thread_lock);
 737         if (!np->enabled) {
 738                 spin_unlock_bh(&np->np_thread_lock);
 739                 isert_dbg("iscsi_np is not enabled, reject connect request\n");
 740                 return rdma_reject(cma_id, NULL, 0);
 741         }
 742         spin_unlock_bh(&np->np_thread_lock);
 743
 744         isert_dbg("cma_id: %p, portal: %p\n",
 745                  cma_id, cma_id->context);
 746
 747         isert_conn = kzalloc(sizeof(struct isert_conn), GFP_KERNEL);
 748         if (!isert_conn)
 749                 return -ENOMEM;
 750
 751         isert_init_conn(isert_conn);
 752         isert_conn->cm_id = cma_id;
 753
 754         ret = isert_alloc_login_buf(isert_conn, cma_id->device);
 755         if (ret)
 756                 goto out;
 757
 758         device = isert_device_get(cma_id);
 759         if (IS_ERR(device)) {
 760                 ret = PTR_ERR(device);
 761                 goto out_rsp_dma_map;
 762         }
 763         isert_conn->device = device;
 764
 765         /* Set max inflight RDMA READ requests */
 766         isert_conn->initiator_depth = min_t(u8,
 767                                 event->param.conn.initiator_depth,
 768                                 device->dev_attr.max_qp_init_rd_atom);
 769         isert_dbg("Using initiator_depth: %u\n", isert_conn->initiator_depth);
 770
 771         ret = isert_conn_setup_qp(isert_conn, cma_id);
 772         if (ret)
 773                 goto out_conn_dev;
 774
 775         ret = isert_rdma_post_recvl(isert_conn);
 776         if (ret)
 777                 goto out_conn_dev;
 778
 779         ret = isert_rdma_accept(isert_conn);
 780         if (ret)
 781                 goto out_conn_dev;
 782
 783         mutex_lock(&isert_np->np_accept_mutex);
 784         list_add_tail(&isert_conn->accept_node, &isert_np->np_accept_list);
 785         mutex_unlock(&isert_np->np_accept_mutex);
 786
 787         isert_info("np %p: Allow accept_np to continue\n", np);
 788         up(&isert_np->np_sem);
 789         return 0;
 790
 791 out_conn_dev:
 792         isert_device_put(device);
 793 out_rsp_dma_map:
 794         isert_free_login_buf(isert_conn);
 795 out:
 796         kfree(isert_conn);
 797         rdma_reject(cma_id, NULL, 0);
 798         return ret;
 799 }
 800
 801 static void
 802 isert_connect_release(struct isert_conn *isert_conn)
 803 {
 804         struct isert_device *device = isert_conn->device;
 805
 806         isert_dbg("conn %p\n", isert_conn);
 807
 808         BUG_ON(!device);
 809
 810         if (device->use_fastreg)
 811                 isert_conn_free_fastreg_pool(isert_conn);
 812
 813         isert_free_rx_descriptors(isert_conn);
 814         if (isert_conn->cm_id)
 815                 rdma_destroy_id(isert_conn->cm_id);
 816
 817         if (isert_conn->qp) {
 818                 struct isert_comp *comp = isert_conn->qp->recv_cq->cq_context;
 819
 820                 isert_comp_put(comp);
 821                 ib_destroy_qp(isert_conn->qp);
 822         }
 823
 824         if (isert_conn->login_buf)
 825                 isert_free_login_buf(isert_conn);
 826
 827         isert_device_put(device);
 828
 829         kfree(isert_conn);
 830 }
 831
 832 static void
 833 isert_connected_handler(struct rdma_cm_id *cma_id)
 834 {
 835         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 836
 837         isert_info("conn %p\n", isert_conn);
 838
 839         if (!kref_get_unless_zero(&isert_conn->kref)) {
 840                 isert_warn("conn %p connect_release is running\n", isert_conn);
 841                 return;
 842         }
 843
 844         mutex_lock(&isert_conn->mutex);
 845         if (isert_conn->state != ISER_CONN_FULL_FEATURE)
 846                 isert_conn->state = ISER_CONN_UP;
 847         mutex_unlock(&isert_conn->mutex);
 848 }
 849
 850 static void
 851 isert_release_kref(struct kref *kref)
 852 {
 853         struct isert_conn *isert_conn = container_of(kref,
 854                                 struct isert_conn, kref);
 855
 856         isert_info("conn %p final kref %s/%d\n", isert_conn, current->comm,
 857                    current->pid);
 858
 859         isert_connect_release(isert_conn);
 860 }
 861
 862 static void
 863 isert_put_conn(struct isert_conn *isert_conn)
 864 {
 865         kref_put(&isert_conn->kref, isert_release_kref);
 866 }
 867
 868 /**
 869  * isert_conn_terminate() - Initiate connection termination
 870  * @isert_conn: isert connection struct
 871  *
 872  * Notes:
 873  * In case the connection state is FULL_FEATURE, move state
 874  * to TEMINATING and start teardown sequence (rdma_disconnect).
 875  * In case the connection state is UP, complete flush as well.
 876  *
 877  * This routine must be called with mutex held. Thus it is
 878  * safe to call multiple times.
 879  */
 880 static void
 881 isert_conn_terminate(struct isert_conn *isert_conn)
 882 {
 883         int err;
 884
 885         switch (isert_conn->state) {
 886         case ISER_CONN_TERMINATING:
 887                 break;
 888         case ISER_CONN_UP:
 889         case ISER_CONN_FULL_FEATURE: /* FALLTHRU */
 890                 isert_info("Terminating conn %p state %d\n",
 891                            isert_conn, isert_conn->state);
 892                 isert_conn->state = ISER_CONN_TERMINATING;
 893                 err = rdma_disconnect(isert_conn->cm_id);
 894                 if (err)
 895                         isert_warn("Failed rdma_disconnect isert_conn %p\n",
 896                                    isert_conn);
 897                 break;
 898         default:
 899                 isert_warn("conn %p teminating in state %d\n",
 900                            isert_conn, isert_conn->state);
 901         }
 902 }
 903
 904 static int
 905 isert_np_cma_handler(struct isert_np *isert_np,
 906                      enum rdma_cm_event_type event)
 907 {
 908         isert_dbg("%s (%d): isert np %p\n",
 909                   rdma_event_msg(event), event, isert_np);
 910
 911         switch (event) {
 912         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 913                 isert_np->np_cm_id = NULL;
 914                 break;
 915         case RDMA_CM_EVENT_ADDR_CHANGE:
 916                 isert_np->np_cm_id = isert_setup_id(isert_np);
 917                 if (IS_ERR(isert_np->np_cm_id)) {
 918                         isert_err("isert np %p setup id failed: %ld\n",
 919                                   isert_np, PTR_ERR(isert_np->np_cm_id));
 920                         isert_np->np_cm_id = NULL;
 921                 }
 922                 break;
 923         default:
 924                 isert_err("isert np %p Unexpected event %d\n",
 925                           isert_np, event);
 926         }
 927
 928         return -1;
 929 }
 930
 931 static int
 932 isert_disconnected_handler(struct rdma_cm_id *cma_id,
 933                            enum rdma_cm_event_type event)
 934 {
 935         struct isert_np *isert_np = cma_id->context;
 936         struct isert_conn *isert_conn;
 937         bool terminating = false;
 938
 939         if (isert_np->np_cm_id == cma_id)
 940                 return isert_np_cma_handler(cma_id->context, event);
 941
 942         isert_conn = cma_id->qp->qp_context;
 943
 944         mutex_lock(&isert_conn->mutex);
 945         terminating = (isert_conn->state == ISER_CONN_TERMINATING);
 946         isert_conn_terminate(isert_conn);
 947         mutex_unlock(&isert_conn->mutex);
 948
 949         isert_info("conn %p completing wait\n", isert_conn);
 950         complete(&isert_conn->wait);
 951
 952         if (terminating)
 953                 goto out;
 954
 955         mutex_lock(&isert_np->np_accept_mutex);
 956         if (!list_empty(&isert_conn->accept_node)) {
 957                 list_del_init(&isert_conn->accept_node);
 958                 isert_put_conn(isert_conn);
 959                 queue_work(isert_release_wq, &isert_conn->release_work);
 960         }
 961         mutex_unlock(&isert_np->np_accept_mutex);
 962
 963 out:
 964         return 0;
 965 }
 966
 967 static int
 968 isert_connect_error(struct rdma_cm_id *cma_id)
 969 {
 970         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 971
 972         isert_conn->cm_id = NULL;
 973         isert_put_conn(isert_conn);
 974
 975         return -1;
 976 }
 977
 978 static int
 979 isert_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 980 {
 981         int ret = 0;
 982
 983         isert_info("%s (%d): status %d id %p np %p\n",
 984                    rdma_event_msg(event->event), event->event,
 985                    event->status, cma_id, cma_id->context);
 986
 987         switch (event->event) {
 988         case RDMA_CM_EVENT_CONNECT_REQUEST:
 989                 ret = isert_connect_request(cma_id, event);
 990                 if (ret)
 991                         isert_err("failed handle connect request %d\n", ret);
 992                 break;
 993         case RDMA_CM_EVENT_ESTABLISHED:
 994                 isert_connected_handler(cma_id);
 995                 break;
 996         case RDMA_CM_EVENT_ADDR_CHANGE:    /* FALLTHRU */
 997         case RDMA_CM_EVENT_DISCONNECTED:   /* FALLTHRU */
 998         case RDMA_CM_EVENT_DEVICE_REMOVAL: /* FALLTHRU */
 999         case RDMA_CM_EVENT_TIMEWAIT_EXIT:  /* FALLTHRU */
1000                 ret = isert_disconnected_handler(cma_id, event->event);
1001                 break;
1002         case RDMA_CM_EVENT_REJECTED:       /* FALLTHRU */
1003         case RDMA_CM_EVENT_UNREACHABLE:    /* FALLTHRU */
1004         case RDMA_CM_EVENT_CONNECT_ERROR:
1005                 ret = isert_connect_error(cma_id);
1006                 break;
1007         default:
1008                 isert_err("Unhandled RDMA CMA event: %d\n", event->event);
1009                 break;
1010         }
1011
1012         return ret;
1013 }
1014
1015 static int
1016 isert_post_recv(struct isert_conn *isert_conn, u32 count)
1017 {
1018         struct ib_recv_wr *rx_wr, *rx_wr_failed;
1019         int i, ret;
1020         unsigned int rx_head = isert_conn->rx_desc_head;
1021         struct iser_rx_desc *rx_desc;
1022
1023         for (rx_wr = isert_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
1024                 rx_desc         = &isert_conn->rx_descs[rx_head];
1025                 rx_wr->wr_id    = (uintptr_t)rx_desc;
1026                 rx_wr->sg_list  = &rx_desc->rx_sg;
1027                 rx_wr->num_sge  = 1;
1028                 rx_wr->next     = rx_wr + 1;
1029                 rx_head = (rx_head + 1) & (ISERT_QP_MAX_RECV_DTOS - 1);
1030         }
1031
1032         rx_wr--;
1033         rx_wr->next = NULL; /* mark end of work requests list */
1034
1035         isert_conn->post_recv_buf_count += count;
1036         ret = ib_post_recv(isert_conn->qp, isert_conn->rx_wr,
1037                                 &rx_wr_failed);
1038         if (ret) {
1039                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
1040                 isert_conn->post_recv_buf_count -= count;
1041         } else {
1042                 isert_dbg("Posted %d RX buffers\n", count);
1043                 isert_conn->rx_desc_head = rx_head;
1044         }
1045         return ret;
1046 }
1047
1048 static int
1049 isert_post_send(struct isert_conn *isert_conn, struct iser_tx_desc *tx_desc)
1050 {
1051         struct ib_device *ib_dev = isert_conn->cm_id->device;
1052         struct ib_send_wr send_wr, *send_wr_failed;
1053         int ret;
1054
1055         ib_dma_sync_single_for_device(ib_dev, tx_desc->dma_addr,
1056                                       ISER_HEADERS_LEN, DMA_TO_DEVICE);
1057
1058         send_wr.next    = NULL;
1059         send_wr.wr_id   = (uintptr_t)tx_desc;
1060         send_wr.sg_list = tx_desc->tx_sg;
1061         send_wr.num_sge = tx_desc->num_sge;
1062         send_wr.opcode  = IB_WR_SEND;
1063         send_wr.send_flags = IB_SEND_SIGNALED;
1064
1065         ret = ib_post_send(isert_conn->qp, &send_wr, &send_wr_failed);
1066         if (ret)
1067                 isert_err("ib_post_send() failed, ret: %d\n", ret);
1068
1069         return ret;
1070 }
1071
1072 static void
1073 isert_create_send_desc(struct isert_conn *isert_conn,
1074                        struct isert_cmd *isert_cmd,
1075                        struct iser_tx_desc *tx_desc)
1076 {
1077         struct isert_device *device = isert_conn->device;
1078         struct ib_device *ib_dev = device->ib_device;
1079
1080         ib_dma_sync_single_for_cpu(ib_dev, tx_desc->dma_addr,
1081                                    ISER_HEADERS_LEN, DMA_TO_DEVICE);
1082
1083         memset(&tx_desc->iser_header, 0, sizeof(struct iser_hdr));
1084         tx_desc->iser_header.flags = ISER_VER;
1085
1086         tx_desc->num_sge = 1;
1087         tx_desc->isert_cmd = isert_cmd;
1088
1089         if (tx_desc->tx_sg[0].lkey != device->mr->lkey) {
1090                 tx_desc->tx_sg[0].lkey = device->mr->lkey;
1091                 isert_dbg("tx_desc %p lkey mismatch, fixing\n", tx_desc);
1092         }
1093 }
1094
1095 static int
1096 isert_init_tx_hdrs(struct isert_conn *isert_conn,
1097                    struct iser_tx_desc *tx_desc)
1098 {
1099         struct isert_device *device = isert_conn->device;
1100         struct ib_device *ib_dev = device->ib_device;
1101         u64 dma_addr;
1102
1103         dma_addr = ib_dma_map_single(ib_dev, (void *)tx_desc,
1104                         ISER_HEADERS_LEN, DMA_TO_DEVICE);
1105         if (ib_dma_mapping_error(ib_dev, dma_addr)) {
1106                 isert_err("ib_dma_mapping_error() failed\n");
1107                 return -ENOMEM;
1108         }
1109
1110         tx_desc->dma_addr = dma_addr;
1111         tx_desc->tx_sg[0].addr  = tx_desc->dma_addr;
1112         tx_desc->tx_sg[0].length = ISER_HEADERS_LEN;
1113         tx_desc->tx_sg[0].lkey = device->mr->lkey;
1114
1115         isert_dbg("Setup tx_sg[0].addr: 0x%llx length: %u lkey: 0x%x\n",
1116                   tx_desc->tx_sg[0].addr, tx_desc->tx_sg[0].length,
1117                   tx_desc->tx_sg[0].lkey);
1118
1119         return 0;
1120 }
1121
1122 static void
1123 isert_init_send_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1124                    struct ib_send_wr *send_wr)
1125 {
1126         struct iser_tx_desc *tx_desc = &isert_cmd->tx_desc;
1127
1128         isert_cmd->rdma_wr.iser_ib_op = ISER_IB_SEND;
1129         send_wr->wr_id = (uintptr_t)&isert_cmd->tx_desc;
1130         send_wr->opcode = IB_WR_SEND;
1131         send_wr->sg_list = &tx_desc->tx_sg[0];
1132         send_wr->num_sge = isert_cmd->tx_desc.num_sge;
1133         send_wr->send_flags = IB_SEND_SIGNALED;
1134 }
1135
1136 static int
1137 isert_rdma_post_recvl(struct isert_conn *isert_conn)
1138 {
1139         struct ib_recv_wr rx_wr, *rx_wr_fail;
1140         struct ib_sge sge;
1141         int ret;
1142
1143         memset(&sge, 0, sizeof(struct ib_sge));
1144         sge.addr = isert_conn->login_req_dma;
1145         sge.length = ISER_RX_LOGIN_SIZE;
1146         sge.lkey = isert_conn->device->mr->lkey;
1147
1148         isert_dbg("Setup sge: addr: %llx length: %d 0x%08x\n",
1149                 sge.addr, sge.length, sge.lkey);
1150
1151         memset(&rx_wr, 0, sizeof(struct ib_recv_wr));
1152         rx_wr.wr_id = (uintptr_t)isert_conn->login_req_buf;
1153         rx_wr.sg_list = &sge;
1154         rx_wr.num_sge = 1;
1155
1156         isert_conn->post_recv_buf_count++;
1157         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_fail);
1158         if (ret) {
1159                 isert_err("ib_post_recv() failed: %d\n", ret);
1160                 isert_conn->post_recv_buf_count--;
1161         }
1162
1163         return ret;
1164 }
1165
1166 static int
1167 isert_put_login_tx(struct iscsi_conn *conn, struct iscsi_login *login,
1168                    u32 length)
1169 {
1170         struct isert_conn *isert_conn = conn->context;
1171         struct isert_device *device = isert_conn->device;
1172         struct ib_device *ib_dev = device->ib_device;
1173         struct iser_tx_desc *tx_desc = &isert_conn->login_tx_desc;
1174         int ret;
1175
1176         isert_create_send_desc(isert_conn, NULL, tx_desc);
1177
1178         memcpy(&tx_desc->iscsi_header, &login->rsp[0],
1179                sizeof(struct iscsi_hdr));
1180
1181         isert_init_tx_hdrs(isert_conn, tx_desc);
1182
1183         if (length > 0) {
1184                 struct ib_sge *tx_dsg = &tx_desc->tx_sg[1];
1185
1186                 ib_dma_sync_single_for_cpu(ib_dev, isert_conn->login_rsp_dma,
1187                                            length, DMA_TO_DEVICE);
1188
1189                 memcpy(isert_conn->login_rsp_buf, login->rsp_buf, length);
1190
1191                 ib_dma_sync_single_for_device(ib_dev, isert_conn->login_rsp_dma,
1192                                               length, DMA_TO_DEVICE);
1193
1194                 tx_dsg->addr    = isert_conn->login_rsp_dma;
1195                 tx_dsg->length  = length;
1196                 tx_dsg->lkey    = isert_conn->device->mr->lkey;
1197                 tx_desc->num_sge = 2;
1198         }
1199         if (!login->login_failed) {
1200                 if (login->login_complete) {
1201                         if (!conn->sess->sess_ops->SessionType &&
1202                             isert_conn->device->use_fastreg) {
1203                                 ret = isert_conn_create_fastreg_pool(isert_conn);
1204                                 if (ret) {
1205                                         isert_err("Conn: %p failed to create"
1206                                                " fastreg pool\n", isert_conn);
1207                                         return ret;
1208                                 }
1209                         }
1210
1211                         ret = isert_alloc_rx_descriptors(isert_conn);
1212                         if (ret)
1213                                 return ret;
1214
1215                         ret = isert_post_recv(isert_conn, ISERT_MIN_POSTED_RX);
1216                         if (ret)
1217                                 return ret;
1218
1219                         /* Now we are in FULL_FEATURE phase */
1220                         mutex_lock(&isert_conn->mutex);
1221                         isert_conn->state = ISER_CONN_FULL_FEATURE;
1222                         mutex_unlock(&isert_conn->mutex);
1223                         goto post_send;
1224                 }
1225
1226                 ret = isert_rdma_post_recvl(isert_conn);
1227                 if (ret)
1228                         return ret;
1229         }
1230 post_send:
1231         ret = isert_post_send(isert_conn, tx_desc);
1232         if (ret)
1233                 return ret;
1234
1235         return 0;
1236 }
1237
1238 static void
1239 isert_rx_login_req(struct isert_conn *isert_conn)
1240 {
1241         struct iser_rx_desc *rx_desc = (void *)isert_conn->login_req_buf;
1242         int rx_buflen = isert_conn->login_req_len;
1243         struct iscsi_conn *conn = isert_conn->conn;
1244         struct iscsi_login *login = conn->conn_login;
1245         int size;
1246
1247         isert_info("conn %p\n", isert_conn);
1248
1249         WARN_ON_ONCE(!login);
1250
1251         if (login->first_request) {
1252                 struct iscsi_login_req *login_req =
1253                         (struct iscsi_login_req *)&rx_desc->iscsi_header;
1254                 /*
1255                  * Setup the initial iscsi_login values from the leading
1256                  * login request PDU.
1257                  */
1258                 login->leading_connection = (!login_req->tsih) ? 1 : 0;
1259                 login->current_stage =
1260                         (login_req->flags & ISCSI_FLAG_LOGIN_CURRENT_STAGE_MASK)
1261                          >> 2;
1262                 login->version_min      = login_req->min_version;
1263                 login->version_max      = login_req->max_version;
1264                 memcpy(login->isid, login_req->isid, 6);
1265                 login->cmd_sn           = be32_to_cpu(login_req->cmdsn);
1266                 login->init_task_tag    = login_req->itt;
1267                 login->initial_exp_statsn = be32_to_cpu(login_req->exp_statsn);
1268                 login->cid              = be16_to_cpu(login_req->cid);
1269                 login->tsih             = be16_to_cpu(login_req->tsih);
1270         }
1271
1272         memcpy(&login->req[0], (void *)&rx_desc->iscsi_header, ISCSI_HDR_LEN);
1273
1274         size = min(rx_buflen, MAX_KEY_VALUE_PAIRS);
1275         isert_dbg("Using login payload size: %d, rx_buflen: %d "
1276                   "MAX_KEY_VALUE_PAIRS: %d\n", size, rx_buflen,
1277                   MAX_KEY_VALUE_PAIRS);
1278         memcpy(login->req_buf, &rx_desc->data[0], size);
1279
1280         if (login->first_request) {
1281                 complete(&isert_conn->login_comp);
1282                 return;
1283         }
1284         schedule_delayed_work(&conn->login_work, 0);
1285 }
1286
1287 static struct iscsi_cmd
1288 *isert_allocate_cmd(struct iscsi_conn *conn)
1289 {
1290         struct isert_conn *isert_conn = conn->context;
1291         struct isert_cmd *isert_cmd;
1292         struct iscsi_cmd *cmd;
1293
1294         cmd = iscsit_allocate_cmd(conn, TASK_INTERRUPTIBLE);
1295         if (!cmd) {
1296                 isert_err("Unable to allocate iscsi_cmd + isert_cmd\n");
1297                 return NULL;
1298         }
1299         isert_cmd = iscsit_priv_cmd(cmd);
1300         isert_cmd->conn = isert_conn;
1301         isert_cmd->iscsi_cmd = cmd;
1302
1303         return cmd;
1304 }
1305
1306 static int
1307 isert_handle_scsi_cmd(struct isert_conn *isert_conn,
1308                       struct isert_cmd *isert_cmd, struct iscsi_cmd *cmd,
1309                       struct iser_rx_desc *rx_desc, unsigned char *buf)
1310 {
1311         struct iscsi_conn *conn = isert_conn->conn;
1312         struct iscsi_scsi_req *hdr = (struct iscsi_scsi_req *)buf;
1313         struct scatterlist *sg;
1314         int imm_data, imm_data_len, unsol_data, sg_nents, rc;
1315         bool dump_payload = false;
1316
1317         rc = iscsit_setup_scsi_cmd(conn, cmd, buf);
1318         if (rc < 0)
1319                 return rc;
1320
1321         imm_data = cmd->immediate_data;
1322         imm_data_len = cmd->first_burst_len;
1323         unsol_data = cmd->unsolicited_data;
1324
1325         rc = iscsit_process_scsi_cmd(conn, cmd, hdr);
1326         if (rc < 0) {
1327                 return 0;
1328         } else if (rc > 0) {
1329                 dump_payload = true;
1330                 goto sequence_cmd;
1331         }
1332
1333         if (!imm_data)
1334                 return 0;
1335
1336         sg = &cmd->se_cmd.t_data_sg[0];
1337         sg_nents = max(1UL, DIV_ROUND_UP(imm_data_len, PAGE_SIZE));
1338
1339         isert_dbg("Copying Immediate SG: %p sg_nents: %u from %p imm_data_len: %d\n",
1340                   sg, sg_nents, &rx_desc->data[0], imm_data_len);
1341
1342         sg_copy_from_buffer(sg, sg_nents, &rx_desc->data[0], imm_data_len);
1343
1344         cmd->write_data_done += imm_data_len;
1345
1346         if (cmd->write_data_done == cmd->se_cmd.data_length) {
1347                 spin_lock_bh(&cmd->istate_lock);
1348                 cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1349                 cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1350                 spin_unlock_bh(&cmd->istate_lock);
1351         }
1352
1353 sequence_cmd:
1354         rc = iscsit_sequence_cmd(conn, cmd, buf, hdr->cmdsn);
1355
1356         if (!rc && dump_payload == false && unsol_data)
1357                 iscsit_set_unsoliticed_dataout(cmd);
1358         else if (dump_payload && imm_data)
1359                 target_put_sess_cmd(&cmd->se_cmd);
1360
1361         return 0;
1362 }
1363
1364 static int
1365 isert_handle_iscsi_dataout(struct isert_conn *isert_conn,
1366                            struct iser_rx_desc *rx_desc, unsigned char *buf)
1367 {
1368         struct scatterlist *sg_start;
1369         struct iscsi_conn *conn = isert_conn->conn;
1370         struct iscsi_cmd *cmd = NULL;
1371         struct iscsi_data *hdr = (struct iscsi_data *)buf;
1372         u32 unsol_data_len = ntoh24(hdr->dlength);
1373         int rc, sg_nents, sg_off, page_off;
1374
1375         rc = iscsit_check_dataout_hdr(conn, buf, &cmd);
1376         if (rc < 0)
1377                 return rc;
1378         else if (!cmd)
1379                 return 0;
1380         /*
1381          * FIXME: Unexpected unsolicited_data out
1382          */
1383         if (!cmd->unsolicited_data) {
1384                 isert_err("Received unexpected solicited data payload\n");
1385                 dump_stack();
1386                 return -1;
1387         }
1388
1389         isert_dbg("Unsolicited DataOut unsol_data_len: %u, "
1390                   "write_data_done: %u, data_length: %u\n",
1391                   unsol_data_len,  cmd->write_data_done,
1392                   cmd->se_cmd.data_length);
1393
1394         sg_off = cmd->write_data_done / PAGE_SIZE;
1395         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
1396         sg_nents = max(1UL, DIV_ROUND_UP(unsol_data_len, PAGE_SIZE));
1397         page_off = cmd->write_data_done % PAGE_SIZE;
1398         /*
1399          * FIXME: Non page-aligned unsolicited_data out
1400          */
1401         if (page_off) {
1402                 isert_err("unexpected non-page aligned data payload\n");
1403                 dump_stack();
1404                 return -1;
1405         }
1406         isert_dbg("Copying DataOut: sg_start: %p, sg_off: %u "
1407                   "sg_nents: %u from %p %u\n", sg_start, sg_off,
1408                   sg_nents, &rx_desc->data[0], unsol_data_len);
1409
1410         sg_copy_from_buffer(sg_start, sg_nents, &rx_desc->data[0],
1411                             unsol_data_len);
1412
1413         rc = iscsit_check_dataout_payload(cmd, hdr, false);
1414         if (rc < 0)
1415                 return rc;
1416
1417         return 0;
1418 }
1419
1420 static int
1421 isert_handle_nop_out(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1422                      struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1423                      unsigned char *buf)
1424 {
1425         struct iscsi_conn *conn = isert_conn->conn;
1426         struct iscsi_nopout *hdr = (struct iscsi_nopout *)buf;
1427         int rc;
1428
1429         rc = iscsit_setup_nop_out(conn, cmd, hdr);
1430         if (rc < 0)
1431                 return rc;
1432         /*
1433          * FIXME: Add support for NOPOUT payload using unsolicited RDMA payload
1434          */
1435
1436         return iscsit_process_nop_out(conn, cmd, hdr);
1437 }
1438
1439 static int
1440 isert_handle_text_cmd(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1441                       struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1442                       struct iscsi_text *hdr)
1443 {
1444         struct iscsi_conn *conn = isert_conn->conn;
1445         u32 payload_length = ntoh24(hdr->dlength);
1446         int rc;
1447         unsigned char *text_in = NULL;
1448
1449         rc = iscsit_setup_text_cmd(conn, cmd, hdr);
1450         if (rc < 0)
1451                 return rc;
1452
1453         if (payload_length) {
1454                 text_in = kzalloc(payload_length, GFP_KERNEL);
1455                 if (!text_in) {
1456                         isert_err("Unable to allocate text_in of payload_length: %u\n",
1457                                   payload_length);
1458                         return -ENOMEM;
1459                 }
1460         }
1461         cmd->text_in_ptr = text_in;
1462
1463         memcpy(cmd->text_in_ptr, &rx_desc->data[0], payload_length);
1464
1465         return iscsit_process_text_cmd(conn, cmd, hdr);
1466 }
1467
1468 static int
1469 isert_rx_opcode(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc,
1470                 uint32_t read_stag, uint64_t read_va,
1471                 uint32_t write_stag, uint64_t write_va)
1472 {
1473         struct iscsi_hdr *hdr = &rx_desc->iscsi_header;
1474         struct iscsi_conn *conn = isert_conn->conn;
1475         struct iscsi_cmd *cmd;
1476         struct isert_cmd *isert_cmd;
1477         int ret = -EINVAL;
1478         u8 opcode = (hdr->opcode & ISCSI_OPCODE_MASK);
1479
1480         if (conn->sess->sess_ops->SessionType &&
1481            (!(opcode & ISCSI_OP_TEXT) || !(opcode & ISCSI_OP_LOGOUT))) {
1482                 isert_err("Got illegal opcode: 0x%02x in SessionType=Discovery,"
1483                           " ignoring\n", opcode);
1484                 return 0;
1485         }
1486
1487         switch (opcode) {
1488         case ISCSI_OP_SCSI_CMD:
1489                 cmd = isert_allocate_cmd(conn);
1490                 if (!cmd)
1491                         break;
1492
1493                 isert_cmd = iscsit_priv_cmd(cmd);
1494                 isert_cmd->read_stag = read_stag;
1495                 isert_cmd->read_va = read_va;
1496                 isert_cmd->write_stag = write_stag;
1497                 isert_cmd->write_va = write_va;
1498
1499                 ret = isert_handle_scsi_cmd(isert_conn, isert_cmd, cmd,
1500                                         rx_desc, (unsigned char *)hdr);
1501                 break;
1502         case ISCSI_OP_NOOP_OUT:
1503                 cmd = isert_allocate_cmd(conn);
1504                 if (!cmd)
1505                         break;
1506
1507                 isert_cmd = iscsit_priv_cmd(cmd);
1508                 ret = isert_handle_nop_out(isert_conn, isert_cmd, cmd,
1509                                            rx_desc, (unsigned char *)hdr);
1510                 break;
1511         case ISCSI_OP_SCSI_DATA_OUT:
1512                 ret = isert_handle_iscsi_dataout(isert_conn, rx_desc,
1513                                                 (unsigned char *)hdr);
1514                 break;
1515         case ISCSI_OP_SCSI_TMFUNC:
1516                 cmd = isert_allocate_cmd(conn);
1517                 if (!cmd)
1518                         break;
1519
1520                 ret = iscsit_handle_task_mgt_cmd(conn, cmd,
1521                                                 (unsigned char *)hdr);
1522                 break;
1523         case ISCSI_OP_LOGOUT:
1524                 cmd = isert_allocate_cmd(conn);
1525                 if (!cmd)
1526                         break;
1527
1528                 ret = iscsit_handle_logout_cmd(conn, cmd, (unsigned char *)hdr);
1529                 break;
1530         case ISCSI_OP_TEXT:
1531                 if (be32_to_cpu(hdr->ttt) != 0xFFFFFFFF) {
1532                         cmd = iscsit_find_cmd_from_itt(conn, hdr->itt);
1533                         if (!cmd)
1534                                 break;
1535                 } else {
1536                         cmd = isert_allocate_cmd(conn);
1537                         if (!cmd)
1538                                 break;
1539                 }
1540
1541                 isert_cmd = iscsit_priv_cmd(cmd);
1542                 ret = isert_handle_text_cmd(isert_conn, isert_cmd, cmd,
1543                                             rx_desc, (struct iscsi_text *)hdr);
1544                 break;
1545         default:
1546                 isert_err("Got unknown iSCSI OpCode: 0x%02x\n", opcode);
1547                 dump_stack();
1548                 break;
1549         }
1550
1551         return ret;
1552 }
1553
1554 static void
1555 isert_rx_do_work(struct iser_rx_desc *rx_desc, struct isert_conn *isert_conn)
1556 {
1557         struct iser_hdr *iser_hdr = &rx_desc->iser_header;
1558         uint64_t read_va = 0, write_va = 0;
1559         uint32_t read_stag = 0, write_stag = 0;
1560         int rc;
1561
1562         switch (iser_hdr->flags & 0xF0) {
1563         case ISCSI_CTRL:
1564                 if (iser_hdr->flags & ISER_RSV) {
1565                         read_stag = be32_to_cpu(iser_hdr->read_stag);
1566                         read_va = be64_to_cpu(iser_hdr->read_va);
1567                         isert_dbg("ISER_RSV: read_stag: 0x%x read_va: 0x%llx\n",
1568                                   read_stag, (unsigned long long)read_va);
1569                 }
1570                 if (iser_hdr->flags & ISER_WSV) {
1571                         write_stag = be32_to_cpu(iser_hdr->write_stag);
1572                         write_va = be64_to_cpu(iser_hdr->write_va);
1573                         isert_dbg("ISER_WSV: write_stag: 0x%x write_va: 0x%llx\n",
1574                                   write_stag, (unsigned long long)write_va);
1575                 }
1576
1577                 isert_dbg("ISER ISCSI_CTRL PDU\n");
1578                 break;
1579         case ISER_HELLO:
1580                 isert_err("iSER Hello message\n");
1581                 break;
1582         default:
1583                 isert_warn("Unknown iSER hdr flags: 0x%02x\n", iser_hdr->flags);
1584                 break;
1585         }
1586
1587         rc = isert_rx_opcode(isert_conn, rx_desc,
1588                              read_stag, read_va, write_stag, write_va);
1589 }
1590
1591 static void
1592 isert_rcv_completion(struct iser_rx_desc *desc,
1593                      struct isert_conn *isert_conn,
1594                      u32 xfer_len)
1595 {
1596         struct ib_device *ib_dev = isert_conn->cm_id->device;
1597         struct iscsi_hdr *hdr;
1598         u64 rx_dma;
1599         int rx_buflen, outstanding;
1600
1601         if ((char *)desc == isert_conn->login_req_buf) {
1602                 rx_dma = isert_conn->login_req_dma;
1603                 rx_buflen = ISER_RX_LOGIN_SIZE;
1604                 isert_dbg("login_buf: Using rx_dma: 0x%llx, rx_buflen: %d\n",
1605                          rx_dma, rx_buflen);
1606         } else {
1607                 rx_dma = desc->dma_addr;
1608                 rx_buflen = ISER_RX_PAYLOAD_SIZE;
1609                 isert_dbg("req_buf: Using rx_dma: 0x%llx, rx_buflen: %d\n",
1610                          rx_dma, rx_buflen);
1611         }
1612
1613         ib_dma_sync_single_for_cpu(ib_dev, rx_dma, rx_buflen, DMA_FROM_DEVICE);
1614
1615         hdr = &desc->iscsi_header;
1616         isert_dbg("iSCSI opcode: 0x%02x, ITT: 0x%08x, flags: 0x%02x dlen: %d\n",
1617                  hdr->opcode, hdr->itt, hdr->flags,
1618                  (int)(xfer_len - ISER_HEADERS_LEN));
1619
1620         if ((char *)desc == isert_conn->login_req_buf) {
1621                 isert_conn->login_req_len = xfer_len - ISER_HEADERS_LEN;
1622                 if (isert_conn->conn) {
1623                         struct iscsi_login *login = isert_conn->conn->conn_login;
1624
1625                         if (login && !login->first_request)
1626                                 isert_rx_login_req(isert_conn);
1627                 }
1628                 mutex_lock(&isert_conn->mutex);
1629                 complete(&isert_conn->login_req_comp);
1630                 mutex_unlock(&isert_conn->mutex);
1631         } else {
1632                 isert_rx_do_work(desc, isert_conn);
1633         }
1634
1635         ib_dma_sync_single_for_device(ib_dev, rx_dma, rx_buflen,
1636                                       DMA_FROM_DEVICE);
1637
1638         isert_conn->post_recv_buf_count--;
1639         isert_dbg("Decremented post_recv_buf_count: %d\n",
1640                   isert_conn->post_recv_buf_count);
1641
1642         if ((char *)desc == isert_conn->login_req_buf)
1643                 return;
1644
1645         outstanding = isert_conn->post_recv_buf_count;
1646         if (outstanding + ISERT_MIN_POSTED_RX <= ISERT_QP_MAX_RECV_DTOS) {
1647                 int err, count = min(ISERT_QP_MAX_RECV_DTOS - outstanding,
1648                                 ISERT_MIN_POSTED_RX);
1649                 err = isert_post_recv(isert_conn, count);
1650                 if (err) {
1651                         isert_err("isert_post_recv() count: %d failed, %d\n",
1652                                count, err);
1653                 }
1654         }
1655 }
1656
1657 static int
1658 isert_map_data_buf(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1659                    struct scatterlist *sg, u32 nents, u32 length, u32 offset,
1660                    enum iser_ib_op_code op, struct isert_data_buf *data)
1661 {
1662         struct ib_device *ib_dev = isert_conn->cm_id->device;
1663
1664         data->dma_dir = op == ISER_IB_RDMA_WRITE ?
1665                               DMA_TO_DEVICE : DMA_FROM_DEVICE;
1666
1667         data->len = length - offset;
1668         data->offset = offset;
1669         data->sg_off = data->offset / PAGE_SIZE;
1670
1671         data->sg = &sg[data->sg_off];
1672         data->nents = min_t(unsigned int, nents - data->sg_off,
1673                                           ISCSI_ISER_SG_TABLESIZE);
1674         data->len = min_t(unsigned int, data->len, ISCSI_ISER_SG_TABLESIZE *
1675                                         PAGE_SIZE);
1676
1677         data->dma_nents = ib_dma_map_sg(ib_dev, data->sg, data->nents,
1678                                         data->dma_dir);
1679         if (unlikely(!data->dma_nents)) {
1680                 isert_err("Cmd: unable to dma map SGs %p\n", sg);
1681                 return -EINVAL;
1682         }
1683
1684         isert_dbg("Mapped cmd: %p count: %u sg: %p sg_nents: %u rdma_len %d\n",
1685                   isert_cmd, data->dma_nents, data->sg, data->nents, data->len);
1686
1687         return 0;
1688 }
1689
1690 static void
1691 isert_unmap_data_buf(struct isert_conn *isert_conn, struct isert_data_buf *data)
1692 {
1693         struct ib_device *ib_dev = isert_conn->cm_id->device;
1694
1695         ib_dma_unmap_sg(ib_dev, data->sg, data->nents, data->dma_dir);
1696         memset(data, 0, sizeof(*data));
1697 }
1698
1699
1700
1701 static void
1702 isert_unmap_cmd(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn)
1703 {
1704         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1705
1706         isert_dbg("Cmd %p\n", isert_cmd);
1707
1708         if (wr->data.sg) {
1709                 isert_dbg("Cmd %p unmap_sg op\n", isert_cmd);
1710                 isert_unmap_data_buf(isert_conn, &wr->data);
1711         }
1712
1713         if (wr->send_wr) {
1714                 isert_dbg("Cmd %p free send_wr\n", isert_cmd);
1715                 kfree(wr->send_wr);
1716                 wr->send_wr = NULL;
1717         }
1718
1719         if (wr->ib_sge) {
1720                 isert_dbg("Cmd %p free ib_sge\n", isert_cmd);
1721                 kfree(wr->ib_sge);
1722                 wr->ib_sge = NULL;
1723         }
1724 }
1725
1726 static void
1727 isert_unreg_rdma(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn)
1728 {
1729         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1730
1731         isert_dbg("Cmd %p\n", isert_cmd);
1732
1733         if (wr->fr_desc) {
1734                 isert_dbg("Cmd %p free fr_desc %p\n", isert_cmd, wr->fr_desc);
1735                 if (wr->fr_desc->ind & ISERT_PROTECTED) {
1736                         isert_unmap_data_buf(isert_conn, &wr->prot);
1737                         wr->fr_desc->ind &= ~ISERT_PROTECTED;
1738                 }
1739                 spin_lock_bh(&isert_conn->pool_lock);
1740                 list_add_tail(&wr->fr_desc->list, &isert_conn->fr_pool);
1741                 spin_unlock_bh(&isert_conn->pool_lock);
1742                 wr->fr_desc = NULL;
1743         }
1744
1745         if (wr->data.sg) {
1746                 isert_dbg("Cmd %p unmap_sg op\n", isert_cmd);
1747                 isert_unmap_data_buf(isert_conn, &wr->data);
1748         }
1749
1750         wr->ib_sge = NULL;
1751         wr->send_wr = NULL;
1752 }
1753
1754 static void
1755 isert_put_cmd(struct isert_cmd *isert_cmd, bool comp_err)
1756 {
1757         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1758         struct isert_conn *isert_conn = isert_cmd->conn;
1759         struct iscsi_conn *conn = isert_conn->conn;
1760         struct isert_device *device = isert_conn->device;
1761         struct iscsi_text_rsp *hdr;
1762
1763         isert_dbg("Cmd %p\n", isert_cmd);
1764
1765         switch (cmd->iscsi_opcode) {
1766         case ISCSI_OP_SCSI_CMD:
1767                 spin_lock_bh(&conn->cmd_lock);
1768                 if (!list_empty(&cmd->i_conn_node))
1769                         list_del_init(&cmd->i_conn_node);
1770                 spin_unlock_bh(&conn->cmd_lock);
1771
1772                 if (cmd->data_direction == DMA_TO_DEVICE) {
1773                         iscsit_stop_dataout_timer(cmd);
1774                         /*
1775                          * Check for special case during comp_err where
1776                          * WRITE_PENDING has been handed off from core,
1777                          * but requires an extra target_put_sess_cmd()
1778                          * before transport_generic_free_cmd() below.
1779                          */
1780                         if (comp_err &&
1781                             cmd->se_cmd.t_state == TRANSPORT_WRITE_PENDING) {
1782                                 struct se_cmd *se_cmd = &cmd->se_cmd;
1783
1784                                 target_put_sess_cmd(se_cmd);
1785                         }
1786                 }
1787
1788                 device->unreg_rdma_mem(isert_cmd, isert_conn);
1789                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1790                 break;
1791         case ISCSI_OP_SCSI_TMFUNC:
1792                 spin_lock_bh(&conn->cmd_lock);
1793                 if (!list_empty(&cmd->i_conn_node))
1794                         list_del_init(&cmd->i_conn_node);
1795                 spin_unlock_bh(&conn->cmd_lock);
1796
1797                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1798                 break;
1799         case ISCSI_OP_REJECT:
1800         case ISCSI_OP_NOOP_OUT:
1801         case ISCSI_OP_TEXT:
1802                 hdr = (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
1803                 /* If the continue bit is on, keep the command alive */
1804                 if (hdr->flags & ISCSI_FLAG_TEXT_CONTINUE)
1805                         break;
1806
1807                 spin_lock_bh(&conn->cmd_lock);
1808                 if (!list_empty(&cmd->i_conn_node))
1809                         list_del_init(&cmd->i_conn_node);
1810                 spin_unlock_bh(&conn->cmd_lock);
1811
1812                 /*
1813                  * Handle special case for REJECT when iscsi_add_reject*() has
1814                  * overwritten the original iscsi_opcode assignment, and the
1815                  * associated cmd->se_cmd needs to be released.
1816                  */
1817                 if (cmd->se_cmd.se_tfo != NULL) {
1818                         isert_dbg("Calling transport_generic_free_cmd for 0x%02x\n",
1819                                  cmd->iscsi_opcode);
1820                         transport_generic_free_cmd(&cmd->se_cmd, 0);
1821                         break;
1822                 }
1823                 /*
1824                  * Fall-through
1825                  */
1826         default:
1827                 iscsit_release_cmd(cmd);
1828                 break;
1829         }
1830 }
1831
1832 static void
1833 isert_unmap_tx_desc(struct iser_tx_desc *tx_desc, struct ib_device *ib_dev)
1834 {
1835         if (tx_desc->dma_addr != 0) {
1836                 isert_dbg("unmap single for tx_desc->dma_addr\n");
1837                 ib_dma_unmap_single(ib_dev, tx_desc->dma_addr,
1838                                     ISER_HEADERS_LEN, DMA_TO_DEVICE);
1839                 tx_desc->dma_addr = 0;
1840         }
1841 }
1842
1843 static void
1844 isert_completion_put(struct iser_tx_desc *tx_desc, struct isert_cmd *isert_cmd,
1845                      struct ib_device *ib_dev, bool comp_err)
1846 {
1847         if (isert_cmd->pdu_buf_dma != 0) {
1848                 isert_dbg("unmap single for isert_cmd->pdu_buf_dma\n");
1849                 ib_dma_unmap_single(ib_dev, isert_cmd->pdu_buf_dma,
1850                                     isert_cmd->pdu_buf_len, DMA_TO_DEVICE);
1851                 isert_cmd->pdu_buf_dma = 0;
1852         }
1853
1854         isert_unmap_tx_desc(tx_desc, ib_dev);
1855         isert_put_cmd(isert_cmd, comp_err);
1856 }
1857
1858 static int
1859 isert_check_pi_status(struct se_cmd *se_cmd, struct ib_mr *sig_mr)
1860 {
1861         struct ib_mr_status mr_status;
1862         int ret;
1863
1864         ret = ib_check_mr_status(sig_mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
1865         if (ret) {
1866                 isert_err("ib_check_mr_status failed, ret %d\n", ret);
1867                 goto fail_mr_status;
1868         }
1869
1870         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1871                 u64 sec_offset_err;
1872                 u32 block_size = se_cmd->se_dev->dev_attrib.block_size + 8;
1873
1874                 switch (mr_status.sig_err.err_type) {
1875                 case IB_SIG_BAD_GUARD:
1876                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_GUARD_CHECK_FAILED;
1877                         break;
1878                 case IB_SIG_BAD_REFTAG:
1879                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_REF_TAG_CHECK_FAILED;
1880                         break;
1881                 case IB_SIG_BAD_APPTAG:
1882                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_APP_TAG_CHECK_FAILED;
1883                         break;
1884                 }
1885                 sec_offset_err = mr_status.sig_err.sig_err_offset;
1886                 do_div(sec_offset_err, block_size);
1887                 se_cmd->bad_sector = sec_offset_err + se_cmd->t_task_lba;
1888
1889                 isert_err("PI error found type %d at sector 0x%llx "
1890                           "expected 0x%x vs actual 0x%x\n",
1891                           mr_status.sig_err.err_type,
1892                           (unsigned long long)se_cmd->bad_sector,
1893                           mr_status.sig_err.expected,
1894                           mr_status.sig_err.actual);
1895                 ret = 1;
1896         }
1897
1898 fail_mr_status:
1899         return ret;
1900 }
1901
1902 static void
1903 isert_completion_rdma_write(struct iser_tx_desc *tx_desc,
1904                             struct isert_cmd *isert_cmd)
1905 {
1906         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1907         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1908         struct se_cmd *se_cmd = &cmd->se_cmd;
1909         struct isert_conn *isert_conn = isert_cmd->conn;
1910         struct isert_device *device = isert_conn->device;
1911         int ret = 0;
1912
1913         if (wr->fr_desc && wr->fr_desc->ind & ISERT_PROTECTED) {
1914                 ret = isert_check_pi_status(se_cmd,
1915                                             wr->fr_desc->pi_ctx->sig_mr);
1916                 wr->fr_desc->ind &= ~ISERT_PROTECTED;
1917         }
1918
1919         device->unreg_rdma_mem(isert_cmd, isert_conn);
1920         wr->send_wr_num = 0;
1921         if (ret)
1922                 transport_send_check_condition_and_sense(se_cmd,
1923                                                          se_cmd->pi_err, 0);
1924         else
1925                 isert_put_response(isert_conn->conn, cmd);
1926 }
1927
1928 static void
1929 isert_completion_rdma_read(struct iser_tx_desc *tx_desc,
1930                            struct isert_cmd *isert_cmd)
1931 {
1932         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
1933         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1934         struct se_cmd *se_cmd = &cmd->se_cmd;
1935         struct isert_conn *isert_conn = isert_cmd->conn;
1936         struct isert_device *device = isert_conn->device;
1937         int ret = 0;
1938
1939         if (wr->fr_desc && wr->fr_desc->ind & ISERT_PROTECTED) {
1940                 ret = isert_check_pi_status(se_cmd,
1941                                             wr->fr_desc->pi_ctx->sig_mr);
1942                 wr->fr_desc->ind &= ~ISERT_PROTECTED;
1943         }
1944
1945         iscsit_stop_dataout_timer(cmd);
1946         device->unreg_rdma_mem(isert_cmd, isert_conn);
1947         cmd->write_data_done = wr->data.len;
1948         wr->send_wr_num = 0;
1949
1950         isert_dbg("Cmd: %p RDMA_READ comp calling execute_cmd\n", isert_cmd);
1951         spin_lock_bh(&cmd->istate_lock);
1952         cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1953         cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1954         spin_unlock_bh(&cmd->istate_lock);
1955
1956         if (ret) {
1957                 target_put_sess_cmd(se_cmd);
1958                 transport_send_check_condition_and_sense(se_cmd,
1959                                                          se_cmd->pi_err, 0);
1960         } else {
1961                 target_execute_cmd(se_cmd);
1962         }
1963 }
1964
1965 static void
1966 isert_do_control_comp(struct work_struct *work)
1967 {
1968         struct isert_cmd *isert_cmd = container_of(work,
1969                         struct isert_cmd, comp_work);
1970         struct isert_conn *isert_conn = isert_cmd->conn;
1971         struct ib_device *ib_dev = isert_conn->cm_id->device;
1972         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1973
1974         isert_dbg("Cmd %p i_state %d\n", isert_cmd, cmd->i_state);
1975
1976         switch (cmd->i_state) {
1977         case ISTATE_SEND_TASKMGTRSP:
1978                 iscsit_tmr_post_handler(cmd, cmd->conn);
1979         case ISTATE_SEND_REJECT:   /* FALLTHRU */
1980         case ISTATE_SEND_TEXTRSP:  /* FALLTHRU */
1981                 cmd->i_state = ISTATE_SENT_STATUS;
1982                 isert_completion_put(&isert_cmd->tx_desc, isert_cmd,
1983                                      ib_dev, false);
1984                 break;
1985         case ISTATE_SEND_LOGOUTRSP:
1986                 iscsit_logout_post_handler(cmd, cmd->conn);
1987                 break;
1988         default:
1989                 isert_err("Unknown i_state %d\n", cmd->i_state);
1990                 dump_stack();
1991                 break;
1992         }
1993 }
1994
1995 static void
1996 isert_response_completion(struct iser_tx_desc *tx_desc,
1997                           struct isert_cmd *isert_cmd,
1998                           struct isert_conn *isert_conn,
1999                           struct ib_device *ib_dev)
2000 {
2001         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
2002
2003         if (cmd->i_state == ISTATE_SEND_TASKMGTRSP ||
2004             cmd->i_state == ISTATE_SEND_LOGOUTRSP ||
2005             cmd->i_state == ISTATE_SEND_REJECT ||
2006             cmd->i_state == ISTATE_SEND_TEXTRSP) {
2007                 isert_unmap_tx_desc(tx_desc, ib_dev);
2008
2009                 INIT_WORK(&isert_cmd->comp_work, isert_do_control_comp);
2010                 queue_work(isert_comp_wq, &isert_cmd->comp_work);
2011                 return;
2012         }
2013
2014         cmd->i_state = ISTATE_SENT_STATUS;
2015         isert_completion_put(tx_desc, isert_cmd, ib_dev, false);
2016 }
2017
2018 static void
2019 isert_snd_completion(struct iser_tx_desc *tx_desc,
2020                       struct isert_conn *isert_conn)
2021 {
2022         struct ib_device *ib_dev = isert_conn->cm_id->device;
2023         struct isert_cmd *isert_cmd = tx_desc->isert_cmd;
2024         struct isert_rdma_wr *wr;
2025
2026         if (!isert_cmd) {
2027                 isert_unmap_tx_desc(tx_desc, ib_dev);
2028                 return;
2029         }
2030         wr = &isert_cmd->rdma_wr;
2031
2032         isert_dbg("Cmd %p iser_ib_op %d\n", isert_cmd, wr->iser_ib_op);
2033
2034         switch (wr->iser_ib_op) {
2035         case ISER_IB_SEND:
2036                 isert_response_completion(tx_desc, isert_cmd,
2037                                           isert_conn, ib_dev);
2038                 break;
2039         case ISER_IB_RDMA_WRITE:
2040                 isert_completion_rdma_write(tx_desc, isert_cmd);
2041                 break;
2042         case ISER_IB_RDMA_READ:
2043                 isert_completion_rdma_read(tx_desc, isert_cmd);
2044                 break;
2045         default:
2046                 isert_err("Unknown wr->iser_ib_op: 0x%x\n", wr->iser_ib_op);
2047                 dump_stack();
2048                 break;
2049         }
2050 }
2051
2052 /**
2053  * is_isert_tx_desc() - Indicate if the completion wr_id
2054  *     is a TX descriptor or not.
2055  * @isert_conn: iser connection
2056  * @wr_id: completion WR identifier
2057  *
2058  * Since we cannot rely on wc opcode in FLUSH errors
2059  * we must work around it by checking if the wr_id address
2060  * falls in the iser connection rx_descs buffer. If so
2061  * it is an RX descriptor, otherwize it is a TX.
2062  */
2063 static inline bool
2064 is_isert_tx_desc(struct isert_conn *isert_conn, void *wr_id)
2065 {
2066         void *start = isert_conn->rx_descs;
2067         int len = ISERT_QP_MAX_RECV_DTOS * sizeof(*isert_conn->rx_descs);
2068
2069         if (wr_id >= start && wr_id < start + len)
2070                 return false;
2071
2072         return true;
2073 }
2074
2075 static void
2076 isert_cq_comp_err(struct isert_conn *isert_conn, struct ib_wc *wc)
2077 {
2078         if (wc->wr_id == ISER_BEACON_WRID) {
2079                 isert_info("conn %p completing wait_comp_err\n",
2080                            isert_conn);
2081                 complete(&isert_conn->wait_comp_err);
2082         } else if (is_isert_tx_desc(isert_conn, (void *)(uintptr_t)wc->wr_id)) {
2083                 struct ib_device *ib_dev = isert_conn->cm_id->device;
2084                 struct isert_cmd *isert_cmd;
2085                 struct iser_tx_desc *desc;
2086
2087                 desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
2088                 isert_cmd = desc->isert_cmd;
2089                 if (!isert_cmd)
2090                         isert_unmap_tx_desc(desc, ib_dev);
2091                 else
2092                         isert_completion_put(desc, isert_cmd, ib_dev, true);
2093         } else {
2094                 isert_conn->post_recv_buf_count--;
2095                 if (!isert_conn->post_recv_buf_count)
2096                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
2097         }
2098 }
2099
2100 static void
2101 isert_handle_wc(struct ib_wc *wc)
2102 {
2103         struct isert_conn *isert_conn;
2104         struct iser_tx_desc *tx_desc;
2105         struct iser_rx_desc *rx_desc;
2106
2107         isert_conn = wc->qp->qp_context;
2108         if (likely(wc->status == IB_WC_SUCCESS)) {
2109                 if (wc->opcode == IB_WC_RECV) {
2110                         rx_desc = (struct iser_rx_desc *)(uintptr_t)wc->wr_id;
2111                         isert_rcv_completion(rx_desc, isert_conn, wc->byte_len);
2112                 } else {
2113                         tx_desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
2114                         isert_snd_completion(tx_desc, isert_conn);
2115                 }
2116         } else {
2117                 if (wc->status != IB_WC_WR_FLUSH_ERR)
2118                         isert_err("%s (%d): wr id %llx vend_err %x\n",
2119                                   ib_wc_status_msg(wc->status), wc->status,
2120                                   wc->wr_id, wc->vendor_err);
2121                 else
2122                         isert_dbg("%s (%d): wr id %llx\n",
2123                                   ib_wc_status_msg(wc->status), wc->status,
2124                                   wc->wr_id);
2125
2126                 if (wc->wr_id != ISER_FASTREG_LI_WRID)
2127                         isert_cq_comp_err(isert_conn, wc);
2128         }
2129 }
2130
2131 static void
2132 isert_cq_work(struct work_struct *work)
2133 {
2134         enum { isert_poll_budget = 65536 };
2135         struct isert_comp *comp = container_of(work, struct isert_comp,
2136                                                work);
2137         struct ib_wc *const wcs = comp->wcs;
2138         int i, n, completed = 0;
2139
2140         while ((n = ib_poll_cq(comp->cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
2141                 for (i = 0; i < n; i++)
2142                         isert_handle_wc(&wcs[i]);
2143
2144                 completed += n;
2145                 if (completed >= isert_poll_budget)
2146                         break;
2147         }
2148
2149         ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP);
2150 }
2151
2152 static void
2153 isert_cq_callback(struct ib_cq *cq, void *context)
2154 {
2155         struct isert_comp *comp = context;
2156
2157         queue_work(isert_comp_wq, &comp->work);
2158 }
2159
2160 static int
2161 isert_post_response(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd)
2162 {
2163         struct ib_send_wr *wr_failed;
2164         int ret;
2165
2166         ret = ib_post_send(isert_conn->qp, &isert_cmd->tx_desc.send_wr,
2167                            &wr_failed);
2168         if (ret) {
2169                 isert_err("ib_post_send failed with %d\n", ret);
2170                 return ret;
2171         }
2172         return ret;
2173 }
2174
2175 static int
2176 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2177 {
2178         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2179         struct isert_conn *isert_conn = conn->context;
2180         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2181         struct iscsi_scsi_rsp *hdr = (struct iscsi_scsi_rsp *)
2182                                 &isert_cmd->tx_desc.iscsi_header;
2183
2184         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2185         iscsit_build_rsp_pdu(cmd, conn, true, hdr);
2186         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2187         /*
2188          * Attach SENSE DATA payload to iSCSI Response PDU
2189          */
2190         if (cmd->se_cmd.sense_buffer &&
2191             ((cmd->se_cmd.se_cmd_flags & SCF_TRANSPORT_TASK_SENSE) ||
2192             (cmd->se_cmd.se_cmd_flags & SCF_EMULATED_TASK_SENSE))) {
2193                 struct isert_device *device = isert_conn->device;
2194                 struct ib_device *ib_dev = device->ib_device;
2195                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2196                 u32 padding, pdu_len;
2197
2198                 put_unaligned_be16(cmd->se_cmd.scsi_sense_length,
2199                                    cmd->sense_buffer);
2200                 cmd->se_cmd.scsi_sense_length += sizeof(__be16);
2201
2202                 padding = -(cmd->se_cmd.scsi_sense_length) & 3;
2203                 hton24(hdr->dlength, (u32)cmd->se_cmd.scsi_sense_length);
2204                 pdu_len = cmd->se_cmd.scsi_sense_length + padding;
2205
2206                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2207                                 (void *)cmd->sense_buffer, pdu_len,
2208                                 DMA_TO_DEVICE);
2209
2210                 isert_cmd->pdu_buf_len = pdu_len;
2211                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2212                 tx_dsg->length  = pdu_len;
2213                 tx_dsg->lkey    = device->mr->lkey;
2214                 isert_cmd->tx_desc.num_sge = 2;
2215         }
2216
2217         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2218
2219         isert_dbg("Posting SCSI Response\n");
2220
2221         return isert_post_response(isert_conn, isert_cmd);
2222 }
2223
2224 static void
2225 isert_aborted_task(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2226 {
2227         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2228         struct isert_conn *isert_conn = conn->context;
2229         struct isert_device *device = isert_conn->device;
2230
2231         spin_lock_bh(&conn->cmd_lock);
2232         if (!list_empty(&cmd->i_conn_node))
2233                 list_del_init(&cmd->i_conn_node);
2234         spin_unlock_bh(&conn->cmd_lock);
2235
2236         if (cmd->data_direction == DMA_TO_DEVICE)
2237                 iscsit_stop_dataout_timer(cmd);
2238
2239         device->unreg_rdma_mem(isert_cmd, isert_conn);
2240 }
2241
2242 static enum target_prot_op
2243 isert_get_sup_prot_ops(struct iscsi_conn *conn)
2244 {
2245         struct isert_conn *isert_conn = conn->context;
2246         struct isert_device *device = isert_conn->device;
2247
2248         if (conn->tpg->tpg_attrib.t10_pi) {
2249                 if (device->pi_capable) {
2250                         isert_info("conn %p PI offload enabled\n", isert_conn);
2251                         isert_conn->pi_support = true;
2252                         return TARGET_PROT_ALL;
2253                 }
2254         }
2255
2256         isert_info("conn %p PI offload disabled\n", isert_conn);
2257         isert_conn->pi_support = false;
2258
2259         return TARGET_PROT_NORMAL;
2260 }
2261
2262 static int
2263 isert_put_nopin(struct iscsi_cmd *cmd, struct iscsi_conn *conn,
2264                 bool nopout_response)
2265 {
2266         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2267         struct isert_conn *isert_conn = conn->context;
2268         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2269
2270         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2271         iscsit_build_nopin_rsp(cmd, conn, (struct iscsi_nopin *)
2272                                &isert_cmd->tx_desc.iscsi_header,
2273                                nopout_response);
2274         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2275         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2276
2277         isert_dbg("conn %p Posting NOPIN Response\n", isert_conn);
2278
2279         return isert_post_response(isert_conn, isert_cmd);
2280 }
2281
2282 static int
2283 isert_put_logout_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2284 {
2285         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2286         struct isert_conn *isert_conn = conn->context;
2287         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2288
2289         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2290         iscsit_build_logout_rsp(cmd, conn, (struct iscsi_logout_rsp *)
2291                                 &isert_cmd->tx_desc.iscsi_header);
2292         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2293         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2294
2295         isert_dbg("conn %p Posting Logout Response\n", isert_conn);
2296
2297         return isert_post_response(isert_conn, isert_cmd);
2298 }
2299
2300 static int
2301 isert_put_tm_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2302 {
2303         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2304         struct isert_conn *isert_conn = conn->context;
2305         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2306
2307         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2308         iscsit_build_task_mgt_rsp(cmd, conn, (struct iscsi_tm_rsp *)
2309                                   &isert_cmd->tx_desc.iscsi_header);
2310         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2311         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2312
2313         isert_dbg("conn %p Posting Task Management Response\n", isert_conn);
2314
2315         return isert_post_response(isert_conn, isert_cmd);
2316 }
2317
2318 static int
2319 isert_put_reject(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2320 {
2321         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2322         struct isert_conn *isert_conn = conn->context;
2323         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2324         struct isert_device *device = isert_conn->device;
2325         struct ib_device *ib_dev = device->ib_device;
2326         struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2327         struct iscsi_reject *hdr =
2328                 (struct iscsi_reject *)&isert_cmd->tx_desc.iscsi_header;
2329
2330         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2331         iscsit_build_reject(cmd, conn, hdr);
2332         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2333
2334         hton24(hdr->dlength, ISCSI_HDR_LEN);
2335         isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2336                         (void *)cmd->buf_ptr, ISCSI_HDR_LEN,
2337                         DMA_TO_DEVICE);
2338         isert_cmd->pdu_buf_len = ISCSI_HDR_LEN;
2339         tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2340         tx_dsg->length  = ISCSI_HDR_LEN;
2341         tx_dsg->lkey    = device->mr->lkey;
2342         isert_cmd->tx_desc.num_sge = 2;
2343
2344         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2345
2346         isert_dbg("conn %p Posting Reject\n", isert_conn);
2347
2348         return isert_post_response(isert_conn, isert_cmd);
2349 }
2350
2351 static int
2352 isert_put_text_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2353 {
2354         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2355         struct isert_conn *isert_conn = conn->context;
2356         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2357         struct iscsi_text_rsp *hdr =
2358                 (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
2359         u32 txt_rsp_len;
2360         int rc;
2361
2362         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2363         rc = iscsit_build_text_rsp(cmd, conn, hdr, ISCSI_INFINIBAND);
2364         if (rc < 0)
2365                 return rc;
2366
2367         txt_rsp_len = rc;
2368         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2369
2370         if (txt_rsp_len) {
2371                 struct isert_device *device = isert_conn->device;
2372                 struct ib_device *ib_dev = device->ib_device;
2373                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2374                 void *txt_rsp_buf = cmd->buf_ptr;
2375
2376                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2377                                 txt_rsp_buf, txt_rsp_len, DMA_TO_DEVICE);
2378
2379                 isert_cmd->pdu_buf_len = txt_rsp_len;
2380                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2381                 tx_dsg->length  = txt_rsp_len;
2382                 tx_dsg->lkey    = device->mr->lkey;
2383                 isert_cmd->tx_desc.num_sge = 2;
2384         }
2385         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2386
2387         isert_dbg("conn %p Text Response\n", isert_conn);
2388
2389         return isert_post_response(isert_conn, isert_cmd);
2390 }
2391
2392 static int
2393 isert_build_rdma_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
2394                     struct ib_sge *ib_sge, struct ib_send_wr *send_wr,
2395                     u32 data_left, u32 offset)
2396 {
2397         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
2398         struct scatterlist *sg_start, *tmp_sg;
2399         struct isert_device *device = isert_conn->device;
2400         struct ib_device *ib_dev = device->ib_device;
2401         u32 sg_off, page_off;
2402         int i = 0, sg_nents;
2403
2404         sg_off = offset / PAGE_SIZE;
2405         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
2406         sg_nents = min(cmd->se_cmd.t_data_nents - sg_off, isert_conn->max_sge);
2407         page_off = offset % PAGE_SIZE;
2408
2409         send_wr->sg_list = ib_sge;
2410         send_wr->wr_id = (uintptr_t)&isert_cmd->tx_desc;
2411         /*
2412          * Perform mapping of TCM scatterlist memory ib_sge dma_addr.
2413          */
2414         for_each_sg(sg_start, tmp_sg, sg_nents, i) {
2415                 isert_dbg("RDMA from SGL dma_addr: 0x%llx dma_len: %u, "
2416                           "page_off: %u\n",
2417                           (unsigned long long)tmp_sg->dma_address,
2418                           tmp_sg->length, page_off);
2419
2420                 ib_sge->addr = ib_sg_dma_address(ib_dev, tmp_sg) + page_off;
2421                 ib_sge->length = min_t(u32, data_left,
2422                                 ib_sg_dma_len(ib_dev, tmp_sg) - page_off);
2423                 ib_sge->lkey = device->mr->lkey;
2424
2425                 isert_dbg("RDMA ib_sge: addr: 0x%llx  length: %u lkey: %x\n",
2426                           ib_sge->addr, ib_sge->length, ib_sge->lkey);
2427                 page_off = 0;
2428                 data_left -= ib_sge->length;
2429                 if (!data_left)
2430                         break;
2431                 ib_sge++;
2432                 isert_dbg("Incrementing ib_sge pointer to %p\n", ib_sge);
2433         }
2434
2435         send_wr->num_sge = ++i;
2436         isert_dbg("Set outgoing sg_list: %p num_sg: %u from TCM SGLs\n",
2437                   send_wr->sg_list, send_wr->num_sge);
2438
2439         return send_wr->num_sge;
2440 }
2441
2442 static int
2443 isert_map_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2444                struct isert_rdma_wr *wr)
2445 {
2446         struct se_cmd *se_cmd = &cmd->se_cmd;
2447         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2448         struct isert_conn *isert_conn = conn->context;
2449         struct isert_data_buf *data = &wr->data;
2450         struct ib_send_wr *send_wr;
2451         struct ib_sge *ib_sge;
2452         u32 offset, data_len, data_left, rdma_write_max, va_offset = 0;
2453         int ret = 0, i, ib_sge_cnt;
2454
2455         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2456
2457         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2458         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2459                                  se_cmd->t_data_nents, se_cmd->data_length,
2460                                  offset, wr->iser_ib_op, &wr->data);
2461         if (ret)
2462                 return ret;
2463
2464         data_left = data->len;
2465         offset = data->offset;
2466
2467         ib_sge = kzalloc(sizeof(struct ib_sge) * data->nents, GFP_KERNEL);
2468         if (!ib_sge) {
2469                 isert_warn("Unable to allocate ib_sge\n");
2470                 ret = -ENOMEM;
2471                 goto unmap_cmd;
2472         }
2473         wr->ib_sge = ib_sge;
2474
2475         wr->send_wr_num = DIV_ROUND_UP(data->nents, isert_conn->max_sge);
2476         wr->send_wr = kzalloc(sizeof(struct ib_send_wr) * wr->send_wr_num,
2477                                 GFP_KERNEL);
2478         if (!wr->send_wr) {
2479                 isert_dbg("Unable to allocate wr->send_wr\n");
2480                 ret = -ENOMEM;
2481                 goto unmap_cmd;
2482         }
2483
2484         wr->isert_cmd = isert_cmd;
2485         rdma_write_max = isert_conn->max_sge * PAGE_SIZE;
2486
2487         for (i = 0; i < wr->send_wr_num; i++) {
2488                 send_wr = &isert_cmd->rdma_wr.send_wr[i];
2489                 data_len = min(data_left, rdma_write_max);
2490
2491                 send_wr->send_flags = 0;
2492                 if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2493                         send_wr->opcode = IB_WR_RDMA_WRITE;
2494                         send_wr->wr.rdma.remote_addr = isert_cmd->read_va + offset;
2495                         send_wr->wr.rdma.rkey = isert_cmd->read_stag;
2496                         if (i + 1 == wr->send_wr_num)
2497                                 send_wr->next = &isert_cmd->tx_desc.send_wr;
2498                         else
2499                                 send_wr->next = &wr->send_wr[i + 1];
2500                 } else {
2501                         send_wr->opcode = IB_WR_RDMA_READ;
2502                         send_wr->wr.rdma.remote_addr = isert_cmd->write_va + va_offset;
2503                         send_wr->wr.rdma.rkey = isert_cmd->write_stag;
2504                         if (i + 1 == wr->send_wr_num)
2505                                 send_wr->send_flags = IB_SEND_SIGNALED;
2506                         else
2507                                 send_wr->next = &wr->send_wr[i + 1];
2508                 }
2509
2510                 ib_sge_cnt = isert_build_rdma_wr(isert_conn, isert_cmd, ib_sge,
2511                                         send_wr, data_len, offset);
2512                 ib_sge += ib_sge_cnt;
2513
2514                 offset += data_len;
2515                 va_offset += data_len;
2516                 data_left -= data_len;
2517         }
2518
2519         return 0;
2520 unmap_cmd:
2521         isert_unmap_data_buf(isert_conn, data);
2522
2523         return ret;
2524 }
2525
2526 static int
2527 isert_map_fr_pagelist(struct ib_device *ib_dev,
2528                       struct scatterlist *sg_start, int sg_nents, u64 *fr_pl)
2529 {
2530         u64 start_addr, end_addr, page, chunk_start = 0;
2531         struct scatterlist *tmp_sg;
2532         int i = 0, new_chunk, last_ent, n_pages;
2533
2534         n_pages = 0;
2535         new_chunk = 1;
2536         last_ent = sg_nents - 1;
2537         for_each_sg(sg_start, tmp_sg, sg_nents, i) {
2538                 start_addr = ib_sg_dma_address(ib_dev, tmp_sg);
2539                 if (new_chunk)
2540                         chunk_start = start_addr;
2541                 end_addr = start_addr + ib_sg_dma_len(ib_dev, tmp_sg);
2542
2543                 isert_dbg("SGL[%d] dma_addr: 0x%llx len: %u\n",
2544                           i, (unsigned long long)tmp_sg->dma_address,
2545                           tmp_sg->length);
2546
2547                 if ((end_addr & ~PAGE_MASK) && i < last_ent) {
2548                         new_chunk = 0;
2549                         continue;
2550                 }
2551                 new_chunk = 1;
2552
2553                 page = chunk_start & PAGE_MASK;
2554                 do {
2555                         fr_pl[n_pages++] = page;
2556                         isert_dbg("Mapped page_list[%d] page_addr: 0x%llx\n",
2557                                   n_pages - 1, page);
2558                         page += PAGE_SIZE;
2559                 } while (page < end_addr);
2560         }
2561
2562         return n_pages;
2563 }
2564
2565 static inline void
2566 isert_inv_rkey(struct ib_send_wr *inv_wr, struct ib_mr *mr)
2567 {
2568         u32 rkey;
2569
2570         memset(inv_wr, 0, sizeof(*inv_wr));
2571         inv_wr->wr_id = ISER_FASTREG_LI_WRID;
2572         inv_wr->opcode = IB_WR_LOCAL_INV;
2573         inv_wr->ex.invalidate_rkey = mr->rkey;
2574
2575         /* Bump the key */
2576         rkey = ib_inc_rkey(mr->rkey);
2577         ib_update_fast_reg_key(mr, rkey);
2578 }
2579
2580 static int
2581 isert_fast_reg_mr(struct isert_conn *isert_conn,
2582                   struct fast_reg_descriptor *fr_desc,
2583                   struct isert_data_buf *mem,
2584                   enum isert_indicator ind,
2585                   struct ib_sge *sge)
2586 {
2587         struct isert_device *device = isert_conn->device;
2588         struct ib_device *ib_dev = device->ib_device;
2589         struct ib_mr *mr;
2590         struct ib_fast_reg_page_list *frpl;
2591         struct ib_send_wr fr_wr, inv_wr;
2592         struct ib_send_wr *bad_wr, *wr = NULL;
2593         int ret, pagelist_len;
2594         u32 page_off;
2595
2596         if (mem->dma_nents == 1) {
2597                 sge->lkey = device->mr->lkey;
2598                 sge->addr = ib_sg_dma_address(ib_dev, &mem->sg[0]);
2599                 sge->length = ib_sg_dma_len(ib_dev, &mem->sg[0]);
2600                 isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2601                          sge->addr, sge->length, sge->lkey);
2602                 return 0;
2603         }
2604
2605         if (ind == ISERT_DATA_KEY_VALID) {
2606                 /* Registering data buffer */
2607                 mr = fr_desc->data_mr;
2608                 frpl = fr_desc->data_frpl;
2609         } else {
2610                 /* Registering protection buffer */
2611                 mr = fr_desc->pi_ctx->prot_mr;
2612                 frpl = fr_desc->pi_ctx->prot_frpl;
2613         }
2614
2615         page_off = mem->offset % PAGE_SIZE;
2616
2617         isert_dbg("Use fr_desc %p sg_nents %d offset %u\n",
2618                   fr_desc, mem->nents, mem->offset);
2619
2620         pagelist_len = isert_map_fr_pagelist(ib_dev, mem->sg, mem->nents,
2621                                              &frpl->page_list[0]);
2622
2623         if (!(fr_desc->ind & ind)) {
2624                 isert_inv_rkey(&inv_wr, mr);
2625                 wr = &inv_wr;
2626         }
2627
2628         /* Prepare FASTREG WR */
2629         memset(&fr_wr, 0, sizeof(fr_wr));
2630         fr_wr.wr_id = ISER_FASTREG_LI_WRID;
2631         fr_wr.opcode = IB_WR_FAST_REG_MR;
2632         fr_wr.wr.fast_reg.iova_start = frpl->page_list[0] + page_off;
2633         fr_wr.wr.fast_reg.page_list = frpl;
2634         fr_wr.wr.fast_reg.page_list_len = pagelist_len;
2635         fr_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
2636         fr_wr.wr.fast_reg.length = mem->len;
2637         fr_wr.wr.fast_reg.rkey = mr->rkey;
2638         fr_wr.wr.fast_reg.access_flags = IB_ACCESS_LOCAL_WRITE;
2639
2640         if (!wr)
2641                 wr = &fr_wr;
2642         else
2643                 wr->next = &fr_wr;
2644
2645         ret = ib_post_send(isert_conn->qp, wr, &bad_wr);
2646         if (ret) {
2647                 isert_err("fast registration failed, ret:%d\n", ret);
2648                 return ret;
2649         }
2650         fr_desc->ind &= ~ind;
2651
2652         sge->lkey = mr->lkey;
2653         sge->addr = frpl->page_list[0] + page_off;
2654         sge->length = mem->len;
2655
2656         isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2657                   sge->addr, sge->length, sge->lkey);
2658
2659         return ret;
2660 }
2661
2662 static inline void
2663 isert_set_dif_domain(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs,
2664                      struct ib_sig_domain *domain)
2665 {
2666         domain->sig_type = IB_SIG_TYPE_T10_DIF;
2667         domain->sig.dif.bg_type = IB_T10DIF_CRC;
2668         domain->sig.dif.pi_interval = se_cmd->se_dev->dev_attrib.block_size;
2669         domain->sig.dif.ref_tag = se_cmd->reftag_seed;
2670         /*
2671          * At the moment we hard code those, but if in the future
2672          * the target core would like to use it, we will take it
2673          * from se_cmd.
2674          */
2675         domain->sig.dif.apptag_check_mask = 0xffff;
2676         domain->sig.dif.app_escape = true;
2677         domain->sig.dif.ref_escape = true;
2678         if (se_cmd->prot_type == TARGET_DIF_TYPE1_PROT ||
2679             se_cmd->prot_type == TARGET_DIF_TYPE2_PROT)
2680                 domain->sig.dif.ref_remap = true;
2681 };
2682
2683 static int
2684 isert_set_sig_attrs(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs)
2685 {
2686         switch (se_cmd->prot_op) {
2687         case TARGET_PROT_DIN_INSERT:
2688         case TARGET_PROT_DOUT_STRIP:
2689                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
2690                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2691                 break;
2692         case TARGET_PROT_DOUT_INSERT:
2693         case TARGET_PROT_DIN_STRIP:
2694                 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
2695                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2696                 break;
2697         case TARGET_PROT_DIN_PASS:
2698         case TARGET_PROT_DOUT_PASS:
2699                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2700                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2701                 break;
2702         default:
2703                 isert_err("Unsupported PI operation %d\n", se_cmd->prot_op);
2704                 return -EINVAL;
2705         }
2706
2707         return 0;
2708 }
2709
2710 static inline u8
2711 isert_set_prot_checks(u8 prot_checks)
2712 {
2713         return (prot_checks & TARGET_DIF_CHECK_GUARD  ? 0xc0 : 0) |
2714                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x30 : 0) |
2715                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x0f : 0);
2716 }
2717
2718 static int
2719 isert_reg_sig_mr(struct isert_conn *isert_conn,
2720                  struct se_cmd *se_cmd,
2721                  struct isert_rdma_wr *rdma_wr,
2722                  struct fast_reg_descriptor *fr_desc)
2723 {
2724         struct ib_send_wr sig_wr, inv_wr;
2725         struct ib_send_wr *bad_wr, *wr = NULL;
2726         struct pi_context *pi_ctx = fr_desc->pi_ctx;
2727         struct ib_sig_attrs sig_attrs;
2728         int ret;
2729
2730         memset(&sig_attrs, 0, sizeof(sig_attrs));
2731         ret = isert_set_sig_attrs(se_cmd, &sig_attrs);
2732         if (ret)
2733                 goto err;
2734
2735         sig_attrs.check_mask = isert_set_prot_checks(se_cmd->prot_checks);
2736
2737         if (!(fr_desc->ind & ISERT_SIG_KEY_VALID)) {
2738                 isert_inv_rkey(&inv_wr, pi_ctx->sig_mr);
2739                 wr = &inv_wr;
2740         }
2741
2742         memset(&sig_wr, 0, sizeof(sig_wr));
2743         sig_wr.opcode = IB_WR_REG_SIG_MR;
2744         sig_wr.wr_id = ISER_FASTREG_LI_WRID;
2745         sig_wr.sg_list = &rdma_wr->ib_sg[DATA];
2746         sig_wr.num_sge = 1;
2747         sig_wr.wr.sig_handover.access_flags = IB_ACCESS_LOCAL_WRITE;
2748         sig_wr.wr.sig_handover.sig_attrs = &sig_attrs;
2749         sig_wr.wr.sig_handover.sig_mr = pi_ctx->sig_mr;
2750         if (se_cmd->t_prot_sg)
2751                 sig_wr.wr.sig_handover.prot = &rdma_wr->ib_sg[PROT];
2752
2753         if (!wr)
2754                 wr = &sig_wr;
2755         else
2756                 wr->next = &sig_wr;
2757
2758         ret = ib_post_send(isert_conn->qp, wr, &bad_wr);
2759         if (ret) {
2760                 isert_err("fast registration failed, ret:%d\n", ret);
2761                 goto err;
2762         }
2763         fr_desc->ind &= ~ISERT_SIG_KEY_VALID;
2764
2765         rdma_wr->ib_sg[SIG].lkey = pi_ctx->sig_mr->lkey;
2766         rdma_wr->ib_sg[SIG].addr = 0;
2767         rdma_wr->ib_sg[SIG].length = se_cmd->data_length;
2768         if (se_cmd->prot_op != TARGET_PROT_DIN_STRIP &&
2769             se_cmd->prot_op != TARGET_PROT_DOUT_INSERT)
2770                 /*
2771                  * We have protection guards on the wire
2772                  * so we need to set a larget transfer
2773                  */
2774                 rdma_wr->ib_sg[SIG].length += se_cmd->prot_length;
2775
2776         isert_dbg("sig_sge: addr: 0x%llx  length: %u lkey: %x\n",
2777                   rdma_wr->ib_sg[SIG].addr, rdma_wr->ib_sg[SIG].length,
2778                   rdma_wr->ib_sg[SIG].lkey);
2779 err:
2780         return ret;
2781 }
2782
2783 static int
2784 isert_handle_prot_cmd(struct isert_conn *isert_conn,
2785                       struct isert_cmd *isert_cmd,
2786                       struct isert_rdma_wr *wr)
2787 {
2788         struct isert_device *device = isert_conn->device;
2789         struct se_cmd *se_cmd = &isert_cmd->iscsi_cmd->se_cmd;
2790         int ret;
2791
2792         if (!wr->fr_desc->pi_ctx) {
2793                 ret = isert_create_pi_ctx(wr->fr_desc,
2794                                           device->ib_device,
2795                                           device->pd);
2796                 if (ret) {
2797                         isert_err("conn %p failed to allocate pi_ctx\n",
2798                                   isert_conn);
2799                         return ret;
2800                 }
2801         }
2802
2803         if (se_cmd->t_prot_sg) {
2804                 ret = isert_map_data_buf(isert_conn, isert_cmd,
2805                                          se_cmd->t_prot_sg,
2806                                          se_cmd->t_prot_nents,
2807                                          se_cmd->prot_length,
2808                                          0, wr->iser_ib_op, &wr->prot);
2809                 if (ret) {
2810                         isert_err("conn %p failed to map protection buffer\n",
2811                                   isert_conn);
2812                         return ret;
2813                 }
2814
2815                 memset(&wr->ib_sg[PROT], 0, sizeof(wr->ib_sg[PROT]));
2816                 ret = isert_fast_reg_mr(isert_conn, wr->fr_desc, &wr->prot,
2817                                         ISERT_PROT_KEY_VALID, &wr->ib_sg[PROT]);
2818                 if (ret) {
2819                         isert_err("conn %p failed to fast reg mr\n",
2820                                   isert_conn);
2821                         goto unmap_prot_cmd;
2822                 }
2823         }
2824
2825         ret = isert_reg_sig_mr(isert_conn, se_cmd, wr, wr->fr_desc);
2826         if (ret) {
2827                 isert_err("conn %p failed to fast reg mr\n",
2828                           isert_conn);
2829                 goto unmap_prot_cmd;
2830         }
2831         wr->fr_desc->ind |= ISERT_PROTECTED;
2832
2833         return 0;
2834
2835 unmap_prot_cmd:
2836         if (se_cmd->t_prot_sg)
2837                 isert_unmap_data_buf(isert_conn, &wr->prot);
2838
2839         return ret;
2840 }
2841
2842 static int
2843 isert_reg_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2844                struct isert_rdma_wr *wr)
2845 {
2846         struct se_cmd *se_cmd = &cmd->se_cmd;
2847         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2848         struct isert_conn *isert_conn = conn->context;
2849         struct fast_reg_descriptor *fr_desc = NULL;
2850         struct ib_send_wr *send_wr;
2851         struct ib_sge *ib_sg;
2852         u32 offset;
2853         int ret = 0;
2854         unsigned long flags;
2855
2856         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2857
2858         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2859         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2860                                  se_cmd->t_data_nents, se_cmd->data_length,
2861                                  offset, wr->iser_ib_op, &wr->data);
2862         if (ret)
2863                 return ret;
2864
2865         if (wr->data.dma_nents != 1 || isert_prot_cmd(isert_conn, se_cmd)) {
2866                 spin_lock_irqsave(&isert_conn->pool_lock, flags);
2867                 fr_desc = list_first_entry(&isert_conn->fr_pool,
2868                                            struct fast_reg_descriptor, list);
2869                 list_del(&fr_desc->list);
2870                 spin_unlock_irqrestore(&isert_conn->pool_lock, flags);
2871                 wr->fr_desc = fr_desc;
2872         }
2873
2874         ret = isert_fast_reg_mr(isert_conn, fr_desc, &wr->data,
2875                                 ISERT_DATA_KEY_VALID, &wr->ib_sg[DATA]);
2876         if (ret)
2877                 goto unmap_cmd;
2878
2879         if (isert_prot_cmd(isert_conn, se_cmd)) {
2880                 ret = isert_handle_prot_cmd(isert_conn, isert_cmd, wr);
2881                 if (ret)
2882                         goto unmap_cmd;
2883
2884                 ib_sg = &wr->ib_sg[SIG];
2885         } else {
2886                 ib_sg = &wr->ib_sg[DATA];
2887         }
2888
2889         memcpy(&wr->s_ib_sge, ib_sg, sizeof(*ib_sg));
2890         wr->ib_sge = &wr->s_ib_sge;
2891         wr->send_wr_num = 1;
2892         memset(&wr->s_send_wr, 0, sizeof(*send_wr));
2893         wr->send_wr = &wr->s_send_wr;
2894         wr->isert_cmd = isert_cmd;
2895
2896         send_wr = &isert_cmd->rdma_wr.s_send_wr;
2897         send_wr->sg_list = &wr->s_ib_sge;
2898         send_wr->num_sge = 1;
2899         send_wr->wr_id = (uintptr_t)&isert_cmd->tx_desc;
2900         if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2901                 send_wr->opcode = IB_WR_RDMA_WRITE;
2902                 send_wr->wr.rdma.remote_addr = isert_cmd->read_va;
2903                 send_wr->wr.rdma.rkey = isert_cmd->read_stag;
2904                 send_wr->send_flags = !isert_prot_cmd(isert_conn, se_cmd) ?
2905                                       0 : IB_SEND_SIGNALED;
2906         } else {
2907                 send_wr->opcode = IB_WR_RDMA_READ;
2908                 send_wr->wr.rdma.remote_addr = isert_cmd->write_va;
2909                 send_wr->wr.rdma.rkey = isert_cmd->write_stag;
2910                 send_wr->send_flags = IB_SEND_SIGNALED;
2911         }
2912
2913         return 0;
2914
2915 unmap_cmd:
2916         if (fr_desc) {
2917                 spin_lock_irqsave(&isert_conn->pool_lock, flags);
2918                 list_add_tail(&fr_desc->list, &isert_conn->fr_pool);
2919                 spin_unlock_irqrestore(&isert_conn->pool_lock, flags);
2920         }
2921         isert_unmap_data_buf(isert_conn, &wr->data);
2922
2923         return ret;
2924 }
2925
2926 static int
2927 isert_put_datain(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2928 {
2929         struct se_cmd *se_cmd = &cmd->se_cmd;
2930         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2931         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2932         struct isert_conn *isert_conn = conn->context;
2933         struct isert_device *device = isert_conn->device;
2934         struct ib_send_wr *wr_failed;
2935         int rc;
2936
2937         isert_dbg("Cmd: %p RDMA_WRITE data_length: %u\n",
2938                  isert_cmd, se_cmd->data_length);
2939
2940         wr->iser_ib_op = ISER_IB_RDMA_WRITE;
2941         rc = device->reg_rdma_mem(conn, cmd, wr);
2942         if (rc) {
2943                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2944                 return rc;
2945         }
2946
2947         if (!isert_prot_cmd(isert_conn, se_cmd)) {
2948                 /*
2949                  * Build isert_conn->tx_desc for iSCSI response PDU and attach
2950                  */
2951                 isert_create_send_desc(isert_conn, isert_cmd,
2952                                        &isert_cmd->tx_desc);
2953                 iscsit_build_rsp_pdu(cmd, conn, true, (struct iscsi_scsi_rsp *)
2954                                      &isert_cmd->tx_desc.iscsi_header);
2955                 isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2956                 isert_init_send_wr(isert_conn, isert_cmd,
2957                                    &isert_cmd->tx_desc.send_wr);
2958                 isert_cmd->rdma_wr.s_send_wr.next = &isert_cmd->tx_desc.send_wr;
2959                 wr->send_wr_num += 1;
2960         }
2961
2962         rc = ib_post_send(isert_conn->qp, wr->send_wr, &wr_failed);
2963         if (rc)
2964                 isert_warn("ib_post_send() failed for IB_WR_RDMA_WRITE\n");
2965
2966         if (!isert_prot_cmd(isert_conn, se_cmd))
2967                 isert_dbg("Cmd: %p posted RDMA_WRITE + Response for iSER Data "
2968                          "READ\n", isert_cmd);
2969         else
2970                 isert_dbg("Cmd: %p posted RDMA_WRITE for iSER Data READ\n",
2971                          isert_cmd);
2972
2973         return 1;
2974 }
2975
2976 static int
2977 isert_get_dataout(struct iscsi_conn *conn, struct iscsi_cmd *cmd, bool recovery)
2978 {
2979         struct se_cmd *se_cmd = &cmd->se_cmd;
2980         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2981         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2982         struct isert_conn *isert_conn = conn->context;
2983         struct isert_device *device = isert_conn->device;
2984         struct ib_send_wr *wr_failed;
2985         int rc;
2986
2987         isert_dbg("Cmd: %p RDMA_READ data_length: %u write_data_done: %u\n",
2988                  isert_cmd, se_cmd->data_length, cmd->write_data_done);
2989         wr->iser_ib_op = ISER_IB_RDMA_READ;
2990         rc = device->reg_rdma_mem(conn, cmd, wr);
2991         if (rc) {
2992                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2993                 return rc;
2994         }
2995
2996         rc = ib_post_send(isert_conn->qp, wr->send_wr, &wr_failed);
2997         if (rc)
2998                 isert_warn("ib_post_send() failed for IB_WR_RDMA_READ\n");
2999
3000         isert_dbg("Cmd: %p posted RDMA_READ memory for ISER Data WRITE\n",
3001                  isert_cmd);
3002
3003         return 0;
3004 }
3005
3006 static int
3007 isert_immediate_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
3008 {
3009         int ret;
3010
3011         switch (state) {
3012         case ISTATE_SEND_NOPIN_WANT_RESPONSE:
3013                 ret = isert_put_nopin(cmd, conn, false);
3014                 break;
3015         default:
3016                 isert_err("Unknown immediate state: 0x%02x\n", state);
3017                 ret = -EINVAL;
3018                 break;
3019         }
3020
3021         return ret;
3022 }
3023
3024 static int
3025 isert_response_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
3026 {
3027         struct isert_conn *isert_conn = conn->context;
3028         int ret;
3029
3030         switch (state) {
3031         case ISTATE_SEND_LOGOUTRSP:
3032                 ret = isert_put_logout_rsp(cmd, conn);
3033                 if (!ret)
3034                         isert_conn->logout_posted = true;
3035                 break;
3036         case ISTATE_SEND_NOPIN:
3037                 ret = isert_put_nopin(cmd, conn, true);
3038                 break;
3039         case ISTATE_SEND_TASKMGTRSP:
3040                 ret = isert_put_tm_rsp(cmd, conn);
3041                 break;
3042         case ISTATE_SEND_REJECT:
3043                 ret = isert_put_reject(cmd, conn);
3044                 break;
3045         case ISTATE_SEND_TEXTRSP:
3046                 ret = isert_put_text_rsp(cmd, conn);
3047                 break;
3048         case ISTATE_SEND_STATUS:
3049                 /*
3050                  * Special case for sending non GOOD SCSI status from TX thread
3051                  * context during pre se_cmd excecution failure.
3052                  */
3053                 ret = isert_put_response(conn, cmd);
3054                 break;
3055         default:
3056                 isert_err("Unknown response state: 0x%02x\n", state);
3057                 ret = -EINVAL;
3058                 break;
3059         }
3060
3061         return ret;
3062 }
3063
3064 struct rdma_cm_id *
3065 isert_setup_id(struct isert_np *isert_np)
3066 {
3067         struct iscsi_np *np = isert_np->np;
3068         struct rdma_cm_id *id;
3069         struct sockaddr *sa;
3070         int ret;
3071
3072         sa = (struct sockaddr *)&np->np_sockaddr;
3073         isert_dbg("ksockaddr: %p, sa: %p\n", &np->np_sockaddr, sa);
3074
3075         id = rdma_create_id(isert_cma_handler, isert_np,
3076                             RDMA_PS_TCP, IB_QPT_RC);
3077         if (IS_ERR(id)) {
3078                 isert_err("rdma_create_id() failed: %ld\n", PTR_ERR(id));
3079                 ret = PTR_ERR(id);
3080                 goto out;
3081         }
3082         isert_dbg("id %p context %p\n", id, id->context);
3083
3084         ret = rdma_bind_addr(id, sa);
3085         if (ret) {
3086                 isert_err("rdma_bind_addr() failed: %d\n", ret);
3087                 goto out_id;
3088         }
3089
3090         ret = rdma_listen(id, 0);
3091         if (ret) {
3092                 isert_err("rdma_listen() failed: %d\n", ret);
3093                 goto out_id;
3094         }
3095
3096         return id;
3097 out_id:
3098         rdma_destroy_id(id);
3099 out:
3100         return ERR_PTR(ret);
3101 }
3102
3103 static int
3104 isert_setup_np(struct iscsi_np *np,
3105                struct __kernel_sockaddr_storage *ksockaddr)
3106 {
3107         struct isert_np *isert_np;
3108         struct rdma_cm_id *isert_lid;
3109         int ret;
3110
3111         isert_np = kzalloc(sizeof(struct isert_np), GFP_KERNEL);
3112         if (!isert_np) {
3113                 isert_err("Unable to allocate struct isert_np\n");
3114                 return -ENOMEM;
3115         }
3116         sema_init(&isert_np->np_sem, 0);
3117         mutex_init(&isert_np->np_accept_mutex);
3118         INIT_LIST_HEAD(&isert_np->np_accept_list);
3119         init_completion(&isert_np->np_login_comp);
3120         isert_np->np = np;
3121
3122         /*
3123          * Setup the np->np_sockaddr from the passed sockaddr setup
3124          * in iscsi_target_configfs.c code..
3125          */
3126         memcpy(&np->np_sockaddr, ksockaddr,
3127                sizeof(struct __kernel_sockaddr_storage));
3128
3129         isert_lid = isert_setup_id(isert_np);
3130         if (IS_ERR(isert_lid)) {
3131                 ret = PTR_ERR(isert_lid);
3132                 goto out;
3133         }
3134
3135         isert_np->np_cm_id = isert_lid;
3136         np->np_context = isert_np;
3137
3138         return 0;
3139
3140 out:
3141         kfree(isert_np);
3142
3143         return ret;
3144 }
3145
3146 static int
3147 isert_rdma_accept(struct isert_conn *isert_conn)
3148 {
3149         struct rdma_cm_id *cm_id = isert_conn->cm_id;
3150         struct rdma_conn_param cp;
3151         int ret;
3152
3153         memset(&cp, 0, sizeof(struct rdma_conn_param));
3154         cp.initiator_depth = isert_conn->initiator_depth;
3155         cp.retry_count = 7;
3156         cp.rnr_retry_count = 7;
3157
3158         ret = rdma_accept(cm_id, &cp);
3159         if (ret) {
3160                 isert_err("rdma_accept() failed with: %d\n", ret);
3161                 return ret;
3162         }
3163
3164         return 0;
3165 }
3166
3167 static int
3168 isert_get_login_rx(struct iscsi_conn *conn, struct iscsi_login *login)
3169 {
3170         struct isert_conn *isert_conn = conn->context;
3171         int ret;
3172
3173         isert_info("before login_req comp conn: %p\n", isert_conn);
3174         ret = wait_for_completion_interruptible(&isert_conn->login_req_comp);
3175         if (ret) {
3176                 isert_err("isert_conn %p interrupted before got login req\n",
3177                           isert_conn);
3178                 return ret;
3179         }
3180         reinit_completion(&isert_conn->login_req_comp);
3181
3182         /*
3183          * For login requests after the first PDU, isert_rx_login_req() will
3184          * kick schedule_delayed_work(&conn->login_work) as the packet is
3185          * received, which turns this callback from iscsi_target_do_login_rx()
3186          * into a NOP.
3187          */
3188         if (!login->first_request)
3189                 return 0;
3190
3191         isert_rx_login_req(isert_conn);
3192
3193         isert_info("before login_comp conn: %p\n", conn);
3194         ret = wait_for_completion_interruptible(&isert_conn->login_comp);
3195         if (ret)
3196                 return ret;
3197
3198         isert_info("processing login->req: %p\n", login->req);
3199
3200         return 0;
3201 }
3202
3203 static void
3204 isert_set_conn_info(struct iscsi_np *np, struct iscsi_conn *conn,
3205                     struct isert_conn *isert_conn)
3206 {
3207         struct rdma_cm_id *cm_id = isert_conn->cm_id;
3208         struct rdma_route *cm_route = &cm_id->route;
3209         struct sockaddr_in *sock_in;
3210         struct sockaddr_in6 *sock_in6;
3211
3212         conn->login_family = np->np_sockaddr.ss_family;
3213
3214         if (np->np_sockaddr.ss_family == AF_INET6) {
3215                 sock_in6 = (struct sockaddr_in6 *)&cm_route->addr.dst_addr;
3216                 snprintf(conn->login_ip, sizeof(conn->login_ip), "%pI6c",
3217                          &sock_in6->sin6_addr.in6_u);
3218                 conn->login_port = ntohs(sock_in6->sin6_port);
3219
3220                 sock_in6 = (struct sockaddr_in6 *)&cm_route->addr.src_addr;
3221                 snprintf(conn->local_ip, sizeof(conn->local_ip), "%pI6c",
3222                          &sock_in6->sin6_addr.in6_u);
3223                 conn->local_port = ntohs(sock_in6->sin6_port);
3224         } else {
3225                 sock_in = (struct sockaddr_in *)&cm_route->addr.dst_addr;
3226                 sprintf(conn->login_ip, "%pI4",
3227                         &sock_in->sin_addr.s_addr);
3228                 conn->login_port = ntohs(sock_in->sin_port);
3229
3230                 sock_in = (struct sockaddr_in *)&cm_route->addr.src_addr;
3231                 sprintf(conn->local_ip, "%pI4",
3232                         &sock_in->sin_addr.s_addr);
3233                 conn->local_port = ntohs(sock_in->sin_port);
3234         }
3235 }
3236
3237 static int
3238 isert_accept_np(struct iscsi_np *np, struct iscsi_conn *conn)
3239 {
3240         struct isert_np *isert_np = np->np_context;
3241         struct isert_conn *isert_conn;
3242         int ret;
3243
3244 accept_wait:
3245         ret = down_interruptible(&isert_np->np_sem);
3246         if (ret)
3247                 return -ENODEV;
3248
3249         spin_lock_bh(&np->np_thread_lock);
3250         if (np->np_thread_state >= ISCSI_NP_THREAD_RESET) {
3251                 spin_unlock_bh(&np->np_thread_lock);
3252                 isert_dbg("np_thread_state %d\n",
3253                          np->np_thread_state);
3254                 /**
3255                  * No point in stalling here when np_thread
3256                  * is in state RESET/SHUTDOWN/EXIT - bail
3257                  **/
3258                 return -ENODEV;
3259         }
3260         spin_unlock_bh(&np->np_thread_lock);
3261
3262         mutex_lock(&isert_np->np_accept_mutex);
3263         if (list_empty(&isert_np->np_accept_list)) {
3264                 mutex_unlock(&isert_np->np_accept_mutex);
3265                 goto accept_wait;
3266         }
3267         isert_conn = list_first_entry(&isert_np->np_accept_list,
3268                         struct isert_conn, accept_node);
3269         list_del_init(&isert_conn->accept_node);
3270         mutex_unlock(&isert_np->np_accept_mutex);
3271
3272         conn->context = isert_conn;
3273         isert_conn->conn = conn;
3274
3275         isert_set_conn_info(np, conn, isert_conn);
3276
3277         isert_dbg("Processing isert_conn: %p\n", isert_conn);
3278
3279         return 0;
3280 }
3281
3282 static void
3283 isert_free_np(struct iscsi_np *np)
3284 {
3285         struct isert_np *isert_np = np->np_context;
3286         struct isert_conn *isert_conn, *n;
3287
3288         if (isert_np->np_cm_id)
3289                 rdma_destroy_id(isert_np->np_cm_id);
3290
3291         /*
3292          * FIXME: At this point we don't have a good way to insure
3293          * that at this point we don't have hanging connections that
3294          * completed RDMA establishment but didn't start iscsi login
3295          * process. So work-around this by cleaning up what ever piled
3296          * up in np_accept_list.
3297          */
3298         mutex_lock(&isert_np->np_accept_mutex);
3299         if (!list_empty(&isert_np->np_accept_list)) {
3300                 isert_info("Still have isert connections, cleaning up...\n");
3301                 list_for_each_entry_safe(isert_conn, n,
3302                                          &isert_np->np_accept_list,
3303                                          accept_node) {
3304                         isert_info("cleaning isert_conn %p state (%d)\n",
3305                                    isert_conn, isert_conn->state);
3306                         isert_connect_release(isert_conn);
3307                 }
3308         }
3309         mutex_unlock(&isert_np->np_accept_mutex);
3310
3311         np->np_context = NULL;
3312         kfree(isert_np);
3313 }
3314
3315 static void isert_release_work(struct work_struct *work)
3316 {
3317         struct isert_conn *isert_conn = container_of(work,
3318                                                      struct isert_conn,
3319                                                      release_work);
3320
3321         isert_info("Starting release conn %p\n", isert_conn);
3322
3323         wait_for_completion(&isert_conn->wait);
3324
3325         mutex_lock(&isert_conn->mutex);
3326         isert_conn->state = ISER_CONN_DOWN;
3327         mutex_unlock(&isert_conn->mutex);
3328
3329         isert_info("Destroying conn %p\n", isert_conn);
3330         isert_put_conn(isert_conn);
3331 }
3332
3333 static void
3334 isert_wait4logout(struct isert_conn *isert_conn)
3335 {
3336         struct iscsi_conn *conn = isert_conn->conn;
3337
3338         isert_info("conn %p\n", isert_conn);
3339
3340         if (isert_conn->logout_posted) {
3341                 isert_info("conn %p wait for conn_logout_comp\n", isert_conn);
3342                 wait_for_completion_timeout(&conn->conn_logout_comp,
3343                                             SECONDS_FOR_LOGOUT_COMP * HZ);
3344         }
3345 }
3346
3347 static void
3348 isert_wait4cmds(struct iscsi_conn *conn)
3349 {
3350         isert_info("iscsi_conn %p\n", conn);
3351
3352         if (conn->sess) {
3353                 target_sess_cmd_list_set_waiting(conn->sess->se_sess);
3354                 target_wait_for_sess_cmds(conn->sess->se_sess);
3355         }
3356 }
3357
3358 static void
3359 isert_wait4flush(struct isert_conn *isert_conn)
3360 {
3361         struct ib_recv_wr *bad_wr;
3362
3363         isert_info("conn %p\n", isert_conn);
3364
3365         init_completion(&isert_conn->wait_comp_err);
3366         isert_conn->beacon.wr_id = ISER_BEACON_WRID;
3367         /* post an indication that all flush errors were consumed */
3368         if (ib_post_recv(isert_conn->qp, &isert_conn->beacon, &bad_wr)) {
3369                 isert_err("conn %p failed to post beacon", isert_conn);
3370                 return;
3371         }
3372
3373         wait_for_completion(&isert_conn->wait_comp_err);
3374 }
3375
3376 static void isert_wait_conn(struct iscsi_conn *conn)
3377 {
3378         struct isert_conn *isert_conn = conn->context;
3379
3380         isert_info("Starting conn %p\n", isert_conn);
3381
3382         mutex_lock(&isert_conn->mutex);
3383         /*
3384          * Only wait for wait_comp_err if the isert_conn made it
3385          * into full feature phase..
3386          */
3387         if (isert_conn->state == ISER_CONN_INIT) {
3388                 mutex_unlock(&isert_conn->mutex);
3389                 return;
3390         }
3391         isert_conn_terminate(isert_conn);
3392         mutex_unlock(&isert_conn->mutex);
3393
3394         isert_wait4cmds(conn);
3395         isert_wait4flush(isert_conn);
3396         isert_wait4logout(isert_conn);
3397
3398         queue_work(isert_release_wq, &isert_conn->release_work);
3399 }
3400
3401 static void isert_free_conn(struct iscsi_conn *conn)
3402 {
3403         struct isert_conn *isert_conn = conn->context;
3404
3405         isert_wait4flush(isert_conn);
3406         isert_put_conn(isert_conn);
3407 }
3408
3409 static struct iscsit_transport iser_target_transport = {
3410         .name                   = "IB/iSER",
3411         .transport_type         = ISCSI_INFINIBAND,
3412         .priv_size              = sizeof(struct isert_cmd),
3413         .owner                  = THIS_MODULE,
3414         .iscsit_setup_np        = isert_setup_np,
3415         .iscsit_accept_np       = isert_accept_np,
3416         .iscsit_free_np         = isert_free_np,
3417         .iscsit_wait_conn       = isert_wait_conn,
3418         .iscsit_free_conn       = isert_free_conn,
3419         .iscsit_get_login_rx    = isert_get_login_rx,
3420         .iscsit_put_login_tx    = isert_put_login_tx,
3421         .iscsit_immediate_queue = isert_immediate_queue,
3422         .iscsit_response_queue  = isert_response_queue,
3423         .iscsit_get_dataout     = isert_get_dataout,
3424         .iscsit_queue_data_in   = isert_put_datain,
3425         .iscsit_queue_status    = isert_put_response,
3426         .iscsit_aborted_task    = isert_aborted_task,
3427         .iscsit_get_sup_prot_ops = isert_get_sup_prot_ops,
3428 };
3429
3430 static int __init isert_init(void)
3431 {
3432         int ret;
3433
3434         isert_comp_wq = alloc_workqueue("isert_comp_wq",
3435                                         WQ_UNBOUND | WQ_HIGHPRI, 0);
3436         if (!isert_comp_wq) {
3437                 isert_err("Unable to allocate isert_comp_wq\n");
3438                 ret = -ENOMEM;
3439                 return -ENOMEM;
3440         }
3441
3442         isert_release_wq = alloc_workqueue("isert_release_wq", WQ_UNBOUND,
3443                                         WQ_UNBOUND_MAX_ACTIVE);
3444         if (!isert_release_wq) {
3445                 isert_err("Unable to allocate isert_release_wq\n");
3446                 ret = -ENOMEM;
3447                 goto destroy_comp_wq;
3448         }
3449
3450         iscsit_register_transport(&iser_target_transport);
3451         isert_info("iSER_TARGET[0] - Loaded iser_target_transport\n");
3452
3453         return 0;
3454
3455 destroy_comp_wq:
3456         destroy_workqueue(isert_comp_wq);
3457
3458         return ret;
3459 }
3460
3461 static void __exit isert_exit(void)
3462 {
3463         flush_scheduled_work();
3464         destroy_workqueue(isert_release_wq);
3465         destroy_workqueue(isert_comp_wq);
3466         iscsit_unregister_transport(&iser_target_transport);
3467         isert_info("iSER_TARGET[0] - Released iser_target_transport\n");
3468 }
3469
3470 MODULE_DESCRIPTION("iSER-Target for mainline target infrastructure");
3471 MODULE_VERSION("1.0");
3472 MODULE_AUTHOR("nab@Linux-iSCSI.org");
3473 MODULE_LICENSE("GPL");
3474
3475 module_init(isert_init);
3476 module_exit(isert_exit);