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_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 = kzalloc(ISERT_QP_MAX_RECV_DTOS *
 185                                 sizeof(struct iser_rx_desc), GFP_KERNEL);
 186         if (!isert_conn->rx_descs)
 187                 return -ENOMEM;
 188
 189         rx_desc = isert_conn->rx_descs;
 190
 191         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 192                 dma_addr = ib_dma_map_single(ib_dev, (void *)rx_desc,
 193                                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 194                 if (ib_dma_mapping_error(ib_dev, dma_addr))
 195                         goto dma_map_fail;
 196
 197                 rx_desc->dma_addr = dma_addr;
 198
 199                 rx_sg = &rx_desc->rx_sg;
 200                 rx_sg->addr = rx_desc->dma_addr;
 201                 rx_sg->length = ISER_RX_PAYLOAD_SIZE;
 202                 rx_sg->lkey = device->pd->local_dma_lkey;
 203                 rx_desc->rx_cqe.done = isert_recv_done;
 204         }
 205
 206         return 0;
 207
 208 dma_map_fail:
 209         rx_desc = isert_conn->rx_descs;
 210         for (j = 0; j < i; j++, rx_desc++) {
 211                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 212                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 213         }
 214         kfree(isert_conn->rx_descs);
 215         isert_conn->rx_descs = NULL;
 216         isert_err("conn %p failed to allocate rx descriptors\n", isert_conn);
 217         return -ENOMEM;
 218 }
 219
 220 static void
 221 isert_free_rx_descriptors(struct isert_conn *isert_conn)
 222 {
 223         struct ib_device *ib_dev = isert_conn->device->ib_device;
 224         struct iser_rx_desc *rx_desc;
 225         int i;
 226
 227         if (!isert_conn->rx_descs)
 228                 return;
 229
 230         rx_desc = isert_conn->rx_descs;
 231         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 232                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 233                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 234         }
 235
 236         kfree(isert_conn->rx_descs);
 237         isert_conn->rx_descs = NULL;
 238 }
 239
 240 static void
 241 isert_free_comps(struct isert_device *device)
 242 {
 243         int i;
 244
 245         for (i = 0; i < device->comps_used; i++) {
 246                 struct isert_comp *comp = &device->comps[i];
 247
 248                 if (comp->cq)
 249                         ib_free_cq(comp->cq);
 250         }
 251         kfree(device->comps);
 252 }
 253
 254 static int
 255 isert_alloc_comps(struct isert_device *device)
 256 {
 257         int i, max_cqe, ret = 0;
 258
 259         device->comps_used = min(ISERT_MAX_CQ, min_t(int, num_online_cpus(),
 260                                  device->ib_device->num_comp_vectors));
 261
 262         isert_info("Using %d CQs, %s supports %d vectors support "
 263                    "pi_capable %d\n",
 264                    device->comps_used, device->ib_device->name,
 265                    device->ib_device->num_comp_vectors,
 266                    device->pi_capable);
 267
 268         device->comps = kcalloc(device->comps_used, sizeof(struct isert_comp),
 269                                 GFP_KERNEL);
 270         if (!device->comps)
 271                 return -ENOMEM;
 272
 273         max_cqe = min(ISER_MAX_CQ_LEN, device->ib_device->attrs.max_cqe);
 274
 275         for (i = 0; i < device->comps_used; i++) {
 276                 struct isert_comp *comp = &device->comps[i];
 277
 278                 comp->device = device;
 279                 comp->cq = ib_alloc_cq(device->ib_device, comp, max_cqe, i,
 280                                 IB_POLL_WORKQUEUE);
 281                 if (IS_ERR(comp->cq)) {
 282                         isert_err("Unable to allocate cq\n");
 283                         ret = PTR_ERR(comp->cq);
 284                         comp->cq = NULL;
 285                         goto out_cq;
 286                 }
 287         }
 288
 289         return 0;
 290 out_cq:
 291         isert_free_comps(device);
 292         return ret;
 293 }
 294
 295 static int
 296 isert_create_device_ib_res(struct isert_device *device)
 297 {
 298         struct ib_device *ib_dev = device->ib_device;
 299         int ret;
 300
 301         isert_dbg("devattr->max_sge: %d\n", ib_dev->attrs.max_sge);
 302         isert_dbg("devattr->max_sge_rd: %d\n", ib_dev->attrs.max_sge_rd);
 303
 304         ret = isert_alloc_comps(device);
 305         if (ret)
 306                 goto out;
 307
 308         device->pd = ib_alloc_pd(ib_dev, 0);
 309         if (IS_ERR(device->pd)) {
 310                 ret = PTR_ERR(device->pd);
 311                 isert_err("failed to allocate pd, device %p, ret=%d\n",
 312                           device, ret);
 313                 goto out_cq;
 314         }
 315
 316         /* Check signature cap */
 317         device->pi_capable = ib_dev->attrs.device_cap_flags &
 318                              IB_DEVICE_SIGNATURE_HANDOVER ? true : false;
 319
 320         return 0;
 321
 322 out_cq:
 323         isert_free_comps(device);
 324 out:
 325         if (ret > 0)
 326                 ret = -EINVAL;
 327         return ret;
 328 }
 329
 330 static void
 331 isert_free_device_ib_res(struct isert_device *device)
 332 {
 333         isert_info("device %p\n", device);
 334
 335         ib_dealloc_pd(device->pd);
 336         isert_free_comps(device);
 337 }
 338
 339 static void
 340 isert_device_put(struct isert_device *device)
 341 {
 342         mutex_lock(&device_list_mutex);
 343         device->refcount--;
 344         isert_info("device %p refcount %d\n", device, device->refcount);
 345         if (!device->refcount) {
 346                 isert_free_device_ib_res(device);
 347                 list_del(&device->dev_node);
 348                 kfree(device);
 349         }
 350         mutex_unlock(&device_list_mutex);
 351 }
 352
 353 static struct isert_device *
 354 isert_device_get(struct rdma_cm_id *cma_id)
 355 {
 356         struct isert_device *device;
 357         int ret;
 358
 359         mutex_lock(&device_list_mutex);
 360         list_for_each_entry(device, &device_list, dev_node) {
 361                 if (device->ib_device->node_guid == cma_id->device->node_guid) {
 362                         device->refcount++;
 363                         isert_info("Found iser device %p refcount %d\n",
 364                                    device, device->refcount);
 365                         mutex_unlock(&device_list_mutex);
 366                         return device;
 367                 }
 368         }
 369
 370         device = kzalloc(sizeof(struct isert_device), GFP_KERNEL);
 371         if (!device) {
 372                 mutex_unlock(&device_list_mutex);
 373                 return ERR_PTR(-ENOMEM);
 374         }
 375
 376         INIT_LIST_HEAD(&device->dev_node);
 377
 378         device->ib_device = cma_id->device;
 379         ret = isert_create_device_ib_res(device);
 380         if (ret) {
 381                 kfree(device);
 382                 mutex_unlock(&device_list_mutex);
 383                 return ERR_PTR(ret);
 384         }
 385
 386         device->refcount++;
 387         list_add_tail(&device->dev_node, &device_list);
 388         isert_info("Created a new iser device %p refcount %d\n",
 389                    device, device->refcount);
 390         mutex_unlock(&device_list_mutex);
 391
 392         return device;
 393 }
 394
 395 static void
 396 isert_init_conn(struct isert_conn *isert_conn)
 397 {
 398         isert_conn->state = ISER_CONN_INIT;
 399         INIT_LIST_HEAD(&isert_conn->node);
 400         init_completion(&isert_conn->login_comp);
 401         init_completion(&isert_conn->login_req_comp);
 402         init_waitqueue_head(&isert_conn->rem_wait);
 403         kref_init(&isert_conn->kref);
 404         mutex_init(&isert_conn->mutex);
 405         INIT_WORK(&isert_conn->release_work, isert_release_work);
 406 }
 407
 408 static void
 409 isert_free_login_buf(struct isert_conn *isert_conn)
 410 {
 411         struct ib_device *ib_dev = isert_conn->device->ib_device;
 412
 413         ib_dma_unmap_single(ib_dev, isert_conn->login_rsp_dma,
 414                             ISER_RX_PAYLOAD_SIZE, DMA_TO_DEVICE);
 415         kfree(isert_conn->login_rsp_buf);
 416
 417         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 418                             ISER_RX_PAYLOAD_SIZE,
 419                             DMA_FROM_DEVICE);
 420         kfree(isert_conn->login_req_buf);
 421 }
 422
 423 static int
 424 isert_alloc_login_buf(struct isert_conn *isert_conn,
 425                       struct ib_device *ib_dev)
 426 {
 427         int ret;
 428
 429         isert_conn->login_req_buf = kzalloc(sizeof(*isert_conn->login_req_buf),
 430                         GFP_KERNEL);
 431         if (!isert_conn->login_req_buf)
 432                 return -ENOMEM;
 433
 434         isert_conn->login_req_dma = ib_dma_map_single(ib_dev,
 435                                 isert_conn->login_req_buf,
 436                                 ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 437         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_req_dma);
 438         if (ret) {
 439                 isert_err("login_req_dma mapping error: %d\n", ret);
 440                 isert_conn->login_req_dma = 0;
 441                 goto out_free_login_req_buf;
 442         }
 443
 444         isert_conn->login_rsp_buf = kzalloc(ISER_RX_PAYLOAD_SIZE, GFP_KERNEL);
 445         if (!isert_conn->login_rsp_buf) {
 446                 ret = -ENOMEM;
 447                 goto out_unmap_login_req_buf;
 448         }
 449
 450         isert_conn->login_rsp_dma = ib_dma_map_single(ib_dev,
 451                                         isert_conn->login_rsp_buf,
 452                                         ISER_RX_PAYLOAD_SIZE, DMA_TO_DEVICE);
 453         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_rsp_dma);
 454         if (ret) {
 455                 isert_err("login_rsp_dma mapping error: %d\n", ret);
 456                 isert_conn->login_rsp_dma = 0;
 457                 goto out_free_login_rsp_buf;
 458         }
 459
 460         return 0;
 461
 462 out_free_login_rsp_buf:
 463         kfree(isert_conn->login_rsp_buf);
 464 out_unmap_login_req_buf:
 465         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 466                             ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 467 out_free_login_req_buf:
 468         kfree(isert_conn->login_req_buf);
 469         return ret;
 470 }
 471
 472 static void
 473 isert_set_nego_params(struct isert_conn *isert_conn,
 474                       struct rdma_conn_param *param)
 475 {
 476         struct ib_device_attr *attr = &isert_conn->device->ib_device->attrs;
 477
 478         /* Set max inflight RDMA READ requests */
 479         isert_conn->initiator_depth = min_t(u8, param->initiator_depth,
 480                                 attr->max_qp_init_rd_atom);
 481         isert_dbg("Using initiator_depth: %u\n", isert_conn->initiator_depth);
 482
 483         if (param->private_data) {
 484                 u8 flags = *(u8 *)param->private_data;
 485
 486                 /*
 487                  * use remote invalidation if the both initiator
 488                  * and the HCA support it
 489                  */
 490                 isert_conn->snd_w_inv = !(flags & ISER_SEND_W_INV_NOT_SUP) &&
 491                                           (attr->device_cap_flags &
 492                                            IB_DEVICE_MEM_MGT_EXTENSIONS);
 493                 if (isert_conn->snd_w_inv)
 494                         isert_info("Using remote invalidation\n");
 495         }
 496 }
 497
 498 static int
 499 isert_connect_request(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 500 {
 501         struct isert_np *isert_np = cma_id->context;
 502         struct iscsi_np *np = isert_np->np;
 503         struct isert_conn *isert_conn;
 504         struct isert_device *device;
 505         int ret = 0;
 506
 507         spin_lock_bh(&np->np_thread_lock);
 508         if (!np->enabled) {
 509                 spin_unlock_bh(&np->np_thread_lock);
 510                 isert_dbg("iscsi_np is not enabled, reject connect request\n");
 511                 return rdma_reject(cma_id, NULL, 0);
 512         }
 513         spin_unlock_bh(&np->np_thread_lock);
 514
 515         isert_dbg("cma_id: %p, portal: %p\n",
 516                  cma_id, cma_id->context);
 517
 518         isert_conn = kzalloc(sizeof(struct isert_conn), GFP_KERNEL);
 519         if (!isert_conn)
 520                 return -ENOMEM;
 521
 522         isert_init_conn(isert_conn);
 523         isert_conn->cm_id = cma_id;
 524
 525         ret = isert_alloc_login_buf(isert_conn, cma_id->device);
 526         if (ret)
 527                 goto out;
 528
 529         device = isert_device_get(cma_id);
 530         if (IS_ERR(device)) {
 531                 ret = PTR_ERR(device);
 532                 goto out_rsp_dma_map;
 533         }
 534         isert_conn->device = device;
 535
 536         isert_set_nego_params(isert_conn, &event->param.conn);
 537
 538         ret = isert_conn_setup_qp(isert_conn, cma_id);
 539         if (ret)
 540                 goto out_conn_dev;
 541
 542         ret = isert_login_post_recv(isert_conn);
 543         if (ret)
 544                 goto out_conn_dev;
 545
 546         ret = isert_rdma_accept(isert_conn);
 547         if (ret)
 548                 goto out_conn_dev;
 549
 550         mutex_lock(&isert_np->mutex);
 551         list_add_tail(&isert_conn->node, &isert_np->accepted);
 552         mutex_unlock(&isert_np->mutex);
 553
 554         return 0;
 555
 556 out_conn_dev:
 557         isert_device_put(device);
 558 out_rsp_dma_map:
 559         isert_free_login_buf(isert_conn);
 560 out:
 561         kfree(isert_conn);
 562         rdma_reject(cma_id, NULL, 0);
 563         return ret;
 564 }
 565
 566 static void
 567 isert_connect_release(struct isert_conn *isert_conn)
 568 {
 569         struct isert_device *device = isert_conn->device;
 570
 571         isert_dbg("conn %p\n", isert_conn);
 572
 573         BUG_ON(!device);
 574
 575         isert_free_rx_descriptors(isert_conn);
 576         if (isert_conn->cm_id &&
 577             !isert_conn->dev_removed)
 578                 rdma_destroy_id(isert_conn->cm_id);
 579
 580         if (isert_conn->qp) {
 581                 struct isert_comp *comp = isert_conn->qp->recv_cq->cq_context;
 582
 583                 isert_comp_put(comp);
 584                 ib_destroy_qp(isert_conn->qp);
 585         }
 586
 587         if (isert_conn->login_req_buf)
 588                 isert_free_login_buf(isert_conn);
 589
 590         isert_device_put(device);
 591
 592         if (isert_conn->dev_removed)
 593                 wake_up_interruptible(&isert_conn->rem_wait);
 594         else
 595                 kfree(isert_conn);
 596 }
 597
 598 static void
 599 isert_connected_handler(struct rdma_cm_id *cma_id)
 600 {
 601         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 602         struct isert_np *isert_np = cma_id->context;
 603
 604         isert_info("conn %p\n", isert_conn);
 605
 606         mutex_lock(&isert_conn->mutex);
 607         isert_conn->state = ISER_CONN_UP;
 608         kref_get(&isert_conn->kref);
 609         mutex_unlock(&isert_conn->mutex);
 610
 611         mutex_lock(&isert_np->mutex);
 612         list_move_tail(&isert_conn->node, &isert_np->pending);
 613         mutex_unlock(&isert_np->mutex);
 614
 615         isert_info("np %p: Allow accept_np to continue\n", isert_np);
 616         up(&isert_np->sem);
 617 }
 618
 619 static void
 620 isert_release_kref(struct kref *kref)
 621 {
 622         struct isert_conn *isert_conn = container_of(kref,
 623                                 struct isert_conn, kref);
 624
 625         isert_info("conn %p final kref %s/%d\n", isert_conn, current->comm,
 626                    current->pid);
 627
 628         isert_connect_release(isert_conn);
 629 }
 630
 631 static void
 632 isert_put_conn(struct isert_conn *isert_conn)
 633 {
 634         kref_put(&isert_conn->kref, isert_release_kref);
 635 }
 636
 637 static void
 638 isert_handle_unbound_conn(struct isert_conn *isert_conn)
 639 {
 640         struct isert_np *isert_np = isert_conn->cm_id->context;
 641
 642         mutex_lock(&isert_np->mutex);
 643         if (!list_empty(&isert_conn->node)) {
 644                 /*
 645                  * This means iscsi doesn't know this connection
 646                  * so schedule a cleanup ourselves
 647                  */
 648                 list_del_init(&isert_conn->node);
 649                 isert_put_conn(isert_conn);
 650                 queue_work(isert_release_wq, &isert_conn->release_work);
 651         }
 652         mutex_unlock(&isert_np->mutex);
 653 }
 654
 655 /**
 656  * isert_conn_terminate() - Initiate connection termination
 657  * @isert_conn: isert connection struct
 658  *
 659  * Notes:
 660  * In case the connection state is BOUND, move state
 661  * to TEMINATING and start teardown sequence (rdma_disconnect).
 662  * In case the connection state is UP, complete flush as well.
 663  *
 664  * This routine must be called with mutex held. Thus it is
 665  * safe to call multiple times.
 666  */
 667 static void
 668 isert_conn_terminate(struct isert_conn *isert_conn)
 669 {
 670         int err;
 671
 672         if (isert_conn->state >= ISER_CONN_TERMINATING)
 673                 return;
 674
 675         isert_info("Terminating conn %p state %d\n",
 676                    isert_conn, isert_conn->state);
 677         isert_conn->state = ISER_CONN_TERMINATING;
 678         err = rdma_disconnect(isert_conn->cm_id);
 679         if (err)
 680                 isert_warn("Failed rdma_disconnect isert_conn %p\n",
 681                            isert_conn);
 682 }
 683
 684 static int
 685 isert_np_cma_handler(struct isert_np *isert_np,
 686                      enum rdma_cm_event_type event)
 687 {
 688         isert_dbg("%s (%d): isert np %p\n",
 689                   rdma_event_msg(event), event, isert_np);
 690
 691         switch (event) {
 692         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 693                 isert_np->cm_id = NULL;
 694                 break;
 695         case RDMA_CM_EVENT_ADDR_CHANGE:
 696                 isert_np->cm_id = isert_setup_id(isert_np);
 697                 if (IS_ERR(isert_np->cm_id)) {
 698                         isert_err("isert np %p setup id failed: %ld\n",
 699                                   isert_np, PTR_ERR(isert_np->cm_id));
 700                         isert_np->cm_id = NULL;
 701                 }
 702                 break;
 703         default:
 704                 isert_err("isert np %p Unexpected event %d\n",
 705                           isert_np, event);
 706         }
 707
 708         return -1;
 709 }
 710
 711 static int
 712 isert_disconnected_handler(struct rdma_cm_id *cma_id,
 713                            enum rdma_cm_event_type event)
 714 {
 715         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 716
 717         mutex_lock(&isert_conn->mutex);
 718         switch (isert_conn->state) {
 719         case ISER_CONN_TERMINATING:
 720                 break;
 721         case ISER_CONN_UP:
 722                 isert_conn_terminate(isert_conn);
 723                 ib_drain_qp(isert_conn->qp);
 724                 isert_handle_unbound_conn(isert_conn);
 725                 break;
 726         case ISER_CONN_BOUND:
 727         case ISER_CONN_FULL_FEATURE: /* FALLTHRU */
 728                 iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
 729                 break;
 730         default:
 731                 isert_warn("conn %p terminating in state %d\n",
 732                            isert_conn, isert_conn->state);
 733         }
 734         mutex_unlock(&isert_conn->mutex);
 735
 736         return 0;
 737 }
 738
 739 static int
 740 isert_connect_error(struct rdma_cm_id *cma_id)
 741 {
 742         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 743
 744         ib_drain_qp(isert_conn->qp);
 745         list_del_init(&isert_conn->node);
 746         isert_conn->cm_id = NULL;
 747         isert_put_conn(isert_conn);
 748
 749         return -1;
 750 }
 751
 752 static int
 753 isert_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 754 {
 755         struct isert_np *isert_np = cma_id->context;
 756         struct isert_conn *isert_conn;
 757         int ret = 0;
 758
 759         isert_info("%s (%d): status %d id %p np %p\n",
 760                    rdma_event_msg(event->event), event->event,
 761                    event->status, cma_id, cma_id->context);
 762
 763         if (isert_np->cm_id == cma_id)
 764                 return isert_np_cma_handler(cma_id->context, event->event);
 765
 766         switch (event->event) {
 767         case RDMA_CM_EVENT_CONNECT_REQUEST:
 768                 ret = isert_connect_request(cma_id, event);
 769                 if (ret)
 770                         isert_err("failed handle connect request %d\n", ret);
 771                 break;
 772         case RDMA_CM_EVENT_ESTABLISHED:
 773                 isert_connected_handler(cma_id);
 774                 break;
 775         case RDMA_CM_EVENT_ADDR_CHANGE:    /* FALLTHRU */
 776         case RDMA_CM_EVENT_DISCONNECTED:   /* FALLTHRU */
 777         case RDMA_CM_EVENT_TIMEWAIT_EXIT:  /* FALLTHRU */
 778                 ret = isert_disconnected_handler(cma_id, event->event);
 779                 break;
 780         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 781                 isert_conn = cma_id->qp->qp_context;
 782                 isert_conn->dev_removed = true;
 783                 isert_disconnected_handler(cma_id, event->event);
 784                 wait_event_interruptible(isert_conn->rem_wait,
 785                                          isert_conn->state == ISER_CONN_DOWN);
 786                 kfree(isert_conn);
 787                 /*
 788                  * return non-zero from the callback to destroy
 789                  * the rdma cm id
 790                  */
 791                 return 1;
 792         case RDMA_CM_EVENT_REJECTED:
 793                 isert_info("Connection rejected: %s\n",
 794                            rdma_reject_msg(cma_id, event->status));
 795                 /* fall through */
 796         case RDMA_CM_EVENT_UNREACHABLE:
 797         case RDMA_CM_EVENT_CONNECT_ERROR:
 798                 ret = isert_connect_error(cma_id);
 799                 break;
 800         default:
 801                 isert_err("Unhandled RDMA CMA event: %d\n", event->event);
 802                 break;
 803         }
 804
 805         return ret;
 806 }
 807
 808 static int
 809 isert_post_recvm(struct isert_conn *isert_conn, u32 count)
 810 {
 811         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 812         int i, ret;
 813         struct iser_rx_desc *rx_desc;
 814
 815         for (rx_wr = isert_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
 816                 rx_desc = &isert_conn->rx_descs[i];
 817
 818                 rx_wr->wr_cqe = &rx_desc->rx_cqe;
 819                 rx_wr->sg_list = &rx_desc->rx_sg;
 820                 rx_wr->num_sge = 1;
 821                 rx_wr->next = rx_wr + 1;
 822                 rx_desc->in_use = false;
 823         }
 824         rx_wr--;
 825         rx_wr->next = NULL; /* mark end of work requests list */
 826
 827         ret = ib_post_recv(isert_conn->qp, isert_conn->rx_wr,
 828                            &rx_wr_failed);
 829         if (ret)
 830                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
 831
 832         return ret;
 833 }
 834
 835 static int
 836 isert_post_recv(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc)
 837 {
 838         struct ib_recv_wr *rx_wr_failed, rx_wr;
 839         int ret;
 840
 841         if (!rx_desc->in_use) {
 842                 /*
 843                  * if the descriptor is not in-use we already reposted it
 844                  * for recv, so just silently return
 845                  */
 846                 return 0;
 847         }
 848
 849         rx_desc->in_use = false;
 850         rx_wr.wr_cqe = &rx_desc->rx_cqe;
 851         rx_wr.sg_list = &rx_desc->rx_sg;
 852         rx_wr.num_sge = 1;
 853         rx_wr.next = NULL;
 854
 855         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_failed);
 856         if (ret)
 857                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
 858
 859         return ret;
 860 }
 861
 862 static int
 863 isert_login_post_send(struct isert_conn *isert_conn, struct iser_tx_desc *tx_desc)
 864 {
 865         struct ib_device *ib_dev = isert_conn->cm_id->device;
 866         struct ib_send_wr send_wr, *send_wr_failed;
 867         int ret;
 868
 869         ib_dma_sync_single_for_device(ib_dev, tx_desc->dma_addr,
 870                                       ISER_HEADERS_LEN, DMA_TO_DEVICE);
 871
 872         tx_desc->tx_cqe.done = isert_login_send_done;
 873
 874         send_wr.next    = NULL;
 875         send_wr.wr_cqe  = &tx_desc->tx_cqe;
 876         send_wr.sg_list = tx_desc->tx_sg;
 877         send_wr.num_sge = tx_desc->num_sge;
 878         send_wr.opcode  = IB_WR_SEND;
 879         send_wr.send_flags = IB_SEND_SIGNALED;
 880
 881         ret = ib_post_send(isert_conn->qp, &send_wr, &send_wr_failed);
 882         if (ret)
 883                 isert_err("ib_post_send() failed, ret: %d\n", ret);
 884
 885         return ret;
 886 }
 887
 888 static void
 889 isert_create_send_desc(struct isert_conn *isert_conn,
 890                        struct isert_cmd *isert_cmd,
 891                        struct iser_tx_desc *tx_desc)
 892 {
 893         struct isert_device *device = isert_conn->device;
 894         struct ib_device *ib_dev = device->ib_device;
 895
 896         ib_dma_sync_single_for_cpu(ib_dev, tx_desc->dma_addr,
 897                                    ISER_HEADERS_LEN, DMA_TO_DEVICE);
 898
 899         memset(&tx_desc->iser_header, 0, sizeof(struct iser_ctrl));
 900         tx_desc->iser_header.flags = ISCSI_CTRL;
 901
 902         tx_desc->num_sge = 1;
 903
 904         if (tx_desc->tx_sg[0].lkey != device->pd->local_dma_lkey) {
 905                 tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
 906                 isert_dbg("tx_desc %p lkey mismatch, fixing\n", tx_desc);
 907         }
 908 }
 909
 910 static int
 911 isert_init_tx_hdrs(struct isert_conn *isert_conn,
 912                    struct iser_tx_desc *tx_desc)
 913 {
 914         struct isert_device *device = isert_conn->device;
 915         struct ib_device *ib_dev = device->ib_device;
 916         u64 dma_addr;
 917
 918         dma_addr = ib_dma_map_single(ib_dev, (void *)tx_desc,
 919                         ISER_HEADERS_LEN, DMA_TO_DEVICE);
 920         if (ib_dma_mapping_error(ib_dev, dma_addr)) {
 921                 isert_err("ib_dma_mapping_error() failed\n");
 922                 return -ENOMEM;
 923         }
 924
 925         tx_desc->dma_addr = dma_addr;
 926         tx_desc->tx_sg[0].addr  = tx_desc->dma_addr;
 927         tx_desc->tx_sg[0].length = ISER_HEADERS_LEN;
 928         tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
 929
 930         isert_dbg("Setup tx_sg[0].addr: 0x%llx length: %u lkey: 0x%x\n",
 931                   tx_desc->tx_sg[0].addr, tx_desc->tx_sg[0].length,
 932                   tx_desc->tx_sg[0].lkey);
 933
 934         return 0;
 935 }
 936
 937 static void
 938 isert_init_send_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
 939                    struct ib_send_wr *send_wr)
 940 {
 941         struct iser_tx_desc *tx_desc = &isert_cmd->tx_desc;
 942
 943         tx_desc->tx_cqe.done = isert_send_done;
 944         send_wr->wr_cqe = &tx_desc->tx_cqe;
 945
 946         if (isert_conn->snd_w_inv && isert_cmd->inv_rkey) {
 947                 send_wr->opcode  = IB_WR_SEND_WITH_INV;
 948                 send_wr->ex.invalidate_rkey = isert_cmd->inv_rkey;
 949         } else {
 950                 send_wr->opcode = IB_WR_SEND;
 951         }
 952
 953         send_wr->sg_list = &tx_desc->tx_sg[0];
 954         send_wr->num_sge = isert_cmd->tx_desc.num_sge;
 955         send_wr->send_flags = IB_SEND_SIGNALED;
 956 }
 957
 958 static int
 959 isert_login_post_recv(struct isert_conn *isert_conn)
 960 {
 961         struct ib_recv_wr rx_wr, *rx_wr_fail;
 962         struct ib_sge sge;
 963         int ret;
 964
 965         memset(&sge, 0, sizeof(struct ib_sge));
 966         sge.addr = isert_conn->login_req_dma;
 967         sge.length = ISER_RX_PAYLOAD_SIZE;
 968         sge.lkey = isert_conn->device->pd->local_dma_lkey;
 969
 970         isert_dbg("Setup sge: addr: %llx length: %d 0x%08x\n",
 971                 sge.addr, sge.length, sge.lkey);
 972
 973         isert_conn->login_req_buf->rx_cqe.done = isert_login_recv_done;
 974
 975         memset(&rx_wr, 0, sizeof(struct ib_recv_wr));
 976         rx_wr.wr_cqe = &isert_conn->login_req_buf->rx_cqe;
 977         rx_wr.sg_list = &sge;
 978         rx_wr.num_sge = 1;
 979
 980         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_fail);
 981         if (ret)
 982                 isert_err("ib_post_recv() failed: %d\n", ret);
 983
 984         return ret;
 985 }
 986
 987 static int
 988 isert_put_login_tx(struct iscsi_conn *conn, struct iscsi_login *login,
 989                    u32 length)
 990 {
 991         struct isert_conn *isert_conn = conn->context;
 992         struct isert_device *device = isert_conn->device;
 993         struct ib_device *ib_dev = device->ib_device;
 994         struct iser_tx_desc *tx_desc = &isert_conn->login_tx_desc;
 995         int ret;
 996
 997         isert_create_send_desc(isert_conn, NULL, tx_desc);
 998
 999         memcpy(&tx_desc->iscsi_header, &login->rsp[0],
1000                sizeof(struct iscsi_hdr));
1001
1002         isert_init_tx_hdrs(isert_conn, tx_desc);
1003
1004         if (length > 0) {
1005                 struct ib_sge *tx_dsg = &tx_desc->tx_sg[1];
1006
1007                 ib_dma_sync_single_for_cpu(ib_dev, isert_conn->login_rsp_dma,
1008                                            length, DMA_TO_DEVICE);
1009
1010                 memcpy(isert_conn->login_rsp_buf, login->rsp_buf, length);
1011
1012                 ib_dma_sync_single_for_device(ib_dev, isert_conn->login_rsp_dma,
1013                                               length, DMA_TO_DEVICE);
1014
1015                 tx_dsg->addr    = isert_conn->login_rsp_dma;
1016                 tx_dsg->length  = length;
1017                 tx_dsg->lkey    = isert_conn->device->pd->local_dma_lkey;
1018                 tx_desc->num_sge = 2;
1019         }
1020         if (!login->login_failed) {
1021                 if (login->login_complete) {
1022                         ret = isert_alloc_rx_descriptors(isert_conn);
1023                         if (ret)
1024                                 return ret;
1025
1026                         ret = isert_post_recvm(isert_conn,
1027                                                ISERT_QP_MAX_RECV_DTOS);
1028                         if (ret)
1029                                 return ret;
1030
1031                         /* Now we are in FULL_FEATURE phase */
1032                         mutex_lock(&isert_conn->mutex);
1033                         isert_conn->state = ISER_CONN_FULL_FEATURE;
1034                         mutex_unlock(&isert_conn->mutex);
1035                         goto post_send;
1036                 }
1037
1038                 ret = isert_login_post_recv(isert_conn);
1039                 if (ret)
1040                         return ret;
1041         }
1042 post_send:
1043         ret = isert_login_post_send(isert_conn, tx_desc);
1044         if (ret)
1045                 return ret;
1046
1047         return 0;
1048 }
1049
1050 static void
1051 isert_rx_login_req(struct isert_conn *isert_conn)
1052 {
1053         struct iser_rx_desc *rx_desc = isert_conn->login_req_buf;
1054         int rx_buflen = isert_conn->login_req_len;
1055         struct iscsi_conn *conn = isert_conn->conn;
1056         struct iscsi_login *login = conn->conn_login;
1057         int size;
1058
1059         isert_info("conn %p\n", isert_conn);
1060
1061         WARN_ON_ONCE(!login);
1062
1063         if (login->first_request) {
1064                 struct iscsi_login_req *login_req =
1065                         (struct iscsi_login_req *)&rx_desc->iscsi_header;
1066                 /*
1067                  * Setup the initial iscsi_login values from the leading
1068                  * login request PDU.
1069                  */
1070                 login->leading_connection = (!login_req->tsih) ? 1 : 0;
1071                 login->current_stage =
1072                         (login_req->flags & ISCSI_FLAG_LOGIN_CURRENT_STAGE_MASK)
1073                          >> 2;
1074                 login->version_min      = login_req->min_version;
1075                 login->version_max      = login_req->max_version;
1076                 memcpy(login->isid, login_req->isid, 6);
1077                 login->cmd_sn           = be32_to_cpu(login_req->cmdsn);
1078                 login->init_task_tag    = login_req->itt;
1079                 login->initial_exp_statsn = be32_to_cpu(login_req->exp_statsn);
1080                 login->cid              = be16_to_cpu(login_req->cid);
1081                 login->tsih             = be16_to_cpu(login_req->tsih);
1082         }
1083
1084         memcpy(&login->req[0], (void *)&rx_desc->iscsi_header, ISCSI_HDR_LEN);
1085
1086         size = min(rx_buflen, MAX_KEY_VALUE_PAIRS);
1087         isert_dbg("Using login payload size: %d, rx_buflen: %d "
1088                   "MAX_KEY_VALUE_PAIRS: %d\n", size, rx_buflen,
1089                   MAX_KEY_VALUE_PAIRS);
1090         memcpy(login->req_buf, &rx_desc->data[0], size);
1091
1092         if (login->first_request) {
1093                 complete(&isert_conn->login_comp);
1094                 return;
1095         }
1096         schedule_delayed_work(&conn->login_work, 0);
1097 }
1098
1099 static struct iscsi_cmd
1100 *isert_allocate_cmd(struct iscsi_conn *conn, struct iser_rx_desc *rx_desc)
1101 {
1102         struct isert_conn *isert_conn = conn->context;
1103         struct isert_cmd *isert_cmd;
1104         struct iscsi_cmd *cmd;
1105
1106         cmd = iscsit_allocate_cmd(conn, TASK_INTERRUPTIBLE);
1107         if (!cmd) {
1108                 isert_err("Unable to allocate iscsi_cmd + isert_cmd\n");
1109                 return NULL;
1110         }
1111         isert_cmd = iscsit_priv_cmd(cmd);
1112         isert_cmd->conn = isert_conn;
1113         isert_cmd->iscsi_cmd = cmd;
1114         isert_cmd->rx_desc = rx_desc;
1115
1116         return cmd;
1117 }
1118
1119 static int
1120 isert_handle_scsi_cmd(struct isert_conn *isert_conn,
1121                       struct isert_cmd *isert_cmd, struct iscsi_cmd *cmd,
1122                       struct iser_rx_desc *rx_desc, unsigned char *buf)
1123 {
1124         struct iscsi_conn *conn = isert_conn->conn;
1125         struct iscsi_scsi_req *hdr = (struct iscsi_scsi_req *)buf;
1126         int imm_data, imm_data_len, unsol_data, sg_nents, rc;
1127         bool dump_payload = false;
1128         unsigned int data_len;
1129
1130         rc = iscsit_setup_scsi_cmd(conn, cmd, buf);
1131         if (rc < 0)
1132                 return rc;
1133
1134         imm_data = cmd->immediate_data;
1135         imm_data_len = cmd->first_burst_len;
1136         unsol_data = cmd->unsolicited_data;
1137         data_len = cmd->se_cmd.data_length;
1138
1139         if (imm_data && imm_data_len == data_len)
1140                 cmd->se_cmd.se_cmd_flags |= SCF_PASSTHROUGH_SG_TO_MEM_NOALLOC;
1141         rc = iscsit_process_scsi_cmd(conn, cmd, hdr);
1142         if (rc < 0) {
1143                 return 0;
1144         } else if (rc > 0) {
1145                 dump_payload = true;
1146                 goto sequence_cmd;
1147         }
1148
1149         if (!imm_data)
1150                 return 0;
1151
1152         if (imm_data_len != data_len) {
1153                 sg_nents = max(1UL, DIV_ROUND_UP(imm_data_len, PAGE_SIZE));
1154                 sg_copy_from_buffer(cmd->se_cmd.t_data_sg, sg_nents,
1155                                     &rx_desc->data[0], imm_data_len);
1156                 isert_dbg("Copy Immediate sg_nents: %u imm_data_len: %d\n",
1157                           sg_nents, imm_data_len);
1158         } else {
1159                 sg_init_table(&isert_cmd->sg, 1);
1160                 cmd->se_cmd.t_data_sg = &isert_cmd->sg;
1161                 cmd->se_cmd.t_data_nents = 1;
1162                 sg_set_buf(&isert_cmd->sg, &rx_desc->data[0], imm_data_len);
1163                 isert_dbg("Transfer Immediate imm_data_len: %d\n",
1164                           imm_data_len);
1165         }
1166
1167         cmd->write_data_done += imm_data_len;
1168
1169         if (cmd->write_data_done == cmd->se_cmd.data_length) {
1170                 spin_lock_bh(&cmd->istate_lock);
1171                 cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1172                 cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1173                 spin_unlock_bh(&cmd->istate_lock);
1174         }
1175
1176 sequence_cmd:
1177         rc = iscsit_sequence_cmd(conn, cmd, buf, hdr->cmdsn);
1178
1179         if (!rc && dump_payload == false && unsol_data)
1180                 iscsit_set_unsoliticed_dataout(cmd);
1181         else if (dump_payload && imm_data)
1182                 target_put_sess_cmd(&cmd->se_cmd);
1183
1184         return 0;
1185 }
1186
1187 static int
1188 isert_handle_iscsi_dataout(struct isert_conn *isert_conn,
1189                            struct iser_rx_desc *rx_desc, unsigned char *buf)
1190 {
1191         struct scatterlist *sg_start;
1192         struct iscsi_conn *conn = isert_conn->conn;
1193         struct iscsi_cmd *cmd = NULL;
1194         struct iscsi_data *hdr = (struct iscsi_data *)buf;
1195         u32 unsol_data_len = ntoh24(hdr->dlength);
1196         int rc, sg_nents, sg_off, page_off;
1197
1198         rc = iscsit_check_dataout_hdr(conn, buf, &cmd);
1199         if (rc < 0)
1200                 return rc;
1201         else if (!cmd)
1202                 return 0;
1203         /*
1204          * FIXME: Unexpected unsolicited_data out
1205          */
1206         if (!cmd->unsolicited_data) {
1207                 isert_err("Received unexpected solicited data payload\n");
1208                 dump_stack();
1209                 return -1;
1210         }
1211
1212         isert_dbg("Unsolicited DataOut unsol_data_len: %u, "
1213                   "write_data_done: %u, data_length: %u\n",
1214                   unsol_data_len,  cmd->write_data_done,
1215                   cmd->se_cmd.data_length);
1216
1217         sg_off = cmd->write_data_done / PAGE_SIZE;
1218         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
1219         sg_nents = max(1UL, DIV_ROUND_UP(unsol_data_len, PAGE_SIZE));
1220         page_off = cmd->write_data_done % PAGE_SIZE;
1221         /*
1222          * FIXME: Non page-aligned unsolicited_data out
1223          */
1224         if (page_off) {
1225                 isert_err("unexpected non-page aligned data payload\n");
1226                 dump_stack();
1227                 return -1;
1228         }
1229         isert_dbg("Copying DataOut: sg_start: %p, sg_off: %u "
1230                   "sg_nents: %u from %p %u\n", sg_start, sg_off,
1231                   sg_nents, &rx_desc->data[0], unsol_data_len);
1232
1233         sg_copy_from_buffer(sg_start, sg_nents, &rx_desc->data[0],
1234                             unsol_data_len);
1235
1236         rc = iscsit_check_dataout_payload(cmd, hdr, false);
1237         if (rc < 0)
1238                 return rc;
1239
1240         /*
1241          * multiple data-outs on the same command can arrive -
1242          * so post the buffer before hand
1243          */
1244         rc = isert_post_recv(isert_conn, rx_desc);
1245         if (rc) {
1246                 isert_err("ib_post_recv failed with %d\n", rc);
1247                 return rc;
1248         }
1249         return 0;
1250 }
1251
1252 static int
1253 isert_handle_nop_out(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1254                      struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1255                      unsigned char *buf)
1256 {
1257         struct iscsi_conn *conn = isert_conn->conn;
1258         struct iscsi_nopout *hdr = (struct iscsi_nopout *)buf;
1259         int rc;
1260
1261         rc = iscsit_setup_nop_out(conn, cmd, hdr);
1262         if (rc < 0)
1263                 return rc;
1264         /*
1265          * FIXME: Add support for NOPOUT payload using unsolicited RDMA payload
1266          */
1267
1268         return iscsit_process_nop_out(conn, cmd, hdr);
1269 }
1270
1271 static int
1272 isert_handle_text_cmd(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1273                       struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1274                       struct iscsi_text *hdr)
1275 {
1276         struct iscsi_conn *conn = isert_conn->conn;
1277         u32 payload_length = ntoh24(hdr->dlength);
1278         int rc;
1279         unsigned char *text_in = NULL;
1280
1281         rc = iscsit_setup_text_cmd(conn, cmd, hdr);
1282         if (rc < 0)
1283                 return rc;
1284
1285         if (payload_length) {
1286                 text_in = kzalloc(payload_length, GFP_KERNEL);
1287                 if (!text_in)
1288                         return -ENOMEM;
1289         }
1290         cmd->text_in_ptr = text_in;
1291
1292         memcpy(cmd->text_in_ptr, &rx_desc->data[0], payload_length);
1293
1294         return iscsit_process_text_cmd(conn, cmd, hdr);
1295 }
1296
1297 static int
1298 isert_rx_opcode(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc,
1299                 uint32_t read_stag, uint64_t read_va,
1300                 uint32_t write_stag, uint64_t write_va)
1301 {
1302         struct iscsi_hdr *hdr = &rx_desc->iscsi_header;
1303         struct iscsi_conn *conn = isert_conn->conn;
1304         struct iscsi_cmd *cmd;
1305         struct isert_cmd *isert_cmd;
1306         int ret = -EINVAL;
1307         u8 opcode = (hdr->opcode & ISCSI_OPCODE_MASK);
1308
1309         if (conn->sess->sess_ops->SessionType &&
1310            (!(opcode & ISCSI_OP_TEXT) || !(opcode & ISCSI_OP_LOGOUT))) {
1311                 isert_err("Got illegal opcode: 0x%02x in SessionType=Discovery,"
1312                           " ignoring\n", opcode);
1313                 return 0;
1314         }
1315
1316         switch (opcode) {
1317         case ISCSI_OP_SCSI_CMD:
1318                 cmd = isert_allocate_cmd(conn, rx_desc);
1319                 if (!cmd)
1320                         break;
1321
1322                 isert_cmd = iscsit_priv_cmd(cmd);
1323                 isert_cmd->read_stag = read_stag;
1324                 isert_cmd->read_va = read_va;
1325                 isert_cmd->write_stag = write_stag;
1326                 isert_cmd->write_va = write_va;
1327                 isert_cmd->inv_rkey = read_stag ? read_stag : write_stag;
1328
1329                 ret = isert_handle_scsi_cmd(isert_conn, isert_cmd, cmd,
1330                                         rx_desc, (unsigned char *)hdr);
1331                 break;
1332         case ISCSI_OP_NOOP_OUT:
1333                 cmd = isert_allocate_cmd(conn, rx_desc);
1334                 if (!cmd)
1335                         break;
1336
1337                 isert_cmd = iscsit_priv_cmd(cmd);
1338                 ret = isert_handle_nop_out(isert_conn, isert_cmd, cmd,
1339                                            rx_desc, (unsigned char *)hdr);
1340                 break;
1341         case ISCSI_OP_SCSI_DATA_OUT:
1342                 ret = isert_handle_iscsi_dataout(isert_conn, rx_desc,
1343                                                 (unsigned char *)hdr);
1344                 break;
1345         case ISCSI_OP_SCSI_TMFUNC:
1346                 cmd = isert_allocate_cmd(conn, rx_desc);
1347                 if (!cmd)
1348                         break;
1349
1350                 ret = iscsit_handle_task_mgt_cmd(conn, cmd,
1351                                                 (unsigned char *)hdr);
1352                 break;
1353         case ISCSI_OP_LOGOUT:
1354                 cmd = isert_allocate_cmd(conn, rx_desc);
1355                 if (!cmd)
1356                         break;
1357
1358                 ret = iscsit_handle_logout_cmd(conn, cmd, (unsigned char *)hdr);
1359                 break;
1360         case ISCSI_OP_TEXT:
1361                 if (be32_to_cpu(hdr->ttt) != 0xFFFFFFFF)
1362                         cmd = iscsit_find_cmd_from_itt(conn, hdr->itt);
1363                 else
1364                         cmd = isert_allocate_cmd(conn, rx_desc);
1365
1366                 if (!cmd)
1367                         break;
1368
1369                 isert_cmd = iscsit_priv_cmd(cmd);
1370                 ret = isert_handle_text_cmd(isert_conn, isert_cmd, cmd,
1371                                             rx_desc, (struct iscsi_text *)hdr);
1372                 break;
1373         default:
1374                 isert_err("Got unknown iSCSI OpCode: 0x%02x\n", opcode);
1375                 dump_stack();
1376                 break;
1377         }
1378
1379         return ret;
1380 }
1381
1382 static void
1383 isert_print_wc(struct ib_wc *wc, const char *type)
1384 {
1385         if (wc->status != IB_WC_WR_FLUSH_ERR)
1386                 isert_err("%s failure: %s (%d) vend_err %x\n", type,
1387                           ib_wc_status_msg(wc->status), wc->status,
1388                           wc->vendor_err);
1389         else
1390                 isert_dbg("%s failure: %s (%d)\n", type,
1391                           ib_wc_status_msg(wc->status), wc->status);
1392 }
1393
1394 static void
1395 isert_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1396 {
1397         struct isert_conn *isert_conn = wc->qp->qp_context;
1398         struct ib_device *ib_dev = isert_conn->cm_id->device;
1399         struct iser_rx_desc *rx_desc = cqe_to_rx_desc(wc->wr_cqe);
1400         struct iscsi_hdr *hdr = &rx_desc->iscsi_header;
1401         struct iser_ctrl *iser_ctrl = &rx_desc->iser_header;
1402         uint64_t read_va = 0, write_va = 0;
1403         uint32_t read_stag = 0, write_stag = 0;
1404
1405         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1406                 isert_print_wc(wc, "recv");
1407                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1408                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1409                 return;
1410         }
1411
1412         rx_desc->in_use = true;
1413
1414         ib_dma_sync_single_for_cpu(ib_dev, rx_desc->dma_addr,
1415                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1416
1417         isert_dbg("DMA: 0x%llx, iSCSI opcode: 0x%02x, ITT: 0x%08x, flags: 0x%02x dlen: %d\n",
1418                  rx_desc->dma_addr, hdr->opcode, hdr->itt, hdr->flags,
1419                  (int)(wc->byte_len - ISER_HEADERS_LEN));
1420
1421         switch (iser_ctrl->flags & 0xF0) {
1422         case ISCSI_CTRL:
1423                 if (iser_ctrl->flags & ISER_RSV) {
1424                         read_stag = be32_to_cpu(iser_ctrl->read_stag);
1425                         read_va = be64_to_cpu(iser_ctrl->read_va);
1426                         isert_dbg("ISER_RSV: read_stag: 0x%x read_va: 0x%llx\n",
1427                                   read_stag, (unsigned long long)read_va);
1428                 }
1429                 if (iser_ctrl->flags & ISER_WSV) {
1430                         write_stag = be32_to_cpu(iser_ctrl->write_stag);
1431                         write_va = be64_to_cpu(iser_ctrl->write_va);
1432                         isert_dbg("ISER_WSV: write_stag: 0x%x write_va: 0x%llx\n",
1433                                   write_stag, (unsigned long long)write_va);
1434                 }
1435
1436                 isert_dbg("ISER ISCSI_CTRL PDU\n");
1437                 break;
1438         case ISER_HELLO:
1439                 isert_err("iSER Hello message\n");
1440                 break;
1441         default:
1442                 isert_warn("Unknown iSER hdr flags: 0x%02x\n", iser_ctrl->flags);
1443                 break;
1444         }
1445
1446         isert_rx_opcode(isert_conn, rx_desc,
1447                         read_stag, read_va, write_stag, write_va);
1448
1449         ib_dma_sync_single_for_device(ib_dev, rx_desc->dma_addr,
1450                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1451 }
1452
1453 static void
1454 isert_login_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1455 {
1456         struct isert_conn *isert_conn = wc->qp->qp_context;
1457         struct ib_device *ib_dev = isert_conn->device->ib_device;
1458
1459         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1460                 isert_print_wc(wc, "login recv");
1461                 return;
1462         }
1463
1464         ib_dma_sync_single_for_cpu(ib_dev, isert_conn->login_req_dma,
1465                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1466
1467         isert_conn->login_req_len = wc->byte_len - ISER_HEADERS_LEN;
1468
1469         if (isert_conn->conn) {
1470                 struct iscsi_login *login = isert_conn->conn->conn_login;
1471
1472                 if (login && !login->first_request)
1473                         isert_rx_login_req(isert_conn);
1474         }
1475
1476         mutex_lock(&isert_conn->mutex);
1477         complete(&isert_conn->login_req_comp);
1478         mutex_unlock(&isert_conn->mutex);
1479
1480         ib_dma_sync_single_for_device(ib_dev, isert_conn->login_req_dma,
1481                                 ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1482 }
1483
1484 static void
1485 isert_rdma_rw_ctx_destroy(struct isert_cmd *cmd, struct isert_conn *conn)
1486 {
1487         struct se_cmd *se_cmd = &cmd->iscsi_cmd->se_cmd;
1488         enum dma_data_direction dir = target_reverse_dma_direction(se_cmd);
1489
1490         if (!cmd->rw.nr_ops)
1491                 return;
1492
1493         if (isert_prot_cmd(conn, se_cmd)) {
1494                 rdma_rw_ctx_destroy_signature(&cmd->rw, conn->qp,
1495                                 conn->cm_id->port_num, se_cmd->t_data_sg,
1496                                 se_cmd->t_data_nents, se_cmd->t_prot_sg,
1497                                 se_cmd->t_prot_nents, dir);
1498         } else {
1499                 rdma_rw_ctx_destroy(&cmd->rw, conn->qp, conn->cm_id->port_num,
1500                                 se_cmd->t_data_sg, se_cmd->t_data_nents, dir);
1501         }
1502
1503         cmd->rw.nr_ops = 0;
1504 }
1505
1506 static void
1507 isert_put_cmd(struct isert_cmd *isert_cmd, bool comp_err)
1508 {
1509         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1510         struct isert_conn *isert_conn = isert_cmd->conn;
1511         struct iscsi_conn *conn = isert_conn->conn;
1512         struct iscsi_text_rsp *hdr;
1513
1514         isert_dbg("Cmd %p\n", isert_cmd);
1515
1516         switch (cmd->iscsi_opcode) {
1517         case ISCSI_OP_SCSI_CMD:
1518                 spin_lock_bh(&conn->cmd_lock);
1519                 if (!list_empty(&cmd->i_conn_node))
1520                         list_del_init(&cmd->i_conn_node);
1521                 spin_unlock_bh(&conn->cmd_lock);
1522
1523                 if (cmd->data_direction == DMA_TO_DEVICE) {
1524                         iscsit_stop_dataout_timer(cmd);
1525                         /*
1526                          * Check for special case during comp_err where
1527                          * WRITE_PENDING has been handed off from core,
1528                          * but requires an extra target_put_sess_cmd()
1529                          * before transport_generic_free_cmd() below.
1530                          */
1531                         if (comp_err &&
1532                             cmd->se_cmd.t_state == TRANSPORT_WRITE_PENDING) {
1533                                 struct se_cmd *se_cmd = &cmd->se_cmd;
1534
1535                                 target_put_sess_cmd(se_cmd);
1536                         }
1537                 }
1538
1539                 isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1540                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1541                 break;
1542         case ISCSI_OP_SCSI_TMFUNC:
1543                 spin_lock_bh(&conn->cmd_lock);
1544                 if (!list_empty(&cmd->i_conn_node))
1545                         list_del_init(&cmd->i_conn_node);
1546                 spin_unlock_bh(&conn->cmd_lock);
1547
1548                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1549                 break;
1550         case ISCSI_OP_REJECT:
1551         case ISCSI_OP_NOOP_OUT:
1552         case ISCSI_OP_TEXT:
1553                 hdr = (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
1554                 /* If the continue bit is on, keep the command alive */
1555                 if (hdr->flags & ISCSI_FLAG_TEXT_CONTINUE)
1556                         break;
1557
1558                 spin_lock_bh(&conn->cmd_lock);
1559                 if (!list_empty(&cmd->i_conn_node))
1560                         list_del_init(&cmd->i_conn_node);
1561                 spin_unlock_bh(&conn->cmd_lock);
1562
1563                 /*
1564                  * Handle special case for REJECT when iscsi_add_reject*() has
1565                  * overwritten the original iscsi_opcode assignment, and the
1566                  * associated cmd->se_cmd needs to be released.
1567                  */
1568                 if (cmd->se_cmd.se_tfo != NULL) {
1569                         isert_dbg("Calling transport_generic_free_cmd for 0x%02x\n",
1570                                  cmd->iscsi_opcode);
1571                         transport_generic_free_cmd(&cmd->se_cmd, 0);
1572                         break;
1573                 }
1574                 /* fall through */
1575         default:
1576                 iscsit_release_cmd(cmd);
1577                 break;
1578         }
1579 }
1580
1581 static void
1582 isert_unmap_tx_desc(struct iser_tx_desc *tx_desc, struct ib_device *ib_dev)
1583 {
1584         if (tx_desc->dma_addr != 0) {
1585                 isert_dbg("unmap single for tx_desc->dma_addr\n");
1586                 ib_dma_unmap_single(ib_dev, tx_desc->dma_addr,
1587                                     ISER_HEADERS_LEN, DMA_TO_DEVICE);
1588                 tx_desc->dma_addr = 0;
1589         }
1590 }
1591
1592 static void
1593 isert_completion_put(struct iser_tx_desc *tx_desc, struct isert_cmd *isert_cmd,
1594                      struct ib_device *ib_dev, bool comp_err)
1595 {
1596         if (isert_cmd->pdu_buf_dma != 0) {
1597                 isert_dbg("unmap single for isert_cmd->pdu_buf_dma\n");
1598                 ib_dma_unmap_single(ib_dev, isert_cmd->pdu_buf_dma,
1599                                     isert_cmd->pdu_buf_len, DMA_TO_DEVICE);
1600                 isert_cmd->pdu_buf_dma = 0;
1601         }
1602
1603         isert_unmap_tx_desc(tx_desc, ib_dev);
1604         isert_put_cmd(isert_cmd, comp_err);
1605 }
1606
1607 static int
1608 isert_check_pi_status(struct se_cmd *se_cmd, struct ib_mr *sig_mr)
1609 {
1610         struct ib_mr_status mr_status;
1611         int ret;
1612
1613         ret = ib_check_mr_status(sig_mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
1614         if (ret) {
1615                 isert_err("ib_check_mr_status failed, ret %d\n", ret);
1616                 goto fail_mr_status;
1617         }
1618
1619         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1620                 u64 sec_offset_err;
1621                 u32 block_size = se_cmd->se_dev->dev_attrib.block_size + 8;
1622
1623                 switch (mr_status.sig_err.err_type) {
1624                 case IB_SIG_BAD_GUARD:
1625                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_GUARD_CHECK_FAILED;
1626                         break;
1627                 case IB_SIG_BAD_REFTAG:
1628                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_REF_TAG_CHECK_FAILED;
1629                         break;
1630                 case IB_SIG_BAD_APPTAG:
1631                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_APP_TAG_CHECK_FAILED;
1632                         break;
1633                 }
1634                 sec_offset_err = mr_status.sig_err.sig_err_offset;
1635                 do_div(sec_offset_err, block_size);
1636                 se_cmd->bad_sector = sec_offset_err + se_cmd->t_task_lba;
1637
1638                 isert_err("PI error found type %d at sector 0x%llx "
1639                           "expected 0x%x vs actual 0x%x\n",
1640                           mr_status.sig_err.err_type,
1641                           (unsigned long long)se_cmd->bad_sector,
1642                           mr_status.sig_err.expected,
1643                           mr_status.sig_err.actual);
1644                 ret = 1;
1645         }
1646
1647 fail_mr_status:
1648         return ret;
1649 }
1650
1651 static void
1652 isert_rdma_write_done(struct ib_cq *cq, struct ib_wc *wc)
1653 {
1654         struct isert_conn *isert_conn = wc->qp->qp_context;
1655         struct isert_device *device = isert_conn->device;
1656         struct iser_tx_desc *desc = cqe_to_tx_desc(wc->wr_cqe);
1657         struct isert_cmd *isert_cmd = tx_desc_to_cmd(desc);
1658         struct se_cmd *cmd = &isert_cmd->iscsi_cmd->se_cmd;
1659         int ret = 0;
1660
1661         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1662                 isert_print_wc(wc, "rdma write");
1663                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1664                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1665                 isert_completion_put(desc, isert_cmd, device->ib_device, true);
1666                 return;
1667         }
1668
1669         isert_dbg("Cmd %p\n", isert_cmd);
1670
1671         ret = isert_check_pi_status(cmd, isert_cmd->rw.sig->sig_mr);
1672         isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1673
1674         if (ret) {
1675                 /*
1676                  * transport_generic_request_failure() expects to have
1677                  * plus two references to handle queue-full, so re-add
1678                  * one here as target-core will have already dropped
1679                  * it after the first isert_put_datain() callback.
1680                  */
1681                 kref_get(&cmd->cmd_kref);
1682                 transport_generic_request_failure(cmd, cmd->pi_err);
1683         } else {
1684                 /*
1685                  * XXX: isert_put_response() failure is not retried.
1686                  */
1687                 ret = isert_put_response(isert_conn->conn, isert_cmd->iscsi_cmd);
1688                 if (ret)
1689                         pr_warn_ratelimited("isert_put_response() ret: %d\n", ret);
1690         }
1691 }
1692
1693 static void
1694 isert_rdma_read_done(struct ib_cq *cq, struct ib_wc *wc)
1695 {
1696         struct isert_conn *isert_conn = wc->qp->qp_context;
1697         struct isert_device *device = isert_conn->device;
1698         struct iser_tx_desc *desc = cqe_to_tx_desc(wc->wr_cqe);
1699         struct isert_cmd *isert_cmd = tx_desc_to_cmd(desc);
1700         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1701         struct se_cmd *se_cmd = &cmd->se_cmd;
1702         int ret = 0;
1703
1704         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1705                 isert_print_wc(wc, "rdma read");
1706                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1707                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1708                 isert_completion_put(desc, isert_cmd, device->ib_device, true);
1709                 return;
1710         }
1711
1712         isert_dbg("Cmd %p\n", isert_cmd);
1713
1714         iscsit_stop_dataout_timer(cmd);
1715
1716         if (isert_prot_cmd(isert_conn, se_cmd))
1717                 ret = isert_check_pi_status(se_cmd, isert_cmd->rw.sig->sig_mr);
1718         isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1719         cmd->write_data_done = 0;
1720
1721         isert_dbg("Cmd: %p RDMA_READ comp calling execute_cmd\n", isert_cmd);
1722         spin_lock_bh(&cmd->istate_lock);
1723         cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1724         cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1725         spin_unlock_bh(&cmd->istate_lock);
1726
1727         /*
1728          * transport_generic_request_failure() will drop the extra
1729          * se_cmd->cmd_kref reference after T10-PI error, and handle
1730          * any non-zero ->queue_status() callback error retries.
1731          */
1732         if (ret)
1733                 transport_generic_request_failure(se_cmd, se_cmd->pi_err);
1734         else
1735                 target_execute_cmd(se_cmd);
1736 }
1737
1738 static void
1739 isert_do_control_comp(struct work_struct *work)
1740 {
1741         struct isert_cmd *isert_cmd = container_of(work,
1742                         struct isert_cmd, comp_work);
1743         struct isert_conn *isert_conn = isert_cmd->conn;
1744         struct ib_device *ib_dev = isert_conn->cm_id->device;
1745         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1746
1747         isert_dbg("Cmd %p i_state %d\n", isert_cmd, cmd->i_state);
1748
1749         switch (cmd->i_state) {
1750         case ISTATE_SEND_TASKMGTRSP:
1751                 iscsit_tmr_post_handler(cmd, cmd->conn);
1752                 /* fall through */
1753         case ISTATE_SEND_REJECT:
1754         case ISTATE_SEND_TEXTRSP:
1755                 cmd->i_state = ISTATE_SENT_STATUS;
1756                 isert_completion_put(&isert_cmd->tx_desc, isert_cmd,
1757                                      ib_dev, false);
1758                 break;
1759         case ISTATE_SEND_LOGOUTRSP:
1760                 iscsit_logout_post_handler(cmd, cmd->conn);
1761                 break;
1762         default:
1763                 isert_err("Unknown i_state %d\n", cmd->i_state);
1764                 dump_stack();
1765                 break;
1766         }
1767 }
1768
1769 static void
1770 isert_login_send_done(struct ib_cq *cq, struct ib_wc *wc)
1771 {
1772         struct isert_conn *isert_conn = wc->qp->qp_context;
1773         struct ib_device *ib_dev = isert_conn->cm_id->device;
1774         struct iser_tx_desc *tx_desc = cqe_to_tx_desc(wc->wr_cqe);
1775
1776         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1777                 isert_print_wc(wc, "login send");
1778                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1779                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1780         }
1781
1782         isert_unmap_tx_desc(tx_desc, ib_dev);
1783 }
1784
1785 static void
1786 isert_send_done(struct ib_cq *cq, struct ib_wc *wc)
1787 {
1788         struct isert_conn *isert_conn = wc->qp->qp_context;
1789         struct ib_device *ib_dev = isert_conn->cm_id->device;
1790         struct iser_tx_desc *tx_desc = cqe_to_tx_desc(wc->wr_cqe);
1791         struct isert_cmd *isert_cmd = tx_desc_to_cmd(tx_desc);
1792
1793         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1794                 isert_print_wc(wc, "send");
1795                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1796                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1797                 isert_completion_put(tx_desc, isert_cmd, ib_dev, true);
1798                 return;
1799         }
1800
1801         isert_dbg("Cmd %p\n", isert_cmd);
1802
1803         switch (isert_cmd->iscsi_cmd->i_state) {
1804         case ISTATE_SEND_TASKMGTRSP:
1805         case ISTATE_SEND_LOGOUTRSP:
1806         case ISTATE_SEND_REJECT:
1807         case ISTATE_SEND_TEXTRSP:
1808                 isert_unmap_tx_desc(tx_desc, ib_dev);
1809
1810                 INIT_WORK(&isert_cmd->comp_work, isert_do_control_comp);
1811                 queue_work(isert_comp_wq, &isert_cmd->comp_work);
1812                 return;
1813         default:
1814                 isert_cmd->iscsi_cmd->i_state = ISTATE_SENT_STATUS;
1815                 isert_completion_put(tx_desc, isert_cmd, ib_dev, false);
1816                 break;
1817         }
1818 }
1819
1820 static int
1821 isert_post_response(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd)
1822 {
1823         struct ib_send_wr *wr_failed;
1824         int ret;
1825
1826         ret = isert_post_recv(isert_conn, isert_cmd->rx_desc);
1827         if (ret) {
1828                 isert_err("ib_post_recv failed with %d\n", ret);
1829                 return ret;
1830         }
1831
1832         ret = ib_post_send(isert_conn->qp, &isert_cmd->tx_desc.send_wr,
1833                            &wr_failed);
1834         if (ret) {
1835                 isert_err("ib_post_send failed with %d\n", ret);
1836                 return ret;
1837         }
1838         return ret;
1839 }
1840
1841 static int
1842 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
1843 {
1844         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1845         struct isert_conn *isert_conn = conn->context;
1846         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1847         struct iscsi_scsi_rsp *hdr = (struct iscsi_scsi_rsp *)
1848                                 &isert_cmd->tx_desc.iscsi_header;
1849
1850         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1851         iscsit_build_rsp_pdu(cmd, conn, true, hdr);
1852         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1853         /*
1854          * Attach SENSE DATA payload to iSCSI Response PDU
1855          */
1856         if (cmd->se_cmd.sense_buffer &&
1857             ((cmd->se_cmd.se_cmd_flags & SCF_TRANSPORT_TASK_SENSE) ||
1858             (cmd->se_cmd.se_cmd_flags & SCF_EMULATED_TASK_SENSE))) {
1859                 struct isert_device *device = isert_conn->device;
1860                 struct ib_device *ib_dev = device->ib_device;
1861                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
1862                 u32 padding, pdu_len;
1863
1864                 put_unaligned_be16(cmd->se_cmd.scsi_sense_length,
1865                                    cmd->sense_buffer);
1866                 cmd->se_cmd.scsi_sense_length += sizeof(__be16);
1867
1868                 padding = -(cmd->se_cmd.scsi_sense_length) & 3;
1869                 hton24(hdr->dlength, (u32)cmd->se_cmd.scsi_sense_length);
1870                 pdu_len = cmd->se_cmd.scsi_sense_length + padding;
1871
1872                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
1873                                 (void *)cmd->sense_buffer, pdu_len,
1874                                 DMA_TO_DEVICE);
1875                 if (ib_dma_mapping_error(ib_dev, isert_cmd->pdu_buf_dma))
1876                         return -ENOMEM;
1877
1878                 isert_cmd->pdu_buf_len = pdu_len;
1879                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
1880                 tx_dsg->length  = pdu_len;
1881                 tx_dsg->lkey    = device->pd->local_dma_lkey;
1882                 isert_cmd->tx_desc.num_sge = 2;
1883         }
1884
1885         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1886
1887         isert_dbg("Posting SCSI Response\n");
1888
1889         return isert_post_response(isert_conn, isert_cmd);
1890 }
1891
1892 static void
1893 isert_aborted_task(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
1894 {
1895         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1896         struct isert_conn *isert_conn = conn->context;
1897
1898         spin_lock_bh(&conn->cmd_lock);
1899         if (!list_empty(&cmd->i_conn_node))
1900                 list_del_init(&cmd->i_conn_node);
1901         spin_unlock_bh(&conn->cmd_lock);
1902
1903         if (cmd->data_direction == DMA_TO_DEVICE)
1904                 iscsit_stop_dataout_timer(cmd);
1905         isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1906 }
1907
1908 static enum target_prot_op
1909 isert_get_sup_prot_ops(struct iscsi_conn *conn)
1910 {
1911         struct isert_conn *isert_conn = conn->context;
1912         struct isert_device *device = isert_conn->device;
1913
1914         if (conn->tpg->tpg_attrib.t10_pi) {
1915                 if (device->pi_capable) {
1916                         isert_info("conn %p PI offload enabled\n", isert_conn);
1917                         isert_conn->pi_support = true;
1918                         return TARGET_PROT_ALL;
1919                 }
1920         }
1921
1922         isert_info("conn %p PI offload disabled\n", isert_conn);
1923         isert_conn->pi_support = false;
1924
1925         return TARGET_PROT_NORMAL;
1926 }
1927
1928 static int
1929 isert_put_nopin(struct iscsi_cmd *cmd, struct iscsi_conn *conn,
1930                 bool nopout_response)
1931 {
1932         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1933         struct isert_conn *isert_conn = conn->context;
1934         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1935
1936         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1937         iscsit_build_nopin_rsp(cmd, conn, (struct iscsi_nopin *)
1938                                &isert_cmd->tx_desc.iscsi_header,
1939                                nopout_response);
1940         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1941         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1942
1943         isert_dbg("conn %p Posting NOPIN Response\n", isert_conn);
1944
1945         return isert_post_response(isert_conn, isert_cmd);
1946 }
1947
1948 static int
1949 isert_put_logout_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
1950 {
1951         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1952         struct isert_conn *isert_conn = conn->context;
1953         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1954
1955         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1956         iscsit_build_logout_rsp(cmd, conn, (struct iscsi_logout_rsp *)
1957                                 &isert_cmd->tx_desc.iscsi_header);
1958         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1959         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1960
1961         isert_dbg("conn %p Posting Logout Response\n", isert_conn);
1962
1963         return isert_post_response(isert_conn, isert_cmd);
1964 }
1965
1966 static int
1967 isert_put_tm_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
1968 {
1969         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1970         struct isert_conn *isert_conn = conn->context;
1971         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1972
1973         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1974         iscsit_build_task_mgt_rsp(cmd, conn, (struct iscsi_tm_rsp *)
1975                                   &isert_cmd->tx_desc.iscsi_header);
1976         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1977         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1978
1979         isert_dbg("conn %p Posting Task Management Response\n", isert_conn);
1980
1981         return isert_post_response(isert_conn, isert_cmd);
1982 }
1983
1984 static int
1985 isert_put_reject(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
1986 {
1987         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1988         struct isert_conn *isert_conn = conn->context;
1989         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1990         struct isert_device *device = isert_conn->device;
1991         struct ib_device *ib_dev = device->ib_device;
1992         struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
1993         struct iscsi_reject *hdr =
1994                 (struct iscsi_reject *)&isert_cmd->tx_desc.iscsi_header;
1995
1996         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1997         iscsit_build_reject(cmd, conn, hdr);
1998         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1999
2000         hton24(hdr->dlength, ISCSI_HDR_LEN);
2001         isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2002                         (void *)cmd->buf_ptr, ISCSI_HDR_LEN,
2003                         DMA_TO_DEVICE);
2004         if (ib_dma_mapping_error(ib_dev, isert_cmd->pdu_buf_dma))
2005                 return -ENOMEM;
2006         isert_cmd->pdu_buf_len = ISCSI_HDR_LEN;
2007         tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2008         tx_dsg->length  = ISCSI_HDR_LEN;
2009         tx_dsg->lkey    = device->pd->local_dma_lkey;
2010         isert_cmd->tx_desc.num_sge = 2;
2011
2012         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2013
2014         isert_dbg("conn %p Posting Reject\n", isert_conn);
2015
2016         return isert_post_response(isert_conn, isert_cmd);
2017 }
2018
2019 static int
2020 isert_put_text_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2021 {
2022         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2023         struct isert_conn *isert_conn = conn->context;
2024         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2025         struct iscsi_text_rsp *hdr =
2026                 (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
2027         u32 txt_rsp_len;
2028         int rc;
2029
2030         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2031         rc = iscsit_build_text_rsp(cmd, conn, hdr, ISCSI_INFINIBAND);
2032         if (rc < 0)
2033                 return rc;
2034
2035         txt_rsp_len = rc;
2036         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2037
2038         if (txt_rsp_len) {
2039                 struct isert_device *device = isert_conn->device;
2040                 struct ib_device *ib_dev = device->ib_device;
2041                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2042                 void *txt_rsp_buf = cmd->buf_ptr;
2043
2044                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2045                                 txt_rsp_buf, txt_rsp_len, DMA_TO_DEVICE);
2046                 if (ib_dma_mapping_error(ib_dev, isert_cmd->pdu_buf_dma))
2047                         return -ENOMEM;
2048
2049                 isert_cmd->pdu_buf_len = txt_rsp_len;
2050                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2051                 tx_dsg->length  = txt_rsp_len;
2052                 tx_dsg->lkey    = device->pd->local_dma_lkey;
2053                 isert_cmd->tx_desc.num_sge = 2;
2054         }
2055         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2056
2057         isert_dbg("conn %p Text Response\n", isert_conn);
2058
2059         return isert_post_response(isert_conn, isert_cmd);
2060 }
2061
2062 static inline void
2063 isert_set_dif_domain(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs,
2064                      struct ib_sig_domain *domain)
2065 {
2066         domain->sig_type = IB_SIG_TYPE_T10_DIF;
2067         domain->sig.dif.bg_type = IB_T10DIF_CRC;
2068         domain->sig.dif.pi_interval = se_cmd->se_dev->dev_attrib.block_size;
2069         domain->sig.dif.ref_tag = se_cmd->reftag_seed;
2070         /*
2071          * At the moment we hard code those, but if in the future
2072          * the target core would like to use it, we will take it
2073          * from se_cmd.
2074          */
2075         domain->sig.dif.apptag_check_mask = 0xffff;
2076         domain->sig.dif.app_escape = true;
2077         domain->sig.dif.ref_escape = true;
2078         if (se_cmd->prot_type == TARGET_DIF_TYPE1_PROT ||
2079             se_cmd->prot_type == TARGET_DIF_TYPE2_PROT)
2080                 domain->sig.dif.ref_remap = true;
2081 };
2082
2083 static int
2084 isert_set_sig_attrs(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs)
2085 {
2086         memset(sig_attrs, 0, sizeof(*sig_attrs));
2087
2088         switch (se_cmd->prot_op) {
2089         case TARGET_PROT_DIN_INSERT:
2090         case TARGET_PROT_DOUT_STRIP:
2091                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
2092                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2093                 break;
2094         case TARGET_PROT_DOUT_INSERT:
2095         case TARGET_PROT_DIN_STRIP:
2096                 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
2097                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2098                 break;
2099         case TARGET_PROT_DIN_PASS:
2100         case TARGET_PROT_DOUT_PASS:
2101                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2102                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2103                 break;
2104         default:
2105                 isert_err("Unsupported PI operation %d\n", se_cmd->prot_op);
2106                 return -EINVAL;
2107         }
2108
2109         sig_attrs->check_mask =
2110                (se_cmd->prot_checks & TARGET_DIF_CHECK_GUARD  ? 0xc0 : 0) |
2111                (se_cmd->prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x30 : 0) |
2112                (se_cmd->prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x0f : 0);
2113         return 0;
2114 }
2115
2116 static int
2117 isert_rdma_rw_ctx_post(struct isert_cmd *cmd, struct isert_conn *conn,
2118                 struct ib_cqe *cqe, struct ib_send_wr *chain_wr)
2119 {
2120         struct se_cmd *se_cmd = &cmd->iscsi_cmd->se_cmd;
2121         enum dma_data_direction dir = target_reverse_dma_direction(se_cmd);
2122         u8 port_num = conn->cm_id->port_num;
2123         u64 addr;
2124         u32 rkey, offset;
2125         int ret;
2126
2127         if (dir == DMA_FROM_DEVICE) {
2128                 addr = cmd->write_va;
2129                 rkey = cmd->write_stag;
2130                 offset = cmd->iscsi_cmd->write_data_done;
2131         } else {
2132                 addr = cmd->read_va;
2133                 rkey = cmd->read_stag;
2134                 offset = 0;
2135         }
2136
2137         if (isert_prot_cmd(conn, se_cmd)) {
2138                 struct ib_sig_attrs sig_attrs;
2139
2140                 ret = isert_set_sig_attrs(se_cmd, &sig_attrs);
2141                 if (ret)
2142                         return ret;
2143
2144                 WARN_ON_ONCE(offset);
2145                 ret = rdma_rw_ctx_signature_init(&cmd->rw, conn->qp, port_num,
2146                                 se_cmd->t_data_sg, se_cmd->t_data_nents,
2147                                 se_cmd->t_prot_sg, se_cmd->t_prot_nents,
2148                                 &sig_attrs, addr, rkey, dir);
2149         } else {
2150                 ret = rdma_rw_ctx_init(&cmd->rw, conn->qp, port_num,
2151                                 se_cmd->t_data_sg, se_cmd->t_data_nents,
2152                                 offset, addr, rkey, dir);
2153         }
2154         if (ret < 0) {
2155                 isert_err("Cmd: %p failed to prepare RDMA res\n", cmd);
2156                 return ret;
2157         }
2158
2159         ret = rdma_rw_ctx_post(&cmd->rw, conn->qp, port_num, cqe, chain_wr);
2160         if (ret < 0)
2161                 isert_err("Cmd: %p failed to post RDMA res\n", cmd);
2162         return ret;
2163 }
2164
2165 static int
2166 isert_put_datain(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2167 {
2168         struct se_cmd *se_cmd = &cmd->se_cmd;
2169         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2170         struct isert_conn *isert_conn = conn->context;
2171         struct ib_cqe *cqe = NULL;
2172         struct ib_send_wr *chain_wr = NULL;
2173         int rc;
2174
2175         isert_dbg("Cmd: %p RDMA_WRITE data_length: %u\n",
2176                  isert_cmd, se_cmd->data_length);
2177
2178         if (isert_prot_cmd(isert_conn, se_cmd)) {
2179                 isert_cmd->tx_desc.tx_cqe.done = isert_rdma_write_done;
2180                 cqe = &isert_cmd->tx_desc.tx_cqe;
2181         } else {
2182                 /*
2183                  * Build isert_conn->tx_desc for iSCSI response PDU and attach
2184                  */
2185                 isert_create_send_desc(isert_conn, isert_cmd,
2186                                        &isert_cmd->tx_desc);
2187                 iscsit_build_rsp_pdu(cmd, conn, true, (struct iscsi_scsi_rsp *)
2188                                      &isert_cmd->tx_desc.iscsi_header);
2189                 isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2190                 isert_init_send_wr(isert_conn, isert_cmd,
2191                                    &isert_cmd->tx_desc.send_wr);
2192
2193                 rc = isert_post_recv(isert_conn, isert_cmd->rx_desc);
2194                 if (rc) {
2195                         isert_err("ib_post_recv failed with %d\n", rc);
2196                         return rc;
2197                 }
2198
2199                 chain_wr = &isert_cmd->tx_desc.send_wr;
2200         }
2201
2202         rc = isert_rdma_rw_ctx_post(isert_cmd, isert_conn, cqe, chain_wr);
2203         isert_dbg("Cmd: %p posted RDMA_WRITE for iSER Data READ rc: %d\n",
2204                   isert_cmd, rc);
2205         return rc;
2206 }
2207
2208 static int
2209 isert_get_dataout(struct iscsi_conn *conn, struct iscsi_cmd *cmd, bool recovery)
2210 {
2211         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2212         int ret;
2213
2214         isert_dbg("Cmd: %p RDMA_READ data_length: %u write_data_done: %u\n",
2215                  isert_cmd, cmd->se_cmd.data_length, cmd->write_data_done);
2216
2217         isert_cmd->tx_desc.tx_cqe.done = isert_rdma_read_done;
2218         ret = isert_rdma_rw_ctx_post(isert_cmd, conn->context,
2219                                      &isert_cmd->tx_desc.tx_cqe, NULL);
2220
2221         isert_dbg("Cmd: %p posted RDMA_READ memory for ISER Data WRITE rc: %d\n",
2222                  isert_cmd, ret);
2223         return ret;
2224 }
2225
2226 static int
2227 isert_immediate_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2228 {
2229         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2230         int ret = 0;
2231
2232         switch (state) {
2233         case ISTATE_REMOVE:
2234                 spin_lock_bh(&conn->cmd_lock);
2235                 list_del_init(&cmd->i_conn_node);
2236                 spin_unlock_bh(&conn->cmd_lock);
2237                 isert_put_cmd(isert_cmd, true);
2238                 break;
2239         case ISTATE_SEND_NOPIN_WANT_RESPONSE:
2240                 ret = isert_put_nopin(cmd, conn, false);
2241                 break;
2242         default:
2243                 isert_err("Unknown immediate state: 0x%02x\n", state);
2244                 ret = -EINVAL;
2245                 break;
2246         }
2247
2248         return ret;
2249 }
2250
2251 static int
2252 isert_response_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2253 {
2254         struct isert_conn *isert_conn = conn->context;
2255         int ret;
2256
2257         switch (state) {
2258         case ISTATE_SEND_LOGOUTRSP:
2259                 ret = isert_put_logout_rsp(cmd, conn);
2260                 if (!ret)
2261                         isert_conn->logout_posted = true;
2262                 break;
2263         case ISTATE_SEND_NOPIN:
2264                 ret = isert_put_nopin(cmd, conn, true);
2265                 break;
2266         case ISTATE_SEND_TASKMGTRSP:
2267                 ret = isert_put_tm_rsp(cmd, conn);
2268                 break;
2269         case ISTATE_SEND_REJECT:
2270                 ret = isert_put_reject(cmd, conn);
2271                 break;
2272         case ISTATE_SEND_TEXTRSP:
2273                 ret = isert_put_text_rsp(cmd, conn);
2274                 break;
2275         case ISTATE_SEND_STATUS:
2276                 /*
2277                  * Special case for sending non GOOD SCSI status from TX thread
2278                  * context during pre se_cmd excecution failure.
2279                  */
2280                 ret = isert_put_response(conn, cmd);
2281                 break;
2282         default:
2283                 isert_err("Unknown response state: 0x%02x\n", state);
2284                 ret = -EINVAL;
2285                 break;
2286         }
2287
2288         return ret;
2289 }
2290
2291 struct rdma_cm_id *
2292 isert_setup_id(struct isert_np *isert_np)
2293 {
2294         struct iscsi_np *np = isert_np->np;
2295         struct rdma_cm_id *id;
2296         struct sockaddr *sa;
2297         int ret;
2298
2299         sa = (struct sockaddr *)&np->np_sockaddr;
2300         isert_dbg("ksockaddr: %p, sa: %p\n", &np->np_sockaddr, sa);
2301
2302         id = rdma_create_id(&init_net, isert_cma_handler, isert_np,
2303                             RDMA_PS_TCP, IB_QPT_RC);
2304         if (IS_ERR(id)) {
2305                 isert_err("rdma_create_id() failed: %ld\n", PTR_ERR(id));
2306                 ret = PTR_ERR(id);
2307                 goto out;
2308         }
2309         isert_dbg("id %p context %p\n", id, id->context);
2310
2311         ret = rdma_bind_addr(id, sa);
2312         if (ret) {
2313                 isert_err("rdma_bind_addr() failed: %d\n", ret);
2314                 goto out_id;
2315         }
2316
2317         ret = rdma_listen(id, 0);
2318         if (ret) {
2319                 isert_err("rdma_listen() failed: %d\n", ret);
2320                 goto out_id;
2321         }
2322
2323         return id;
2324 out_id:
2325         rdma_destroy_id(id);
2326 out:
2327         return ERR_PTR(ret);
2328 }
2329
2330 static int
2331 isert_setup_np(struct iscsi_np *np,
2332                struct sockaddr_storage *ksockaddr)
2333 {
2334         struct isert_np *isert_np;
2335         struct rdma_cm_id *isert_lid;
2336         int ret;
2337
2338         isert_np = kzalloc(sizeof(struct isert_np), GFP_KERNEL);
2339         if (!isert_np)
2340                 return -ENOMEM;
2341
2342         sema_init(&isert_np->sem, 0);
2343         mutex_init(&isert_np->mutex);
2344         INIT_LIST_HEAD(&isert_np->accepted);
2345         INIT_LIST_HEAD(&isert_np->pending);
2346         isert_np->np = np;
2347
2348         /*
2349          * Setup the np->np_sockaddr from the passed sockaddr setup
2350          * in iscsi_target_configfs.c code..
2351          */
2352         memcpy(&np->np_sockaddr, ksockaddr,
2353                sizeof(struct sockaddr_storage));
2354
2355         isert_lid = isert_setup_id(isert_np);
2356         if (IS_ERR(isert_lid)) {
2357                 ret = PTR_ERR(isert_lid);
2358                 goto out;
2359         }
2360
2361         isert_np->cm_id = isert_lid;
2362         np->np_context = isert_np;
2363
2364         return 0;
2365
2366 out:
2367         kfree(isert_np);
2368
2369         return ret;
2370 }
2371
2372 static int
2373 isert_rdma_accept(struct isert_conn *isert_conn)
2374 {
2375         struct rdma_cm_id *cm_id = isert_conn->cm_id;
2376         struct rdma_conn_param cp;
2377         int ret;
2378         struct iser_cm_hdr rsp_hdr;
2379
2380         memset(&cp, 0, sizeof(struct rdma_conn_param));
2381         cp.initiator_depth = isert_conn->initiator_depth;
2382         cp.retry_count = 7;
2383         cp.rnr_retry_count = 7;
2384
2385         memset(&rsp_hdr, 0, sizeof(rsp_hdr));
2386         rsp_hdr.flags = ISERT_ZBVA_NOT_USED;
2387         if (!isert_conn->snd_w_inv)
2388                 rsp_hdr.flags = rsp_hdr.flags | ISERT_SEND_W_INV_NOT_USED;
2389         cp.private_data = (void *)&rsp_hdr;
2390         cp.private_data_len = sizeof(rsp_hdr);
2391
2392         ret = rdma_accept(cm_id, &cp);
2393         if (ret) {
2394                 isert_err("rdma_accept() failed with: %d\n", ret);
2395                 return ret;
2396         }
2397
2398         return 0;
2399 }
2400
2401 static int
2402 isert_get_login_rx(struct iscsi_conn *conn, struct iscsi_login *login)
2403 {
2404         struct isert_conn *isert_conn = conn->context;
2405         int ret;
2406
2407         isert_info("before login_req comp conn: %p\n", isert_conn);
2408         ret = wait_for_completion_interruptible(&isert_conn->login_req_comp);
2409         if (ret) {
2410                 isert_err("isert_conn %p interrupted before got login req\n",
2411                           isert_conn);
2412                 return ret;
2413         }
2414         reinit_completion(&isert_conn->login_req_comp);
2415
2416         /*
2417          * For login requests after the first PDU, isert_rx_login_req() will
2418          * kick schedule_delayed_work(&conn->login_work) as the packet is
2419          * received, which turns this callback from iscsi_target_do_login_rx()
2420          * into a NOP.
2421          */
2422         if (!login->first_request)
2423                 return 0;
2424
2425         isert_rx_login_req(isert_conn);
2426
2427         isert_info("before login_comp conn: %p\n", conn);
2428         ret = wait_for_completion_interruptible(&isert_conn->login_comp);
2429         if (ret)
2430                 return ret;
2431
2432         isert_info("processing login->req: %p\n", login->req);
2433
2434         return 0;
2435 }
2436
2437 static void
2438 isert_set_conn_info(struct iscsi_np *np, struct iscsi_conn *conn,
2439                     struct isert_conn *isert_conn)
2440 {
2441         struct rdma_cm_id *cm_id = isert_conn->cm_id;
2442         struct rdma_route *cm_route = &cm_id->route;
2443
2444         conn->login_family = np->np_sockaddr.ss_family;
2445
2446         conn->login_sockaddr = cm_route->addr.dst_addr;
2447         conn->local_sockaddr = cm_route->addr.src_addr;
2448 }
2449
2450 static int
2451 isert_accept_np(struct iscsi_np *np, struct iscsi_conn *conn)
2452 {
2453         struct isert_np *isert_np = np->np_context;
2454         struct isert_conn *isert_conn;
2455         int ret;
2456
2457 accept_wait:
2458         ret = down_interruptible(&isert_np->sem);
2459         if (ret)
2460                 return -ENODEV;
2461
2462         spin_lock_bh(&np->np_thread_lock);
2463         if (np->np_thread_state >= ISCSI_NP_THREAD_RESET) {
2464                 spin_unlock_bh(&np->np_thread_lock);
2465                 isert_dbg("np_thread_state %d\n",
2466                          np->np_thread_state);
2467                 /**
2468                  * No point in stalling here when np_thread
2469                  * is in state RESET/SHUTDOWN/EXIT - bail
2470                  **/
2471                 return -ENODEV;
2472         }
2473         spin_unlock_bh(&np->np_thread_lock);
2474
2475         mutex_lock(&isert_np->mutex);
2476         if (list_empty(&isert_np->pending)) {
2477                 mutex_unlock(&isert_np->mutex);
2478                 goto accept_wait;
2479         }
2480         isert_conn = list_first_entry(&isert_np->pending,
2481                         struct isert_conn, node);
2482         list_del_init(&isert_conn->node);
2483         mutex_unlock(&isert_np->mutex);
2484
2485         conn->context = isert_conn;
2486         isert_conn->conn = conn;
2487         isert_conn->state = ISER_CONN_BOUND;
2488
2489         isert_set_conn_info(np, conn, isert_conn);
2490
2491         isert_dbg("Processing isert_conn: %p\n", isert_conn);
2492
2493         return 0;
2494 }
2495
2496 static void
2497 isert_free_np(struct iscsi_np *np)
2498 {
2499         struct isert_np *isert_np = np->np_context;
2500         struct isert_conn *isert_conn, *n;
2501
2502         if (isert_np->cm_id)
2503                 rdma_destroy_id(isert_np->cm_id);
2504
2505         /*
2506          * FIXME: At this point we don't have a good way to insure
2507          * that at this point we don't have hanging connections that
2508          * completed RDMA establishment but didn't start iscsi login
2509          * process. So work-around this by cleaning up what ever piled
2510          * up in accepted and pending lists.
2511          */
2512         mutex_lock(&isert_np->mutex);
2513         if (!list_empty(&isert_np->pending)) {
2514                 isert_info("Still have isert pending connections\n");
2515                 list_for_each_entry_safe(isert_conn, n,
2516                                          &isert_np->pending,
2517                                          node) {
2518                         isert_info("cleaning isert_conn %p state (%d)\n",
2519                                    isert_conn, isert_conn->state);
2520                         isert_connect_release(isert_conn);
2521                 }
2522         }
2523
2524         if (!list_empty(&isert_np->accepted)) {
2525                 isert_info("Still have isert accepted connections\n");
2526                 list_for_each_entry_safe(isert_conn, n,
2527                                          &isert_np->accepted,
2528                                          node) {
2529                         isert_info("cleaning isert_conn %p state (%d)\n",
2530                                    isert_conn, isert_conn->state);
2531                         isert_connect_release(isert_conn);
2532                 }
2533         }
2534         mutex_unlock(&isert_np->mutex);
2535
2536         np->np_context = NULL;
2537         kfree(isert_np);
2538 }
2539
2540 static void isert_release_work(struct work_struct *work)
2541 {
2542         struct isert_conn *isert_conn = container_of(work,
2543                                                      struct isert_conn,
2544                                                      release_work);
2545
2546         isert_info("Starting release conn %p\n", isert_conn);
2547
2548         mutex_lock(&isert_conn->mutex);
2549         isert_conn->state = ISER_CONN_DOWN;
2550         mutex_unlock(&isert_conn->mutex);
2551
2552         isert_info("Destroying conn %p\n", isert_conn);
2553         isert_put_conn(isert_conn);
2554 }
2555
2556 static void
2557 isert_wait4logout(struct isert_conn *isert_conn)
2558 {
2559         struct iscsi_conn *conn = isert_conn->conn;
2560
2561         isert_info("conn %p\n", isert_conn);
2562
2563         if (isert_conn->logout_posted) {
2564                 isert_info("conn %p wait for conn_logout_comp\n", isert_conn);
2565                 wait_for_completion_timeout(&conn->conn_logout_comp,
2566                                             SECONDS_FOR_LOGOUT_COMP * HZ);
2567         }
2568 }
2569
2570 static void
2571 isert_wait4cmds(struct iscsi_conn *conn)
2572 {
2573         isert_info("iscsi_conn %p\n", conn);
2574
2575         if (conn->sess) {
2576                 target_sess_cmd_list_set_waiting(conn->sess->se_sess);
2577                 target_wait_for_sess_cmds(conn->sess->se_sess);
2578         }
2579 }
2580
2581 /**
2582  * isert_put_unsol_pending_cmds() - Drop commands waiting for
2583  *     unsolicitate dataout
2584  * @conn:    iscsi connection
2585  *
2586  * We might still have commands that are waiting for unsolicited
2587  * dataouts messages. We must put the extra reference on those
2588  * before blocking on the target_wait_for_session_cmds
2589  */
2590 static void
2591 isert_put_unsol_pending_cmds(struct iscsi_conn *conn)
2592 {
2593         struct iscsi_cmd *cmd, *tmp;
2594         static LIST_HEAD(drop_cmd_list);
2595
2596         spin_lock_bh(&conn->cmd_lock);
2597         list_for_each_entry_safe(cmd, tmp, &conn->conn_cmd_list, i_conn_node) {
2598                 if ((cmd->cmd_flags & ICF_NON_IMMEDIATE_UNSOLICITED_DATA) &&
2599                     (cmd->write_data_done < conn->sess->sess_ops->FirstBurstLength) &&
2600                     (cmd->write_data_done < cmd->se_cmd.data_length))
2601                         list_move_tail(&cmd->i_conn_node, &drop_cmd_list);
2602         }
2603         spin_unlock_bh(&conn->cmd_lock);
2604
2605         list_for_each_entry_safe(cmd, tmp, &drop_cmd_list, i_conn_node) {
2606                 list_del_init(&cmd->i_conn_node);
2607                 if (cmd->i_state != ISTATE_REMOVE) {
2608                         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2609
2610                         isert_info("conn %p dropping cmd %p\n", conn, cmd);
2611                         isert_put_cmd(isert_cmd, true);
2612                 }
2613         }
2614 }
2615
2616 static void isert_wait_conn(struct iscsi_conn *conn)
2617 {
2618         struct isert_conn *isert_conn = conn->context;
2619
2620         isert_info("Starting conn %p\n", isert_conn);
2621
2622         mutex_lock(&isert_conn->mutex);
2623         isert_conn_terminate(isert_conn);
2624         mutex_unlock(&isert_conn->mutex);
2625
2626         ib_drain_qp(isert_conn->qp);
2627         isert_put_unsol_pending_cmds(conn);
2628         isert_wait4cmds(conn);
2629         isert_wait4logout(isert_conn);
2630
2631         queue_work(isert_release_wq, &isert_conn->release_work);
2632 }
2633
2634 static void isert_free_conn(struct iscsi_conn *conn)
2635 {
2636         struct isert_conn *isert_conn = conn->context;
2637
2638         ib_drain_qp(isert_conn->qp);
2639         isert_put_conn(isert_conn);
2640 }
2641
2642 static void isert_get_rx_pdu(struct iscsi_conn *conn)
2643 {
2644         struct completion comp;
2645
2646         init_completion(&comp);
2647
2648         wait_for_completion_interruptible(&comp);
2649 }
2650
2651 static struct iscsit_transport iser_target_transport = {
2652         .name                   = "IB/iSER",
2653         .transport_type         = ISCSI_INFINIBAND,
2654         .rdma_shutdown          = true,
2655         .priv_size              = sizeof(struct isert_cmd),
2656         .owner                  = THIS_MODULE,
2657         .iscsit_setup_np        = isert_setup_np,
2658         .iscsit_accept_np       = isert_accept_np,
2659         .iscsit_free_np         = isert_free_np,
2660         .iscsit_wait_conn       = isert_wait_conn,
2661         .iscsit_free_conn       = isert_free_conn,
2662         .iscsit_get_login_rx    = isert_get_login_rx,
2663         .iscsit_put_login_tx    = isert_put_login_tx,
2664         .iscsit_immediate_queue = isert_immediate_queue,
2665         .iscsit_response_queue  = isert_response_queue,
2666         .iscsit_get_dataout     = isert_get_dataout,
2667         .iscsit_queue_data_in   = isert_put_datain,
2668         .iscsit_queue_status    = isert_put_response,
2669         .iscsit_aborted_task    = isert_aborted_task,
2670         .iscsit_get_rx_pdu      = isert_get_rx_pdu,
2671         .iscsit_get_sup_prot_ops = isert_get_sup_prot_ops,
2672 };
2673
2674 static int __init isert_init(void)
2675 {
2676         int ret;
2677
2678         isert_comp_wq = alloc_workqueue("isert_comp_wq",
2679                                         WQ_UNBOUND | WQ_HIGHPRI, 0);
2680         if (!isert_comp_wq) {
2681                 isert_err("Unable to allocate isert_comp_wq\n");
2682                 return -ENOMEM;
2683         }
2684
2685         isert_release_wq = alloc_workqueue("isert_release_wq", WQ_UNBOUND,
2686                                         WQ_UNBOUND_MAX_ACTIVE);
2687         if (!isert_release_wq) {
2688                 isert_err("Unable to allocate isert_release_wq\n");
2689                 ret = -ENOMEM;
2690                 goto destroy_comp_wq;
2691         }
2692
2693         iscsit_register_transport(&iser_target_transport);
2694         isert_info("iSER_TARGET[0] - Loaded iser_target_transport\n");
2695
2696         return 0;
2697
2698 destroy_comp_wq:
2699         destroy_workqueue(isert_comp_wq);
2700
2701         return ret;
2702 }
2703
2704 static void __exit isert_exit(void)
2705 {
2706         flush_scheduled_work();
2707         destroy_workqueue(isert_release_wq);
2708         destroy_workqueue(isert_comp_wq);
2709         iscsit_unregister_transport(&iser_target_transport);
2710         isert_info("iSER_TARGET[0] - Released iser_target_transport\n");
2711 }
2712
2713 MODULE_DESCRIPTION("iSER-Target for mainline target infrastructure");
2714 MODULE_AUTHOR("nab@Linux-iSCSI.org");
2715 MODULE_LICENSE("GPL");
2716
2717 module_init(isert_init);
2718 module_exit(isert_exit);