drivers/infiniband/ulp/isert/ib_isert.c

   1 /*******************************************************************************
   2  * This file contains iSCSI extentions for RDMA (iSER) Verbs
   3  *
   4  * (c) Copyright 2013 Datera, Inc.
   5  *
   6  * Nicholas A. Bellinger <nab@linux-iscsi.org>
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License as published by
  10  * the Free Software Foundation; either version 2 of the License, or
  11  * (at your option) any later version.
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  ****************************************************************************/
  18
  19 #include <linux/string.h>
  20 #include <linux/module.h>
  21 #include <linux/scatterlist.h>
  22 #include <linux/socket.h>
  23 #include <linux/in.h>
  24 #include <linux/in6.h>
  25 #include <rdma/ib_verbs.h>
  26 #include <rdma/rdma_cm.h>
  27 #include <target/target_core_base.h>
  28 #include <target/target_core_fabric.h>
  29 #include <target/iscsi/iscsi_transport.h>
  30 #include <linux/semaphore.h>
  31
  32 #include "isert_proto.h"
  33 #include "ib_isert.h"
  34
  35 #define ISERT_MAX_CONN          8
  36 #define ISER_MAX_RX_CQ_LEN      (ISERT_QP_MAX_RECV_DTOS * ISERT_MAX_CONN)
  37 #define ISER_MAX_TX_CQ_LEN      (ISERT_QP_MAX_REQ_DTOS  * ISERT_MAX_CONN)
  38 #define ISER_MAX_CQ_LEN         (ISER_MAX_RX_CQ_LEN + ISER_MAX_TX_CQ_LEN + \
  39                                  ISERT_MAX_CONN)
  40
  41 static int isert_debug_level;
  42 module_param_named(debug_level, isert_debug_level, int, 0644);
  43 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0 (default:0)");
  44
  45 static DEFINE_MUTEX(device_list_mutex);
  46 static LIST_HEAD(device_list);
  47 static struct workqueue_struct *isert_comp_wq;
  48 static struct workqueue_struct *isert_release_wq;
  49
  50 static void
  51 isert_unmap_cmd(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn);
  52 static int
  53 isert_map_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
  54                struct isert_rdma_wr *wr);
  55 static void
  56 isert_unreg_rdma(struct isert_cmd *isert_cmd, struct isert_conn *isert_conn);
  57 static int
  58 isert_reg_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
  59                struct isert_rdma_wr *wr);
  60 static int
  61 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd);
  62 static int
  63 isert_rdma_post_recvl(struct isert_conn *isert_conn);
  64 static int
  65 isert_rdma_accept(struct isert_conn *isert_conn);
  66 struct rdma_cm_id *isert_setup_id(struct isert_np *isert_np);
  67
  68 static void isert_release_work(struct work_struct *work);
  69
  70 static inline bool
  71 isert_prot_cmd(struct isert_conn *conn, struct se_cmd *cmd)
  72 {
  73         return (conn->pi_support &&
  74                 cmd->prot_op != TARGET_PROT_NORMAL);
  75 }
  76
  77
  78 static void
  79 isert_qp_event_callback(struct ib_event *e, void *context)
  80 {
  81         struct isert_conn *isert_conn = context;
  82
  83         isert_err("%s (%d): conn %p\n",
  84                   ib_event_msg(e->event), e->event, isert_conn);
  85
  86         switch (e->event) {
  87         case IB_EVENT_COMM_EST:
  88                 rdma_notify(isert_conn->cm_id, IB_EVENT_COMM_EST);
  89                 break;
  90         case IB_EVENT_QP_LAST_WQE_REACHED:
  91                 isert_warn("Reached TX IB_EVENT_QP_LAST_WQE_REACHED\n");
  92                 break;
  93         default:
  94                 break;
  95         }
  96 }
  97
  98 static int
  99 isert_query_device(struct ib_device *ib_dev, struct ib_device_attr *devattr)
 100 {
 101         int ret;
 102
 103         ret = ib_query_device(ib_dev, devattr);
 104         if (ret) {
 105                 isert_err("ib_query_device() failed: %d\n", ret);
 106                 return ret;
 107         }
 108         isert_dbg("devattr->max_sge: %d\n", devattr->max_sge);
 109         isert_dbg("devattr->max_sge_rd: %d\n", devattr->max_sge_rd);
 110
 111         return 0;
 112 }
 113
 114 static struct isert_comp *
 115 isert_comp_get(struct isert_conn *isert_conn)
 116 {
 117         struct isert_device *device = isert_conn->device;
 118         struct isert_comp *comp;
 119         int i, min = 0;
 120
 121         mutex_lock(&device_list_mutex);
 122         for (i = 0; i < device->comps_used; i++)
 123                 if (device->comps[i].active_qps <
 124                     device->comps[min].active_qps)
 125                         min = i;
 126         comp = &device->comps[min];
 127         comp->active_qps++;
 128         mutex_unlock(&device_list_mutex);
 129
 130         isert_info("conn %p, using comp %p min_index: %d\n",
 131                    isert_conn, comp, min);
 132
 133         return comp;
 134 }
 135
 136 static void
 137 isert_comp_put(struct isert_comp *comp)
 138 {
 139         mutex_lock(&device_list_mutex);
 140         comp->active_qps--;
 141         mutex_unlock(&device_list_mutex);
 142 }
 143
 144 static struct ib_qp *
 145 isert_create_qp(struct isert_conn *isert_conn,
 146                 struct isert_comp *comp,
 147                 struct rdma_cm_id *cma_id)
 148 {
 149         struct isert_device *device = isert_conn->device;
 150         struct ib_qp_init_attr attr;
 151         int ret;
 152
 153         memset(&attr, 0, sizeof(struct ib_qp_init_attr));
 154         attr.event_handler = isert_qp_event_callback;
 155         attr.qp_context = isert_conn;
 156         attr.send_cq = comp->cq;
 157         attr.recv_cq = comp->cq;
 158         attr.cap.max_send_wr = ISERT_QP_MAX_REQ_DTOS;
 159         attr.cap.max_recv_wr = ISERT_QP_MAX_RECV_DTOS + 1;
 160         /*
 161          * FIXME: Use devattr.max_sge - 2 for max_send_sge as
 162          * work-around for RDMA_READs with ConnectX-2.
 163          *
 164          * Also, still make sure to have at least two SGEs for
 165          * outgoing control PDU responses.
 166          */
 167         attr.cap.max_send_sge = max(2, device->dev_attr.max_sge - 2);
 168         isert_conn->max_sge = attr.cap.max_send_sge;
 169
 170         attr.cap.max_recv_sge = 1;
 171         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 172         attr.qp_type = IB_QPT_RC;
 173         if (device->pi_capable)
 174                 attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
 175
 176         ret = rdma_create_qp(cma_id, device->pd, &attr);
 177         if (ret) {
 178                 isert_err("rdma_create_qp failed for cma_id %d\n", ret);
 179                 return ERR_PTR(ret);
 180         }
 181
 182         return cma_id->qp;
 183 }
 184
 185 static int
 186 isert_conn_setup_qp(struct isert_conn *isert_conn, struct rdma_cm_id *cma_id)
 187 {
 188         struct isert_comp *comp;
 189         int ret;
 190
 191         comp = isert_comp_get(isert_conn);
 192         isert_conn->qp = isert_create_qp(isert_conn, comp, cma_id);
 193         if (IS_ERR(isert_conn->qp)) {
 194                 ret = PTR_ERR(isert_conn->qp);
 195                 goto err;
 196         }
 197
 198         return 0;
 199 err:
 200         isert_comp_put(comp);
 201         return ret;
 202 }
 203
 204 static void
 205 isert_cq_event_callback(struct ib_event *e, void *context)
 206 {
 207         isert_dbg("event: %d\n", e->event);
 208 }
 209
 210 static int
 211 isert_alloc_rx_descriptors(struct isert_conn *isert_conn)
 212 {
 213         struct isert_device *device = isert_conn->device;
 214         struct ib_device *ib_dev = device->ib_device;
 215         struct iser_rx_desc *rx_desc;
 216         struct ib_sge *rx_sg;
 217         u64 dma_addr;
 218         int i, j;
 219
 220         isert_conn->rx_descs = kzalloc(ISERT_QP_MAX_RECV_DTOS *
 221                                 sizeof(struct iser_rx_desc), GFP_KERNEL);
 222         if (!isert_conn->rx_descs)
 223                 goto fail;
 224
 225         rx_desc = isert_conn->rx_descs;
 226
 227         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 228                 dma_addr = ib_dma_map_single(ib_dev, (void *)rx_desc,
 229                                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 230                 if (ib_dma_mapping_error(ib_dev, dma_addr))
 231                         goto dma_map_fail;
 232
 233                 rx_desc->dma_addr = dma_addr;
 234
 235                 rx_sg = &rx_desc->rx_sg;
 236                 rx_sg->addr = rx_desc->dma_addr;
 237                 rx_sg->length = ISER_RX_PAYLOAD_SIZE;
 238                 rx_sg->lkey = device->pd->local_dma_lkey;
 239         }
 240
 241         isert_conn->rx_desc_head = 0;
 242
 243         return 0;
 244
 245 dma_map_fail:
 246         rx_desc = isert_conn->rx_descs;
 247         for (j = 0; j < i; j++, rx_desc++) {
 248                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 249                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 250         }
 251         kfree(isert_conn->rx_descs);
 252         isert_conn->rx_descs = NULL;
 253 fail:
 254         isert_err("conn %p failed to allocate rx descriptors\n", isert_conn);
 255
 256         return -ENOMEM;
 257 }
 258
 259 static void
 260 isert_free_rx_descriptors(struct isert_conn *isert_conn)
 261 {
 262         struct ib_device *ib_dev = isert_conn->device->ib_device;
 263         struct iser_rx_desc *rx_desc;
 264         int i;
 265
 266         if (!isert_conn->rx_descs)
 267                 return;
 268
 269         rx_desc = isert_conn->rx_descs;
 270         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 271                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 272                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 273         }
 274
 275         kfree(isert_conn->rx_descs);
 276         isert_conn->rx_descs = NULL;
 277 }
 278
 279 static void isert_cq_work(struct work_struct *);
 280 static void isert_cq_callback(struct ib_cq *, void *);
 281
 282 static void
 283 isert_free_comps(struct isert_device *device)
 284 {
 285         int i;
 286
 287         for (i = 0; i < device->comps_used; i++) {
 288                 struct isert_comp *comp = &device->comps[i];
 289
 290                 if (comp->cq) {
 291                         cancel_work_sync(&comp->work);
 292                         ib_destroy_cq(comp->cq);
 293                 }
 294         }
 295         kfree(device->comps);
 296 }
 297
 298 static int
 299 isert_alloc_comps(struct isert_device *device,
 300                   struct ib_device_attr *attr)
 301 {
 302         int i, max_cqe, ret = 0;
 303
 304         device->comps_used = min(ISERT_MAX_CQ, min_t(int, num_online_cpus(),
 305                                  device->ib_device->num_comp_vectors));
 306
 307         isert_info("Using %d CQs, %s supports %d vectors support "
 308                    "Fast registration %d pi_capable %d\n",
 309                    device->comps_used, device->ib_device->name,
 310                    device->ib_device->num_comp_vectors, device->use_fastreg,
 311                    device->pi_capable);
 312
 313         device->comps = kcalloc(device->comps_used, sizeof(struct isert_comp),
 314                                 GFP_KERNEL);
 315         if (!device->comps) {
 316                 isert_err("Unable to allocate completion contexts\n");
 317                 return -ENOMEM;
 318         }
 319
 320         max_cqe = min(ISER_MAX_CQ_LEN, attr->max_cqe);
 321
 322         for (i = 0; i < device->comps_used; i++) {
 323                 struct ib_cq_init_attr cq_attr = {};
 324                 struct isert_comp *comp = &device->comps[i];
 325
 326                 comp->device = device;
 327                 INIT_WORK(&comp->work, isert_cq_work);
 328                 cq_attr.cqe = max_cqe;
 329                 cq_attr.comp_vector = i;
 330                 comp->cq = ib_create_cq(device->ib_device,
 331                                         isert_cq_callback,
 332                                         isert_cq_event_callback,
 333                                         (void *)comp,
 334                                         &cq_attr);
 335                 if (IS_ERR(comp->cq)) {
 336                         isert_err("Unable to allocate cq\n");
 337                         ret = PTR_ERR(comp->cq);
 338                         comp->cq = NULL;
 339                         goto out_cq;
 340                 }
 341
 342                 ret = ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP);
 343                 if (ret)
 344                         goto out_cq;
 345         }
 346
 347         return 0;
 348 out_cq:
 349         isert_free_comps(device);
 350         return ret;
 351 }
 352
 353 static int
 354 isert_create_device_ib_res(struct isert_device *device)
 355 {
 356         struct ib_device_attr *dev_attr;
 357         int ret;
 358
 359         dev_attr = &device->dev_attr;
 360         ret = isert_query_device(device->ib_device, dev_attr);
 361         if (ret)
 362                 return ret;
 363
 364         /* asign function handlers */
 365         if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS &&
 366             dev_attr->device_cap_flags & IB_DEVICE_SIGNATURE_HANDOVER) {
 367                 device->use_fastreg = 1;
 368                 device->reg_rdma_mem = isert_reg_rdma;
 369                 device->unreg_rdma_mem = isert_unreg_rdma;
 370         } else {
 371                 device->use_fastreg = 0;
 372                 device->reg_rdma_mem = isert_map_rdma;
 373                 device->unreg_rdma_mem = isert_unmap_cmd;
 374         }
 375
 376         ret = isert_alloc_comps(device, dev_attr);
 377         if (ret)
 378                 return ret;
 379
 380         device->pd = ib_alloc_pd(device->ib_device);
 381         if (IS_ERR(device->pd)) {
 382                 ret = PTR_ERR(device->pd);
 383                 isert_err("failed to allocate pd, device %p, ret=%d\n",
 384                           device, ret);
 385                 goto out_cq;
 386         }
 387
 388         /* Check signature cap */
 389         device->pi_capable = dev_attr->device_cap_flags &
 390                              IB_DEVICE_SIGNATURE_HANDOVER ? true : false;
 391
 392         return 0;
 393
 394 out_cq:
 395         isert_free_comps(device);
 396         return ret;
 397 }
 398
 399 static void
 400 isert_free_device_ib_res(struct isert_device *device)
 401 {
 402         isert_info("device %p\n", device);
 403
 404         ib_dealloc_pd(device->pd);
 405         isert_free_comps(device);
 406 }
 407
 408 static void
 409 isert_device_put(struct isert_device *device)
 410 {
 411         mutex_lock(&device_list_mutex);
 412         device->refcount--;
 413         isert_info("device %p refcount %d\n", device, device->refcount);
 414         if (!device->refcount) {
 415                 isert_free_device_ib_res(device);
 416                 list_del(&device->dev_node);
 417                 kfree(device);
 418         }
 419         mutex_unlock(&device_list_mutex);
 420 }
 421
 422 static struct isert_device *
 423 isert_device_get(struct rdma_cm_id *cma_id)
 424 {
 425         struct isert_device *device;
 426         int ret;
 427
 428         mutex_lock(&device_list_mutex);
 429         list_for_each_entry(device, &device_list, dev_node) {
 430                 if (device->ib_device->node_guid == cma_id->device->node_guid) {
 431                         device->refcount++;
 432                         isert_info("Found iser device %p refcount %d\n",
 433                                    device, device->refcount);
 434                         mutex_unlock(&device_list_mutex);
 435                         return device;
 436                 }
 437         }
 438
 439         device = kzalloc(sizeof(struct isert_device), GFP_KERNEL);
 440         if (!device) {
 441                 mutex_unlock(&device_list_mutex);
 442                 return ERR_PTR(-ENOMEM);
 443         }
 444
 445         INIT_LIST_HEAD(&device->dev_node);
 446
 447         device->ib_device = cma_id->device;
 448         ret = isert_create_device_ib_res(device);
 449         if (ret) {
 450                 kfree(device);
 451                 mutex_unlock(&device_list_mutex);
 452                 return ERR_PTR(ret);
 453         }
 454
 455         device->refcount++;
 456         list_add_tail(&device->dev_node, &device_list);
 457         isert_info("Created a new iser device %p refcount %d\n",
 458                    device, device->refcount);
 459         mutex_unlock(&device_list_mutex);
 460
 461         return device;
 462 }
 463
 464 static void
 465 isert_conn_free_fastreg_pool(struct isert_conn *isert_conn)
 466 {
 467         struct fast_reg_descriptor *fr_desc, *tmp;
 468         int i = 0;
 469
 470         if (list_empty(&isert_conn->fr_pool))
 471                 return;
 472
 473         isert_info("Freeing conn %p fastreg pool", isert_conn);
 474
 475         list_for_each_entry_safe(fr_desc, tmp,
 476                                  &isert_conn->fr_pool, list) {
 477                 list_del(&fr_desc->list);
 478                 ib_free_fast_reg_page_list(fr_desc->data_frpl);
 479                 ib_dereg_mr(fr_desc->data_mr);
 480                 if (fr_desc->pi_ctx) {
 481                         ib_free_fast_reg_page_list(fr_desc->pi_ctx->prot_frpl);
 482                         ib_dereg_mr(fr_desc->pi_ctx->prot_mr);
 483                         ib_dereg_mr(fr_desc->pi_ctx->sig_mr);
 484                         kfree(fr_desc->pi_ctx);
 485                 }
 486                 kfree(fr_desc);
 487                 ++i;
 488         }
 489
 490         if (i < isert_conn->fr_pool_size)
 491                 isert_warn("Pool still has %d regions registered\n",
 492                         isert_conn->fr_pool_size - i);
 493 }
 494
 495 static int
 496 isert_create_pi_ctx(struct fast_reg_descriptor *desc,
 497                     struct ib_device *device,
 498                     struct ib_pd *pd)
 499 {
 500         struct pi_context *pi_ctx;
 501         int ret;
 502
 503         pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
 504         if (!pi_ctx) {
 505                 isert_err("Failed to allocate pi context\n");
 506                 return -ENOMEM;
 507         }
 508
 509         pi_ctx->prot_frpl = ib_alloc_fast_reg_page_list(device,
 510                                             ISCSI_ISER_SG_TABLESIZE);
 511         if (IS_ERR(pi_ctx->prot_frpl)) {
 512                 isert_err("Failed to allocate prot frpl err=%ld\n",
 513                           PTR_ERR(pi_ctx->prot_frpl));
 514                 ret = PTR_ERR(pi_ctx->prot_frpl);
 515                 goto err_pi_ctx;
 516         }
 517
 518         pi_ctx->prot_mr = ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG,
 519                                       ISCSI_ISER_SG_TABLESIZE);
 520         if (IS_ERR(pi_ctx->prot_mr)) {
 521                 isert_err("Failed to allocate prot frmr err=%ld\n",
 522                           PTR_ERR(pi_ctx->prot_mr));
 523                 ret = PTR_ERR(pi_ctx->prot_mr);
 524                 goto err_prot_frpl;
 525         }
 526         desc->ind |= ISERT_PROT_KEY_VALID;
 527
 528         pi_ctx->sig_mr = ib_alloc_mr(pd, IB_MR_TYPE_SIGNATURE, 2);
 529         if (IS_ERR(pi_ctx->sig_mr)) {
 530                 isert_err("Failed to allocate signature enabled mr err=%ld\n",
 531                           PTR_ERR(pi_ctx->sig_mr));
 532                 ret = PTR_ERR(pi_ctx->sig_mr);
 533                 goto err_prot_mr;
 534         }
 535
 536         desc->pi_ctx = pi_ctx;
 537         desc->ind |= ISERT_SIG_KEY_VALID;
 538         desc->ind &= ~ISERT_PROTECTED;
 539
 540         return 0;
 541
 542 err_prot_mr:
 543         ib_dereg_mr(pi_ctx->prot_mr);
 544 err_prot_frpl:
 545         ib_free_fast_reg_page_list(pi_ctx->prot_frpl);
 546 err_pi_ctx:
 547         kfree(pi_ctx);
 548
 549         return ret;
 550 }
 551
 552 static int
 553 isert_create_fr_desc(struct ib_device *ib_device, struct ib_pd *pd,
 554                      struct fast_reg_descriptor *fr_desc)
 555 {
 556         int ret;
 557
 558         fr_desc->data_frpl = ib_alloc_fast_reg_page_list(ib_device,
 559                                                          ISCSI_ISER_SG_TABLESIZE);
 560         if (IS_ERR(fr_desc->data_frpl)) {
 561                 isert_err("Failed to allocate data frpl err=%ld\n",
 562                           PTR_ERR(fr_desc->data_frpl));
 563                 return PTR_ERR(fr_desc->data_frpl);
 564         }
 565
 566         fr_desc->data_mr = ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG,
 567                                        ISCSI_ISER_SG_TABLESIZE);
 568         if (IS_ERR(fr_desc->data_mr)) {
 569                 isert_err("Failed to allocate data frmr err=%ld\n",
 570                           PTR_ERR(fr_desc->data_mr));
 571                 ret = PTR_ERR(fr_desc->data_mr);
 572                 goto err_data_frpl;
 573         }
 574         fr_desc->ind |= ISERT_DATA_KEY_VALID;
 575
 576         isert_dbg("Created fr_desc %p\n", fr_desc);
 577
 578         return 0;
 579
 580 err_data_frpl:
 581         ib_free_fast_reg_page_list(fr_desc->data_frpl);
 582
 583         return ret;
 584 }
 585
 586 static int
 587 isert_conn_create_fastreg_pool(struct isert_conn *isert_conn)
 588 {
 589         struct fast_reg_descriptor *fr_desc;
 590         struct isert_device *device = isert_conn->device;
 591         struct se_session *se_sess = isert_conn->conn->sess->se_sess;
 592         struct se_node_acl *se_nacl = se_sess->se_node_acl;
 593         int i, ret, tag_num;
 594         /*
 595          * Setup the number of FRMRs based upon the number of tags
 596          * available to session in iscsi_target_locate_portal().
 597          */
 598         tag_num = max_t(u32, ISCSIT_MIN_TAGS, se_nacl->queue_depth);
 599         tag_num = (tag_num * 2) + ISCSIT_EXTRA_TAGS;
 600
 601         isert_conn->fr_pool_size = 0;
 602         for (i = 0; i < tag_num; i++) {
 603                 fr_desc = kzalloc(sizeof(*fr_desc), GFP_KERNEL);
 604                 if (!fr_desc) {
 605                         isert_err("Failed to allocate fast_reg descriptor\n");
 606                         ret = -ENOMEM;
 607                         goto err;
 608                 }
 609
 610                 ret = isert_create_fr_desc(device->ib_device,
 611                                            device->pd, fr_desc);
 612                 if (ret) {
 613                         isert_err("Failed to create fastreg descriptor err=%d\n",
 614                                ret);
 615                         kfree(fr_desc);
 616                         goto err;
 617                 }
 618
 619                 list_add_tail(&fr_desc->list, &isert_conn->fr_pool);
 620                 isert_conn->fr_pool_size++;
 621         }
 622
 623         isert_dbg("Creating conn %p fastreg pool size=%d",
 624                  isert_conn, isert_conn->fr_pool_size);
 625
 626         return 0;
 627
 628 err:
 629         isert_conn_free_fastreg_pool(isert_conn);
 630         return ret;
 631 }
 632
 633 static void
 634 isert_init_conn(struct isert_conn *isert_conn)
 635 {
 636         isert_conn->state = ISER_CONN_INIT;
 637         INIT_LIST_HEAD(&isert_conn->accept_node);
 638         init_completion(&isert_conn->login_comp);
 639         init_completion(&isert_conn->login_req_comp);
 640         init_completion(&isert_conn->wait);
 641         kref_init(&isert_conn->kref);
 642         mutex_init(&isert_conn->mutex);
 643         spin_lock_init(&isert_conn->pool_lock);
 644         INIT_LIST_HEAD(&isert_conn->fr_pool);
 645         INIT_WORK(&isert_conn->release_work, isert_release_work);
 646 }
 647
 648 static void
 649 isert_free_login_buf(struct isert_conn *isert_conn)
 650 {
 651         struct ib_device *ib_dev = isert_conn->device->ib_device;
 652
 653         ib_dma_unmap_single(ib_dev, isert_conn->login_rsp_dma,
 654                             ISER_RX_LOGIN_SIZE, DMA_TO_DEVICE);
 655         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 656                             ISCSI_DEF_MAX_RECV_SEG_LEN,
 657                             DMA_FROM_DEVICE);
 658         kfree(isert_conn->login_buf);
 659 }
 660
 661 static int
 662 isert_alloc_login_buf(struct isert_conn *isert_conn,
 663                       struct ib_device *ib_dev)
 664 {
 665         int ret;
 666
 667         isert_conn->login_buf = kzalloc(ISCSI_DEF_MAX_RECV_SEG_LEN +
 668                                         ISER_RX_LOGIN_SIZE, GFP_KERNEL);
 669         if (!isert_conn->login_buf) {
 670                 isert_err("Unable to allocate isert_conn->login_buf\n");
 671                 return -ENOMEM;
 672         }
 673
 674         isert_conn->login_req_buf = isert_conn->login_buf;
 675         isert_conn->login_rsp_buf = isert_conn->login_buf +
 676                                     ISCSI_DEF_MAX_RECV_SEG_LEN;
 677
 678         isert_dbg("Set login_buf: %p login_req_buf: %p login_rsp_buf: %p\n",
 679                  isert_conn->login_buf, isert_conn->login_req_buf,
 680                  isert_conn->login_rsp_buf);
 681
 682         isert_conn->login_req_dma = ib_dma_map_single(ib_dev,
 683                                 (void *)isert_conn->login_req_buf,
 684                                 ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_FROM_DEVICE);
 685
 686         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_req_dma);
 687         if (ret) {
 688                 isert_err("login_req_dma mapping error: %d\n", ret);
 689                 isert_conn->login_req_dma = 0;
 690                 goto out_login_buf;
 691         }
 692
 693         isert_conn->login_rsp_dma = ib_dma_map_single(ib_dev,
 694                                         (void *)isert_conn->login_rsp_buf,
 695                                         ISER_RX_LOGIN_SIZE, DMA_TO_DEVICE);
 696
 697         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_rsp_dma);
 698         if (ret) {
 699                 isert_err("login_rsp_dma mapping error: %d\n", ret);
 700                 isert_conn->login_rsp_dma = 0;
 701                 goto out_req_dma_map;
 702         }
 703
 704         return 0;
 705
 706 out_req_dma_map:
 707         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 708                             ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_FROM_DEVICE);
 709 out_login_buf:
 710         kfree(isert_conn->login_buf);
 711         return ret;
 712 }
 713
 714 static int
 715 isert_connect_request(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 716 {
 717         struct isert_np *isert_np = cma_id->context;
 718         struct iscsi_np *np = isert_np->np;
 719         struct isert_conn *isert_conn;
 720         struct isert_device *device;
 721         int ret = 0;
 722
 723         spin_lock_bh(&np->np_thread_lock);
 724         if (!np->enabled) {
 725                 spin_unlock_bh(&np->np_thread_lock);
 726                 isert_dbg("iscsi_np is not enabled, reject connect request\n");
 727                 return rdma_reject(cma_id, NULL, 0);
 728         }
 729         spin_unlock_bh(&np->np_thread_lock);
 730
 731         isert_dbg("cma_id: %p, portal: %p\n",
 732                  cma_id, cma_id->context);
 733
 734         isert_conn = kzalloc(sizeof(struct isert_conn), GFP_KERNEL);
 735         if (!isert_conn)
 736                 return -ENOMEM;
 737
 738         isert_init_conn(isert_conn);
 739         isert_conn->cm_id = cma_id;
 740
 741         ret = isert_alloc_login_buf(isert_conn, cma_id->device);
 742         if (ret)
 743                 goto out;
 744
 745         device = isert_device_get(cma_id);
 746         if (IS_ERR(device)) {
 747                 ret = PTR_ERR(device);
 748                 goto out_rsp_dma_map;
 749         }
 750         isert_conn->device = device;
 751
 752         /* Set max inflight RDMA READ requests */
 753         isert_conn->initiator_depth = min_t(u8,
 754                                 event->param.conn.initiator_depth,
 755                                 device->dev_attr.max_qp_init_rd_atom);
 756         isert_dbg("Using initiator_depth: %u\n", isert_conn->initiator_depth);
 757
 758         ret = isert_conn_setup_qp(isert_conn, cma_id);
 759         if (ret)
 760                 goto out_conn_dev;
 761
 762         ret = isert_rdma_post_recvl(isert_conn);
 763         if (ret)
 764                 goto out_conn_dev;
 765         /*
 766          * Obtain the second reference now before isert_rdma_accept() to
 767          * ensure that any initiator generated REJECT CM event that occurs
 768          * asynchronously won't drop the last reference until the error path
 769          * in iscsi_target_login_sess_out() does it's ->iscsit_free_conn() ->
 770          * isert_free_conn() -> isert_put_conn() -> kref_put().
 771          */
 772         if (!kref_get_unless_zero(&isert_conn->kref)) {
 773                 isert_warn("conn %p connect_release is running\n", isert_conn);
 774                 goto out_conn_dev;
 775         }
 776
 777         ret = isert_rdma_accept(isert_conn);
 778         if (ret)
 779                 goto out_conn_dev;
 780
 781         mutex_lock(&isert_np->np_accept_mutex);
 782         list_add_tail(&isert_conn->accept_node, &isert_np->np_accept_list);
 783         mutex_unlock(&isert_np->np_accept_mutex);
 784
 785         isert_info("np %p: Allow accept_np to continue\n", np);
 786         up(&isert_np->np_sem);
 787         return 0;
 788
 789 out_conn_dev:
 790         isert_device_put(device);
 791 out_rsp_dma_map:
 792         isert_free_login_buf(isert_conn);
 793 out:
 794         kfree(isert_conn);
 795         rdma_reject(cma_id, NULL, 0);
 796         return ret;
 797 }
 798
 799 static void
 800 isert_connect_release(struct isert_conn *isert_conn)
 801 {
 802         struct isert_device *device = isert_conn->device;
 803
 804         isert_dbg("conn %p\n", isert_conn);
 805
 806         BUG_ON(!device);
 807
 808         if (device->use_fastreg)
 809                 isert_conn_free_fastreg_pool(isert_conn);
 810
 811         isert_free_rx_descriptors(isert_conn);
 812         if (isert_conn->cm_id)
 813                 rdma_destroy_id(isert_conn->cm_id);
 814
 815         if (isert_conn->qp) {
 816                 struct isert_comp *comp = isert_conn->qp->recv_cq->cq_context;
 817
 818                 isert_comp_put(comp);
 819                 ib_destroy_qp(isert_conn->qp);
 820         }
 821
 822         if (isert_conn->login_buf)
 823                 isert_free_login_buf(isert_conn);
 824
 825         isert_device_put(device);
 826
 827         kfree(isert_conn);
 828 }
 829
 830 static void
 831 isert_connected_handler(struct rdma_cm_id *cma_id)
 832 {
 833         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 834
 835         isert_info("conn %p\n", isert_conn);
 836
 837         mutex_lock(&isert_conn->mutex);
 838         if (isert_conn->state != ISER_CONN_FULL_FEATURE)
 839                 isert_conn->state = ISER_CONN_UP;
 840         mutex_unlock(&isert_conn->mutex);
 841 }
 842
 843 static void
 844 isert_release_kref(struct kref *kref)
 845 {
 846         struct isert_conn *isert_conn = container_of(kref,
 847                                 struct isert_conn, kref);
 848
 849         isert_info("conn %p final kref %s/%d\n", isert_conn, current->comm,
 850                    current->pid);
 851
 852         isert_connect_release(isert_conn);
 853 }
 854
 855 static void
 856 isert_put_conn(struct isert_conn *isert_conn)
 857 {
 858         kref_put(&isert_conn->kref, isert_release_kref);
 859 }
 860
 861 /**
 862  * isert_conn_terminate() - Initiate connection termination
 863  * @isert_conn: isert connection struct
 864  *
 865  * Notes:
 866  * In case the connection state is FULL_FEATURE, move state
 867  * to TEMINATING and start teardown sequence (rdma_disconnect).
 868  * In case the connection state is UP, complete flush as well.
 869  *
 870  * This routine must be called with mutex held. Thus it is
 871  * safe to call multiple times.
 872  */
 873 static void
 874 isert_conn_terminate(struct isert_conn *isert_conn)
 875 {
 876         int err;
 877
 878         switch (isert_conn->state) {
 879         case ISER_CONN_TERMINATING:
 880                 break;
 881         case ISER_CONN_UP:
 882         case ISER_CONN_FULL_FEATURE: /* FALLTHRU */
 883                 isert_info("Terminating conn %p state %d\n",
 884                            isert_conn, isert_conn->state);
 885                 isert_conn->state = ISER_CONN_TERMINATING;
 886                 err = rdma_disconnect(isert_conn->cm_id);
 887                 if (err)
 888                         isert_warn("Failed rdma_disconnect isert_conn %p\n",
 889                                    isert_conn);
 890                 break;
 891         default:
 892                 isert_warn("conn %p teminating in state %d\n",
 893                            isert_conn, isert_conn->state);
 894         }
 895 }
 896
 897 static int
 898 isert_np_cma_handler(struct isert_np *isert_np,
 899                      enum rdma_cm_event_type event)
 900 {
 901         isert_dbg("%s (%d): isert np %p\n",
 902                   rdma_event_msg(event), event, isert_np);
 903
 904         switch (event) {
 905         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 906                 isert_np->np_cm_id = NULL;
 907                 break;
 908         case RDMA_CM_EVENT_ADDR_CHANGE:
 909                 isert_np->np_cm_id = isert_setup_id(isert_np);
 910                 if (IS_ERR(isert_np->np_cm_id)) {
 911                         isert_err("isert np %p setup id failed: %ld\n",
 912                                   isert_np, PTR_ERR(isert_np->np_cm_id));
 913                         isert_np->np_cm_id = NULL;
 914                 }
 915                 break;
 916         default:
 917                 isert_err("isert np %p Unexpected event %d\n",
 918                           isert_np, event);
 919         }
 920
 921         return -1;
 922 }
 923
 924 static int
 925 isert_disconnected_handler(struct rdma_cm_id *cma_id,
 926                            enum rdma_cm_event_type event)
 927 {
 928         struct isert_np *isert_np = cma_id->context;
 929         struct isert_conn *isert_conn;
 930         bool terminating = false;
 931
 932         if (isert_np->np_cm_id == cma_id)
 933                 return isert_np_cma_handler(cma_id->context, event);
 934
 935         isert_conn = cma_id->qp->qp_context;
 936
 937         mutex_lock(&isert_conn->mutex);
 938         terminating = (isert_conn->state == ISER_CONN_TERMINATING);
 939         isert_conn_terminate(isert_conn);
 940         mutex_unlock(&isert_conn->mutex);
 941
 942         isert_info("conn %p completing wait\n", isert_conn);
 943         complete(&isert_conn->wait);
 944
 945         if (terminating)
 946                 goto out;
 947
 948         mutex_lock(&isert_np->np_accept_mutex);
 949         if (!list_empty(&isert_conn->accept_node)) {
 950                 list_del_init(&isert_conn->accept_node);
 951                 isert_put_conn(isert_conn);
 952                 queue_work(isert_release_wq, &isert_conn->release_work);
 953         }
 954         mutex_unlock(&isert_np->np_accept_mutex);
 955
 956 out:
 957         return 0;
 958 }
 959
 960 static int
 961 isert_connect_error(struct rdma_cm_id *cma_id)
 962 {
 963         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 964
 965         isert_conn->cm_id = NULL;
 966         isert_put_conn(isert_conn);
 967
 968         return -1;
 969 }
 970
 971 static int
 972 isert_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 973 {
 974         int ret = 0;
 975
 976         isert_info("%s (%d): status %d id %p np %p\n",
 977                    rdma_event_msg(event->event), 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->pd->local_dma_lkey) {
1083                 tx_desc->tx_sg[0].lkey = device->pd->local_dma_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->pd->local_dma_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->pd->local_dma_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->pd->local_dma_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(&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);
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);
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("%s (%d): wr id %llx vend_err %x\n",
2112                                   ib_wc_status_msg(wc->status), wc->status,
2113                                   wc->wr_id, wc->vendor_err);
2114                 else
2115                         isert_dbg("%s (%d): wr id %llx\n",
2116                                   ib_wc_status_msg(wc->status), wc->status,
2117                                   wc->wr_id);
2118
2119                 if (wc->wr_id != ISER_FASTREG_LI_WRID)
2120                         isert_cq_comp_err(isert_conn, wc);
2121         }
2122 }
2123
2124 static void
2125 isert_cq_work(struct work_struct *work)
2126 {
2127         enum { isert_poll_budget = 65536 };
2128         struct isert_comp *comp = container_of(work, struct isert_comp,
2129                                                work);
2130         struct ib_wc *const wcs = comp->wcs;
2131         int i, n, completed = 0;
2132
2133         while ((n = ib_poll_cq(comp->cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
2134                 for (i = 0; i < n; i++)
2135                         isert_handle_wc(&wcs[i]);
2136
2137                 completed += n;
2138                 if (completed >= isert_poll_budget)
2139                         break;
2140         }
2141
2142         ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP);
2143 }
2144
2145 static void
2146 isert_cq_callback(struct ib_cq *cq, void *context)
2147 {
2148         struct isert_comp *comp = context;
2149
2150         queue_work(isert_comp_wq, &comp->work);
2151 }
2152
2153 static int
2154 isert_post_response(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd)
2155 {
2156         struct ib_send_wr *wr_failed;
2157         int ret;
2158
2159         ret = ib_post_send(isert_conn->qp, &isert_cmd->tx_desc.send_wr,
2160                            &wr_failed);
2161         if (ret) {
2162                 isert_err("ib_post_send failed with %d\n", ret);
2163                 return ret;
2164         }
2165         return ret;
2166 }
2167
2168 static int
2169 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2170 {
2171         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2172         struct isert_conn *isert_conn = conn->context;
2173         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2174         struct iscsi_scsi_rsp *hdr = (struct iscsi_scsi_rsp *)
2175                                 &isert_cmd->tx_desc.iscsi_header;
2176
2177         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2178         iscsit_build_rsp_pdu(cmd, conn, true, hdr);
2179         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2180         /*
2181          * Attach SENSE DATA payload to iSCSI Response PDU
2182          */
2183         if (cmd->se_cmd.sense_buffer &&
2184             ((cmd->se_cmd.se_cmd_flags & SCF_TRANSPORT_TASK_SENSE) ||
2185             (cmd->se_cmd.se_cmd_flags & SCF_EMULATED_TASK_SENSE))) {
2186                 struct isert_device *device = isert_conn->device;
2187                 struct ib_device *ib_dev = device->ib_device;
2188                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2189                 u32 padding, pdu_len;
2190
2191                 put_unaligned_be16(cmd->se_cmd.scsi_sense_length,
2192                                    cmd->sense_buffer);
2193                 cmd->se_cmd.scsi_sense_length += sizeof(__be16);
2194
2195                 padding = -(cmd->se_cmd.scsi_sense_length) & 3;
2196                 hton24(hdr->dlength, (u32)cmd->se_cmd.scsi_sense_length);
2197                 pdu_len = cmd->se_cmd.scsi_sense_length + padding;
2198
2199                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2200                                 (void *)cmd->sense_buffer, pdu_len,
2201                                 DMA_TO_DEVICE);
2202
2203                 isert_cmd->pdu_buf_len = pdu_len;
2204                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2205                 tx_dsg->length  = pdu_len;
2206                 tx_dsg->lkey    = device->pd->local_dma_lkey;
2207                 isert_cmd->tx_desc.num_sge = 2;
2208         }
2209
2210         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2211
2212         isert_dbg("Posting SCSI Response\n");
2213
2214         return isert_post_response(isert_conn, isert_cmd);
2215 }
2216
2217 static void
2218 isert_aborted_task(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2219 {
2220         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2221         struct isert_conn *isert_conn = conn->context;
2222         struct isert_device *device = isert_conn->device;
2223
2224         spin_lock_bh(&conn->cmd_lock);
2225         if (!list_empty(&cmd->i_conn_node))
2226                 list_del_init(&cmd->i_conn_node);
2227         spin_unlock_bh(&conn->cmd_lock);
2228
2229         if (cmd->data_direction == DMA_TO_DEVICE)
2230                 iscsit_stop_dataout_timer(cmd);
2231
2232         device->unreg_rdma_mem(isert_cmd, isert_conn);
2233 }
2234
2235 static enum target_prot_op
2236 isert_get_sup_prot_ops(struct iscsi_conn *conn)
2237 {
2238         struct isert_conn *isert_conn = conn->context;
2239         struct isert_device *device = isert_conn->device;
2240
2241         if (conn->tpg->tpg_attrib.t10_pi) {
2242                 if (device->pi_capable) {
2243                         isert_info("conn %p PI offload enabled\n", isert_conn);
2244                         isert_conn->pi_support = true;
2245                         return TARGET_PROT_ALL;
2246                 }
2247         }
2248
2249         isert_info("conn %p PI offload disabled\n", isert_conn);
2250         isert_conn->pi_support = false;
2251
2252         return TARGET_PROT_NORMAL;
2253 }
2254
2255 static int
2256 isert_put_nopin(struct iscsi_cmd *cmd, struct iscsi_conn *conn,
2257                 bool nopout_response)
2258 {
2259         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2260         struct isert_conn *isert_conn = conn->context;
2261         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2262
2263         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2264         iscsit_build_nopin_rsp(cmd, conn, (struct iscsi_nopin *)
2265                                &isert_cmd->tx_desc.iscsi_header,
2266                                nopout_response);
2267         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2268         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2269
2270         isert_dbg("conn %p Posting NOPIN Response\n", isert_conn);
2271
2272         return isert_post_response(isert_conn, isert_cmd);
2273 }
2274
2275 static int
2276 isert_put_logout_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2277 {
2278         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2279         struct isert_conn *isert_conn = conn->context;
2280         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2281
2282         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2283         iscsit_build_logout_rsp(cmd, conn, (struct iscsi_logout_rsp *)
2284                                 &isert_cmd->tx_desc.iscsi_header);
2285         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2286         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2287
2288         isert_dbg("conn %p Posting Logout Response\n", isert_conn);
2289
2290         return isert_post_response(isert_conn, isert_cmd);
2291 }
2292
2293 static int
2294 isert_put_tm_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2295 {
2296         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2297         struct isert_conn *isert_conn = conn->context;
2298         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2299
2300         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2301         iscsit_build_task_mgt_rsp(cmd, conn, (struct iscsi_tm_rsp *)
2302                                   &isert_cmd->tx_desc.iscsi_header);
2303         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2304         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2305
2306         isert_dbg("conn %p Posting Task Management Response\n", isert_conn);
2307
2308         return isert_post_response(isert_conn, isert_cmd);
2309 }
2310
2311 static int
2312 isert_put_reject(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2313 {
2314         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2315         struct isert_conn *isert_conn = conn->context;
2316         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2317         struct isert_device *device = isert_conn->device;
2318         struct ib_device *ib_dev = device->ib_device;
2319         struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2320         struct iscsi_reject *hdr =
2321                 (struct iscsi_reject *)&isert_cmd->tx_desc.iscsi_header;
2322
2323         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2324         iscsit_build_reject(cmd, conn, hdr);
2325         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2326
2327         hton24(hdr->dlength, ISCSI_HDR_LEN);
2328         isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2329                         (void *)cmd->buf_ptr, ISCSI_HDR_LEN,
2330                         DMA_TO_DEVICE);
2331         isert_cmd->pdu_buf_len = ISCSI_HDR_LEN;
2332         tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2333         tx_dsg->length  = ISCSI_HDR_LEN;
2334         tx_dsg->lkey    = device->pd->local_dma_lkey;
2335         isert_cmd->tx_desc.num_sge = 2;
2336
2337         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2338
2339         isert_dbg("conn %p Posting Reject\n", isert_conn);
2340
2341         return isert_post_response(isert_conn, isert_cmd);
2342 }
2343
2344 static int
2345 isert_put_text_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2346 {
2347         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2348         struct isert_conn *isert_conn = conn->context;
2349         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2350         struct iscsi_text_rsp *hdr =
2351                 (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
2352         u32 txt_rsp_len;
2353         int rc;
2354
2355         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2356         rc = iscsit_build_text_rsp(cmd, conn, hdr, ISCSI_INFINIBAND);
2357         if (rc < 0)
2358                 return rc;
2359
2360         txt_rsp_len = rc;
2361         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2362
2363         if (txt_rsp_len) {
2364                 struct isert_device *device = isert_conn->device;
2365                 struct ib_device *ib_dev = device->ib_device;
2366                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2367                 void *txt_rsp_buf = cmd->buf_ptr;
2368
2369                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2370                                 txt_rsp_buf, txt_rsp_len, DMA_TO_DEVICE);
2371
2372                 isert_cmd->pdu_buf_len = txt_rsp_len;
2373                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2374                 tx_dsg->length  = txt_rsp_len;
2375                 tx_dsg->lkey    = device->pd->local_dma_lkey;
2376                 isert_cmd->tx_desc.num_sge = 2;
2377         }
2378         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2379
2380         isert_dbg("conn %p Text Response\n", isert_conn);
2381
2382         return isert_post_response(isert_conn, isert_cmd);
2383 }
2384
2385 static int
2386 isert_build_rdma_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
2387                     struct ib_sge *ib_sge, struct ib_send_wr *send_wr,
2388                     u32 data_left, u32 offset)
2389 {
2390         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
2391         struct scatterlist *sg_start, *tmp_sg;
2392         struct isert_device *device = isert_conn->device;
2393         struct ib_device *ib_dev = device->ib_device;
2394         u32 sg_off, page_off;
2395         int i = 0, sg_nents;
2396
2397         sg_off = offset / PAGE_SIZE;
2398         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
2399         sg_nents = min(cmd->se_cmd.t_data_nents - sg_off, isert_conn->max_sge);
2400         page_off = offset % PAGE_SIZE;
2401
2402         send_wr->sg_list = ib_sge;
2403         send_wr->wr_id = (uintptr_t)&isert_cmd->tx_desc;
2404         /*
2405          * Perform mapping of TCM scatterlist memory ib_sge dma_addr.
2406          */
2407         for_each_sg(sg_start, tmp_sg, sg_nents, i) {
2408                 isert_dbg("RDMA from SGL dma_addr: 0x%llx dma_len: %u, "
2409                           "page_off: %u\n",
2410                           (unsigned long long)tmp_sg->dma_address,
2411                           tmp_sg->length, page_off);
2412
2413                 ib_sge->addr = ib_sg_dma_address(ib_dev, tmp_sg) + page_off;
2414                 ib_sge->length = min_t(u32, data_left,
2415                                 ib_sg_dma_len(ib_dev, tmp_sg) - page_off);
2416                 ib_sge->lkey = device->pd->local_dma_lkey;
2417
2418                 isert_dbg("RDMA ib_sge: addr: 0x%llx  length: %u lkey: %x\n",
2419                           ib_sge->addr, ib_sge->length, ib_sge->lkey);
2420                 page_off = 0;
2421                 data_left -= ib_sge->length;
2422                 if (!data_left)
2423                         break;
2424                 ib_sge++;
2425                 isert_dbg("Incrementing ib_sge pointer to %p\n", ib_sge);
2426         }
2427
2428         send_wr->num_sge = ++i;
2429         isert_dbg("Set outgoing sg_list: %p num_sg: %u from TCM SGLs\n",
2430                   send_wr->sg_list, send_wr->num_sge);
2431
2432         return send_wr->num_sge;
2433 }
2434
2435 static int
2436 isert_map_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2437                struct isert_rdma_wr *wr)
2438 {
2439         struct se_cmd *se_cmd = &cmd->se_cmd;
2440         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2441         struct isert_conn *isert_conn = conn->context;
2442         struct isert_data_buf *data = &wr->data;
2443         struct ib_send_wr *send_wr;
2444         struct ib_sge *ib_sge;
2445         u32 offset, data_len, data_left, rdma_write_max, va_offset = 0;
2446         int ret = 0, i, ib_sge_cnt;
2447
2448         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2449
2450         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2451         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2452                                  se_cmd->t_data_nents, se_cmd->data_length,
2453                                  offset, wr->iser_ib_op, &wr->data);
2454         if (ret)
2455                 return ret;
2456
2457         data_left = data->len;
2458         offset = data->offset;
2459
2460         ib_sge = kzalloc(sizeof(struct ib_sge) * data->nents, GFP_KERNEL);
2461         if (!ib_sge) {
2462                 isert_warn("Unable to allocate ib_sge\n");
2463                 ret = -ENOMEM;
2464                 goto unmap_cmd;
2465         }
2466         wr->ib_sge = ib_sge;
2467
2468         wr->send_wr_num = DIV_ROUND_UP(data->nents, isert_conn->max_sge);
2469         wr->send_wr = kzalloc(sizeof(struct ib_send_wr) * wr->send_wr_num,
2470                                 GFP_KERNEL);
2471         if (!wr->send_wr) {
2472                 isert_dbg("Unable to allocate wr->send_wr\n");
2473                 ret = -ENOMEM;
2474                 goto unmap_cmd;
2475         }
2476
2477         wr->isert_cmd = isert_cmd;
2478         rdma_write_max = isert_conn->max_sge * PAGE_SIZE;
2479
2480         for (i = 0; i < wr->send_wr_num; i++) {
2481                 send_wr = &isert_cmd->rdma_wr.send_wr[i];
2482                 data_len = min(data_left, rdma_write_max);
2483
2484                 send_wr->send_flags = 0;
2485                 if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2486                         send_wr->opcode = IB_WR_RDMA_WRITE;
2487                         send_wr->wr.rdma.remote_addr = isert_cmd->read_va + offset;
2488                         send_wr->wr.rdma.rkey = isert_cmd->read_stag;
2489                         if (i + 1 == wr->send_wr_num)
2490                                 send_wr->next = &isert_cmd->tx_desc.send_wr;
2491                         else
2492                                 send_wr->next = &wr->send_wr[i + 1];
2493                 } else {
2494                         send_wr->opcode = IB_WR_RDMA_READ;
2495                         send_wr->wr.rdma.remote_addr = isert_cmd->write_va + va_offset;
2496                         send_wr->wr.rdma.rkey = isert_cmd->write_stag;
2497                         if (i + 1 == wr->send_wr_num)
2498                                 send_wr->send_flags = IB_SEND_SIGNALED;
2499                         else
2500                                 send_wr->next = &wr->send_wr[i + 1];
2501                 }
2502
2503                 ib_sge_cnt = isert_build_rdma_wr(isert_conn, isert_cmd, ib_sge,
2504                                         send_wr, data_len, offset);
2505                 ib_sge += ib_sge_cnt;
2506
2507                 offset += data_len;
2508                 va_offset += data_len;
2509                 data_left -= data_len;
2510         }
2511
2512         return 0;
2513 unmap_cmd:
2514         isert_unmap_data_buf(isert_conn, data);
2515
2516         return ret;
2517 }
2518
2519 static int
2520 isert_map_fr_pagelist(struct ib_device *ib_dev,
2521                       struct scatterlist *sg_start, int sg_nents, u64 *fr_pl)
2522 {
2523         u64 start_addr, end_addr, page, chunk_start = 0;
2524         struct scatterlist *tmp_sg;
2525         int i = 0, new_chunk, last_ent, n_pages;
2526
2527         n_pages = 0;
2528         new_chunk = 1;
2529         last_ent = sg_nents - 1;
2530         for_each_sg(sg_start, tmp_sg, sg_nents, i) {
2531                 start_addr = ib_sg_dma_address(ib_dev, tmp_sg);
2532                 if (new_chunk)
2533                         chunk_start = start_addr;
2534                 end_addr = start_addr + ib_sg_dma_len(ib_dev, tmp_sg);
2535
2536                 isert_dbg("SGL[%d] dma_addr: 0x%llx len: %u\n",
2537                           i, (unsigned long long)tmp_sg->dma_address,
2538                           tmp_sg->length);
2539
2540                 if ((end_addr & ~PAGE_MASK) && i < last_ent) {
2541                         new_chunk = 0;
2542                         continue;
2543                 }
2544                 new_chunk = 1;
2545
2546                 page = chunk_start & PAGE_MASK;
2547                 do {
2548                         fr_pl[n_pages++] = page;
2549                         isert_dbg("Mapped page_list[%d] page_addr: 0x%llx\n",
2550                                   n_pages - 1, page);
2551                         page += PAGE_SIZE;
2552                 } while (page < end_addr);
2553         }
2554
2555         return n_pages;
2556 }
2557
2558 static inline void
2559 isert_inv_rkey(struct ib_send_wr *inv_wr, struct ib_mr *mr)
2560 {
2561         u32 rkey;
2562
2563         memset(inv_wr, 0, sizeof(*inv_wr));
2564         inv_wr->wr_id = ISER_FASTREG_LI_WRID;
2565         inv_wr->opcode = IB_WR_LOCAL_INV;
2566         inv_wr->ex.invalidate_rkey = mr->rkey;
2567
2568         /* Bump the key */
2569         rkey = ib_inc_rkey(mr->rkey);
2570         ib_update_fast_reg_key(mr, rkey);
2571 }
2572
2573 static int
2574 isert_fast_reg_mr(struct isert_conn *isert_conn,
2575                   struct fast_reg_descriptor *fr_desc,
2576                   struct isert_data_buf *mem,
2577                   enum isert_indicator ind,
2578                   struct ib_sge *sge)
2579 {
2580         struct isert_device *device = isert_conn->device;
2581         struct ib_device *ib_dev = device->ib_device;
2582         struct ib_mr *mr;
2583         struct ib_fast_reg_page_list *frpl;
2584         struct ib_send_wr fr_wr, inv_wr;
2585         struct ib_send_wr *bad_wr, *wr = NULL;
2586         int ret, pagelist_len;
2587         u32 page_off;
2588
2589         if (mem->dma_nents == 1) {
2590                 sge->lkey = device->pd->local_dma_lkey;
2591                 sge->addr = ib_sg_dma_address(ib_dev, &mem->sg[0]);
2592                 sge->length = ib_sg_dma_len(ib_dev, &mem->sg[0]);
2593                 isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2594                          sge->addr, sge->length, sge->lkey);
2595                 return 0;
2596         }
2597
2598         if (ind == ISERT_DATA_KEY_VALID) {
2599                 /* Registering data buffer */
2600                 mr = fr_desc->data_mr;
2601                 frpl = fr_desc->data_frpl;
2602         } else {
2603                 /* Registering protection buffer */
2604                 mr = fr_desc->pi_ctx->prot_mr;
2605                 frpl = fr_desc->pi_ctx->prot_frpl;
2606         }
2607
2608         page_off = mem->offset % PAGE_SIZE;
2609
2610         isert_dbg("Use fr_desc %p sg_nents %d offset %u\n",
2611                   fr_desc, mem->nents, mem->offset);
2612
2613         pagelist_len = isert_map_fr_pagelist(ib_dev, mem->sg, mem->nents,
2614                                              &frpl->page_list[0]);
2615
2616         if (!(fr_desc->ind & ind)) {
2617                 isert_inv_rkey(&inv_wr, mr);
2618                 wr = &inv_wr;
2619         }
2620
2621         /* Prepare FASTREG WR */
2622         memset(&fr_wr, 0, sizeof(fr_wr));
2623         fr_wr.wr_id = ISER_FASTREG_LI_WRID;
2624         fr_wr.opcode = IB_WR_FAST_REG_MR;
2625         fr_wr.wr.fast_reg.iova_start = frpl->page_list[0] + page_off;
2626         fr_wr.wr.fast_reg.page_list = frpl;
2627         fr_wr.wr.fast_reg.page_list_len = pagelist_len;
2628         fr_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
2629         fr_wr.wr.fast_reg.length = mem->len;
2630         fr_wr.wr.fast_reg.rkey = mr->rkey;
2631         fr_wr.wr.fast_reg.access_flags = IB_ACCESS_LOCAL_WRITE;
2632
2633         if (!wr)
2634                 wr = &fr_wr;
2635         else
2636                 wr->next = &fr_wr;
2637
2638         ret = ib_post_send(isert_conn->qp, wr, &bad_wr);
2639         if (ret) {
2640                 isert_err("fast registration failed, ret:%d\n", ret);
2641                 return ret;
2642         }
2643         fr_desc->ind &= ~ind;
2644
2645         sge->lkey = mr->lkey;
2646         sge->addr = frpl->page_list[0] + page_off;
2647         sge->length = mem->len;
2648
2649         isert_dbg("sge: addr: 0x%llx  length: %u lkey: %x\n",
2650                   sge->addr, sge->length, sge->lkey);
2651
2652         return ret;
2653 }
2654
2655 static inline void
2656 isert_set_dif_domain(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs,
2657                      struct ib_sig_domain *domain)
2658 {
2659         domain->sig_type = IB_SIG_TYPE_T10_DIF;
2660         domain->sig.dif.bg_type = IB_T10DIF_CRC;
2661         domain->sig.dif.pi_interval = se_cmd->se_dev->dev_attrib.block_size;
2662         domain->sig.dif.ref_tag = se_cmd->reftag_seed;
2663         /*
2664          * At the moment we hard code those, but if in the future
2665          * the target core would like to use it, we will take it
2666          * from se_cmd.
2667          */
2668         domain->sig.dif.apptag_check_mask = 0xffff;
2669         domain->sig.dif.app_escape = true;
2670         domain->sig.dif.ref_escape = true;
2671         if (se_cmd->prot_type == TARGET_DIF_TYPE1_PROT ||
2672             se_cmd->prot_type == TARGET_DIF_TYPE2_PROT)
2673                 domain->sig.dif.ref_remap = true;
2674 };
2675
2676 static int
2677 isert_set_sig_attrs(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs)
2678 {
2679         switch (se_cmd->prot_op) {
2680         case TARGET_PROT_DIN_INSERT:
2681         case TARGET_PROT_DOUT_STRIP:
2682                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
2683                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2684                 break;
2685         case TARGET_PROT_DOUT_INSERT:
2686         case TARGET_PROT_DIN_STRIP:
2687                 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
2688                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2689                 break;
2690         case TARGET_PROT_DIN_PASS:
2691         case TARGET_PROT_DOUT_PASS:
2692                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2693                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2694                 break;
2695         default:
2696                 isert_err("Unsupported PI operation %d\n", se_cmd->prot_op);
2697                 return -EINVAL;
2698         }
2699
2700         return 0;
2701 }
2702
2703 static inline u8
2704 isert_set_prot_checks(u8 prot_checks)
2705 {
2706         return (prot_checks & TARGET_DIF_CHECK_GUARD  ? 0xc0 : 0) |
2707                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x30 : 0) |
2708                (prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x0f : 0);
2709 }
2710
2711 static int
2712 isert_reg_sig_mr(struct isert_conn *isert_conn,
2713                  struct se_cmd *se_cmd,
2714                  struct isert_rdma_wr *rdma_wr,
2715                  struct fast_reg_descriptor *fr_desc)
2716 {
2717         struct ib_send_wr sig_wr, inv_wr;
2718         struct ib_send_wr *bad_wr, *wr = NULL;
2719         struct pi_context *pi_ctx = fr_desc->pi_ctx;
2720         struct ib_sig_attrs sig_attrs;
2721         int ret;
2722
2723         memset(&sig_attrs, 0, sizeof(sig_attrs));
2724         ret = isert_set_sig_attrs(se_cmd, &sig_attrs);
2725         if (ret)
2726                 goto err;
2727
2728         sig_attrs.check_mask = isert_set_prot_checks(se_cmd->prot_checks);
2729
2730         if (!(fr_desc->ind & ISERT_SIG_KEY_VALID)) {
2731                 isert_inv_rkey(&inv_wr, pi_ctx->sig_mr);
2732                 wr = &inv_wr;
2733         }
2734
2735         memset(&sig_wr, 0, sizeof(sig_wr));
2736         sig_wr.opcode = IB_WR_REG_SIG_MR;
2737         sig_wr.wr_id = ISER_FASTREG_LI_WRID;
2738         sig_wr.sg_list = &rdma_wr->ib_sg[DATA];
2739         sig_wr.num_sge = 1;
2740         sig_wr.wr.sig_handover.access_flags = IB_ACCESS_LOCAL_WRITE;
2741         sig_wr.wr.sig_handover.sig_attrs = &sig_attrs;
2742         sig_wr.wr.sig_handover.sig_mr = pi_ctx->sig_mr;
2743         if (se_cmd->t_prot_sg)
2744                 sig_wr.wr.sig_handover.prot = &rdma_wr->ib_sg[PROT];
2745
2746         if (!wr)
2747                 wr = &sig_wr;
2748         else
2749                 wr->next = &sig_wr;
2750
2751         ret = ib_post_send(isert_conn->qp, wr, &bad_wr);
2752         if (ret) {
2753                 isert_err("fast registration failed, ret:%d\n", ret);
2754                 goto err;
2755         }
2756         fr_desc->ind &= ~ISERT_SIG_KEY_VALID;
2757
2758         rdma_wr->ib_sg[SIG].lkey = pi_ctx->sig_mr->lkey;
2759         rdma_wr->ib_sg[SIG].addr = 0;
2760         rdma_wr->ib_sg[SIG].length = se_cmd->data_length;
2761         if (se_cmd->prot_op != TARGET_PROT_DIN_STRIP &&
2762             se_cmd->prot_op != TARGET_PROT_DOUT_INSERT)
2763                 /*
2764                  * We have protection guards on the wire
2765                  * so we need to set a larget transfer
2766                  */
2767                 rdma_wr->ib_sg[SIG].length += se_cmd->prot_length;
2768
2769         isert_dbg("sig_sge: addr: 0x%llx  length: %u lkey: %x\n",
2770                   rdma_wr->ib_sg[SIG].addr, rdma_wr->ib_sg[SIG].length,
2771                   rdma_wr->ib_sg[SIG].lkey);
2772 err:
2773         return ret;
2774 }
2775
2776 static int
2777 isert_handle_prot_cmd(struct isert_conn *isert_conn,
2778                       struct isert_cmd *isert_cmd,
2779                       struct isert_rdma_wr *wr)
2780 {
2781         struct isert_device *device = isert_conn->device;
2782         struct se_cmd *se_cmd = &isert_cmd->iscsi_cmd->se_cmd;
2783         int ret;
2784
2785         if (!wr->fr_desc->pi_ctx) {
2786                 ret = isert_create_pi_ctx(wr->fr_desc,
2787                                           device->ib_device,
2788                                           device->pd);
2789                 if (ret) {
2790                         isert_err("conn %p failed to allocate pi_ctx\n",
2791                                   isert_conn);
2792                         return ret;
2793                 }
2794         }
2795
2796         if (se_cmd->t_prot_sg) {
2797                 ret = isert_map_data_buf(isert_conn, isert_cmd,
2798                                          se_cmd->t_prot_sg,
2799                                          se_cmd->t_prot_nents,
2800                                          se_cmd->prot_length,
2801                                          0, wr->iser_ib_op, &wr->prot);
2802                 if (ret) {
2803                         isert_err("conn %p failed to map protection buffer\n",
2804                                   isert_conn);
2805                         return ret;
2806                 }
2807
2808                 memset(&wr->ib_sg[PROT], 0, sizeof(wr->ib_sg[PROT]));
2809                 ret = isert_fast_reg_mr(isert_conn, wr->fr_desc, &wr->prot,
2810                                         ISERT_PROT_KEY_VALID, &wr->ib_sg[PROT]);
2811                 if (ret) {
2812                         isert_err("conn %p failed to fast reg mr\n",
2813                                   isert_conn);
2814                         goto unmap_prot_cmd;
2815                 }
2816         }
2817
2818         ret = isert_reg_sig_mr(isert_conn, se_cmd, wr, wr->fr_desc);
2819         if (ret) {
2820                 isert_err("conn %p failed to fast reg mr\n",
2821                           isert_conn);
2822                 goto unmap_prot_cmd;
2823         }
2824         wr->fr_desc->ind |= ISERT_PROTECTED;
2825
2826         return 0;
2827
2828 unmap_prot_cmd:
2829         if (se_cmd->t_prot_sg)
2830                 isert_unmap_data_buf(isert_conn, &wr->prot);
2831
2832         return ret;
2833 }
2834
2835 static int
2836 isert_reg_rdma(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
2837                struct isert_rdma_wr *wr)
2838 {
2839         struct se_cmd *se_cmd = &cmd->se_cmd;
2840         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2841         struct isert_conn *isert_conn = conn->context;
2842         struct fast_reg_descriptor *fr_desc = NULL;
2843         struct ib_send_wr *send_wr;
2844         struct ib_sge *ib_sg;
2845         u32 offset;
2846         int ret = 0;
2847         unsigned long flags;
2848
2849         isert_cmd->tx_desc.isert_cmd = isert_cmd;
2850
2851         offset = wr->iser_ib_op == ISER_IB_RDMA_READ ? cmd->write_data_done : 0;
2852         ret = isert_map_data_buf(isert_conn, isert_cmd, se_cmd->t_data_sg,
2853                                  se_cmd->t_data_nents, se_cmd->data_length,
2854                                  offset, wr->iser_ib_op, &wr->data);
2855         if (ret)
2856                 return ret;
2857
2858         if (wr->data.dma_nents != 1 || isert_prot_cmd(isert_conn, se_cmd)) {
2859                 spin_lock_irqsave(&isert_conn->pool_lock, flags);
2860                 fr_desc = list_first_entry(&isert_conn->fr_pool,
2861                                            struct fast_reg_descriptor, list);
2862                 list_del(&fr_desc->list);
2863                 spin_unlock_irqrestore(&isert_conn->pool_lock, flags);
2864                 wr->fr_desc = fr_desc;
2865         }
2866
2867         ret = isert_fast_reg_mr(isert_conn, fr_desc, &wr->data,
2868                                 ISERT_DATA_KEY_VALID, &wr->ib_sg[DATA]);
2869         if (ret)
2870                 goto unmap_cmd;
2871
2872         if (isert_prot_cmd(isert_conn, se_cmd)) {
2873                 ret = isert_handle_prot_cmd(isert_conn, isert_cmd, wr);
2874                 if (ret)
2875                         goto unmap_cmd;
2876
2877                 ib_sg = &wr->ib_sg[SIG];
2878         } else {
2879                 ib_sg = &wr->ib_sg[DATA];
2880         }
2881
2882         memcpy(&wr->s_ib_sge, ib_sg, sizeof(*ib_sg));
2883         wr->ib_sge = &wr->s_ib_sge;
2884         wr->send_wr_num = 1;
2885         memset(&wr->s_send_wr, 0, sizeof(*send_wr));
2886         wr->send_wr = &wr->s_send_wr;
2887         wr->isert_cmd = isert_cmd;
2888
2889         send_wr = &isert_cmd->rdma_wr.s_send_wr;
2890         send_wr->sg_list = &wr->s_ib_sge;
2891         send_wr->num_sge = 1;
2892         send_wr->wr_id = (uintptr_t)&isert_cmd->tx_desc;
2893         if (wr->iser_ib_op == ISER_IB_RDMA_WRITE) {
2894                 send_wr->opcode = IB_WR_RDMA_WRITE;
2895                 send_wr->wr.rdma.remote_addr = isert_cmd->read_va;
2896                 send_wr->wr.rdma.rkey = isert_cmd->read_stag;
2897                 send_wr->send_flags = !isert_prot_cmd(isert_conn, se_cmd) ?
2898                                       0 : IB_SEND_SIGNALED;
2899         } else {
2900                 send_wr->opcode = IB_WR_RDMA_READ;
2901                 send_wr->wr.rdma.remote_addr = isert_cmd->write_va;
2902                 send_wr->wr.rdma.rkey = isert_cmd->write_stag;
2903                 send_wr->send_flags = IB_SEND_SIGNALED;
2904         }
2905
2906         return 0;
2907
2908 unmap_cmd:
2909         if (fr_desc) {
2910                 spin_lock_irqsave(&isert_conn->pool_lock, flags);
2911                 list_add_tail(&fr_desc->list, &isert_conn->fr_pool);
2912                 spin_unlock_irqrestore(&isert_conn->pool_lock, flags);
2913         }
2914         isert_unmap_data_buf(isert_conn, &wr->data);
2915
2916         return ret;
2917 }
2918
2919 static int
2920 isert_put_datain(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2921 {
2922         struct se_cmd *se_cmd = &cmd->se_cmd;
2923         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2924         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2925         struct isert_conn *isert_conn = conn->context;
2926         struct isert_device *device = isert_conn->device;
2927         struct ib_send_wr *wr_failed;
2928         int rc;
2929
2930         isert_dbg("Cmd: %p RDMA_WRITE data_length: %u\n",
2931                  isert_cmd, se_cmd->data_length);
2932
2933         wr->iser_ib_op = ISER_IB_RDMA_WRITE;
2934         rc = device->reg_rdma_mem(conn, cmd, wr);
2935         if (rc) {
2936                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2937                 return rc;
2938         }
2939
2940         if (!isert_prot_cmd(isert_conn, se_cmd)) {
2941                 /*
2942                  * Build isert_conn->tx_desc for iSCSI response PDU and attach
2943                  */
2944                 isert_create_send_desc(isert_conn, isert_cmd,
2945                                        &isert_cmd->tx_desc);
2946                 iscsit_build_rsp_pdu(cmd, conn, true, (struct iscsi_scsi_rsp *)
2947                                      &isert_cmd->tx_desc.iscsi_header);
2948                 isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2949                 isert_init_send_wr(isert_conn, isert_cmd,
2950                                    &isert_cmd->tx_desc.send_wr);
2951                 isert_cmd->rdma_wr.s_send_wr.next = &isert_cmd->tx_desc.send_wr;
2952                 wr->send_wr_num += 1;
2953         }
2954
2955         rc = ib_post_send(isert_conn->qp, wr->send_wr, &wr_failed);
2956         if (rc)
2957                 isert_warn("ib_post_send() failed for IB_WR_RDMA_WRITE\n");
2958
2959         if (!isert_prot_cmd(isert_conn, se_cmd))
2960                 isert_dbg("Cmd: %p posted RDMA_WRITE + Response for iSER Data "
2961                          "READ\n", isert_cmd);
2962         else
2963                 isert_dbg("Cmd: %p posted RDMA_WRITE for iSER Data READ\n",
2964                          isert_cmd);
2965
2966         return 1;
2967 }
2968
2969 static int
2970 isert_get_dataout(struct iscsi_conn *conn, struct iscsi_cmd *cmd, bool recovery)
2971 {
2972         struct se_cmd *se_cmd = &cmd->se_cmd;
2973         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2974         struct isert_rdma_wr *wr = &isert_cmd->rdma_wr;
2975         struct isert_conn *isert_conn = conn->context;
2976         struct isert_device *device = isert_conn->device;
2977         struct ib_send_wr *wr_failed;
2978         int rc;
2979
2980         isert_dbg("Cmd: %p RDMA_READ data_length: %u write_data_done: %u\n",
2981                  isert_cmd, se_cmd->data_length, cmd->write_data_done);
2982         wr->iser_ib_op = ISER_IB_RDMA_READ;
2983         rc = device->reg_rdma_mem(conn, cmd, wr);
2984         if (rc) {
2985                 isert_err("Cmd: %p failed to prepare RDMA res\n", isert_cmd);
2986                 return rc;
2987         }
2988
2989         rc = ib_post_send(isert_conn->qp, wr->send_wr, &wr_failed);
2990         if (rc)
2991                 isert_warn("ib_post_send() failed for IB_WR_RDMA_READ\n");
2992
2993         isert_dbg("Cmd: %p posted RDMA_READ memory for ISER Data WRITE\n",
2994                  isert_cmd);
2995
2996         return 0;
2997 }
2998
2999 static int
3000 isert_immediate_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
3001 {
3002         int ret;
3003
3004         switch (state) {
3005         case ISTATE_SEND_NOPIN_WANT_RESPONSE:
3006                 ret = isert_put_nopin(cmd, conn, false);
3007                 break;
3008         default:
3009                 isert_err("Unknown immediate state: 0x%02x\n", state);
3010                 ret = -EINVAL;
3011                 break;
3012         }
3013
3014         return ret;
3015 }
3016
3017 static int
3018 isert_response_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
3019 {
3020         struct isert_conn *isert_conn = conn->context;
3021         int ret;
3022
3023         switch (state) {
3024         case ISTATE_SEND_LOGOUTRSP:
3025                 ret = isert_put_logout_rsp(cmd, conn);
3026                 if (!ret)
3027                         isert_conn->logout_posted = true;
3028                 break;
3029         case ISTATE_SEND_NOPIN:
3030                 ret = isert_put_nopin(cmd, conn, true);
3031                 break;
3032         case ISTATE_SEND_TASKMGTRSP:
3033                 ret = isert_put_tm_rsp(cmd, conn);
3034                 break;
3035         case ISTATE_SEND_REJECT:
3036                 ret = isert_put_reject(cmd, conn);
3037                 break;
3038         case ISTATE_SEND_TEXTRSP:
3039                 ret = isert_put_text_rsp(cmd, conn);
3040                 break;
3041         case ISTATE_SEND_STATUS:
3042                 /*
3043                  * Special case for sending non GOOD SCSI status from TX thread
3044                  * context during pre se_cmd excecution failure.
3045                  */
3046                 ret = isert_put_response(conn, cmd);
3047                 break;
3048         default:
3049                 isert_err("Unknown response state: 0x%02x\n", state);
3050                 ret = -EINVAL;
3051                 break;
3052         }
3053
3054         return ret;
3055 }
3056
3057 struct rdma_cm_id *
3058 isert_setup_id(struct isert_np *isert_np)
3059 {
3060         struct iscsi_np *np = isert_np->np;
3061         struct rdma_cm_id *id;
3062         struct sockaddr *sa;
3063         int ret;
3064
3065         sa = (struct sockaddr *)&np->np_sockaddr;
3066         isert_dbg("ksockaddr: %p, sa: %p\n", &np->np_sockaddr, sa);
3067
3068         id = rdma_create_id(isert_cma_handler, isert_np,
3069                             RDMA_PS_TCP, IB_QPT_RC);
3070         if (IS_ERR(id)) {
3071                 isert_err("rdma_create_id() failed: %ld\n", PTR_ERR(id));
3072                 ret = PTR_ERR(id);
3073                 goto out;
3074         }
3075         isert_dbg("id %p context %p\n", id, id->context);
3076
3077         ret = rdma_bind_addr(id, sa);
3078         if (ret) {
3079                 isert_err("rdma_bind_addr() failed: %d\n", ret);
3080                 goto out_id;
3081         }
3082
3083         ret = rdma_listen(id, 0);
3084         if (ret) {
3085                 isert_err("rdma_listen() failed: %d\n", ret);
3086                 goto out_id;
3087         }
3088
3089         return id;
3090 out_id:
3091         rdma_destroy_id(id);
3092 out:
3093         return ERR_PTR(ret);
3094 }
3095
3096 static int
3097 isert_setup_np(struct iscsi_np *np,
3098                struct sockaddr_storage *ksockaddr)
3099 {
3100         struct isert_np *isert_np;
3101         struct rdma_cm_id *isert_lid;
3102         int ret;
3103
3104         isert_np = kzalloc(sizeof(struct isert_np), GFP_KERNEL);
3105         if (!isert_np) {
3106                 isert_err("Unable to allocate struct isert_np\n");
3107                 return -ENOMEM;
3108         }
3109         sema_init(&isert_np->np_sem, 0);
3110         mutex_init(&isert_np->np_accept_mutex);
3111         INIT_LIST_HEAD(&isert_np->np_accept_list);
3112         init_completion(&isert_np->np_login_comp);
3113         isert_np->np = np;
3114
3115         /*
3116          * Setup the np->np_sockaddr from the passed sockaddr setup
3117          * in iscsi_target_configfs.c code..
3118          */
3119         memcpy(&np->np_sockaddr, ksockaddr,
3120                sizeof(struct sockaddr_storage));
3121
3122         isert_lid = isert_setup_id(isert_np);
3123         if (IS_ERR(isert_lid)) {
3124                 ret = PTR_ERR(isert_lid);
3125                 goto out;
3126         }
3127
3128         isert_np->np_cm_id = isert_lid;
3129         np->np_context = isert_np;
3130
3131         return 0;
3132
3133 out:
3134         kfree(isert_np);
3135
3136         return ret;
3137 }
3138
3139 static int
3140 isert_rdma_accept(struct isert_conn *isert_conn)
3141 {
3142         struct rdma_cm_id *cm_id = isert_conn->cm_id;
3143         struct rdma_conn_param cp;
3144         int ret;
3145
3146         memset(&cp, 0, sizeof(struct rdma_conn_param));
3147         cp.initiator_depth = isert_conn->initiator_depth;
3148         cp.retry_count = 7;
3149         cp.rnr_retry_count = 7;
3150
3151         ret = rdma_accept(cm_id, &cp);
3152         if (ret) {
3153                 isert_err("rdma_accept() failed with: %d\n", ret);
3154                 return ret;
3155         }
3156
3157         return 0;
3158 }
3159
3160 static int
3161 isert_get_login_rx(struct iscsi_conn *conn, struct iscsi_login *login)
3162 {
3163         struct isert_conn *isert_conn = conn->context;
3164         int ret;
3165
3166         isert_info("before login_req comp conn: %p\n", isert_conn);
3167         ret = wait_for_completion_interruptible(&isert_conn->login_req_comp);
3168         if (ret) {
3169                 isert_err("isert_conn %p interrupted before got login req\n",
3170                           isert_conn);
3171                 return ret;
3172         }
3173         reinit_completion(&isert_conn->login_req_comp);
3174
3175         /*
3176          * For login requests after the first PDU, isert_rx_login_req() will
3177          * kick schedule_delayed_work(&conn->login_work) as the packet is
3178          * received, which turns this callback from iscsi_target_do_login_rx()
3179          * into a NOP.
3180          */
3181         if (!login->first_request)
3182                 return 0;
3183
3184         isert_rx_login_req(isert_conn);
3185
3186         isert_info("before login_comp conn: %p\n", conn);
3187         ret = wait_for_completion_interruptible(&isert_conn->login_comp);
3188         if (ret)
3189                 return ret;
3190
3191         isert_info("processing login->req: %p\n", login->req);
3192
3193         return 0;
3194 }
3195
3196 static void
3197 isert_set_conn_info(struct iscsi_np *np, struct iscsi_conn *conn,
3198                     struct isert_conn *isert_conn)
3199 {
3200         struct rdma_cm_id *cm_id = isert_conn->cm_id;
3201         struct rdma_route *cm_route = &cm_id->route;
3202
3203         conn->login_family = np->np_sockaddr.ss_family;
3204
3205         conn->login_sockaddr = cm_route->addr.dst_addr;
3206         conn->local_sockaddr = cm_route->addr.src_addr;
3207 }
3208
3209 static int
3210 isert_accept_np(struct iscsi_np *np, struct iscsi_conn *conn)
3211 {
3212         struct isert_np *isert_np = np->np_context;
3213         struct isert_conn *isert_conn;
3214         int ret;
3215
3216 accept_wait:
3217         ret = down_interruptible(&isert_np->np_sem);
3218         if (ret)
3219                 return -ENODEV;
3220
3221         spin_lock_bh(&np->np_thread_lock);
3222         if (np->np_thread_state >= ISCSI_NP_THREAD_RESET) {
3223                 spin_unlock_bh(&np->np_thread_lock);
3224                 isert_dbg("np_thread_state %d\n",
3225                          np->np_thread_state);
3226                 /**
3227                  * No point in stalling here when np_thread
3228                  * is in state RESET/SHUTDOWN/EXIT - bail
3229                  **/
3230                 return -ENODEV;
3231         }
3232         spin_unlock_bh(&np->np_thread_lock);
3233
3234         mutex_lock(&isert_np->np_accept_mutex);
3235         if (list_empty(&isert_np->np_accept_list)) {
3236                 mutex_unlock(&isert_np->np_accept_mutex);
3237                 goto accept_wait;
3238         }
3239         isert_conn = list_first_entry(&isert_np->np_accept_list,
3240                         struct isert_conn, accept_node);
3241         list_del_init(&isert_conn->accept_node);
3242         mutex_unlock(&isert_np->np_accept_mutex);
3243
3244         conn->context = isert_conn;
3245         isert_conn->conn = conn;
3246
3247         isert_set_conn_info(np, conn, isert_conn);
3248
3249         isert_dbg("Processing isert_conn: %p\n", isert_conn);
3250
3251         return 0;
3252 }
3253
3254 static void
3255 isert_free_np(struct iscsi_np *np)
3256 {
3257         struct isert_np *isert_np = np->np_context;
3258         struct isert_conn *isert_conn, *n;
3259
3260         if (isert_np->np_cm_id)
3261                 rdma_destroy_id(isert_np->np_cm_id);
3262
3263         /*
3264          * FIXME: At this point we don't have a good way to insure
3265          * that at this point we don't have hanging connections that
3266          * completed RDMA establishment but didn't start iscsi login
3267          * process. So work-around this by cleaning up what ever piled
3268          * up in np_accept_list.
3269          */
3270         mutex_lock(&isert_np->np_accept_mutex);
3271         if (!list_empty(&isert_np->np_accept_list)) {
3272                 isert_info("Still have isert connections, cleaning up...\n");
3273                 list_for_each_entry_safe(isert_conn, n,
3274                                          &isert_np->np_accept_list,
3275                                          accept_node) {
3276                         isert_info("cleaning isert_conn %p state (%d)\n",
3277                                    isert_conn, isert_conn->state);
3278                         isert_connect_release(isert_conn);
3279                 }
3280         }
3281         mutex_unlock(&isert_np->np_accept_mutex);
3282
3283         np->np_context = NULL;
3284         kfree(isert_np);
3285 }
3286
3287 static void isert_release_work(struct work_struct *work)
3288 {
3289         struct isert_conn *isert_conn = container_of(work,
3290                                                      struct isert_conn,
3291                                                      release_work);
3292
3293         isert_info("Starting release conn %p\n", isert_conn);
3294
3295         wait_for_completion(&isert_conn->wait);
3296
3297         mutex_lock(&isert_conn->mutex);
3298         isert_conn->state = ISER_CONN_DOWN;
3299         mutex_unlock(&isert_conn->mutex);
3300
3301         isert_info("Destroying conn %p\n", isert_conn);
3302         isert_put_conn(isert_conn);
3303 }
3304
3305 static void
3306 isert_wait4logout(struct isert_conn *isert_conn)
3307 {
3308         struct iscsi_conn *conn = isert_conn->conn;
3309
3310         isert_info("conn %p\n", isert_conn);
3311
3312         if (isert_conn->logout_posted) {
3313                 isert_info("conn %p wait for conn_logout_comp\n", isert_conn);
3314                 wait_for_completion_timeout(&conn->conn_logout_comp,
3315                                             SECONDS_FOR_LOGOUT_COMP * HZ);
3316         }
3317 }
3318
3319 static void
3320 isert_wait4cmds(struct iscsi_conn *conn)
3321 {
3322         isert_info("iscsi_conn %p\n", conn);
3323
3324         if (conn->sess) {
3325                 target_sess_cmd_list_set_waiting(conn->sess->se_sess);
3326                 target_wait_for_sess_cmds(conn->sess->se_sess);
3327         }
3328 }
3329
3330 static void
3331 isert_wait4flush(struct isert_conn *isert_conn)
3332 {
3333         struct ib_recv_wr *bad_wr;
3334
3335         isert_info("conn %p\n", isert_conn);
3336
3337         init_completion(&isert_conn->wait_comp_err);
3338         isert_conn->beacon.wr_id = ISER_BEACON_WRID;
3339         /* post an indication that all flush errors were consumed */
3340         if (ib_post_recv(isert_conn->qp, &isert_conn->beacon, &bad_wr)) {
3341                 isert_err("conn %p failed to post beacon", isert_conn);
3342                 return;
3343         }
3344
3345         wait_for_completion(&isert_conn->wait_comp_err);
3346 }
3347
3348 static void isert_wait_conn(struct iscsi_conn *conn)
3349 {
3350         struct isert_conn *isert_conn = conn->context;
3351
3352         isert_info("Starting conn %p\n", isert_conn);
3353
3354         mutex_lock(&isert_conn->mutex);
3355         /*
3356          * Only wait for wait_comp_err if the isert_conn made it
3357          * into full feature phase..
3358          */
3359         if (isert_conn->state == ISER_CONN_INIT) {
3360                 mutex_unlock(&isert_conn->mutex);
3361                 return;
3362         }
3363         isert_conn_terminate(isert_conn);
3364         mutex_unlock(&isert_conn->mutex);
3365
3366         isert_wait4cmds(conn);
3367         isert_wait4flush(isert_conn);
3368         isert_wait4logout(isert_conn);
3369
3370         queue_work(isert_release_wq, &isert_conn->release_work);
3371 }
3372
3373 static void isert_free_conn(struct iscsi_conn *conn)
3374 {
3375         struct isert_conn *isert_conn = conn->context;
3376
3377         isert_wait4flush(isert_conn);
3378         isert_put_conn(isert_conn);
3379 }
3380
3381 static struct iscsit_transport iser_target_transport = {
3382         .name                   = "IB/iSER",
3383         .transport_type         = ISCSI_INFINIBAND,
3384         .priv_size              = sizeof(struct isert_cmd),
3385         .owner                  = THIS_MODULE,
3386         .iscsit_setup_np        = isert_setup_np,
3387         .iscsit_accept_np       = isert_accept_np,
3388         .iscsit_free_np         = isert_free_np,
3389         .iscsit_wait_conn       = isert_wait_conn,
3390         .iscsit_free_conn       = isert_free_conn,
3391         .iscsit_get_login_rx    = isert_get_login_rx,
3392         .iscsit_put_login_tx    = isert_put_login_tx,
3393         .iscsit_immediate_queue = isert_immediate_queue,
3394         .iscsit_response_queue  = isert_response_queue,
3395         .iscsit_get_dataout     = isert_get_dataout,
3396         .iscsit_queue_data_in   = isert_put_datain,
3397         .iscsit_queue_status    = isert_put_response,
3398         .iscsit_aborted_task    = isert_aborted_task,
3399         .iscsit_get_sup_prot_ops = isert_get_sup_prot_ops,
3400 };
3401
3402 static int __init isert_init(void)
3403 {
3404         int ret;
3405
3406         isert_comp_wq = alloc_workqueue("isert_comp_wq",
3407                                         WQ_UNBOUND | WQ_HIGHPRI, 0);
3408         if (!isert_comp_wq) {
3409                 isert_err("Unable to allocate isert_comp_wq\n");
3410                 ret = -ENOMEM;
3411                 return -ENOMEM;
3412         }
3413
3414         isert_release_wq = alloc_workqueue("isert_release_wq", WQ_UNBOUND,
3415                                         WQ_UNBOUND_MAX_ACTIVE);
3416         if (!isert_release_wq) {
3417                 isert_err("Unable to allocate isert_release_wq\n");
3418                 ret = -ENOMEM;
3419                 goto destroy_comp_wq;
3420         }
3421
3422         iscsit_register_transport(&iser_target_transport);
3423         isert_info("iSER_TARGET[0] - Loaded iser_target_transport\n");
3424
3425         return 0;
3426
3427 destroy_comp_wq:
3428         destroy_workqueue(isert_comp_wq);
3429
3430         return ret;
3431 }
3432
3433 static void __exit isert_exit(void)
3434 {
3435         flush_scheduled_work();
3436         destroy_workqueue(isert_release_wq);
3437         destroy_workqueue(isert_comp_wq);
3438         iscsit_unregister_transport(&iser_target_transport);
3439         isert_info("iSER_TARGET[0] - Released iser_target_transport\n");
3440 }
3441
3442 MODULE_DESCRIPTION("iSER-Target for mainline target infrastructure");
3443 MODULE_VERSION("1.0");
3444 MODULE_AUTHOR("nab@Linux-iSCSI.org");
3445 MODULE_LICENSE("GPL");
3446
3447 module_init(isert_init);
3448 module_exit(isert_exit);