kernel/drivers/infiniband/ulp/isert/ib_isert.c

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