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 "ib_isert.h"
  33
  34 #define ISERT_MAX_CONN          8
  35 #define ISER_MAX_RX_CQ_LEN      (ISERT_QP_MAX_RECV_DTOS * ISERT_MAX_CONN)
  36 #define ISER_MAX_TX_CQ_LEN \
  37         ((ISERT_QP_MAX_REQ_DTOS + ISCSI_DEF_XMIT_CMDS_MAX) * 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 int
  51 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd);
  52 static int
  53 isert_login_post_recv(struct isert_conn *isert_conn);
  54 static int
  55 isert_rdma_accept(struct isert_conn *isert_conn);
  56 struct rdma_cm_id *isert_setup_id(struct isert_np *isert_np);
  57
  58 static void isert_release_work(struct work_struct *work);
  59 static void isert_recv_done(struct ib_cq *cq, struct ib_wc *wc);
  60 static void isert_send_done(struct ib_cq *cq, struct ib_wc *wc);
  61 static void isert_login_recv_done(struct ib_cq *cq, struct ib_wc *wc);
  62 static void isert_login_send_done(struct ib_cq *cq, struct ib_wc *wc);
  63
  64 static inline bool
  65 isert_prot_cmd(struct isert_conn *conn, struct se_cmd *cmd)
  66 {
  67         return (conn->pi_support &&
  68                 cmd->prot_op != TARGET_PROT_NORMAL);
  69 }
  70
  71
  72 static void
  73 isert_qp_event_callback(struct ib_event *e, void *context)
  74 {
  75         struct isert_conn *isert_conn = context;
  76
  77         isert_err("%s (%d): conn %p\n",
  78                   ib_event_msg(e->event), e->event, isert_conn);
  79
  80         switch (e->event) {
  81         case IB_EVENT_COMM_EST:
  82                 rdma_notify(isert_conn->cm_id, IB_EVENT_COMM_EST);
  83                 break;
  84         case IB_EVENT_QP_LAST_WQE_REACHED:
  85                 isert_warn("Reached TX IB_EVENT_QP_LAST_WQE_REACHED\n");
  86                 break;
  87         default:
  88                 break;
  89         }
  90 }
  91
  92 static struct isert_comp *
  93 isert_comp_get(struct isert_conn *isert_conn)
  94 {
  95         struct isert_device *device = isert_conn->device;
  96         struct isert_comp *comp;
  97         int i, min = 0;
  98
  99         mutex_lock(&device_list_mutex);
 100         for (i = 0; i < device->comps_used; i++)
 101                 if (device->comps[i].active_qps <
 102                     device->comps[min].active_qps)
 103                         min = i;
 104         comp = &device->comps[min];
 105         comp->active_qps++;
 106         mutex_unlock(&device_list_mutex);
 107
 108         isert_info("conn %p, using comp %p min_index: %d\n",
 109                    isert_conn, comp, min);
 110
 111         return comp;
 112 }
 113
 114 static void
 115 isert_comp_put(struct isert_comp *comp)
 116 {
 117         mutex_lock(&device_list_mutex);
 118         comp->active_qps--;
 119         mutex_unlock(&device_list_mutex);
 120 }
 121
 122 static struct ib_qp *
 123 isert_create_qp(struct isert_conn *isert_conn,
 124                 struct isert_comp *comp,
 125                 struct rdma_cm_id *cma_id)
 126 {
 127         struct isert_device *device = isert_conn->device;
 128         struct ib_qp_init_attr attr;
 129         int ret;
 130
 131         memset(&attr, 0, sizeof(struct ib_qp_init_attr));
 132         attr.event_handler = isert_qp_event_callback;
 133         attr.qp_context = isert_conn;
 134         attr.send_cq = comp->cq;
 135         attr.recv_cq = comp->cq;
 136         attr.cap.max_send_wr = ISERT_QP_MAX_REQ_DTOS + 1;
 137         attr.cap.max_recv_wr = ISERT_QP_MAX_RECV_DTOS + 1;
 138         attr.cap.max_rdma_ctxs = ISCSI_DEF_XMIT_CMDS_MAX;
 139         attr.cap.max_send_sge = device->ib_device->attrs.max_send_sge;
 140         attr.cap.max_recv_sge = 1;
 141         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 142         attr.qp_type = IB_QPT_RC;
 143         if (device->pi_capable)
 144                 attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
 145
 146         ret = rdma_create_qp(cma_id, device->pd, &attr);
 147         if (ret) {
 148                 isert_err("rdma_create_qp failed for cma_id %d\n", ret);
 149                 return ERR_PTR(ret);
 150         }
 151
 152         return cma_id->qp;
 153 }
 154
 155 static int
 156 isert_conn_setup_qp(struct isert_conn *isert_conn, struct rdma_cm_id *cma_id)
 157 {
 158         struct isert_comp *comp;
 159         int ret;
 160
 161         comp = isert_comp_get(isert_conn);
 162         isert_conn->qp = isert_create_qp(isert_conn, comp, cma_id);
 163         if (IS_ERR(isert_conn->qp)) {
 164                 ret = PTR_ERR(isert_conn->qp);
 165                 goto err;
 166         }
 167
 168         return 0;
 169 err:
 170         isert_comp_put(comp);
 171         return ret;
 172 }
 173
 174 static int
 175 isert_alloc_rx_descriptors(struct isert_conn *isert_conn)
 176 {
 177         struct isert_device *device = isert_conn->device;
 178         struct ib_device *ib_dev = device->ib_device;
 179         struct iser_rx_desc *rx_desc;
 180         struct ib_sge *rx_sg;
 181         u64 dma_addr;
 182         int i, j;
 183
 184         isert_conn->rx_descs = kcalloc(ISERT_QP_MAX_RECV_DTOS,
 185                                        sizeof(struct iser_rx_desc),
 186                                        GFP_KERNEL);
 187         if (!isert_conn->rx_descs)
 188                 return -ENOMEM;
 189
 190         rx_desc = isert_conn->rx_descs;
 191
 192         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 193                 dma_addr = ib_dma_map_single(ib_dev, (void *)rx_desc,
 194                                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 195                 if (ib_dma_mapping_error(ib_dev, dma_addr))
 196                         goto dma_map_fail;
 197
 198                 rx_desc->dma_addr = dma_addr;
 199
 200                 rx_sg = &rx_desc->rx_sg;
 201                 rx_sg->addr = rx_desc->dma_addr;
 202                 rx_sg->length = ISER_RX_PAYLOAD_SIZE;
 203                 rx_sg->lkey = device->pd->local_dma_lkey;
 204                 rx_desc->rx_cqe.done = isert_recv_done;
 205         }
 206
 207         return 0;
 208
 209 dma_map_fail:
 210         rx_desc = isert_conn->rx_descs;
 211         for (j = 0; j < i; j++, rx_desc++) {
 212                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 213                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 214         }
 215         kfree(isert_conn->rx_descs);
 216         isert_conn->rx_descs = NULL;
 217         isert_err("conn %p failed to allocate rx descriptors\n", isert_conn);
 218         return -ENOMEM;
 219 }
 220
 221 static void
 222 isert_free_rx_descriptors(struct isert_conn *isert_conn)
 223 {
 224         struct ib_device *ib_dev = isert_conn->device->ib_device;
 225         struct iser_rx_desc *rx_desc;
 226         int i;
 227
 228         if (!isert_conn->rx_descs)
 229                 return;
 230
 231         rx_desc = isert_conn->rx_descs;
 232         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 233                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 234                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 235         }
 236
 237         kfree(isert_conn->rx_descs);
 238         isert_conn->rx_descs = NULL;
 239 }
 240
 241 static void
 242 isert_free_comps(struct isert_device *device)
 243 {
 244         int i;
 245
 246         for (i = 0; i < device->comps_used; i++) {
 247                 struct isert_comp *comp = &device->comps[i];
 248
 249                 if (comp->cq)
 250                         ib_free_cq(comp->cq);
 251         }
 252         kfree(device->comps);
 253 }
 254
 255 static int
 256 isert_alloc_comps(struct isert_device *device)
 257 {
 258         int i, max_cqe, ret = 0;
 259
 260         device->comps_used = min(ISERT_MAX_CQ, min_t(int, num_online_cpus(),
 261                                  device->ib_device->num_comp_vectors));
 262
 263         isert_info("Using %d CQs, %s supports %d vectors support "
 264                    "pi_capable %d\n",
 265                    device->comps_used, device->ib_device->name,
 266                    device->ib_device->num_comp_vectors,
 267                    device->pi_capable);
 268
 269         device->comps = kcalloc(device->comps_used, sizeof(struct isert_comp),
 270                                 GFP_KERNEL);
 271         if (!device->comps)
 272                 return -ENOMEM;
 273
 274         max_cqe = min(ISER_MAX_CQ_LEN, device->ib_device->attrs.max_cqe);
 275
 276         for (i = 0; i < device->comps_used; i++) {
 277                 struct isert_comp *comp = &device->comps[i];
 278
 279                 comp->device = device;
 280                 comp->cq = ib_alloc_cq(device->ib_device, comp, max_cqe, i,
 281                                 IB_POLL_WORKQUEUE);
 282                 if (IS_ERR(comp->cq)) {
 283                         isert_err("Unable to allocate cq\n");
 284                         ret = PTR_ERR(comp->cq);
 285                         comp->cq = NULL;
 286                         goto out_cq;
 287                 }
 288         }
 289
 290         return 0;
 291 out_cq:
 292         isert_free_comps(device);
 293         return ret;
 294 }
 295
 296 static int
 297 isert_create_device_ib_res(struct isert_device *device)
 298 {
 299         struct ib_device *ib_dev = device->ib_device;
 300         int ret;
 301
 302         isert_dbg("devattr->max_send_sge: %d devattr->max_recv_sge %d\n",
 303                   ib_dev->attrs.max_send_sge, ib_dev->attrs.max_recv_sge);
 304         isert_dbg("devattr->max_sge_rd: %d\n", ib_dev->attrs.max_sge_rd);
 305
 306         ret = isert_alloc_comps(device);
 307         if (ret)
 308                 goto out;
 309
 310         device->pd = ib_alloc_pd(ib_dev, 0);
 311         if (IS_ERR(device->pd)) {
 312                 ret = PTR_ERR(device->pd);
 313                 isert_err("failed to allocate pd, device %p, ret=%d\n",
 314                           device, ret);
 315                 goto out_cq;
 316         }
 317
 318         /* Check signature cap */
 319         device->pi_capable = ib_dev->attrs.device_cap_flags &
 320                              IB_DEVICE_SIGNATURE_HANDOVER ? true : false;
 321
 322         return 0;
 323
 324 out_cq:
 325         isert_free_comps(device);
 326 out:
 327         if (ret > 0)
 328                 ret = -EINVAL;
 329         return ret;
 330 }
 331
 332 static void
 333 isert_free_device_ib_res(struct isert_device *device)
 334 {
 335         isert_info("device %p\n", device);
 336
 337         ib_dealloc_pd(device->pd);
 338         isert_free_comps(device);
 339 }
 340
 341 static void
 342 isert_device_put(struct isert_device *device)
 343 {
 344         mutex_lock(&device_list_mutex);
 345         device->refcount--;
 346         isert_info("device %p refcount %d\n", device, device->refcount);
 347         if (!device->refcount) {
 348                 isert_free_device_ib_res(device);
 349                 list_del(&device->dev_node);
 350                 kfree(device);
 351         }
 352         mutex_unlock(&device_list_mutex);
 353 }
 354
 355 static struct isert_device *
 356 isert_device_get(struct rdma_cm_id *cma_id)
 357 {
 358         struct isert_device *device;
 359         int ret;
 360
 361         mutex_lock(&device_list_mutex);
 362         list_for_each_entry(device, &device_list, dev_node) {
 363                 if (device->ib_device->node_guid == cma_id->device->node_guid) {
 364                         device->refcount++;
 365                         isert_info("Found iser device %p refcount %d\n",
 366                                    device, device->refcount);
 367                         mutex_unlock(&device_list_mutex);
 368                         return device;
 369                 }
 370         }
 371
 372         device = kzalloc(sizeof(struct isert_device), GFP_KERNEL);
 373         if (!device) {
 374                 mutex_unlock(&device_list_mutex);
 375                 return ERR_PTR(-ENOMEM);
 376         }
 377
 378         INIT_LIST_HEAD(&device->dev_node);
 379
 380         device->ib_device = cma_id->device;
 381         ret = isert_create_device_ib_res(device);
 382         if (ret) {
 383                 kfree(device);
 384                 mutex_unlock(&device_list_mutex);
 385                 return ERR_PTR(ret);
 386         }
 387
 388         device->refcount++;
 389         list_add_tail(&device->dev_node, &device_list);
 390         isert_info("Created a new iser device %p refcount %d\n",
 391                    device, device->refcount);
 392         mutex_unlock(&device_list_mutex);
 393
 394         return device;
 395 }
 396
 397 static void
 398 isert_init_conn(struct isert_conn *isert_conn)
 399 {
 400         isert_conn->state = ISER_CONN_INIT;
 401         INIT_LIST_HEAD(&isert_conn->node);
 402         init_completion(&isert_conn->login_comp);
 403         init_completion(&isert_conn->login_req_comp);
 404         init_waitqueue_head(&isert_conn->rem_wait);
 405         kref_init(&isert_conn->kref);
 406         mutex_init(&isert_conn->mutex);
 407         INIT_WORK(&isert_conn->release_work, isert_release_work);
 408 }
 409
 410 static void
 411 isert_free_login_buf(struct isert_conn *isert_conn)
 412 {
 413         struct ib_device *ib_dev = isert_conn->device->ib_device;
 414
 415         ib_dma_unmap_single(ib_dev, isert_conn->login_rsp_dma,
 416                             ISER_RX_PAYLOAD_SIZE, DMA_TO_DEVICE);
 417         kfree(isert_conn->login_rsp_buf);
 418
 419         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 420                             ISER_RX_PAYLOAD_SIZE,
 421                             DMA_FROM_DEVICE);
 422         kfree(isert_conn->login_req_buf);
 423 }
 424
 425 static int
 426 isert_alloc_login_buf(struct isert_conn *isert_conn,
 427                       struct ib_device *ib_dev)
 428 {
 429         int ret;
 430
 431         isert_conn->login_req_buf = kzalloc(sizeof(*isert_conn->login_req_buf),
 432                         GFP_KERNEL);
 433         if (!isert_conn->login_req_buf)
 434                 return -ENOMEM;
 435
 436         isert_conn->login_req_dma = ib_dma_map_single(ib_dev,
 437                                 isert_conn->login_req_buf,
 438                                 ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 439         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_req_dma);
 440         if (ret) {
 441                 isert_err("login_req_dma mapping error: %d\n", ret);
 442                 isert_conn->login_req_dma = 0;
 443                 goto out_free_login_req_buf;
 444         }
 445
 446         isert_conn->login_rsp_buf = kzalloc(ISER_RX_PAYLOAD_SIZE, GFP_KERNEL);
 447         if (!isert_conn->login_rsp_buf) {
 448                 ret = -ENOMEM;
 449                 goto out_unmap_login_req_buf;
 450         }
 451
 452         isert_conn->login_rsp_dma = ib_dma_map_single(ib_dev,
 453                                         isert_conn->login_rsp_buf,
 454                                         ISER_RX_PAYLOAD_SIZE, DMA_TO_DEVICE);
 455         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_rsp_dma);
 456         if (ret) {
 457                 isert_err("login_rsp_dma mapping error: %d\n", ret);
 458                 isert_conn->login_rsp_dma = 0;
 459                 goto out_free_login_rsp_buf;
 460         }
 461
 462         return 0;
 463
 464 out_free_login_rsp_buf:
 465         kfree(isert_conn->login_rsp_buf);
 466 out_unmap_login_req_buf:
 467         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 468                             ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 469 out_free_login_req_buf:
 470         kfree(isert_conn->login_req_buf);
 471         return ret;
 472 }
 473
 474 static void
 475 isert_set_nego_params(struct isert_conn *isert_conn,
 476                       struct rdma_conn_param *param)
 477 {
 478         struct ib_device_attr *attr = &isert_conn->device->ib_device->attrs;
 479
 480         /* Set max inflight RDMA READ requests */
 481         isert_conn->initiator_depth = min_t(u8, param->initiator_depth,
 482                                 attr->max_qp_init_rd_atom);
 483         isert_dbg("Using initiator_depth: %u\n", isert_conn->initiator_depth);
 484
 485         if (param->private_data) {
 486                 u8 flags = *(u8 *)param->private_data;
 487
 488                 /*
 489                  * use remote invalidation if the both initiator
 490                  * and the HCA support it
 491                  */
 492                 isert_conn->snd_w_inv = !(flags & ISER_SEND_W_INV_NOT_SUP) &&
 493                                           (attr->device_cap_flags &
 494                                            IB_DEVICE_MEM_MGT_EXTENSIONS);
 495                 if (isert_conn->snd_w_inv)
 496                         isert_info("Using remote invalidation\n");
 497         }
 498 }
 499
 500 static int
 501 isert_connect_request(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 502 {
 503         struct isert_np *isert_np = cma_id->context;
 504         struct iscsi_np *np = isert_np->np;
 505         struct isert_conn *isert_conn;
 506         struct isert_device *device;
 507         int ret = 0;
 508
 509         spin_lock_bh(&np->np_thread_lock);
 510         if (!np->enabled) {
 511                 spin_unlock_bh(&np->np_thread_lock);
 512                 isert_dbg("iscsi_np is not enabled, reject connect request\n");
 513                 return rdma_reject(cma_id, NULL, 0);
 514         }
 515         spin_unlock_bh(&np->np_thread_lock);
 516
 517         isert_dbg("cma_id: %p, portal: %p\n",
 518                  cma_id, cma_id->context);
 519
 520         isert_conn = kzalloc(sizeof(struct isert_conn), GFP_KERNEL);
 521         if (!isert_conn)
 522                 return -ENOMEM;
 523
 524         isert_init_conn(isert_conn);
 525         isert_conn->cm_id = cma_id;
 526
 527         ret = isert_alloc_login_buf(isert_conn, cma_id->device);
 528         if (ret)
 529                 goto out;
 530
 531         device = isert_device_get(cma_id);
 532         if (IS_ERR(device)) {
 533                 ret = PTR_ERR(device);
 534                 goto out_rsp_dma_map;
 535         }
 536         isert_conn->device = device;
 537
 538         isert_set_nego_params(isert_conn, &event->param.conn);
 539
 540         ret = isert_conn_setup_qp(isert_conn, cma_id);
 541         if (ret)
 542                 goto out_conn_dev;
 543
 544         ret = isert_login_post_recv(isert_conn);
 545         if (ret)
 546                 goto out_conn_dev;
 547
 548         ret = isert_rdma_accept(isert_conn);
 549         if (ret)
 550                 goto out_conn_dev;
 551
 552         mutex_lock(&isert_np->mutex);
 553         list_add_tail(&isert_conn->node, &isert_np->accepted);
 554         mutex_unlock(&isert_np->mutex);
 555
 556         return 0;
 557
 558 out_conn_dev:
 559         isert_device_put(device);
 560 out_rsp_dma_map:
 561         isert_free_login_buf(isert_conn);
 562 out:
 563         kfree(isert_conn);
 564         rdma_reject(cma_id, NULL, 0);
 565         return ret;
 566 }
 567
 568 static void
 569 isert_connect_release(struct isert_conn *isert_conn)
 570 {
 571         struct isert_device *device = isert_conn->device;
 572
 573         isert_dbg("conn %p\n", isert_conn);
 574
 575         BUG_ON(!device);
 576
 577         isert_free_rx_descriptors(isert_conn);
 578         if (isert_conn->cm_id &&
 579             !isert_conn->dev_removed)
 580                 rdma_destroy_id(isert_conn->cm_id);
 581
 582         if (isert_conn->qp) {
 583                 struct isert_comp *comp = isert_conn->qp->recv_cq->cq_context;
 584
 585                 isert_comp_put(comp);
 586                 ib_destroy_qp(isert_conn->qp);
 587         }
 588
 589         if (isert_conn->login_req_buf)
 590                 isert_free_login_buf(isert_conn);
 591
 592         isert_device_put(device);
 593
 594         if (isert_conn->dev_removed)
 595                 wake_up_interruptible(&isert_conn->rem_wait);
 596         else
 597                 kfree(isert_conn);
 598 }
 599
 600 static void
 601 isert_connected_handler(struct rdma_cm_id *cma_id)
 602 {
 603         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 604         struct isert_np *isert_np = cma_id->context;
 605
 606         isert_info("conn %p\n", isert_conn);
 607
 608         mutex_lock(&isert_conn->mutex);
 609         isert_conn->state = ISER_CONN_UP;
 610         kref_get(&isert_conn->kref);
 611         mutex_unlock(&isert_conn->mutex);
 612
 613         mutex_lock(&isert_np->mutex);
 614         list_move_tail(&isert_conn->node, &isert_np->pending);
 615         mutex_unlock(&isert_np->mutex);
 616
 617         isert_info("np %p: Allow accept_np to continue\n", isert_np);
 618         up(&isert_np->sem);
 619 }
 620
 621 static void
 622 isert_release_kref(struct kref *kref)
 623 {
 624         struct isert_conn *isert_conn = container_of(kref,
 625                                 struct isert_conn, kref);
 626
 627         isert_info("conn %p final kref %s/%d\n", isert_conn, current->comm,
 628                    current->pid);
 629
 630         isert_connect_release(isert_conn);
 631 }
 632
 633 static void
 634 isert_put_conn(struct isert_conn *isert_conn)
 635 {
 636         kref_put(&isert_conn->kref, isert_release_kref);
 637 }
 638
 639 static void
 640 isert_handle_unbound_conn(struct isert_conn *isert_conn)
 641 {
 642         struct isert_np *isert_np = isert_conn->cm_id->context;
 643
 644         mutex_lock(&isert_np->mutex);
 645         if (!list_empty(&isert_conn->node)) {
 646                 /*
 647                  * This means iscsi doesn't know this connection
 648                  * so schedule a cleanup ourselves
 649                  */
 650                 list_del_init(&isert_conn->node);
 651                 isert_put_conn(isert_conn);
 652                 queue_work(isert_release_wq, &isert_conn->release_work);
 653         }
 654         mutex_unlock(&isert_np->mutex);
 655 }
 656
 657 /**
 658  * isert_conn_terminate() - Initiate connection termination
 659  * @isert_conn: isert connection struct
 660  *
 661  * Notes:
 662  * In case the connection state is BOUND, move state
 663  * to TEMINATING and start teardown sequence (rdma_disconnect).
 664  * In case the connection state is UP, complete flush as well.
 665  *
 666  * This routine must be called with mutex held. Thus it is
 667  * safe to call multiple times.
 668  */
 669 static void
 670 isert_conn_terminate(struct isert_conn *isert_conn)
 671 {
 672         int err;
 673
 674         if (isert_conn->state >= ISER_CONN_TERMINATING)
 675                 return;
 676
 677         isert_info("Terminating conn %p state %d\n",
 678                    isert_conn, isert_conn->state);
 679         isert_conn->state = ISER_CONN_TERMINATING;
 680         err = rdma_disconnect(isert_conn->cm_id);
 681         if (err)
 682                 isert_warn("Failed rdma_disconnect isert_conn %p\n",
 683                            isert_conn);
 684 }
 685
 686 static int
 687 isert_np_cma_handler(struct isert_np *isert_np,
 688                      enum rdma_cm_event_type event)
 689 {
 690         isert_dbg("%s (%d): isert np %p\n",
 691                   rdma_event_msg(event), event, isert_np);
 692
 693         switch (event) {
 694         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 695                 isert_np->cm_id = NULL;
 696                 break;
 697         case RDMA_CM_EVENT_ADDR_CHANGE:
 698                 isert_np->cm_id = isert_setup_id(isert_np);
 699                 if (IS_ERR(isert_np->cm_id)) {
 700                         isert_err("isert np %p setup id failed: %ld\n",
 701                                   isert_np, PTR_ERR(isert_np->cm_id));
 702                         isert_np->cm_id = NULL;
 703                 }
 704                 break;
 705         default:
 706                 isert_err("isert np %p Unexpected event %d\n",
 707                           isert_np, event);
 708         }
 709
 710         return -1;
 711 }
 712
 713 static int
 714 isert_disconnected_handler(struct rdma_cm_id *cma_id,
 715                            enum rdma_cm_event_type event)
 716 {
 717         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 718
 719         mutex_lock(&isert_conn->mutex);
 720         switch (isert_conn->state) {
 721         case ISER_CONN_TERMINATING:
 722                 break;
 723         case ISER_CONN_UP:
 724                 isert_conn_terminate(isert_conn);
 725                 ib_drain_qp(isert_conn->qp);
 726                 isert_handle_unbound_conn(isert_conn);
 727                 break;
 728         case ISER_CONN_BOUND:
 729         case ISER_CONN_FULL_FEATURE: /* FALLTHRU */
 730                 iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
 731                 break;
 732         default:
 733                 isert_warn("conn %p terminating in state %d\n",
 734                            isert_conn, isert_conn->state);
 735         }
 736         mutex_unlock(&isert_conn->mutex);
 737
 738         return 0;
 739 }
 740
 741 static int
 742 isert_connect_error(struct rdma_cm_id *cma_id)
 743 {
 744         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 745
 746         ib_drain_qp(isert_conn->qp);
 747         list_del_init(&isert_conn->node);
 748         isert_conn->cm_id = NULL;
 749         isert_put_conn(isert_conn);
 750
 751         return -1;
 752 }
 753
 754 static int
 755 isert_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 756 {
 757         struct isert_np *isert_np = cma_id->context;
 758         struct isert_conn *isert_conn;
 759         int ret = 0;
 760
 761         isert_info("%s (%d): status %d id %p np %p\n",
 762                    rdma_event_msg(event->event), event->event,
 763                    event->status, cma_id, cma_id->context);
 764
 765         if (isert_np->cm_id == cma_id)
 766                 return isert_np_cma_handler(cma_id->context, event->event);
 767
 768         switch (event->event) {
 769         case RDMA_CM_EVENT_CONNECT_REQUEST:
 770                 ret = isert_connect_request(cma_id, event);
 771                 if (ret)
 772                         isert_err("failed handle connect request %d\n", ret);
 773                 break;
 774         case RDMA_CM_EVENT_ESTABLISHED:
 775                 isert_connected_handler(cma_id);
 776                 break;
 777         case RDMA_CM_EVENT_ADDR_CHANGE:    /* FALLTHRU */
 778         case RDMA_CM_EVENT_DISCONNECTED:   /* FALLTHRU */
 779         case RDMA_CM_EVENT_TIMEWAIT_EXIT:  /* FALLTHRU */
 780                 ret = isert_disconnected_handler(cma_id, event->event);
 781                 break;
 782         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 783                 isert_conn = cma_id->qp->qp_context;
 784                 isert_conn->dev_removed = true;
 785                 isert_disconnected_handler(cma_id, event->event);
 786                 wait_event_interruptible(isert_conn->rem_wait,
 787                                          isert_conn->state == ISER_CONN_DOWN);
 788                 kfree(isert_conn);
 789                 /*
 790                  * return non-zero from the callback to destroy
 791                  * the rdma cm id
 792                  */
 793                 return 1;
 794         case RDMA_CM_EVENT_REJECTED:
 795                 isert_info("Connection rejected: %s\n",
 796                            rdma_reject_msg(cma_id, event->status));
 797                 /* fall through */
 798         case RDMA_CM_EVENT_UNREACHABLE:
 799         case RDMA_CM_EVENT_CONNECT_ERROR:
 800                 ret = isert_connect_error(cma_id);
 801                 break;
 802         default:
 803                 isert_err("Unhandled RDMA CMA event: %d\n", event->event);
 804                 break;
 805         }
 806
 807         return ret;
 808 }
 809
 810 static int
 811 isert_post_recvm(struct isert_conn *isert_conn, u32 count)
 812 {
 813         struct ib_recv_wr *rx_wr;
 814         int i, ret;
 815         struct iser_rx_desc *rx_desc;
 816
 817         for (rx_wr = isert_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
 818                 rx_desc = &isert_conn->rx_descs[i];
 819
 820                 rx_wr->wr_cqe = &rx_desc->rx_cqe;
 821                 rx_wr->sg_list = &rx_desc->rx_sg;
 822                 rx_wr->num_sge = 1;
 823                 rx_wr->next = rx_wr + 1;
 824                 rx_desc->in_use = false;
 825         }
 826         rx_wr--;
 827         rx_wr->next = NULL; /* mark end of work requests list */
 828
 829         ret = ib_post_recv(isert_conn->qp, isert_conn->rx_wr, NULL);
 830         if (ret)
 831                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
 832
 833         return ret;
 834 }
 835
 836 static int
 837 isert_post_recv(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc)
 838 {
 839         struct ib_recv_wr rx_wr;
 840         int ret;
 841
 842         if (!rx_desc->in_use) {
 843                 /*
 844                  * if the descriptor is not in-use we already reposted it
 845                  * for recv, so just silently return
 846                  */
 847                 return 0;
 848         }
 849
 850         rx_desc->in_use = false;
 851         rx_wr.wr_cqe = &rx_desc->rx_cqe;
 852         rx_wr.sg_list = &rx_desc->rx_sg;
 853         rx_wr.num_sge = 1;
 854         rx_wr.next = NULL;
 855
 856         ret = ib_post_recv(isert_conn->qp, &rx_wr, NULL);
 857         if (ret)
 858                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
 859
 860         return ret;
 861 }
 862
 863 static int
 864 isert_login_post_send(struct isert_conn *isert_conn, struct iser_tx_desc *tx_desc)
 865 {
 866         struct ib_device *ib_dev = isert_conn->cm_id->device;
 867         struct ib_send_wr send_wr;
 868         int ret;
 869
 870         ib_dma_sync_single_for_device(ib_dev, tx_desc->dma_addr,
 871                                       ISER_HEADERS_LEN, DMA_TO_DEVICE);
 872
 873         tx_desc->tx_cqe.done = isert_login_send_done;
 874
 875         send_wr.next    = NULL;
 876         send_wr.wr_cqe  = &tx_desc->tx_cqe;
 877         send_wr.sg_list = tx_desc->tx_sg;
 878         send_wr.num_sge = tx_desc->num_sge;
 879         send_wr.opcode  = IB_WR_SEND;
 880         send_wr.send_flags = IB_SEND_SIGNALED;
 881
 882         ret = ib_post_send(isert_conn->qp, &send_wr, NULL);
 883         if (ret)
 884                 isert_err("ib_post_send() failed, ret: %d\n", ret);
 885
 886         return ret;
 887 }
 888
 889 static void
 890 __isert_create_send_desc(struct isert_device *device,
 891                          struct iser_tx_desc *tx_desc)
 892 {
 893
 894         memset(&tx_desc->iser_header, 0, sizeof(struct iser_ctrl));
 895         tx_desc->iser_header.flags = ISCSI_CTRL;
 896
 897         tx_desc->num_sge = 1;
 898
 899         if (tx_desc->tx_sg[0].lkey != device->pd->local_dma_lkey) {
 900                 tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
 901                 isert_dbg("tx_desc %p lkey mismatch, fixing\n", tx_desc);
 902         }
 903 }
 904
 905 static void
 906 isert_create_send_desc(struct isert_conn *isert_conn,
 907                        struct isert_cmd *isert_cmd,
 908                        struct iser_tx_desc *tx_desc)
 909 {
 910         struct isert_device *device = isert_conn->device;
 911         struct ib_device *ib_dev = device->ib_device;
 912
 913         ib_dma_sync_single_for_cpu(ib_dev, tx_desc->dma_addr,
 914                                    ISER_HEADERS_LEN, DMA_TO_DEVICE);
 915
 916         __isert_create_send_desc(device, tx_desc);
 917 }
 918
 919 static int
 920 isert_init_tx_hdrs(struct isert_conn *isert_conn,
 921                    struct iser_tx_desc *tx_desc)
 922 {
 923         struct isert_device *device = isert_conn->device;
 924         struct ib_device *ib_dev = device->ib_device;
 925         u64 dma_addr;
 926
 927         dma_addr = ib_dma_map_single(ib_dev, (void *)tx_desc,
 928                         ISER_HEADERS_LEN, DMA_TO_DEVICE);
 929         if (ib_dma_mapping_error(ib_dev, dma_addr)) {
 930                 isert_err("ib_dma_mapping_error() failed\n");
 931                 return -ENOMEM;
 932         }
 933
 934         tx_desc->dma_addr = dma_addr;
 935         tx_desc->tx_sg[0].addr  = tx_desc->dma_addr;
 936         tx_desc->tx_sg[0].length = ISER_HEADERS_LEN;
 937         tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
 938
 939         isert_dbg("Setup tx_sg[0].addr: 0x%llx length: %u lkey: 0x%x\n",
 940                   tx_desc->tx_sg[0].addr, tx_desc->tx_sg[0].length,
 941                   tx_desc->tx_sg[0].lkey);
 942
 943         return 0;
 944 }
 945
 946 static void
 947 isert_init_send_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
 948                    struct ib_send_wr *send_wr)
 949 {
 950         struct iser_tx_desc *tx_desc = &isert_cmd->tx_desc;
 951
 952         tx_desc->tx_cqe.done = isert_send_done;
 953         send_wr->wr_cqe = &tx_desc->tx_cqe;
 954
 955         if (isert_conn->snd_w_inv && isert_cmd->inv_rkey) {
 956                 send_wr->opcode  = IB_WR_SEND_WITH_INV;
 957                 send_wr->ex.invalidate_rkey = isert_cmd->inv_rkey;
 958         } else {
 959                 send_wr->opcode = IB_WR_SEND;
 960         }
 961
 962         send_wr->sg_list = &tx_desc->tx_sg[0];
 963         send_wr->num_sge = isert_cmd->tx_desc.num_sge;
 964         send_wr->send_flags = IB_SEND_SIGNALED;
 965 }
 966
 967 static int
 968 isert_login_post_recv(struct isert_conn *isert_conn)
 969 {
 970         struct ib_recv_wr rx_wr;
 971         struct ib_sge sge;
 972         int ret;
 973
 974         memset(&sge, 0, sizeof(struct ib_sge));
 975         sge.addr = isert_conn->login_req_dma;
 976         sge.length = ISER_RX_PAYLOAD_SIZE;
 977         sge.lkey = isert_conn->device->pd->local_dma_lkey;
 978
 979         isert_dbg("Setup sge: addr: %llx length: %d 0x%08x\n",
 980                 sge.addr, sge.length, sge.lkey);
 981
 982         isert_conn->login_req_buf->rx_cqe.done = isert_login_recv_done;
 983
 984         memset(&rx_wr, 0, sizeof(struct ib_recv_wr));
 985         rx_wr.wr_cqe = &isert_conn->login_req_buf->rx_cqe;
 986         rx_wr.sg_list = &sge;
 987         rx_wr.num_sge = 1;
 988
 989         ret = ib_post_recv(isert_conn->qp, &rx_wr, NULL);
 990         if (ret)
 991                 isert_err("ib_post_recv() failed: %d\n", ret);
 992
 993         return ret;
 994 }
 995
 996 static int
 997 isert_put_login_tx(struct iscsi_conn *conn, struct iscsi_login *login,
 998                    u32 length)
 999 {
1000         struct isert_conn *isert_conn = conn->context;
1001         struct isert_device *device = isert_conn->device;
1002         struct ib_device *ib_dev = device->ib_device;
1003         struct iser_tx_desc *tx_desc = &isert_conn->login_tx_desc;
1004         int ret;
1005
1006         __isert_create_send_desc(device, tx_desc);
1007
1008         memcpy(&tx_desc->iscsi_header, &login->rsp[0],
1009                sizeof(struct iscsi_hdr));
1010
1011         isert_init_tx_hdrs(isert_conn, tx_desc);
1012
1013         if (length > 0) {
1014                 struct ib_sge *tx_dsg = &tx_desc->tx_sg[1];
1015
1016                 ib_dma_sync_single_for_cpu(ib_dev, isert_conn->login_rsp_dma,
1017                                            length, DMA_TO_DEVICE);
1018
1019                 memcpy(isert_conn->login_rsp_buf, login->rsp_buf, length);
1020
1021                 ib_dma_sync_single_for_device(ib_dev, isert_conn->login_rsp_dma,
1022                                               length, DMA_TO_DEVICE);
1023
1024                 tx_dsg->addr    = isert_conn->login_rsp_dma;
1025                 tx_dsg->length  = length;
1026                 tx_dsg->lkey    = isert_conn->device->pd->local_dma_lkey;
1027                 tx_desc->num_sge = 2;
1028         }
1029         if (!login->login_failed) {
1030                 if (login->login_complete) {
1031                         ret = isert_alloc_rx_descriptors(isert_conn);
1032                         if (ret)
1033                                 return ret;
1034
1035                         ret = isert_post_recvm(isert_conn,
1036                                                ISERT_QP_MAX_RECV_DTOS);
1037                         if (ret)
1038                                 return ret;
1039
1040                         /* Now we are in FULL_FEATURE phase */
1041                         mutex_lock(&isert_conn->mutex);
1042                         isert_conn->state = ISER_CONN_FULL_FEATURE;
1043                         mutex_unlock(&isert_conn->mutex);
1044                         goto post_send;
1045                 }
1046
1047                 ret = isert_login_post_recv(isert_conn);
1048                 if (ret)
1049                         return ret;
1050         }
1051 post_send:
1052         ret = isert_login_post_send(isert_conn, tx_desc);
1053         if (ret)
1054                 return ret;
1055
1056         return 0;
1057 }
1058
1059 static void
1060 isert_rx_login_req(struct isert_conn *isert_conn)
1061 {
1062         struct iser_rx_desc *rx_desc = isert_conn->login_req_buf;
1063         int rx_buflen = isert_conn->login_req_len;
1064         struct iscsi_conn *conn = isert_conn->conn;
1065         struct iscsi_login *login = conn->conn_login;
1066         int size;
1067
1068         isert_info("conn %p\n", isert_conn);
1069
1070         WARN_ON_ONCE(!login);
1071
1072         if (login->first_request) {
1073                 struct iscsi_login_req *login_req =
1074                         (struct iscsi_login_req *)&rx_desc->iscsi_header;
1075                 /*
1076                  * Setup the initial iscsi_login values from the leading
1077                  * login request PDU.
1078                  */
1079                 login->leading_connection = (!login_req->tsih) ? 1 : 0;
1080                 login->current_stage =
1081                         (login_req->flags & ISCSI_FLAG_LOGIN_CURRENT_STAGE_MASK)
1082                          >> 2;
1083                 login->version_min      = login_req->min_version;
1084                 login->version_max      = login_req->max_version;
1085                 memcpy(login->isid, login_req->isid, 6);
1086                 login->cmd_sn           = be32_to_cpu(login_req->cmdsn);
1087                 login->init_task_tag    = login_req->itt;
1088                 login->initial_exp_statsn = be32_to_cpu(login_req->exp_statsn);
1089                 login->cid              = be16_to_cpu(login_req->cid);
1090                 login->tsih             = be16_to_cpu(login_req->tsih);
1091         }
1092
1093         memcpy(&login->req[0], (void *)&rx_desc->iscsi_header, ISCSI_HDR_LEN);
1094
1095         size = min(rx_buflen, MAX_KEY_VALUE_PAIRS);
1096         isert_dbg("Using login payload size: %d, rx_buflen: %d "
1097                   "MAX_KEY_VALUE_PAIRS: %d\n", size, rx_buflen,
1098                   MAX_KEY_VALUE_PAIRS);
1099         memcpy(login->req_buf, &rx_desc->data[0], size);
1100
1101         if (login->first_request) {
1102                 complete(&isert_conn->login_comp);
1103                 return;
1104         }
1105         schedule_delayed_work(&conn->login_work, 0);
1106 }
1107
1108 static struct iscsi_cmd
1109 *isert_allocate_cmd(struct iscsi_conn *conn, struct iser_rx_desc *rx_desc)
1110 {
1111         struct isert_conn *isert_conn = conn->context;
1112         struct isert_cmd *isert_cmd;
1113         struct iscsi_cmd *cmd;
1114
1115         cmd = iscsit_allocate_cmd(conn, TASK_INTERRUPTIBLE);
1116         if (!cmd) {
1117                 isert_err("Unable to allocate iscsi_cmd + isert_cmd\n");
1118                 return NULL;
1119         }
1120         isert_cmd = iscsit_priv_cmd(cmd);
1121         isert_cmd->conn = isert_conn;
1122         isert_cmd->iscsi_cmd = cmd;
1123         isert_cmd->rx_desc = rx_desc;
1124
1125         return cmd;
1126 }
1127
1128 static int
1129 isert_handle_scsi_cmd(struct isert_conn *isert_conn,
1130                       struct isert_cmd *isert_cmd, struct iscsi_cmd *cmd,
1131                       struct iser_rx_desc *rx_desc, unsigned char *buf)
1132 {
1133         struct iscsi_conn *conn = isert_conn->conn;
1134         struct iscsi_scsi_req *hdr = (struct iscsi_scsi_req *)buf;
1135         int imm_data, imm_data_len, unsol_data, sg_nents, rc;
1136         bool dump_payload = false;
1137         unsigned int data_len;
1138
1139         rc = iscsit_setup_scsi_cmd(conn, cmd, buf);
1140         if (rc < 0)
1141                 return rc;
1142
1143         imm_data = cmd->immediate_data;
1144         imm_data_len = cmd->first_burst_len;
1145         unsol_data = cmd->unsolicited_data;
1146         data_len = cmd->se_cmd.data_length;
1147
1148         if (imm_data && imm_data_len == data_len)
1149                 cmd->se_cmd.se_cmd_flags |= SCF_PASSTHROUGH_SG_TO_MEM_NOALLOC;
1150         rc = iscsit_process_scsi_cmd(conn, cmd, hdr);
1151         if (rc < 0) {
1152                 return 0;
1153         } else if (rc > 0) {
1154                 dump_payload = true;
1155                 goto sequence_cmd;
1156         }
1157
1158         if (!imm_data)
1159                 return 0;
1160
1161         if (imm_data_len != data_len) {
1162                 sg_nents = max(1UL, DIV_ROUND_UP(imm_data_len, PAGE_SIZE));
1163                 sg_copy_from_buffer(cmd->se_cmd.t_data_sg, sg_nents,
1164                                     &rx_desc->data[0], imm_data_len);
1165                 isert_dbg("Copy Immediate sg_nents: %u imm_data_len: %d\n",
1166                           sg_nents, imm_data_len);
1167         } else {
1168                 sg_init_table(&isert_cmd->sg, 1);
1169                 cmd->se_cmd.t_data_sg = &isert_cmd->sg;
1170                 cmd->se_cmd.t_data_nents = 1;
1171                 sg_set_buf(&isert_cmd->sg, &rx_desc->data[0], imm_data_len);
1172                 isert_dbg("Transfer Immediate imm_data_len: %d\n",
1173                           imm_data_len);
1174         }
1175
1176         cmd->write_data_done += imm_data_len;
1177
1178         if (cmd->write_data_done == cmd->se_cmd.data_length) {
1179                 spin_lock_bh(&cmd->istate_lock);
1180                 cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1181                 cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1182                 spin_unlock_bh(&cmd->istate_lock);
1183         }
1184
1185 sequence_cmd:
1186         rc = iscsit_sequence_cmd(conn, cmd, buf, hdr->cmdsn);
1187
1188         if (!rc && dump_payload == false && unsol_data)
1189                 iscsit_set_unsoliticed_dataout(cmd);
1190         else if (dump_payload && imm_data)
1191                 target_put_sess_cmd(&cmd->se_cmd);
1192
1193         return 0;
1194 }
1195
1196 static int
1197 isert_handle_iscsi_dataout(struct isert_conn *isert_conn,
1198                            struct iser_rx_desc *rx_desc, unsigned char *buf)
1199 {
1200         struct scatterlist *sg_start;
1201         struct iscsi_conn *conn = isert_conn->conn;
1202         struct iscsi_cmd *cmd = NULL;
1203         struct iscsi_data *hdr = (struct iscsi_data *)buf;
1204         u32 unsol_data_len = ntoh24(hdr->dlength);
1205         int rc, sg_nents, sg_off, page_off;
1206
1207         rc = iscsit_check_dataout_hdr(conn, buf, &cmd);
1208         if (rc < 0)
1209                 return rc;
1210         else if (!cmd)
1211                 return 0;
1212         /*
1213          * FIXME: Unexpected unsolicited_data out
1214          */
1215         if (!cmd->unsolicited_data) {
1216                 isert_err("Received unexpected solicited data payload\n");
1217                 dump_stack();
1218                 return -1;
1219         }
1220
1221         isert_dbg("Unsolicited DataOut unsol_data_len: %u, "
1222                   "write_data_done: %u, data_length: %u\n",
1223                   unsol_data_len,  cmd->write_data_done,
1224                   cmd->se_cmd.data_length);
1225
1226         sg_off = cmd->write_data_done / PAGE_SIZE;
1227         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
1228         sg_nents = max(1UL, DIV_ROUND_UP(unsol_data_len, PAGE_SIZE));
1229         page_off = cmd->write_data_done % PAGE_SIZE;
1230         /*
1231          * FIXME: Non page-aligned unsolicited_data out
1232          */
1233         if (page_off) {
1234                 isert_err("unexpected non-page aligned data payload\n");
1235                 dump_stack();
1236                 return -1;
1237         }
1238         isert_dbg("Copying DataOut: sg_start: %p, sg_off: %u "
1239                   "sg_nents: %u from %p %u\n", sg_start, sg_off,
1240                   sg_nents, &rx_desc->data[0], unsol_data_len);
1241
1242         sg_copy_from_buffer(sg_start, sg_nents, &rx_desc->data[0],
1243                             unsol_data_len);
1244
1245         rc = iscsit_check_dataout_payload(cmd, hdr, false);
1246         if (rc < 0)
1247                 return rc;
1248
1249         /*
1250          * multiple data-outs on the same command can arrive -
1251          * so post the buffer before hand
1252          */
1253         rc = isert_post_recv(isert_conn, rx_desc);
1254         if (rc) {
1255                 isert_err("ib_post_recv failed with %d\n", rc);
1256                 return rc;
1257         }
1258         return 0;
1259 }
1260
1261 static int
1262 isert_handle_nop_out(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1263                      struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1264                      unsigned char *buf)
1265 {
1266         struct iscsi_conn *conn = isert_conn->conn;
1267         struct iscsi_nopout *hdr = (struct iscsi_nopout *)buf;
1268         int rc;
1269
1270         rc = iscsit_setup_nop_out(conn, cmd, hdr);
1271         if (rc < 0)
1272                 return rc;
1273         /*
1274          * FIXME: Add support for NOPOUT payload using unsolicited RDMA payload
1275          */
1276
1277         return iscsit_process_nop_out(conn, cmd, hdr);
1278 }
1279
1280 static int
1281 isert_handle_text_cmd(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1282                       struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1283                       struct iscsi_text *hdr)
1284 {
1285         struct iscsi_conn *conn = isert_conn->conn;
1286         u32 payload_length = ntoh24(hdr->dlength);
1287         int rc;
1288         unsigned char *text_in = NULL;
1289
1290         rc = iscsit_setup_text_cmd(conn, cmd, hdr);
1291         if (rc < 0)
1292                 return rc;
1293
1294         if (payload_length) {
1295                 text_in = kzalloc(payload_length, GFP_KERNEL);
1296                 if (!text_in)
1297                         return -ENOMEM;
1298         }
1299         cmd->text_in_ptr = text_in;
1300
1301         memcpy(cmd->text_in_ptr, &rx_desc->data[0], payload_length);
1302
1303         return iscsit_process_text_cmd(conn, cmd, hdr);
1304 }
1305
1306 static int
1307 isert_rx_opcode(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc,
1308                 uint32_t read_stag, uint64_t read_va,
1309                 uint32_t write_stag, uint64_t write_va)
1310 {
1311         struct iscsi_hdr *hdr = &rx_desc->iscsi_header;
1312         struct iscsi_conn *conn = isert_conn->conn;
1313         struct iscsi_cmd *cmd;
1314         struct isert_cmd *isert_cmd;
1315         int ret = -EINVAL;
1316         u8 opcode = (hdr->opcode & ISCSI_OPCODE_MASK);
1317
1318         if (conn->sess->sess_ops->SessionType &&
1319            (!(opcode & ISCSI_OP_TEXT) || !(opcode & ISCSI_OP_LOGOUT))) {
1320                 isert_err("Got illegal opcode: 0x%02x in SessionType=Discovery,"
1321                           " ignoring\n", opcode);
1322                 return 0;
1323         }
1324
1325         switch (opcode) {
1326         case ISCSI_OP_SCSI_CMD:
1327                 cmd = isert_allocate_cmd(conn, rx_desc);
1328                 if (!cmd)
1329                         break;
1330
1331                 isert_cmd = iscsit_priv_cmd(cmd);
1332                 isert_cmd->read_stag = read_stag;
1333                 isert_cmd->read_va = read_va;
1334                 isert_cmd->write_stag = write_stag;
1335                 isert_cmd->write_va = write_va;
1336                 isert_cmd->inv_rkey = read_stag ? read_stag : write_stag;
1337
1338                 ret = isert_handle_scsi_cmd(isert_conn, isert_cmd, cmd,
1339                                         rx_desc, (unsigned char *)hdr);
1340                 break;
1341         case ISCSI_OP_NOOP_OUT:
1342                 cmd = isert_allocate_cmd(conn, rx_desc);
1343                 if (!cmd)
1344                         break;
1345
1346                 isert_cmd = iscsit_priv_cmd(cmd);
1347                 ret = isert_handle_nop_out(isert_conn, isert_cmd, cmd,
1348                                            rx_desc, (unsigned char *)hdr);
1349                 break;
1350         case ISCSI_OP_SCSI_DATA_OUT:
1351                 ret = isert_handle_iscsi_dataout(isert_conn, rx_desc,
1352                                                 (unsigned char *)hdr);
1353                 break;
1354         case ISCSI_OP_SCSI_TMFUNC:
1355                 cmd = isert_allocate_cmd(conn, rx_desc);
1356                 if (!cmd)
1357                         break;
1358
1359                 ret = iscsit_handle_task_mgt_cmd(conn, cmd,
1360                                                 (unsigned char *)hdr);
1361                 break;
1362         case ISCSI_OP_LOGOUT:
1363                 cmd = isert_allocate_cmd(conn, rx_desc);
1364                 if (!cmd)
1365                         break;
1366
1367                 ret = iscsit_handle_logout_cmd(conn, cmd, (unsigned char *)hdr);
1368                 break;
1369         case ISCSI_OP_TEXT:
1370                 if (be32_to_cpu(hdr->ttt) != 0xFFFFFFFF)
1371                         cmd = iscsit_find_cmd_from_itt(conn, hdr->itt);
1372                 else
1373                         cmd = isert_allocate_cmd(conn, rx_desc);
1374
1375                 if (!cmd)
1376                         break;
1377
1378                 isert_cmd = iscsit_priv_cmd(cmd);
1379                 ret = isert_handle_text_cmd(isert_conn, isert_cmd, cmd,
1380                                             rx_desc, (struct iscsi_text *)hdr);
1381                 break;
1382         default:
1383                 isert_err("Got unknown iSCSI OpCode: 0x%02x\n", opcode);
1384                 dump_stack();
1385                 break;
1386         }
1387
1388         return ret;
1389 }
1390
1391 static void
1392 isert_print_wc(struct ib_wc *wc, const char *type)
1393 {
1394         if (wc->status != IB_WC_WR_FLUSH_ERR)
1395                 isert_err("%s failure: %s (%d) vend_err %x\n", type,
1396                           ib_wc_status_msg(wc->status), wc->status,
1397                           wc->vendor_err);
1398         else
1399                 isert_dbg("%s failure: %s (%d)\n", type,
1400                           ib_wc_status_msg(wc->status), wc->status);
1401 }
1402
1403 static void
1404 isert_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1405 {
1406         struct isert_conn *isert_conn = wc->qp->qp_context;
1407         struct ib_device *ib_dev = isert_conn->cm_id->device;
1408         struct iser_rx_desc *rx_desc = cqe_to_rx_desc(wc->wr_cqe);
1409         struct iscsi_hdr *hdr = &rx_desc->iscsi_header;
1410         struct iser_ctrl *iser_ctrl = &rx_desc->iser_header;
1411         uint64_t read_va = 0, write_va = 0;
1412         uint32_t read_stag = 0, write_stag = 0;
1413
1414         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1415                 isert_print_wc(wc, "recv");
1416                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1417                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1418                 return;
1419         }
1420
1421         rx_desc->in_use = true;
1422
1423         ib_dma_sync_single_for_cpu(ib_dev, rx_desc->dma_addr,
1424                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1425
1426         isert_dbg("DMA: 0x%llx, iSCSI opcode: 0x%02x, ITT: 0x%08x, flags: 0x%02x dlen: %d\n",
1427                  rx_desc->dma_addr, hdr->opcode, hdr->itt, hdr->flags,
1428                  (int)(wc->byte_len - ISER_HEADERS_LEN));
1429
1430         switch (iser_ctrl->flags & 0xF0) {
1431         case ISCSI_CTRL:
1432                 if (iser_ctrl->flags & ISER_RSV) {
1433                         read_stag = be32_to_cpu(iser_ctrl->read_stag);
1434                         read_va = be64_to_cpu(iser_ctrl->read_va);
1435                         isert_dbg("ISER_RSV: read_stag: 0x%x read_va: 0x%llx\n",
1436                                   read_stag, (unsigned long long)read_va);
1437                 }
1438                 if (iser_ctrl->flags & ISER_WSV) {
1439                         write_stag = be32_to_cpu(iser_ctrl->write_stag);
1440                         write_va = be64_to_cpu(iser_ctrl->write_va);
1441                         isert_dbg("ISER_WSV: write_stag: 0x%x write_va: 0x%llx\n",
1442                                   write_stag, (unsigned long long)write_va);
1443                 }
1444
1445                 isert_dbg("ISER ISCSI_CTRL PDU\n");
1446                 break;
1447         case ISER_HELLO:
1448                 isert_err("iSER Hello message\n");
1449                 break;
1450         default:
1451                 isert_warn("Unknown iSER hdr flags: 0x%02x\n", iser_ctrl->flags);
1452                 break;
1453         }
1454
1455         isert_rx_opcode(isert_conn, rx_desc,
1456                         read_stag, read_va, write_stag, write_va);
1457
1458         ib_dma_sync_single_for_device(ib_dev, rx_desc->dma_addr,
1459                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1460 }
1461
1462 static void
1463 isert_login_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1464 {
1465         struct isert_conn *isert_conn = wc->qp->qp_context;
1466         struct ib_device *ib_dev = isert_conn->device->ib_device;
1467
1468         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1469                 isert_print_wc(wc, "login recv");
1470                 return;
1471         }
1472
1473         ib_dma_sync_single_for_cpu(ib_dev, isert_conn->login_req_dma,
1474                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1475
1476         isert_conn->login_req_len = wc->byte_len - ISER_HEADERS_LEN;
1477
1478         if (isert_conn->conn) {
1479                 struct iscsi_login *login = isert_conn->conn->conn_login;
1480
1481                 if (login && !login->first_request)
1482                         isert_rx_login_req(isert_conn);
1483         }
1484
1485         mutex_lock(&isert_conn->mutex);
1486         complete(&isert_conn->login_req_comp);
1487         mutex_unlock(&isert_conn->mutex);
1488
1489         ib_dma_sync_single_for_device(ib_dev, isert_conn->login_req_dma,
1490                                 ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1491 }
1492
1493 static void
1494 isert_rdma_rw_ctx_destroy(struct isert_cmd *cmd, struct isert_conn *conn)
1495 {
1496         struct se_cmd *se_cmd = &cmd->iscsi_cmd->se_cmd;
1497         enum dma_data_direction dir = target_reverse_dma_direction(se_cmd);
1498
1499         if (!cmd->rw.nr_ops)
1500                 return;
1501
1502         if (isert_prot_cmd(conn, se_cmd)) {
1503                 rdma_rw_ctx_destroy_signature(&cmd->rw, conn->qp,
1504                                 conn->cm_id->port_num, se_cmd->t_data_sg,
1505                                 se_cmd->t_data_nents, se_cmd->t_prot_sg,
1506                                 se_cmd->t_prot_nents, dir);
1507         } else {
1508                 rdma_rw_ctx_destroy(&cmd->rw, conn->qp, conn->cm_id->port_num,
1509                                 se_cmd->t_data_sg, se_cmd->t_data_nents, dir);
1510         }
1511
1512         cmd->rw.nr_ops = 0;
1513 }
1514
1515 static void
1516 isert_put_cmd(struct isert_cmd *isert_cmd, bool comp_err)
1517 {
1518         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1519         struct isert_conn *isert_conn = isert_cmd->conn;
1520         struct iscsi_conn *conn = isert_conn->conn;
1521         struct iscsi_text_rsp *hdr;
1522
1523         isert_dbg("Cmd %p\n", isert_cmd);
1524
1525         switch (cmd->iscsi_opcode) {
1526         case ISCSI_OP_SCSI_CMD:
1527                 spin_lock_bh(&conn->cmd_lock);
1528                 if (!list_empty(&cmd->i_conn_node))
1529                         list_del_init(&cmd->i_conn_node);
1530                 spin_unlock_bh(&conn->cmd_lock);
1531
1532                 if (cmd->data_direction == DMA_TO_DEVICE) {
1533                         iscsit_stop_dataout_timer(cmd);
1534                         /*
1535                          * Check for special case during comp_err where
1536                          * WRITE_PENDING has been handed off from core,
1537                          * but requires an extra target_put_sess_cmd()
1538                          * before transport_generic_free_cmd() below.
1539                          */
1540                         if (comp_err &&
1541                             cmd->se_cmd.t_state == TRANSPORT_WRITE_PENDING) {
1542                                 struct se_cmd *se_cmd = &cmd->se_cmd;
1543
1544                                 target_put_sess_cmd(se_cmd);
1545                         }
1546                 }
1547
1548                 isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1549                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1550                 break;
1551         case ISCSI_OP_SCSI_TMFUNC:
1552                 spin_lock_bh(&conn->cmd_lock);
1553                 if (!list_empty(&cmd->i_conn_node))
1554                         list_del_init(&cmd->i_conn_node);
1555                 spin_unlock_bh(&conn->cmd_lock);
1556
1557                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1558                 break;
1559         case ISCSI_OP_REJECT:
1560         case ISCSI_OP_NOOP_OUT:
1561         case ISCSI_OP_TEXT:
1562                 hdr = (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
1563                 /* If the continue bit is on, keep the command alive */
1564                 if (hdr->flags & ISCSI_FLAG_TEXT_CONTINUE)
1565                         break;
1566
1567                 spin_lock_bh(&conn->cmd_lock);
1568                 if (!list_empty(&cmd->i_conn_node))
1569                         list_del_init(&cmd->i_conn_node);
1570                 spin_unlock_bh(&conn->cmd_lock);
1571
1572                 /*
1573                  * Handle special case for REJECT when iscsi_add_reject*() has
1574                  * overwritten the original iscsi_opcode assignment, and the
1575                  * associated cmd->se_cmd needs to be released.
1576                  */
1577                 if (cmd->se_cmd.se_tfo != NULL) {
1578                         isert_dbg("Calling transport_generic_free_cmd for 0x%02x\n",
1579                                  cmd->iscsi_opcode);
1580                         transport_generic_free_cmd(&cmd->se_cmd, 0);
1581                         break;
1582                 }
1583                 /* fall through */
1584         default:
1585                 iscsit_release_cmd(cmd);
1586                 break;
1587         }
1588 }
1589
1590 static void
1591 isert_unmap_tx_desc(struct iser_tx_desc *tx_desc, struct ib_device *ib_dev)
1592 {
1593         if (tx_desc->dma_addr != 0) {
1594                 isert_dbg("unmap single for tx_desc->dma_addr\n");
1595                 ib_dma_unmap_single(ib_dev, tx_desc->dma_addr,
1596                                     ISER_HEADERS_LEN, DMA_TO_DEVICE);
1597                 tx_desc->dma_addr = 0;
1598         }
1599 }
1600
1601 static void
1602 isert_completion_put(struct iser_tx_desc *tx_desc, struct isert_cmd *isert_cmd,
1603                      struct ib_device *ib_dev, bool comp_err)
1604 {
1605         if (isert_cmd->pdu_buf_dma != 0) {
1606                 isert_dbg("unmap single for isert_cmd->pdu_buf_dma\n");
1607                 ib_dma_unmap_single(ib_dev, isert_cmd->pdu_buf_dma,
1608                                     isert_cmd->pdu_buf_len, DMA_TO_DEVICE);
1609                 isert_cmd->pdu_buf_dma = 0;
1610         }
1611
1612         isert_unmap_tx_desc(tx_desc, ib_dev);
1613         isert_put_cmd(isert_cmd, comp_err);
1614 }
1615
1616 static int
1617 isert_check_pi_status(struct se_cmd *se_cmd, struct ib_mr *sig_mr)
1618 {
1619         struct ib_mr_status mr_status;
1620         int ret;
1621
1622         ret = ib_check_mr_status(sig_mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
1623         if (ret) {
1624                 isert_err("ib_check_mr_status failed, ret %d\n", ret);
1625                 goto fail_mr_status;
1626         }
1627
1628         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1629                 u64 sec_offset_err;
1630                 u32 block_size = se_cmd->se_dev->dev_attrib.block_size + 8;
1631
1632                 switch (mr_status.sig_err.err_type) {
1633                 case IB_SIG_BAD_GUARD:
1634                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_GUARD_CHECK_FAILED;
1635                         break;
1636                 case IB_SIG_BAD_REFTAG:
1637                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_REF_TAG_CHECK_FAILED;
1638                         break;
1639                 case IB_SIG_BAD_APPTAG:
1640                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_APP_TAG_CHECK_FAILED;
1641                         break;
1642                 }
1643                 sec_offset_err = mr_status.sig_err.sig_err_offset;
1644                 do_div(sec_offset_err, block_size);
1645                 se_cmd->bad_sector = sec_offset_err + se_cmd->t_task_lba;
1646
1647                 isert_err("PI error found type %d at sector 0x%llx "
1648                           "expected 0x%x vs actual 0x%x\n",
1649                           mr_status.sig_err.err_type,
1650                           (unsigned long long)se_cmd->bad_sector,
1651                           mr_status.sig_err.expected,
1652                           mr_status.sig_err.actual);
1653                 ret = 1;
1654         }
1655
1656 fail_mr_status:
1657         return ret;
1658 }
1659
1660 static void
1661 isert_rdma_write_done(struct ib_cq *cq, struct ib_wc *wc)
1662 {
1663         struct isert_conn *isert_conn = wc->qp->qp_context;
1664         struct isert_device *device = isert_conn->device;
1665         struct iser_tx_desc *desc = cqe_to_tx_desc(wc->wr_cqe);
1666         struct isert_cmd *isert_cmd = tx_desc_to_cmd(desc);
1667         struct se_cmd *cmd = &isert_cmd->iscsi_cmd->se_cmd;
1668         int ret = 0;
1669
1670         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1671                 isert_print_wc(wc, "rdma write");
1672                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1673                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1674                 isert_completion_put(desc, isert_cmd, device->ib_device, true);
1675                 return;
1676         }
1677
1678         isert_dbg("Cmd %p\n", isert_cmd);
1679
1680         ret = isert_check_pi_status(cmd, isert_cmd->rw.sig->sig_mr);
1681         isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1682
1683         if (ret) {
1684                 /*
1685                  * transport_generic_request_failure() expects to have
1686                  * plus two references to handle queue-full, so re-add
1687                  * one here as target-core will have already dropped
1688                  * it after the first isert_put_datain() callback.
1689                  */
1690                 kref_get(&cmd->cmd_kref);
1691                 transport_generic_request_failure(cmd, cmd->pi_err);
1692         } else {
1693                 /*
1694                  * XXX: isert_put_response() failure is not retried.
1695                  */
1696                 ret = isert_put_response(isert_conn->conn, isert_cmd->iscsi_cmd);
1697                 if (ret)
1698                         pr_warn_ratelimited("isert_put_response() ret: %d\n", ret);
1699         }
1700 }
1701
1702 static void
1703 isert_rdma_read_done(struct ib_cq *cq, struct ib_wc *wc)
1704 {
1705         struct isert_conn *isert_conn = wc->qp->qp_context;
1706         struct isert_device *device = isert_conn->device;
1707         struct iser_tx_desc *desc = cqe_to_tx_desc(wc->wr_cqe);
1708         struct isert_cmd *isert_cmd = tx_desc_to_cmd(desc);
1709         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1710         struct se_cmd *se_cmd = &cmd->se_cmd;
1711         int ret = 0;
1712
1713         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1714                 isert_print_wc(wc, "rdma read");
1715                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1716                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1717                 isert_completion_put(desc, isert_cmd, device->ib_device, true);
1718                 return;
1719         }
1720
1721         isert_dbg("Cmd %p\n", isert_cmd);
1722
1723         iscsit_stop_dataout_timer(cmd);
1724
1725         if (isert_prot_cmd(isert_conn, se_cmd))
1726                 ret = isert_check_pi_status(se_cmd, isert_cmd->rw.sig->sig_mr);
1727         isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1728         cmd->write_data_done = 0;
1729
1730         isert_dbg("Cmd: %p RDMA_READ comp calling execute_cmd\n", isert_cmd);
1731         spin_lock_bh(&cmd->istate_lock);
1732         cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1733         cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1734         spin_unlock_bh(&cmd->istate_lock);
1735
1736         /*
1737          * transport_generic_request_failure() will drop the extra
1738          * se_cmd->cmd_kref reference after T10-PI error, and handle
1739          * any non-zero ->queue_status() callback error retries.
1740          */
1741         if (ret)
1742                 transport_generic_request_failure(se_cmd, se_cmd->pi_err);
1743         else
1744                 target_execute_cmd(se_cmd);
1745 }
1746
1747 static void
1748 isert_do_control_comp(struct work_struct *work)
1749 {
1750         struct isert_cmd *isert_cmd = container_of(work,
1751                         struct isert_cmd, comp_work);
1752         struct isert_conn *isert_conn = isert_cmd->conn;
1753         struct ib_device *ib_dev = isert_conn->cm_id->device;
1754         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1755
1756         isert_dbg("Cmd %p i_state %d\n", isert_cmd, cmd->i_state);
1757
1758         switch (cmd->i_state) {
1759         case ISTATE_SEND_TASKMGTRSP:
1760                 iscsit_tmr_post_handler(cmd, cmd->conn);
1761                 /* fall through */
1762         case ISTATE_SEND_REJECT:
1763         case ISTATE_SEND_TEXTRSP:
1764                 cmd->i_state = ISTATE_SENT_STATUS;
1765                 isert_completion_put(&isert_cmd->tx_desc, isert_cmd,
1766                                      ib_dev, false);
1767                 break;
1768         case ISTATE_SEND_LOGOUTRSP:
1769                 iscsit_logout_post_handler(cmd, cmd->conn);
1770                 break;
1771         default:
1772                 isert_err("Unknown i_state %d\n", cmd->i_state);
1773                 dump_stack();
1774                 break;
1775         }
1776 }
1777
1778 static void
1779 isert_login_send_done(struct ib_cq *cq, struct ib_wc *wc)
1780 {
1781         struct isert_conn *isert_conn = wc->qp->qp_context;
1782         struct ib_device *ib_dev = isert_conn->cm_id->device;
1783         struct iser_tx_desc *tx_desc = cqe_to_tx_desc(wc->wr_cqe);
1784
1785         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1786                 isert_print_wc(wc, "login send");
1787                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1788                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1789         }
1790
1791         isert_unmap_tx_desc(tx_desc, ib_dev);
1792 }
1793
1794 static void
1795 isert_send_done(struct ib_cq *cq, struct ib_wc *wc)
1796 {
1797         struct isert_conn *isert_conn = wc->qp->qp_context;
1798         struct ib_device *ib_dev = isert_conn->cm_id->device;
1799         struct iser_tx_desc *tx_desc = cqe_to_tx_desc(wc->wr_cqe);
1800         struct isert_cmd *isert_cmd = tx_desc_to_cmd(tx_desc);
1801
1802         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1803                 isert_print_wc(wc, "send");
1804                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1805                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1806                 isert_completion_put(tx_desc, isert_cmd, ib_dev, true);
1807                 return;
1808         }
1809
1810         isert_dbg("Cmd %p\n", isert_cmd);
1811
1812         switch (isert_cmd->iscsi_cmd->i_state) {
1813         case ISTATE_SEND_TASKMGTRSP:
1814         case ISTATE_SEND_LOGOUTRSP:
1815         case ISTATE_SEND_REJECT:
1816         case ISTATE_SEND_TEXTRSP:
1817                 isert_unmap_tx_desc(tx_desc, ib_dev);
1818
1819                 INIT_WORK(&isert_cmd->comp_work, isert_do_control_comp);
1820                 queue_work(isert_comp_wq, &isert_cmd->comp_work);
1821                 return;
1822         default:
1823                 isert_cmd->iscsi_cmd->i_state = ISTATE_SENT_STATUS;
1824                 isert_completion_put(tx_desc, isert_cmd, ib_dev, false);
1825                 break;
1826         }
1827 }
1828
1829 static int
1830 isert_post_response(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd)
1831 {
1832         int ret;
1833
1834         ret = isert_post_recv(isert_conn, isert_cmd->rx_desc);
1835         if (ret) {
1836                 isert_err("ib_post_recv failed with %d\n", ret);
1837                 return ret;
1838         }
1839
1840         ret = ib_post_send(isert_conn->qp, &isert_cmd->tx_desc.send_wr, NULL);
1841         if (ret) {
1842                 isert_err("ib_post_send failed with %d\n", ret);
1843                 return ret;
1844         }
1845         return ret;
1846 }
1847
1848 static int
1849 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
1850 {
1851         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1852         struct isert_conn *isert_conn = conn->context;
1853         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1854         struct iscsi_scsi_rsp *hdr = (struct iscsi_scsi_rsp *)
1855                                 &isert_cmd->tx_desc.iscsi_header;
1856
1857         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1858         iscsit_build_rsp_pdu(cmd, conn, true, hdr);
1859         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1860         /*
1861          * Attach SENSE DATA payload to iSCSI Response PDU
1862          */
1863         if (cmd->se_cmd.sense_buffer &&
1864             ((cmd->se_cmd.se_cmd_flags & SCF_TRANSPORT_TASK_SENSE) ||
1865             (cmd->se_cmd.se_cmd_flags & SCF_EMULATED_TASK_SENSE))) {
1866                 struct isert_device *device = isert_conn->device;
1867                 struct ib_device *ib_dev = device->ib_device;
1868                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
1869                 u32 padding, pdu_len;
1870
1871                 put_unaligned_be16(cmd->se_cmd.scsi_sense_length,
1872                                    cmd->sense_buffer);
1873                 cmd->se_cmd.scsi_sense_length += sizeof(__be16);
1874
1875                 padding = -(cmd->se_cmd.scsi_sense_length) & 3;
1876                 hton24(hdr->dlength, (u32)cmd->se_cmd.scsi_sense_length);
1877                 pdu_len = cmd->se_cmd.scsi_sense_length + padding;
1878
1879                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
1880                                 (void *)cmd->sense_buffer, pdu_len,
1881                                 DMA_TO_DEVICE);
1882                 if (ib_dma_mapping_error(ib_dev, isert_cmd->pdu_buf_dma))
1883                         return -ENOMEM;
1884
1885                 isert_cmd->pdu_buf_len = pdu_len;
1886                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
1887                 tx_dsg->length  = pdu_len;
1888                 tx_dsg->lkey    = device->pd->local_dma_lkey;
1889                 isert_cmd->tx_desc.num_sge = 2;
1890         }
1891
1892         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1893
1894         isert_dbg("Posting SCSI Response\n");
1895
1896         return isert_post_response(isert_conn, isert_cmd);
1897 }
1898
1899 static void
1900 isert_aborted_task(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
1901 {
1902         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1903         struct isert_conn *isert_conn = conn->context;
1904
1905         spin_lock_bh(&conn->cmd_lock);
1906         if (!list_empty(&cmd->i_conn_node))
1907                 list_del_init(&cmd->i_conn_node);
1908         spin_unlock_bh(&conn->cmd_lock);
1909
1910         if (cmd->data_direction == DMA_TO_DEVICE)
1911                 iscsit_stop_dataout_timer(cmd);
1912         isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1913 }
1914
1915 static enum target_prot_op
1916 isert_get_sup_prot_ops(struct iscsi_conn *conn)
1917 {
1918         struct isert_conn *isert_conn = conn->context;
1919         struct isert_device *device = isert_conn->device;
1920
1921         if (conn->tpg->tpg_attrib.t10_pi) {
1922                 if (device->pi_capable) {
1923                         isert_info("conn %p PI offload enabled\n", isert_conn);
1924                         isert_conn->pi_support = true;
1925                         return TARGET_PROT_ALL;
1926                 }
1927         }
1928
1929         isert_info("conn %p PI offload disabled\n", isert_conn);
1930         isert_conn->pi_support = false;
1931
1932         return TARGET_PROT_NORMAL;
1933 }
1934
1935 static int
1936 isert_put_nopin(struct iscsi_cmd *cmd, struct iscsi_conn *conn,
1937                 bool nopout_response)
1938 {
1939         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1940         struct isert_conn *isert_conn = conn->context;
1941         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1942
1943         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1944         iscsit_build_nopin_rsp(cmd, conn, (struct iscsi_nopin *)
1945                                &isert_cmd->tx_desc.iscsi_header,
1946                                nopout_response);
1947         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1948         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1949
1950         isert_dbg("conn %p Posting NOPIN Response\n", isert_conn);
1951
1952         return isert_post_response(isert_conn, isert_cmd);
1953 }
1954
1955 static int
1956 isert_put_logout_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
1957 {
1958         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1959         struct isert_conn *isert_conn = conn->context;
1960         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1961
1962         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1963         iscsit_build_logout_rsp(cmd, conn, (struct iscsi_logout_rsp *)
1964                                 &isert_cmd->tx_desc.iscsi_header);
1965         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1966         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1967
1968         isert_dbg("conn %p Posting Logout Response\n", isert_conn);
1969
1970         return isert_post_response(isert_conn, isert_cmd);
1971 }
1972
1973 static int
1974 isert_put_tm_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
1975 {
1976         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1977         struct isert_conn *isert_conn = conn->context;
1978         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1979
1980         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1981         iscsit_build_task_mgt_rsp(cmd, conn, (struct iscsi_tm_rsp *)
1982                                   &isert_cmd->tx_desc.iscsi_header);
1983         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1984         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1985
1986         isert_dbg("conn %p Posting Task Management Response\n", isert_conn);
1987
1988         return isert_post_response(isert_conn, isert_cmd);
1989 }
1990
1991 static int
1992 isert_put_reject(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
1993 {
1994         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1995         struct isert_conn *isert_conn = conn->context;
1996         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1997         struct isert_device *device = isert_conn->device;
1998         struct ib_device *ib_dev = device->ib_device;
1999         struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2000         struct iscsi_reject *hdr =
2001                 (struct iscsi_reject *)&isert_cmd->tx_desc.iscsi_header;
2002
2003         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2004         iscsit_build_reject(cmd, conn, hdr);
2005         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2006
2007         hton24(hdr->dlength, ISCSI_HDR_LEN);
2008         isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2009                         (void *)cmd->buf_ptr, ISCSI_HDR_LEN,
2010                         DMA_TO_DEVICE);
2011         if (ib_dma_mapping_error(ib_dev, isert_cmd->pdu_buf_dma))
2012                 return -ENOMEM;
2013         isert_cmd->pdu_buf_len = ISCSI_HDR_LEN;
2014         tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2015         tx_dsg->length  = ISCSI_HDR_LEN;
2016         tx_dsg->lkey    = device->pd->local_dma_lkey;
2017         isert_cmd->tx_desc.num_sge = 2;
2018
2019         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2020
2021         isert_dbg("conn %p Posting Reject\n", isert_conn);
2022
2023         return isert_post_response(isert_conn, isert_cmd);
2024 }
2025
2026 static int
2027 isert_put_text_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2028 {
2029         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2030         struct isert_conn *isert_conn = conn->context;
2031         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2032         struct iscsi_text_rsp *hdr =
2033                 (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
2034         u32 txt_rsp_len;
2035         int rc;
2036
2037         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2038         rc = iscsit_build_text_rsp(cmd, conn, hdr, ISCSI_INFINIBAND);
2039         if (rc < 0)
2040                 return rc;
2041
2042         txt_rsp_len = rc;
2043         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2044
2045         if (txt_rsp_len) {
2046                 struct isert_device *device = isert_conn->device;
2047                 struct ib_device *ib_dev = device->ib_device;
2048                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2049                 void *txt_rsp_buf = cmd->buf_ptr;
2050
2051                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2052                                 txt_rsp_buf, txt_rsp_len, DMA_TO_DEVICE);
2053                 if (ib_dma_mapping_error(ib_dev, isert_cmd->pdu_buf_dma))
2054                         return -ENOMEM;
2055
2056                 isert_cmd->pdu_buf_len = txt_rsp_len;
2057                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2058                 tx_dsg->length  = txt_rsp_len;
2059                 tx_dsg->lkey    = device->pd->local_dma_lkey;
2060                 isert_cmd->tx_desc.num_sge = 2;
2061         }
2062         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2063
2064         isert_dbg("conn %p Text Response\n", isert_conn);
2065
2066         return isert_post_response(isert_conn, isert_cmd);
2067 }
2068
2069 static inline void
2070 isert_set_dif_domain(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs,
2071                      struct ib_sig_domain *domain)
2072 {
2073         domain->sig_type = IB_SIG_TYPE_T10_DIF;
2074         domain->sig.dif.bg_type = IB_T10DIF_CRC;
2075         domain->sig.dif.pi_interval = se_cmd->se_dev->dev_attrib.block_size;
2076         domain->sig.dif.ref_tag = se_cmd->reftag_seed;
2077         /*
2078          * At the moment we hard code those, but if in the future
2079          * the target core would like to use it, we will take it
2080          * from se_cmd.
2081          */
2082         domain->sig.dif.apptag_check_mask = 0xffff;
2083         domain->sig.dif.app_escape = true;
2084         domain->sig.dif.ref_escape = true;
2085         if (se_cmd->prot_type == TARGET_DIF_TYPE1_PROT ||
2086             se_cmd->prot_type == TARGET_DIF_TYPE2_PROT)
2087                 domain->sig.dif.ref_remap = true;
2088 };
2089
2090 static int
2091 isert_set_sig_attrs(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs)
2092 {
2093         memset(sig_attrs, 0, sizeof(*sig_attrs));
2094
2095         switch (se_cmd->prot_op) {
2096         case TARGET_PROT_DIN_INSERT:
2097         case TARGET_PROT_DOUT_STRIP:
2098                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
2099                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2100                 break;
2101         case TARGET_PROT_DOUT_INSERT:
2102         case TARGET_PROT_DIN_STRIP:
2103                 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
2104                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2105                 break;
2106         case TARGET_PROT_DIN_PASS:
2107         case TARGET_PROT_DOUT_PASS:
2108                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2109                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2110                 break;
2111         default:
2112                 isert_err("Unsupported PI operation %d\n", se_cmd->prot_op);
2113                 return -EINVAL;
2114         }
2115
2116         if (se_cmd->prot_checks & TARGET_DIF_CHECK_GUARD)
2117                 sig_attrs->check_mask |= IB_SIG_CHECK_GUARD;
2118         if (se_cmd->prot_checks & TARGET_DIF_CHECK_APPTAG)
2119                 sig_attrs->check_mask |= IB_SIG_CHECK_APPTAG;
2120         if (se_cmd->prot_checks & TARGET_DIF_CHECK_REFTAG)
2121                 sig_attrs->check_mask |= IB_SIG_CHECK_REFTAG;
2122
2123         return 0;
2124 }
2125
2126 static int
2127 isert_rdma_rw_ctx_post(struct isert_cmd *cmd, struct isert_conn *conn,
2128                 struct ib_cqe *cqe, struct ib_send_wr *chain_wr)
2129 {
2130         struct se_cmd *se_cmd = &cmd->iscsi_cmd->se_cmd;
2131         enum dma_data_direction dir = target_reverse_dma_direction(se_cmd);
2132         u8 port_num = conn->cm_id->port_num;
2133         u64 addr;
2134         u32 rkey, offset;
2135         int ret;
2136
2137         if (cmd->ctx_init_done)
2138                 goto rdma_ctx_post;
2139
2140         if (dir == DMA_FROM_DEVICE) {
2141                 addr = cmd->write_va;
2142                 rkey = cmd->write_stag;
2143                 offset = cmd->iscsi_cmd->write_data_done;
2144         } else {
2145                 addr = cmd->read_va;
2146                 rkey = cmd->read_stag;
2147                 offset = 0;
2148         }
2149
2150         if (isert_prot_cmd(conn, se_cmd)) {
2151                 struct ib_sig_attrs sig_attrs;
2152
2153                 ret = isert_set_sig_attrs(se_cmd, &sig_attrs);
2154                 if (ret)
2155                         return ret;
2156
2157                 WARN_ON_ONCE(offset);
2158                 ret = rdma_rw_ctx_signature_init(&cmd->rw, conn->qp, port_num,
2159                                 se_cmd->t_data_sg, se_cmd->t_data_nents,
2160                                 se_cmd->t_prot_sg, se_cmd->t_prot_nents,
2161                                 &sig_attrs, addr, rkey, dir);
2162         } else {
2163                 ret = rdma_rw_ctx_init(&cmd->rw, conn->qp, port_num,
2164                                 se_cmd->t_data_sg, se_cmd->t_data_nents,
2165                                 offset, addr, rkey, dir);
2166         }
2167
2168         if (ret < 0) {
2169                 isert_err("Cmd: %p failed to prepare RDMA res\n", cmd);
2170                 return ret;
2171         }
2172
2173         cmd->ctx_init_done = true;
2174
2175 rdma_ctx_post:
2176         ret = rdma_rw_ctx_post(&cmd->rw, conn->qp, port_num, cqe, chain_wr);
2177         if (ret < 0)
2178                 isert_err("Cmd: %p failed to post RDMA res\n", cmd);
2179         return ret;
2180 }
2181
2182 static int
2183 isert_put_datain(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2184 {
2185         struct se_cmd *se_cmd = &cmd->se_cmd;
2186         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2187         struct isert_conn *isert_conn = conn->context;
2188         struct ib_cqe *cqe = NULL;
2189         struct ib_send_wr *chain_wr = NULL;
2190         int rc;
2191
2192         isert_dbg("Cmd: %p RDMA_WRITE data_length: %u\n",
2193                  isert_cmd, se_cmd->data_length);
2194
2195         if (isert_prot_cmd(isert_conn, se_cmd)) {
2196                 isert_cmd->tx_desc.tx_cqe.done = isert_rdma_write_done;
2197                 cqe = &isert_cmd->tx_desc.tx_cqe;
2198         } else {
2199                 /*
2200                  * Build isert_conn->tx_desc for iSCSI response PDU and attach
2201                  */
2202                 isert_create_send_desc(isert_conn, isert_cmd,
2203                                        &isert_cmd->tx_desc);
2204                 iscsit_build_rsp_pdu(cmd, conn, true, (struct iscsi_scsi_rsp *)
2205                                      &isert_cmd->tx_desc.iscsi_header);
2206                 isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2207                 isert_init_send_wr(isert_conn, isert_cmd,
2208                                    &isert_cmd->tx_desc.send_wr);
2209
2210                 rc = isert_post_recv(isert_conn, isert_cmd->rx_desc);
2211                 if (rc) {
2212                         isert_err("ib_post_recv failed with %d\n", rc);
2213                         return rc;
2214                 }
2215
2216                 chain_wr = &isert_cmd->tx_desc.send_wr;
2217         }
2218
2219         rc = isert_rdma_rw_ctx_post(isert_cmd, isert_conn, cqe, chain_wr);
2220         isert_dbg("Cmd: %p posted RDMA_WRITE for iSER Data READ rc: %d\n",
2221                   isert_cmd, rc);
2222         return rc;
2223 }
2224
2225 static int
2226 isert_get_dataout(struct iscsi_conn *conn, struct iscsi_cmd *cmd, bool recovery)
2227 {
2228         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2229         int ret;
2230
2231         isert_dbg("Cmd: %p RDMA_READ data_length: %u write_data_done: %u\n",
2232                  isert_cmd, cmd->se_cmd.data_length, cmd->write_data_done);
2233
2234         isert_cmd->tx_desc.tx_cqe.done = isert_rdma_read_done;
2235         ret = isert_rdma_rw_ctx_post(isert_cmd, conn->context,
2236                                      &isert_cmd->tx_desc.tx_cqe, NULL);
2237
2238         isert_dbg("Cmd: %p posted RDMA_READ memory for ISER Data WRITE rc: %d\n",
2239                  isert_cmd, ret);
2240         return ret;
2241 }
2242
2243 static int
2244 isert_immediate_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2245 {
2246         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2247         int ret = 0;
2248
2249         switch (state) {
2250         case ISTATE_REMOVE:
2251                 spin_lock_bh(&conn->cmd_lock);
2252                 list_del_init(&cmd->i_conn_node);
2253                 spin_unlock_bh(&conn->cmd_lock);
2254                 isert_put_cmd(isert_cmd, true);
2255                 break;
2256         case ISTATE_SEND_NOPIN_WANT_RESPONSE:
2257                 ret = isert_put_nopin(cmd, conn, false);
2258                 break;
2259         default:
2260                 isert_err("Unknown immediate state: 0x%02x\n", state);
2261                 ret = -EINVAL;
2262                 break;
2263         }
2264
2265         return ret;
2266 }
2267
2268 static int
2269 isert_response_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2270 {
2271         struct isert_conn *isert_conn = conn->context;
2272         int ret;
2273
2274         switch (state) {
2275         case ISTATE_SEND_LOGOUTRSP:
2276                 ret = isert_put_logout_rsp(cmd, conn);
2277                 if (!ret)
2278                         isert_conn->logout_posted = true;
2279                 break;
2280         case ISTATE_SEND_NOPIN:
2281                 ret = isert_put_nopin(cmd, conn, true);
2282                 break;
2283         case ISTATE_SEND_TASKMGTRSP:
2284                 ret = isert_put_tm_rsp(cmd, conn);
2285                 break;
2286         case ISTATE_SEND_REJECT:
2287                 ret = isert_put_reject(cmd, conn);
2288                 break;
2289         case ISTATE_SEND_TEXTRSP:
2290                 ret = isert_put_text_rsp(cmd, conn);
2291                 break;
2292         case ISTATE_SEND_STATUS:
2293                 /*
2294                  * Special case for sending non GOOD SCSI status from TX thread
2295                  * context during pre se_cmd excecution failure.
2296                  */
2297                 ret = isert_put_response(conn, cmd);
2298                 break;
2299         default:
2300                 isert_err("Unknown response state: 0x%02x\n", state);
2301                 ret = -EINVAL;
2302                 break;
2303         }
2304
2305         return ret;
2306 }
2307
2308 struct rdma_cm_id *
2309 isert_setup_id(struct isert_np *isert_np)
2310 {
2311         struct iscsi_np *np = isert_np->np;
2312         struct rdma_cm_id *id;
2313         struct sockaddr *sa;
2314         int ret;
2315
2316         sa = (struct sockaddr *)&np->np_sockaddr;
2317         isert_dbg("ksockaddr: %p, sa: %p\n", &np->np_sockaddr, sa);
2318
2319         id = rdma_create_id(&init_net, isert_cma_handler, isert_np,
2320                             RDMA_PS_TCP, IB_QPT_RC);
2321         if (IS_ERR(id)) {
2322                 isert_err("rdma_create_id() failed: %ld\n", PTR_ERR(id));
2323                 ret = PTR_ERR(id);
2324                 goto out;
2325         }
2326         isert_dbg("id %p context %p\n", id, id->context);
2327
2328         ret = rdma_bind_addr(id, sa);
2329         if (ret) {
2330                 isert_err("rdma_bind_addr() failed: %d\n", ret);
2331                 goto out_id;
2332         }
2333
2334         ret = rdma_listen(id, 0);
2335         if (ret) {
2336                 isert_err("rdma_listen() failed: %d\n", ret);
2337                 goto out_id;
2338         }
2339
2340         return id;
2341 out_id:
2342         rdma_destroy_id(id);
2343 out:
2344         return ERR_PTR(ret);
2345 }
2346
2347 static int
2348 isert_setup_np(struct iscsi_np *np,
2349                struct sockaddr_storage *ksockaddr)
2350 {
2351         struct isert_np *isert_np;
2352         struct rdma_cm_id *isert_lid;
2353         int ret;
2354
2355         isert_np = kzalloc(sizeof(struct isert_np), GFP_KERNEL);
2356         if (!isert_np)
2357                 return -ENOMEM;
2358
2359         sema_init(&isert_np->sem, 0);
2360         mutex_init(&isert_np->mutex);
2361         INIT_LIST_HEAD(&isert_np->accepted);
2362         INIT_LIST_HEAD(&isert_np->pending);
2363         isert_np->np = np;
2364
2365         /*
2366          * Setup the np->np_sockaddr from the passed sockaddr setup
2367          * in iscsi_target_configfs.c code..
2368          */
2369         memcpy(&np->np_sockaddr, ksockaddr,
2370                sizeof(struct sockaddr_storage));
2371
2372         isert_lid = isert_setup_id(isert_np);
2373         if (IS_ERR(isert_lid)) {
2374                 ret = PTR_ERR(isert_lid);
2375                 goto out;
2376         }
2377
2378         isert_np->cm_id = isert_lid;
2379         np->np_context = isert_np;
2380
2381         return 0;
2382
2383 out:
2384         kfree(isert_np);
2385
2386         return ret;
2387 }
2388
2389 static int
2390 isert_rdma_accept(struct isert_conn *isert_conn)
2391 {
2392         struct rdma_cm_id *cm_id = isert_conn->cm_id;
2393         struct rdma_conn_param cp;
2394         int ret;
2395         struct iser_cm_hdr rsp_hdr;
2396
2397         memset(&cp, 0, sizeof(struct rdma_conn_param));
2398         cp.initiator_depth = isert_conn->initiator_depth;
2399         cp.retry_count = 7;
2400         cp.rnr_retry_count = 7;
2401
2402         memset(&rsp_hdr, 0, sizeof(rsp_hdr));
2403         rsp_hdr.flags = ISERT_ZBVA_NOT_USED;
2404         if (!isert_conn->snd_w_inv)
2405                 rsp_hdr.flags = rsp_hdr.flags | ISERT_SEND_W_INV_NOT_USED;
2406         cp.private_data = (void *)&rsp_hdr;
2407         cp.private_data_len = sizeof(rsp_hdr);
2408
2409         ret = rdma_accept(cm_id, &cp);
2410         if (ret) {
2411                 isert_err("rdma_accept() failed with: %d\n", ret);
2412                 return ret;
2413         }
2414
2415         return 0;
2416 }
2417
2418 static int
2419 isert_get_login_rx(struct iscsi_conn *conn, struct iscsi_login *login)
2420 {
2421         struct isert_conn *isert_conn = conn->context;
2422         int ret;
2423
2424         isert_info("before login_req comp conn: %p\n", isert_conn);
2425         ret = wait_for_completion_interruptible(&isert_conn->login_req_comp);
2426         if (ret) {
2427                 isert_err("isert_conn %p interrupted before got login req\n",
2428                           isert_conn);
2429                 return ret;
2430         }
2431         reinit_completion(&isert_conn->login_req_comp);
2432
2433         /*
2434          * For login requests after the first PDU, isert_rx_login_req() will
2435          * kick schedule_delayed_work(&conn->login_work) as the packet is
2436          * received, which turns this callback from iscsi_target_do_login_rx()
2437          * into a NOP.
2438          */
2439         if (!login->first_request)
2440                 return 0;
2441
2442         isert_rx_login_req(isert_conn);
2443
2444         isert_info("before login_comp conn: %p\n", conn);
2445         ret = wait_for_completion_interruptible(&isert_conn->login_comp);
2446         if (ret)
2447                 return ret;
2448
2449         isert_info("processing login->req: %p\n", login->req);
2450
2451         return 0;
2452 }
2453
2454 static void
2455 isert_set_conn_info(struct iscsi_np *np, struct iscsi_conn *conn,
2456                     struct isert_conn *isert_conn)
2457 {
2458         struct rdma_cm_id *cm_id = isert_conn->cm_id;
2459         struct rdma_route *cm_route = &cm_id->route;
2460
2461         conn->login_family = np->np_sockaddr.ss_family;
2462
2463         conn->login_sockaddr = cm_route->addr.dst_addr;
2464         conn->local_sockaddr = cm_route->addr.src_addr;
2465 }
2466
2467 static int
2468 isert_accept_np(struct iscsi_np *np, struct iscsi_conn *conn)
2469 {
2470         struct isert_np *isert_np = np->np_context;
2471         struct isert_conn *isert_conn;
2472         int ret;
2473
2474 accept_wait:
2475         ret = down_interruptible(&isert_np->sem);
2476         if (ret)
2477                 return -ENODEV;
2478
2479         spin_lock_bh(&np->np_thread_lock);
2480         if (np->np_thread_state >= ISCSI_NP_THREAD_RESET) {
2481                 spin_unlock_bh(&np->np_thread_lock);
2482                 isert_dbg("np_thread_state %d\n",
2483                          np->np_thread_state);
2484                 /**
2485                  * No point in stalling here when np_thread
2486                  * is in state RESET/SHUTDOWN/EXIT - bail
2487                  **/
2488                 return -ENODEV;
2489         }
2490         spin_unlock_bh(&np->np_thread_lock);
2491
2492         mutex_lock(&isert_np->mutex);
2493         if (list_empty(&isert_np->pending)) {
2494                 mutex_unlock(&isert_np->mutex);
2495                 goto accept_wait;
2496         }
2497         isert_conn = list_first_entry(&isert_np->pending,
2498                         struct isert_conn, node);
2499         list_del_init(&isert_conn->node);
2500         mutex_unlock(&isert_np->mutex);
2501
2502         conn->context = isert_conn;
2503         isert_conn->conn = conn;
2504         isert_conn->state = ISER_CONN_BOUND;
2505
2506         isert_set_conn_info(np, conn, isert_conn);
2507
2508         isert_dbg("Processing isert_conn: %p\n", isert_conn);
2509
2510         return 0;
2511 }
2512
2513 static void
2514 isert_free_np(struct iscsi_np *np)
2515 {
2516         struct isert_np *isert_np = np->np_context;
2517         struct isert_conn *isert_conn, *n;
2518
2519         if (isert_np->cm_id)
2520                 rdma_destroy_id(isert_np->cm_id);
2521
2522         /*
2523          * FIXME: At this point we don't have a good way to insure
2524          * that at this point we don't have hanging connections that
2525          * completed RDMA establishment but didn't start iscsi login
2526          * process. So work-around this by cleaning up what ever piled
2527          * up in accepted and pending lists.
2528          */
2529         mutex_lock(&isert_np->mutex);
2530         if (!list_empty(&isert_np->pending)) {
2531                 isert_info("Still have isert pending connections\n");
2532                 list_for_each_entry_safe(isert_conn, n,
2533                                          &isert_np->pending,
2534                                          node) {
2535                         isert_info("cleaning isert_conn %p state (%d)\n",
2536                                    isert_conn, isert_conn->state);
2537                         isert_connect_release(isert_conn);
2538                 }
2539         }
2540
2541         if (!list_empty(&isert_np->accepted)) {
2542                 isert_info("Still have isert accepted connections\n");
2543                 list_for_each_entry_safe(isert_conn, n,
2544                                          &isert_np->accepted,
2545                                          node) {
2546                         isert_info("cleaning isert_conn %p state (%d)\n",
2547                                    isert_conn, isert_conn->state);
2548                         isert_connect_release(isert_conn);
2549                 }
2550         }
2551         mutex_unlock(&isert_np->mutex);
2552
2553         np->np_context = NULL;
2554         kfree(isert_np);
2555 }
2556
2557 static void isert_release_work(struct work_struct *work)
2558 {
2559         struct isert_conn *isert_conn = container_of(work,
2560                                                      struct isert_conn,
2561                                                      release_work);
2562
2563         isert_info("Starting release conn %p\n", isert_conn);
2564
2565         mutex_lock(&isert_conn->mutex);
2566         isert_conn->state = ISER_CONN_DOWN;
2567         mutex_unlock(&isert_conn->mutex);
2568
2569         isert_info("Destroying conn %p\n", isert_conn);
2570         isert_put_conn(isert_conn);
2571 }
2572
2573 static void
2574 isert_wait4logout(struct isert_conn *isert_conn)
2575 {
2576         struct iscsi_conn *conn = isert_conn->conn;
2577
2578         isert_info("conn %p\n", isert_conn);
2579
2580         if (isert_conn->logout_posted) {
2581                 isert_info("conn %p wait for conn_logout_comp\n", isert_conn);
2582                 wait_for_completion_timeout(&conn->conn_logout_comp,
2583                                             SECONDS_FOR_LOGOUT_COMP * HZ);
2584         }
2585 }
2586
2587 static void
2588 isert_wait4cmds(struct iscsi_conn *conn)
2589 {
2590         isert_info("iscsi_conn %p\n", conn);
2591
2592         if (conn->sess) {
2593                 target_sess_cmd_list_set_waiting(conn->sess->se_sess);
2594                 target_wait_for_sess_cmds(conn->sess->se_sess);
2595         }
2596 }
2597
2598 /**
2599  * isert_put_unsol_pending_cmds() - Drop commands waiting for
2600  *     unsolicitate dataout
2601  * @conn:    iscsi connection
2602  *
2603  * We might still have commands that are waiting for unsolicited
2604  * dataouts messages. We must put the extra reference on those
2605  * before blocking on the target_wait_for_session_cmds
2606  */
2607 static void
2608 isert_put_unsol_pending_cmds(struct iscsi_conn *conn)
2609 {
2610         struct iscsi_cmd *cmd, *tmp;
2611         static LIST_HEAD(drop_cmd_list);
2612
2613         spin_lock_bh(&conn->cmd_lock);
2614         list_for_each_entry_safe(cmd, tmp, &conn->conn_cmd_list, i_conn_node) {
2615                 if ((cmd->cmd_flags & ICF_NON_IMMEDIATE_UNSOLICITED_DATA) &&
2616                     (cmd->write_data_done < conn->sess->sess_ops->FirstBurstLength) &&
2617                     (cmd->write_data_done < cmd->se_cmd.data_length))
2618                         list_move_tail(&cmd->i_conn_node, &drop_cmd_list);
2619         }
2620         spin_unlock_bh(&conn->cmd_lock);
2621
2622         list_for_each_entry_safe(cmd, tmp, &drop_cmd_list, i_conn_node) {
2623                 list_del_init(&cmd->i_conn_node);
2624                 if (cmd->i_state != ISTATE_REMOVE) {
2625                         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2626
2627                         isert_info("conn %p dropping cmd %p\n", conn, cmd);
2628                         isert_put_cmd(isert_cmd, true);
2629                 }
2630         }
2631 }
2632
2633 static void isert_wait_conn(struct iscsi_conn *conn)
2634 {
2635         struct isert_conn *isert_conn = conn->context;
2636
2637         isert_info("Starting conn %p\n", isert_conn);
2638
2639         mutex_lock(&isert_conn->mutex);
2640         isert_conn_terminate(isert_conn);
2641         mutex_unlock(&isert_conn->mutex);
2642
2643         ib_drain_qp(isert_conn->qp);
2644         isert_put_unsol_pending_cmds(conn);
2645         isert_wait4cmds(conn);
2646         isert_wait4logout(isert_conn);
2647
2648         queue_work(isert_release_wq, &isert_conn->release_work);
2649 }
2650
2651 static void isert_free_conn(struct iscsi_conn *conn)
2652 {
2653         struct isert_conn *isert_conn = conn->context;
2654
2655         ib_drain_qp(isert_conn->qp);
2656         isert_put_conn(isert_conn);
2657 }
2658
2659 static void isert_get_rx_pdu(struct iscsi_conn *conn)
2660 {
2661         struct completion comp;
2662
2663         init_completion(&comp);
2664
2665         wait_for_completion_interruptible(&comp);
2666 }
2667
2668 static struct iscsit_transport iser_target_transport = {
2669         .name                   = "IB/iSER",
2670         .transport_type         = ISCSI_INFINIBAND,
2671         .rdma_shutdown          = true,
2672         .priv_size              = sizeof(struct isert_cmd),
2673         .owner                  = THIS_MODULE,
2674         .iscsit_setup_np        = isert_setup_np,
2675         .iscsit_accept_np       = isert_accept_np,
2676         .iscsit_free_np         = isert_free_np,
2677         .iscsit_wait_conn       = isert_wait_conn,
2678         .iscsit_free_conn       = isert_free_conn,
2679         .iscsit_get_login_rx    = isert_get_login_rx,
2680         .iscsit_put_login_tx    = isert_put_login_tx,
2681         .iscsit_immediate_queue = isert_immediate_queue,
2682         .iscsit_response_queue  = isert_response_queue,
2683         .iscsit_get_dataout     = isert_get_dataout,
2684         .iscsit_queue_data_in   = isert_put_datain,
2685         .iscsit_queue_status    = isert_put_response,
2686         .iscsit_aborted_task    = isert_aborted_task,
2687         .iscsit_get_rx_pdu      = isert_get_rx_pdu,
2688         .iscsit_get_sup_prot_ops = isert_get_sup_prot_ops,
2689 };
2690
2691 static int __init isert_init(void)
2692 {
2693         int ret;
2694
2695         isert_comp_wq = alloc_workqueue("isert_comp_wq",
2696                                         WQ_UNBOUND | WQ_HIGHPRI, 0);
2697         if (!isert_comp_wq) {
2698                 isert_err("Unable to allocate isert_comp_wq\n");
2699                 return -ENOMEM;
2700         }
2701
2702         isert_release_wq = alloc_workqueue("isert_release_wq", WQ_UNBOUND,
2703                                         WQ_UNBOUND_MAX_ACTIVE);
2704         if (!isert_release_wq) {
2705                 isert_err("Unable to allocate isert_release_wq\n");
2706                 ret = -ENOMEM;
2707                 goto destroy_comp_wq;
2708         }
2709
2710         iscsit_register_transport(&iser_target_transport);
2711         isert_info("iSER_TARGET[0] - Loaded iser_target_transport\n");
2712
2713         return 0;
2714
2715 destroy_comp_wq:
2716         destroy_workqueue(isert_comp_wq);
2717
2718         return ret;
2719 }
2720
2721 static void __exit isert_exit(void)
2722 {
2723         flush_scheduled_work();
2724         destroy_workqueue(isert_release_wq);
2725         destroy_workqueue(isert_comp_wq);
2726         iscsit_unregister_transport(&iser_target_transport);
2727         isert_info("iSER_TARGET[0] - Released iser_target_transport\n");
2728 }
2729
2730 MODULE_DESCRIPTION("iSER-Target for mainline target infrastructure");
2731 MODULE_AUTHOR("nab@Linux-iSCSI.org");
2732 MODULE_LICENSE("GPL");
2733
2734 module_init(isert_init);
2735 module_exit(isert_exit);