drivers/infiniband/ulp/isert/ib_isert.c

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