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         struct isert_np *isert_np = cma_id->context;
 744
 745         ib_drain_qp(isert_conn->qp);
 746
 747         mutex_lock(&isert_np->mutex);
 748         list_del_init(&isert_conn->node);
 749         mutex_unlock(&isert_np->mutex);
 750         isert_conn->cm_id = NULL;
 751         isert_put_conn(isert_conn);
 752
 753         return -1;
 754 }
 755
 756 static int
 757 isert_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 758 {
 759         struct isert_np *isert_np = cma_id->context;
 760         struct isert_conn *isert_conn;
 761         int ret = 0;
 762
 763         isert_info("%s (%d): status %d id %p np %p\n",
 764                    rdma_event_msg(event->event), event->event,
 765                    event->status, cma_id, cma_id->context);
 766
 767         if (isert_np->cm_id == cma_id)
 768                 return isert_np_cma_handler(cma_id->context, event->event);
 769
 770         switch (event->event) {
 771         case RDMA_CM_EVENT_CONNECT_REQUEST:
 772                 ret = isert_connect_request(cma_id, event);
 773                 if (ret)
 774                         isert_err("failed handle connect request %d\n", ret);
 775                 break;
 776         case RDMA_CM_EVENT_ESTABLISHED:
 777                 isert_connected_handler(cma_id);
 778                 break;
 779         case RDMA_CM_EVENT_ADDR_CHANGE:    /* FALLTHRU */
 780         case RDMA_CM_EVENT_DISCONNECTED:   /* FALLTHRU */
 781         case RDMA_CM_EVENT_TIMEWAIT_EXIT:  /* FALLTHRU */
 782                 ret = isert_disconnected_handler(cma_id, event->event);
 783                 break;
 784         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 785                 isert_conn = cma_id->qp->qp_context;
 786                 isert_conn->dev_removed = true;
 787                 isert_disconnected_handler(cma_id, event->event);
 788                 wait_event_interruptible(isert_conn->rem_wait,
 789                                          isert_conn->state == ISER_CONN_DOWN);
 790                 kfree(isert_conn);
 791                 /*
 792                  * return non-zero from the callback to destroy
 793                  * the rdma cm id
 794                  */
 795                 return 1;
 796         case RDMA_CM_EVENT_REJECTED:       /* FALLTHRU */
 797                 isert_info("Connection rejected: %s\n",
 798                            rdma_reject_msg(cma_id, event->status));
 799         case RDMA_CM_EVENT_UNREACHABLE:    /* FALLTHRU */
 800         case RDMA_CM_EVENT_CONNECT_ERROR:
 801                 ret = isert_connect_error(cma_id);
 802                 break;
 803         default:
 804                 isert_err("Unhandled RDMA CMA event: %d\n", event->event);
 805                 break;
 806         }
 807
 808         return ret;
 809 }
 810
 811 static int
 812 isert_post_recvm(struct isert_conn *isert_conn, u32 count)
 813 {
 814         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 815         int i, ret;
 816         struct iser_rx_desc *rx_desc;
 817
 818         for (rx_wr = isert_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
 819                 rx_desc = &isert_conn->rx_descs[i];
 820
 821                 rx_wr->wr_cqe = &rx_desc->rx_cqe;
 822                 rx_wr->sg_list = &rx_desc->rx_sg;
 823                 rx_wr->num_sge = 1;
 824                 rx_wr->next = rx_wr + 1;
 825                 rx_desc->in_use = false;
 826         }
 827         rx_wr--;
 828         rx_wr->next = NULL; /* mark end of work requests list */
 829
 830         ret = ib_post_recv(isert_conn->qp, isert_conn->rx_wr,
 831                            &rx_wr_failed);
 832         if (ret)
 833                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
 834
 835         return ret;
 836 }
 837
 838 static int
 839 isert_post_recv(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc)
 840 {
 841         struct ib_recv_wr *rx_wr_failed, rx_wr;
 842         int ret;
 843
 844         if (!rx_desc->in_use) {
 845                 /*
 846                  * if the descriptor is not in-use we already reposted it
 847                  * for recv, so just silently return
 848                  */
 849                 return 0;
 850         }
 851
 852         rx_desc->in_use = false;
 853         rx_wr.wr_cqe = &rx_desc->rx_cqe;
 854         rx_wr.sg_list = &rx_desc->rx_sg;
 855         rx_wr.num_sge = 1;
 856         rx_wr.next = NULL;
 857
 858         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_failed);
 859         if (ret)
 860                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
 861
 862         return ret;
 863 }
 864
 865 static int
 866 isert_login_post_send(struct isert_conn *isert_conn, struct iser_tx_desc *tx_desc)
 867 {
 868         struct ib_device *ib_dev = isert_conn->cm_id->device;
 869         struct ib_send_wr send_wr, *send_wr_failed;
 870         int ret;
 871
 872         ib_dma_sync_single_for_device(ib_dev, tx_desc->dma_addr,
 873                                       ISER_HEADERS_LEN, DMA_TO_DEVICE);
 874
 875         tx_desc->tx_cqe.done = isert_login_send_done;
 876
 877         send_wr.next    = NULL;
 878         send_wr.wr_cqe  = &tx_desc->tx_cqe;
 879         send_wr.sg_list = tx_desc->tx_sg;
 880         send_wr.num_sge = tx_desc->num_sge;
 881         send_wr.opcode  = IB_WR_SEND;
 882         send_wr.send_flags = IB_SEND_SIGNALED;
 883
 884         ret = ib_post_send(isert_conn->qp, &send_wr, &send_wr_failed);
 885         if (ret)
 886                 isert_err("ib_post_send() failed, ret: %d\n", ret);
 887
 888         return ret;
 889 }
 890
 891 static void
 892 __isert_create_send_desc(struct isert_device *device,
 893                          struct iser_tx_desc *tx_desc)
 894 {
 895
 896         memset(&tx_desc->iser_header, 0, sizeof(struct iser_ctrl));
 897         tx_desc->iser_header.flags = ISCSI_CTRL;
 898
 899         tx_desc->num_sge = 1;
 900
 901         if (tx_desc->tx_sg[0].lkey != device->pd->local_dma_lkey) {
 902                 tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
 903                 isert_dbg("tx_desc %p lkey mismatch, fixing\n", tx_desc);
 904         }
 905 }
 906
 907 static void
 908 isert_create_send_desc(struct isert_conn *isert_conn,
 909                        struct isert_cmd *isert_cmd,
 910                        struct iser_tx_desc *tx_desc)
 911 {
 912         struct isert_device *device = isert_conn->device;
 913         struct ib_device *ib_dev = device->ib_device;
 914
 915         ib_dma_sync_single_for_cpu(ib_dev, tx_desc->dma_addr,
 916                                    ISER_HEADERS_LEN, DMA_TO_DEVICE);
 917
 918         __isert_create_send_desc(device, tx_desc);
 919 }
 920
 921 static int
 922 isert_init_tx_hdrs(struct isert_conn *isert_conn,
 923                    struct iser_tx_desc *tx_desc)
 924 {
 925         struct isert_device *device = isert_conn->device;
 926         struct ib_device *ib_dev = device->ib_device;
 927         u64 dma_addr;
 928
 929         dma_addr = ib_dma_map_single(ib_dev, (void *)tx_desc,
 930                         ISER_HEADERS_LEN, DMA_TO_DEVICE);
 931         if (ib_dma_mapping_error(ib_dev, dma_addr)) {
 932                 isert_err("ib_dma_mapping_error() failed\n");
 933                 return -ENOMEM;
 934         }
 935
 936         tx_desc->dma_addr = dma_addr;
 937         tx_desc->tx_sg[0].addr  = tx_desc->dma_addr;
 938         tx_desc->tx_sg[0].length = ISER_HEADERS_LEN;
 939         tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
 940
 941         isert_dbg("Setup tx_sg[0].addr: 0x%llx length: %u lkey: 0x%x\n",
 942                   tx_desc->tx_sg[0].addr, tx_desc->tx_sg[0].length,
 943                   tx_desc->tx_sg[0].lkey);
 944
 945         return 0;
 946 }
 947
 948 static void
 949 isert_init_send_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
 950                    struct ib_send_wr *send_wr)
 951 {
 952         struct iser_tx_desc *tx_desc = &isert_cmd->tx_desc;
 953
 954         tx_desc->tx_cqe.done = isert_send_done;
 955         send_wr->wr_cqe = &tx_desc->tx_cqe;
 956
 957         if (isert_conn->snd_w_inv && isert_cmd->inv_rkey) {
 958                 send_wr->opcode  = IB_WR_SEND_WITH_INV;
 959                 send_wr->ex.invalidate_rkey = isert_cmd->inv_rkey;
 960         } else {
 961                 send_wr->opcode = IB_WR_SEND;
 962         }
 963
 964         send_wr->sg_list = &tx_desc->tx_sg[0];
 965         send_wr->num_sge = isert_cmd->tx_desc.num_sge;
 966         send_wr->send_flags = IB_SEND_SIGNALED;
 967 }
 968
 969 static int
 970 isert_login_post_recv(struct isert_conn *isert_conn)
 971 {
 972         struct ib_recv_wr rx_wr, *rx_wr_fail;
 973         struct ib_sge sge;
 974         int ret;
 975
 976         memset(&sge, 0, sizeof(struct ib_sge));
 977         sge.addr = isert_conn->login_req_dma;
 978         sge.length = ISER_RX_PAYLOAD_SIZE;
 979         sge.lkey = isert_conn->device->pd->local_dma_lkey;
 980
 981         isert_dbg("Setup sge: addr: %llx length: %d 0x%08x\n",
 982                 sge.addr, sge.length, sge.lkey);
 983
 984         isert_conn->login_req_buf->rx_cqe.done = isert_login_recv_done;
 985
 986         memset(&rx_wr, 0, sizeof(struct ib_recv_wr));
 987         rx_wr.wr_cqe = &isert_conn->login_req_buf->rx_cqe;
 988         rx_wr.sg_list = &sge;
 989         rx_wr.num_sge = 1;
 990
 991         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_fail);
 992         if (ret)
 993                 isert_err("ib_post_recv() failed: %d\n", ret);
 994
 995         return ret;
 996 }
 997
 998 static int
 999 isert_put_login_tx(struct iscsi_conn *conn, struct iscsi_login *login,
1000                    u32 length)
1001 {
1002         struct isert_conn *isert_conn = conn->context;
1003         struct isert_device *device = isert_conn->device;
1004         struct ib_device *ib_dev = device->ib_device;
1005         struct iser_tx_desc *tx_desc = &isert_conn->login_tx_desc;
1006         int ret;
1007
1008         __isert_create_send_desc(device, tx_desc);
1009
1010         memcpy(&tx_desc->iscsi_header, &login->rsp[0],
1011                sizeof(struct iscsi_hdr));
1012
1013         isert_init_tx_hdrs(isert_conn, tx_desc);
1014
1015         if (length > 0) {
1016                 struct ib_sge *tx_dsg = &tx_desc->tx_sg[1];
1017
1018                 ib_dma_sync_single_for_cpu(ib_dev, isert_conn->login_rsp_dma,
1019                                            length, DMA_TO_DEVICE);
1020
1021                 memcpy(isert_conn->login_rsp_buf, login->rsp_buf, length);
1022
1023                 ib_dma_sync_single_for_device(ib_dev, isert_conn->login_rsp_dma,
1024                                               length, DMA_TO_DEVICE);
1025
1026                 tx_dsg->addr    = isert_conn->login_rsp_dma;
1027                 tx_dsg->length  = length;
1028                 tx_dsg->lkey    = isert_conn->device->pd->local_dma_lkey;
1029                 tx_desc->num_sge = 2;
1030         }
1031         if (!login->login_failed) {
1032                 if (login->login_complete) {
1033                         ret = isert_alloc_rx_descriptors(isert_conn);
1034                         if (ret)
1035                                 return ret;
1036
1037                         ret = isert_post_recvm(isert_conn,
1038                                                ISERT_QP_MAX_RECV_DTOS);
1039                         if (ret)
1040                                 return ret;
1041
1042                         /* Now we are in FULL_FEATURE phase */
1043                         mutex_lock(&isert_conn->mutex);
1044                         isert_conn->state = ISER_CONN_FULL_FEATURE;
1045                         mutex_unlock(&isert_conn->mutex);
1046                         goto post_send;
1047                 }
1048
1049                 ret = isert_login_post_recv(isert_conn);
1050                 if (ret)
1051                         return ret;
1052         }
1053 post_send:
1054         ret = isert_login_post_send(isert_conn, tx_desc);
1055         if (ret)
1056                 return ret;
1057
1058         return 0;
1059 }
1060
1061 static void
1062 isert_rx_login_req(struct isert_conn *isert_conn)
1063 {
1064         struct iser_rx_desc *rx_desc = isert_conn->login_req_buf;
1065         int rx_buflen = isert_conn->login_req_len;
1066         struct iscsi_conn *conn = isert_conn->conn;
1067         struct iscsi_login *login = conn->conn_login;
1068         int size;
1069
1070         isert_info("conn %p\n", isert_conn);
1071
1072         WARN_ON_ONCE(!login);
1073
1074         if (login->first_request) {
1075                 struct iscsi_login_req *login_req =
1076                         (struct iscsi_login_req *)&rx_desc->iscsi_header;
1077                 /*
1078                  * Setup the initial iscsi_login values from the leading
1079                  * login request PDU.
1080                  */
1081                 login->leading_connection = (!login_req->tsih) ? 1 : 0;
1082                 login->current_stage =
1083                         (login_req->flags & ISCSI_FLAG_LOGIN_CURRENT_STAGE_MASK)
1084                          >> 2;
1085                 login->version_min      = login_req->min_version;
1086                 login->version_max      = login_req->max_version;
1087                 memcpy(login->isid, login_req->isid, 6);
1088                 login->cmd_sn           = be32_to_cpu(login_req->cmdsn);
1089                 login->init_task_tag    = login_req->itt;
1090                 login->initial_exp_statsn = be32_to_cpu(login_req->exp_statsn);
1091                 login->cid              = be16_to_cpu(login_req->cid);
1092                 login->tsih             = be16_to_cpu(login_req->tsih);
1093         }
1094
1095         memcpy(&login->req[0], (void *)&rx_desc->iscsi_header, ISCSI_HDR_LEN);
1096
1097         size = min(rx_buflen, MAX_KEY_VALUE_PAIRS);
1098         isert_dbg("Using login payload size: %d, rx_buflen: %d "
1099                   "MAX_KEY_VALUE_PAIRS: %d\n", size, rx_buflen,
1100                   MAX_KEY_VALUE_PAIRS);
1101         memcpy(login->req_buf, &rx_desc->data[0], size);
1102
1103         if (login->first_request) {
1104                 complete(&isert_conn->login_comp);
1105                 return;
1106         }
1107         schedule_delayed_work(&conn->login_work, 0);
1108 }
1109
1110 static struct iscsi_cmd
1111 *isert_allocate_cmd(struct iscsi_conn *conn, struct iser_rx_desc *rx_desc)
1112 {
1113         struct isert_conn *isert_conn = conn->context;
1114         struct isert_cmd *isert_cmd;
1115         struct iscsi_cmd *cmd;
1116
1117         cmd = iscsit_allocate_cmd(conn, TASK_INTERRUPTIBLE);
1118         if (!cmd) {
1119                 isert_err("Unable to allocate iscsi_cmd + isert_cmd\n");
1120                 return NULL;
1121         }
1122         isert_cmd = iscsit_priv_cmd(cmd);
1123         isert_cmd->conn = isert_conn;
1124         isert_cmd->iscsi_cmd = cmd;
1125         isert_cmd->rx_desc = rx_desc;
1126
1127         return cmd;
1128 }
1129
1130 static int
1131 isert_handle_scsi_cmd(struct isert_conn *isert_conn,
1132                       struct isert_cmd *isert_cmd, struct iscsi_cmd *cmd,
1133                       struct iser_rx_desc *rx_desc, unsigned char *buf)
1134 {
1135         struct iscsi_conn *conn = isert_conn->conn;
1136         struct iscsi_scsi_req *hdr = (struct iscsi_scsi_req *)buf;
1137         int imm_data, imm_data_len, unsol_data, sg_nents, rc;
1138         bool dump_payload = false;
1139         unsigned int data_len;
1140
1141         rc = iscsit_setup_scsi_cmd(conn, cmd, buf);
1142         if (rc < 0)
1143                 return rc;
1144
1145         imm_data = cmd->immediate_data;
1146         imm_data_len = cmd->first_burst_len;
1147         unsol_data = cmd->unsolicited_data;
1148         data_len = cmd->se_cmd.data_length;
1149
1150         if (imm_data && imm_data_len == data_len)
1151                 cmd->se_cmd.se_cmd_flags |= SCF_PASSTHROUGH_SG_TO_MEM_NOALLOC;
1152         rc = iscsit_process_scsi_cmd(conn, cmd, hdr);
1153         if (rc < 0) {
1154                 return 0;
1155         } else if (rc > 0) {
1156                 dump_payload = true;
1157                 goto sequence_cmd;
1158         }
1159
1160         if (!imm_data)
1161                 return 0;
1162
1163         if (imm_data_len != data_len) {
1164                 sg_nents = max(1UL, DIV_ROUND_UP(imm_data_len, PAGE_SIZE));
1165                 sg_copy_from_buffer(cmd->se_cmd.t_data_sg, sg_nents,
1166                                     &rx_desc->data[0], imm_data_len);
1167                 isert_dbg("Copy Immediate sg_nents: %u imm_data_len: %d\n",
1168                           sg_nents, imm_data_len);
1169         } else {
1170                 sg_init_table(&isert_cmd->sg, 1);
1171                 cmd->se_cmd.t_data_sg = &isert_cmd->sg;
1172                 cmd->se_cmd.t_data_nents = 1;
1173                 sg_set_buf(&isert_cmd->sg, &rx_desc->data[0], imm_data_len);
1174                 isert_dbg("Transfer Immediate imm_data_len: %d\n",
1175                           imm_data_len);
1176         }
1177
1178         cmd->write_data_done += imm_data_len;
1179
1180         if (cmd->write_data_done == cmd->se_cmd.data_length) {
1181                 spin_lock_bh(&cmd->istate_lock);
1182                 cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1183                 cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1184                 spin_unlock_bh(&cmd->istate_lock);
1185         }
1186
1187 sequence_cmd:
1188         rc = iscsit_sequence_cmd(conn, cmd, buf, hdr->cmdsn);
1189
1190         if (!rc && dump_payload == false && unsol_data)
1191                 iscsit_set_unsoliticed_dataout(cmd);
1192         else if (dump_payload && imm_data)
1193                 target_put_sess_cmd(&cmd->se_cmd);
1194
1195         return 0;
1196 }
1197
1198 static int
1199 isert_handle_iscsi_dataout(struct isert_conn *isert_conn,
1200                            struct iser_rx_desc *rx_desc, unsigned char *buf)
1201 {
1202         struct scatterlist *sg_start;
1203         struct iscsi_conn *conn = isert_conn->conn;
1204         struct iscsi_cmd *cmd = NULL;
1205         struct iscsi_data *hdr = (struct iscsi_data *)buf;
1206         u32 unsol_data_len = ntoh24(hdr->dlength);
1207         int rc, sg_nents, sg_off, page_off;
1208
1209         rc = iscsit_check_dataout_hdr(conn, buf, &cmd);
1210         if (rc < 0)
1211                 return rc;
1212         else if (!cmd)
1213                 return 0;
1214         /*
1215          * FIXME: Unexpected unsolicited_data out
1216          */
1217         if (!cmd->unsolicited_data) {
1218                 isert_err("Received unexpected solicited data payload\n");
1219                 dump_stack();
1220                 return -1;
1221         }
1222
1223         isert_dbg("Unsolicited DataOut unsol_data_len: %u, "
1224                   "write_data_done: %u, data_length: %u\n",
1225                   unsol_data_len,  cmd->write_data_done,
1226                   cmd->se_cmd.data_length);
1227
1228         sg_off = cmd->write_data_done / PAGE_SIZE;
1229         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
1230         sg_nents = max(1UL, DIV_ROUND_UP(unsol_data_len, PAGE_SIZE));
1231         page_off = cmd->write_data_done % PAGE_SIZE;
1232         /*
1233          * FIXME: Non page-aligned unsolicited_data out
1234          */
1235         if (page_off) {
1236                 isert_err("unexpected non-page aligned data payload\n");
1237                 dump_stack();
1238                 return -1;
1239         }
1240         isert_dbg("Copying DataOut: sg_start: %p, sg_off: %u "
1241                   "sg_nents: %u from %p %u\n", sg_start, sg_off,
1242                   sg_nents, &rx_desc->data[0], unsol_data_len);
1243
1244         sg_copy_from_buffer(sg_start, sg_nents, &rx_desc->data[0],
1245                             unsol_data_len);
1246
1247         rc = iscsit_check_dataout_payload(cmd, hdr, false);
1248         if (rc < 0)
1249                 return rc;
1250
1251         /*
1252          * multiple data-outs on the same command can arrive -
1253          * so post the buffer before hand
1254          */
1255         rc = isert_post_recv(isert_conn, rx_desc);
1256         if (rc) {
1257                 isert_err("ib_post_recv failed with %d\n", rc);
1258                 return rc;
1259         }
1260         return 0;
1261 }
1262
1263 static int
1264 isert_handle_nop_out(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1265                      struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1266                      unsigned char *buf)
1267 {
1268         struct iscsi_conn *conn = isert_conn->conn;
1269         struct iscsi_nopout *hdr = (struct iscsi_nopout *)buf;
1270         int rc;
1271
1272         rc = iscsit_setup_nop_out(conn, cmd, hdr);
1273         if (rc < 0)
1274                 return rc;
1275         /*
1276          * FIXME: Add support for NOPOUT payload using unsolicited RDMA payload
1277          */
1278
1279         return iscsit_process_nop_out(conn, cmd, hdr);
1280 }
1281
1282 static int
1283 isert_handle_text_cmd(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1284                       struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1285                       struct iscsi_text *hdr)
1286 {
1287         struct iscsi_conn *conn = isert_conn->conn;
1288         u32 payload_length = ntoh24(hdr->dlength);
1289         int rc;
1290         unsigned char *text_in = NULL;
1291
1292         rc = iscsit_setup_text_cmd(conn, cmd, hdr);
1293         if (rc < 0)
1294                 return rc;
1295
1296         if (payload_length) {
1297                 text_in = kzalloc(payload_length, GFP_KERNEL);
1298                 if (!text_in)
1299                         return -ENOMEM;
1300         }
1301         cmd->text_in_ptr = text_in;
1302
1303         memcpy(cmd->text_in_ptr, &rx_desc->data[0], payload_length);
1304
1305         return iscsit_process_text_cmd(conn, cmd, hdr);
1306 }
1307
1308 static int
1309 isert_rx_opcode(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc,
1310                 uint32_t read_stag, uint64_t read_va,
1311                 uint32_t write_stag, uint64_t write_va)
1312 {
1313         struct iscsi_hdr *hdr = &rx_desc->iscsi_header;
1314         struct iscsi_conn *conn = isert_conn->conn;
1315         struct iscsi_cmd *cmd;
1316         struct isert_cmd *isert_cmd;
1317         int ret = -EINVAL;
1318         u8 opcode = (hdr->opcode & ISCSI_OPCODE_MASK);
1319
1320         if (conn->sess->sess_ops->SessionType &&
1321            (!(opcode & ISCSI_OP_TEXT) || !(opcode & ISCSI_OP_LOGOUT))) {
1322                 isert_err("Got illegal opcode: 0x%02x in SessionType=Discovery,"
1323                           " ignoring\n", opcode);
1324                 return 0;
1325         }
1326
1327         switch (opcode) {
1328         case ISCSI_OP_SCSI_CMD:
1329                 cmd = isert_allocate_cmd(conn, rx_desc);
1330                 if (!cmd)
1331                         break;
1332
1333                 isert_cmd = iscsit_priv_cmd(cmd);
1334                 isert_cmd->read_stag = read_stag;
1335                 isert_cmd->read_va = read_va;
1336                 isert_cmd->write_stag = write_stag;
1337                 isert_cmd->write_va = write_va;
1338                 isert_cmd->inv_rkey = read_stag ? read_stag : write_stag;
1339
1340                 ret = isert_handle_scsi_cmd(isert_conn, isert_cmd, cmd,
1341                                         rx_desc, (unsigned char *)hdr);
1342                 break;
1343         case ISCSI_OP_NOOP_OUT:
1344                 cmd = isert_allocate_cmd(conn, rx_desc);
1345                 if (!cmd)
1346                         break;
1347
1348                 isert_cmd = iscsit_priv_cmd(cmd);
1349                 ret = isert_handle_nop_out(isert_conn, isert_cmd, cmd,
1350                                            rx_desc, (unsigned char *)hdr);
1351                 break;
1352         case ISCSI_OP_SCSI_DATA_OUT:
1353                 ret = isert_handle_iscsi_dataout(isert_conn, rx_desc,
1354                                                 (unsigned char *)hdr);
1355                 break;
1356         case ISCSI_OP_SCSI_TMFUNC:
1357                 cmd = isert_allocate_cmd(conn, rx_desc);
1358                 if (!cmd)
1359                         break;
1360
1361                 ret = iscsit_handle_task_mgt_cmd(conn, cmd,
1362                                                 (unsigned char *)hdr);
1363                 break;
1364         case ISCSI_OP_LOGOUT:
1365                 cmd = isert_allocate_cmd(conn, rx_desc);
1366                 if (!cmd)
1367                         break;
1368
1369                 ret = iscsit_handle_logout_cmd(conn, cmd, (unsigned char *)hdr);
1370                 break;
1371         case ISCSI_OP_TEXT:
1372                 if (be32_to_cpu(hdr->ttt) != 0xFFFFFFFF)
1373                         cmd = iscsit_find_cmd_from_itt(conn, hdr->itt);
1374                 else
1375                         cmd = isert_allocate_cmd(conn, rx_desc);
1376
1377                 if (!cmd)
1378                         break;
1379
1380                 isert_cmd = iscsit_priv_cmd(cmd);
1381                 ret = isert_handle_text_cmd(isert_conn, isert_cmd, cmd,
1382                                             rx_desc, (struct iscsi_text *)hdr);
1383                 break;
1384         default:
1385                 isert_err("Got unknown iSCSI OpCode: 0x%02x\n", opcode);
1386                 dump_stack();
1387                 break;
1388         }
1389
1390         return ret;
1391 }
1392
1393 static void
1394 isert_print_wc(struct ib_wc *wc, const char *type)
1395 {
1396         if (wc->status != IB_WC_WR_FLUSH_ERR)
1397                 isert_err("%s failure: %s (%d) vend_err %x\n", type,
1398                           ib_wc_status_msg(wc->status), wc->status,
1399                           wc->vendor_err);
1400         else
1401                 isert_dbg("%s failure: %s (%d)\n", type,
1402                           ib_wc_status_msg(wc->status), wc->status);
1403 }
1404
1405 static void
1406 isert_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1407 {
1408         struct isert_conn *isert_conn = wc->qp->qp_context;
1409         struct ib_device *ib_dev = isert_conn->cm_id->device;
1410         struct iser_rx_desc *rx_desc = cqe_to_rx_desc(wc->wr_cqe);
1411         struct iscsi_hdr *hdr = &rx_desc->iscsi_header;
1412         struct iser_ctrl *iser_ctrl = &rx_desc->iser_header;
1413         uint64_t read_va = 0, write_va = 0;
1414         uint32_t read_stag = 0, write_stag = 0;
1415
1416         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1417                 isert_print_wc(wc, "recv");
1418                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1419                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1420                 return;
1421         }
1422
1423         rx_desc->in_use = true;
1424
1425         ib_dma_sync_single_for_cpu(ib_dev, rx_desc->dma_addr,
1426                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1427
1428         isert_dbg("DMA: 0x%llx, iSCSI opcode: 0x%02x, ITT: 0x%08x, flags: 0x%02x dlen: %d\n",
1429                  rx_desc->dma_addr, hdr->opcode, hdr->itt, hdr->flags,
1430                  (int)(wc->byte_len - ISER_HEADERS_LEN));
1431
1432         switch (iser_ctrl->flags & 0xF0) {
1433         case ISCSI_CTRL:
1434                 if (iser_ctrl->flags & ISER_RSV) {
1435                         read_stag = be32_to_cpu(iser_ctrl->read_stag);
1436                         read_va = be64_to_cpu(iser_ctrl->read_va);
1437                         isert_dbg("ISER_RSV: read_stag: 0x%x read_va: 0x%llx\n",
1438                                   read_stag, (unsigned long long)read_va);
1439                 }
1440                 if (iser_ctrl->flags & ISER_WSV) {
1441                         write_stag = be32_to_cpu(iser_ctrl->write_stag);
1442                         write_va = be64_to_cpu(iser_ctrl->write_va);
1443                         isert_dbg("ISER_WSV: write_stag: 0x%x write_va: 0x%llx\n",
1444                                   write_stag, (unsigned long long)write_va);
1445                 }
1446
1447                 isert_dbg("ISER ISCSI_CTRL PDU\n");
1448                 break;
1449         case ISER_HELLO:
1450                 isert_err("iSER Hello message\n");
1451                 break;
1452         default:
1453                 isert_warn("Unknown iSER hdr flags: 0x%02x\n", iser_ctrl->flags);
1454                 break;
1455         }
1456
1457         isert_rx_opcode(isert_conn, rx_desc,
1458                         read_stag, read_va, write_stag, write_va);
1459
1460         ib_dma_sync_single_for_device(ib_dev, rx_desc->dma_addr,
1461                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1462 }
1463
1464 static void
1465 isert_login_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1466 {
1467         struct isert_conn *isert_conn = wc->qp->qp_context;
1468         struct ib_device *ib_dev = isert_conn->device->ib_device;
1469
1470         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1471                 isert_print_wc(wc, "login recv");
1472                 return;
1473         }
1474
1475         ib_dma_sync_single_for_cpu(ib_dev, isert_conn->login_req_dma,
1476                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1477
1478         isert_conn->login_req_len = wc->byte_len - ISER_HEADERS_LEN;
1479
1480         if (isert_conn->conn) {
1481                 struct iscsi_login *login = isert_conn->conn->conn_login;
1482
1483                 if (login && !login->first_request)
1484                         isert_rx_login_req(isert_conn);
1485         }
1486
1487         mutex_lock(&isert_conn->mutex);
1488         complete(&isert_conn->login_req_comp);
1489         mutex_unlock(&isert_conn->mutex);
1490
1491         ib_dma_sync_single_for_device(ib_dev, isert_conn->login_req_dma,
1492                                 ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1493 }
1494
1495 static void
1496 isert_rdma_rw_ctx_destroy(struct isert_cmd *cmd, struct isert_conn *conn)
1497 {
1498         struct se_cmd *se_cmd = &cmd->iscsi_cmd->se_cmd;
1499         enum dma_data_direction dir = target_reverse_dma_direction(se_cmd);
1500
1501         if (!cmd->rw.nr_ops)
1502                 return;
1503
1504         if (isert_prot_cmd(conn, se_cmd)) {
1505                 rdma_rw_ctx_destroy_signature(&cmd->rw, conn->qp,
1506                                 conn->cm_id->port_num, se_cmd->t_data_sg,
1507                                 se_cmd->t_data_nents, se_cmd->t_prot_sg,
1508                                 se_cmd->t_prot_nents, dir);
1509         } else {
1510                 rdma_rw_ctx_destroy(&cmd->rw, conn->qp, conn->cm_id->port_num,
1511                                 se_cmd->t_data_sg, se_cmd->t_data_nents, dir);
1512         }
1513
1514         cmd->rw.nr_ops = 0;
1515 }
1516
1517 static void
1518 isert_put_cmd(struct isert_cmd *isert_cmd, bool comp_err)
1519 {
1520         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1521         struct isert_conn *isert_conn = isert_cmd->conn;
1522         struct iscsi_conn *conn = isert_conn->conn;
1523         struct iscsi_text_rsp *hdr;
1524
1525         isert_dbg("Cmd %p\n", isert_cmd);
1526
1527         switch (cmd->iscsi_opcode) {
1528         case ISCSI_OP_SCSI_CMD:
1529                 spin_lock_bh(&conn->cmd_lock);
1530                 if (!list_empty(&cmd->i_conn_node))
1531                         list_del_init(&cmd->i_conn_node);
1532                 spin_unlock_bh(&conn->cmd_lock);
1533
1534                 if (cmd->data_direction == DMA_TO_DEVICE) {
1535                         iscsit_stop_dataout_timer(cmd);
1536                         /*
1537                          * Check for special case during comp_err where
1538                          * WRITE_PENDING has been handed off from core,
1539                          * but requires an extra target_put_sess_cmd()
1540                          * before transport_generic_free_cmd() below.
1541                          */
1542                         if (comp_err &&
1543                             cmd->se_cmd.t_state == TRANSPORT_WRITE_PENDING) {
1544                                 struct se_cmd *se_cmd = &cmd->se_cmd;
1545
1546                                 target_put_sess_cmd(se_cmd);
1547                         }
1548                 }
1549
1550                 isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1551                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1552                 break;
1553         case ISCSI_OP_SCSI_TMFUNC:
1554                 spin_lock_bh(&conn->cmd_lock);
1555                 if (!list_empty(&cmd->i_conn_node))
1556                         list_del_init(&cmd->i_conn_node);
1557                 spin_unlock_bh(&conn->cmd_lock);
1558
1559                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1560                 break;
1561         case ISCSI_OP_REJECT:
1562         case ISCSI_OP_NOOP_OUT:
1563         case ISCSI_OP_TEXT:
1564                 hdr = (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
1565                 /* If the continue bit is on, keep the command alive */
1566                 if (hdr->flags & ISCSI_FLAG_TEXT_CONTINUE)
1567                         break;
1568
1569                 spin_lock_bh(&conn->cmd_lock);
1570                 if (!list_empty(&cmd->i_conn_node))
1571                         list_del_init(&cmd->i_conn_node);
1572                 spin_unlock_bh(&conn->cmd_lock);
1573
1574                 /*
1575                  * Handle special case for REJECT when iscsi_add_reject*() has
1576                  * overwritten the original iscsi_opcode assignment, and the
1577                  * associated cmd->se_cmd needs to be released.
1578                  */
1579                 if (cmd->se_cmd.se_tfo != NULL) {
1580                         isert_dbg("Calling transport_generic_free_cmd for 0x%02x\n",
1581                                  cmd->iscsi_opcode);
1582                         transport_generic_free_cmd(&cmd->se_cmd, 0);
1583                         break;
1584                 }
1585                 /*
1586                  * Fall-through
1587                  */
1588         default:
1589                 iscsit_release_cmd(cmd);
1590                 break;
1591         }
1592 }
1593
1594 static void
1595 isert_unmap_tx_desc(struct iser_tx_desc *tx_desc, struct ib_device *ib_dev)
1596 {
1597         if (tx_desc->dma_addr != 0) {
1598                 isert_dbg("unmap single for tx_desc->dma_addr\n");
1599                 ib_dma_unmap_single(ib_dev, tx_desc->dma_addr,
1600                                     ISER_HEADERS_LEN, DMA_TO_DEVICE);
1601                 tx_desc->dma_addr = 0;
1602         }
1603 }
1604
1605 static void
1606 isert_completion_put(struct iser_tx_desc *tx_desc, struct isert_cmd *isert_cmd,
1607                      struct ib_device *ib_dev, bool comp_err)
1608 {
1609         if (isert_cmd->pdu_buf_dma != 0) {
1610                 isert_dbg("unmap single for isert_cmd->pdu_buf_dma\n");
1611                 ib_dma_unmap_single(ib_dev, isert_cmd->pdu_buf_dma,
1612                                     isert_cmd->pdu_buf_len, DMA_TO_DEVICE);
1613                 isert_cmd->pdu_buf_dma = 0;
1614         }
1615
1616         isert_unmap_tx_desc(tx_desc, ib_dev);
1617         isert_put_cmd(isert_cmd, comp_err);
1618 }
1619
1620 static int
1621 isert_check_pi_status(struct se_cmd *se_cmd, struct ib_mr *sig_mr)
1622 {
1623         struct ib_mr_status mr_status;
1624         int ret;
1625
1626         ret = ib_check_mr_status(sig_mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
1627         if (ret) {
1628                 isert_err("ib_check_mr_status failed, ret %d\n", ret);
1629                 goto fail_mr_status;
1630         }
1631
1632         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1633                 u64 sec_offset_err;
1634                 u32 block_size = se_cmd->se_dev->dev_attrib.block_size + 8;
1635
1636                 switch (mr_status.sig_err.err_type) {
1637                 case IB_SIG_BAD_GUARD:
1638                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_GUARD_CHECK_FAILED;
1639                         break;
1640                 case IB_SIG_BAD_REFTAG:
1641                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_REF_TAG_CHECK_FAILED;
1642                         break;
1643                 case IB_SIG_BAD_APPTAG:
1644                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_APP_TAG_CHECK_FAILED;
1645                         break;
1646                 }
1647                 sec_offset_err = mr_status.sig_err.sig_err_offset;
1648                 do_div(sec_offset_err, block_size);
1649                 se_cmd->bad_sector = sec_offset_err + se_cmd->t_task_lba;
1650
1651                 isert_err("PI error found type %d at sector 0x%llx "
1652                           "expected 0x%x vs actual 0x%x\n",
1653                           mr_status.sig_err.err_type,
1654                           (unsigned long long)se_cmd->bad_sector,
1655                           mr_status.sig_err.expected,
1656                           mr_status.sig_err.actual);
1657                 ret = 1;
1658         }
1659
1660 fail_mr_status:
1661         return ret;
1662 }
1663
1664 static void
1665 isert_rdma_write_done(struct ib_cq *cq, struct ib_wc *wc)
1666 {
1667         struct isert_conn *isert_conn = wc->qp->qp_context;
1668         struct isert_device *device = isert_conn->device;
1669         struct iser_tx_desc *desc = cqe_to_tx_desc(wc->wr_cqe);
1670         struct isert_cmd *isert_cmd = tx_desc_to_cmd(desc);
1671         struct se_cmd *cmd = &isert_cmd->iscsi_cmd->se_cmd;
1672         int ret = 0;
1673
1674         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1675                 isert_print_wc(wc, "rdma write");
1676                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1677                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1678                 isert_completion_put(desc, isert_cmd, device->ib_device, true);
1679                 return;
1680         }
1681
1682         isert_dbg("Cmd %p\n", isert_cmd);
1683
1684         ret = isert_check_pi_status(cmd, isert_cmd->rw.sig->sig_mr);
1685         isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1686
1687         if (ret) {
1688                 /*
1689                  * transport_generic_request_failure() expects to have
1690                  * plus two references to handle queue-full, so re-add
1691                  * one here as target-core will have already dropped
1692                  * it after the first isert_put_datain() callback.
1693                  */
1694                 kref_get(&cmd->cmd_kref);
1695                 transport_generic_request_failure(cmd, cmd->pi_err);
1696         } else {
1697                 /*
1698                  * XXX: isert_put_response() failure is not retried.
1699                  */
1700                 ret = isert_put_response(isert_conn->conn, isert_cmd->iscsi_cmd);
1701                 if (ret)
1702                         pr_warn_ratelimited("isert_put_response() ret: %d\n", ret);
1703         }
1704 }
1705
1706 static void
1707 isert_rdma_read_done(struct ib_cq *cq, struct ib_wc *wc)
1708 {
1709         struct isert_conn *isert_conn = wc->qp->qp_context;
1710         struct isert_device *device = isert_conn->device;
1711         struct iser_tx_desc *desc = cqe_to_tx_desc(wc->wr_cqe);
1712         struct isert_cmd *isert_cmd = tx_desc_to_cmd(desc);
1713         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1714         struct se_cmd *se_cmd = &cmd->se_cmd;
1715         int ret = 0;
1716
1717         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1718                 isert_print_wc(wc, "rdma read");
1719                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1720                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1721                 isert_completion_put(desc, isert_cmd, device->ib_device, true);
1722                 return;
1723         }
1724
1725         isert_dbg("Cmd %p\n", isert_cmd);
1726
1727         iscsit_stop_dataout_timer(cmd);
1728
1729         if (isert_prot_cmd(isert_conn, se_cmd))
1730                 ret = isert_check_pi_status(se_cmd, isert_cmd->rw.sig->sig_mr);
1731         isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1732         cmd->write_data_done = 0;
1733
1734         isert_dbg("Cmd: %p RDMA_READ comp calling execute_cmd\n", isert_cmd);
1735         spin_lock_bh(&cmd->istate_lock);
1736         cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1737         cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1738         spin_unlock_bh(&cmd->istate_lock);
1739
1740         /*
1741          * transport_generic_request_failure() will drop the extra
1742          * se_cmd->cmd_kref reference after T10-PI error, and handle
1743          * any non-zero ->queue_status() callback error retries.
1744          */
1745         if (ret)
1746                 transport_generic_request_failure(se_cmd, se_cmd->pi_err);
1747         else
1748                 target_execute_cmd(se_cmd);
1749 }
1750
1751 static void
1752 isert_do_control_comp(struct work_struct *work)
1753 {
1754         struct isert_cmd *isert_cmd = container_of(work,
1755                         struct isert_cmd, comp_work);
1756         struct isert_conn *isert_conn = isert_cmd->conn;
1757         struct ib_device *ib_dev = isert_conn->cm_id->device;
1758         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1759
1760         isert_dbg("Cmd %p i_state %d\n", isert_cmd, cmd->i_state);
1761
1762         switch (cmd->i_state) {
1763         case ISTATE_SEND_TASKMGTRSP:
1764                 iscsit_tmr_post_handler(cmd, cmd->conn);
1765         case ISTATE_SEND_REJECT:   /* FALLTHRU */
1766         case ISTATE_SEND_TEXTRSP:  /* FALLTHRU */
1767                 cmd->i_state = ISTATE_SENT_STATUS;
1768                 isert_completion_put(&isert_cmd->tx_desc, isert_cmd,
1769                                      ib_dev, false);
1770                 break;
1771         case ISTATE_SEND_LOGOUTRSP:
1772                 iscsit_logout_post_handler(cmd, cmd->conn);
1773                 break;
1774         default:
1775                 isert_err("Unknown i_state %d\n", cmd->i_state);
1776                 dump_stack();
1777                 break;
1778         }
1779 }
1780
1781 static void
1782 isert_login_send_done(struct ib_cq *cq, struct ib_wc *wc)
1783 {
1784         struct isert_conn *isert_conn = wc->qp->qp_context;
1785         struct ib_device *ib_dev = isert_conn->cm_id->device;
1786         struct iser_tx_desc *tx_desc = cqe_to_tx_desc(wc->wr_cqe);
1787
1788         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1789                 isert_print_wc(wc, "login send");
1790                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1791                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1792         }
1793
1794         isert_unmap_tx_desc(tx_desc, ib_dev);
1795 }
1796
1797 static void
1798 isert_send_done(struct ib_cq *cq, struct ib_wc *wc)
1799 {
1800         struct isert_conn *isert_conn = wc->qp->qp_context;
1801         struct ib_device *ib_dev = isert_conn->cm_id->device;
1802         struct iser_tx_desc *tx_desc = cqe_to_tx_desc(wc->wr_cqe);
1803         struct isert_cmd *isert_cmd = tx_desc_to_cmd(tx_desc);
1804
1805         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1806                 isert_print_wc(wc, "send");
1807                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1808                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1809                 isert_completion_put(tx_desc, isert_cmd, ib_dev, true);
1810                 return;
1811         }
1812
1813         isert_dbg("Cmd %p\n", isert_cmd);
1814
1815         switch (isert_cmd->iscsi_cmd->i_state) {
1816         case ISTATE_SEND_TASKMGTRSP:
1817         case ISTATE_SEND_LOGOUTRSP:
1818         case ISTATE_SEND_REJECT:
1819         case ISTATE_SEND_TEXTRSP:
1820                 isert_unmap_tx_desc(tx_desc, ib_dev);
1821
1822                 INIT_WORK(&isert_cmd->comp_work, isert_do_control_comp);
1823                 queue_work(isert_comp_wq, &isert_cmd->comp_work);
1824                 return;
1825         default:
1826                 isert_cmd->iscsi_cmd->i_state = ISTATE_SENT_STATUS;
1827                 isert_completion_put(tx_desc, isert_cmd, ib_dev, false);
1828                 break;
1829         }
1830 }
1831
1832 static int
1833 isert_post_response(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd)
1834 {
1835         struct ib_send_wr *wr_failed;
1836         int ret;
1837
1838         ret = isert_post_recv(isert_conn, isert_cmd->rx_desc);
1839         if (ret) {
1840                 isert_err("ib_post_recv failed with %d\n", ret);
1841                 return ret;
1842         }
1843
1844         ret = ib_post_send(isert_conn->qp, &isert_cmd->tx_desc.send_wr,
1845                            &wr_failed);
1846         if (ret) {
1847                 isert_err("ib_post_send failed with %d\n", ret);
1848                 return ret;
1849         }
1850         return ret;
1851 }
1852
1853 static int
1854 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
1855 {
1856         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1857         struct isert_conn *isert_conn = conn->context;
1858         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1859         struct iscsi_scsi_rsp *hdr = (struct iscsi_scsi_rsp *)
1860                                 &isert_cmd->tx_desc.iscsi_header;
1861
1862         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1863         iscsit_build_rsp_pdu(cmd, conn, true, hdr);
1864         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1865         /*
1866          * Attach SENSE DATA payload to iSCSI Response PDU
1867          */
1868         if (cmd->se_cmd.sense_buffer &&
1869             ((cmd->se_cmd.se_cmd_flags & SCF_TRANSPORT_TASK_SENSE) ||
1870             (cmd->se_cmd.se_cmd_flags & SCF_EMULATED_TASK_SENSE))) {
1871                 struct isert_device *device = isert_conn->device;
1872                 struct ib_device *ib_dev = device->ib_device;
1873                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
1874                 u32 padding, pdu_len;
1875
1876                 put_unaligned_be16(cmd->se_cmd.scsi_sense_length,
1877                                    cmd->sense_buffer);
1878                 cmd->se_cmd.scsi_sense_length += sizeof(__be16);
1879
1880                 padding = -(cmd->se_cmd.scsi_sense_length) & 3;
1881                 hton24(hdr->dlength, (u32)cmd->se_cmd.scsi_sense_length);
1882                 pdu_len = cmd->se_cmd.scsi_sense_length + padding;
1883
1884                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
1885                                 (void *)cmd->sense_buffer, pdu_len,
1886                                 DMA_TO_DEVICE);
1887                 if (ib_dma_mapping_error(ib_dev, isert_cmd->pdu_buf_dma))
1888                         return -ENOMEM;
1889
1890                 isert_cmd->pdu_buf_len = pdu_len;
1891                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
1892                 tx_dsg->length  = pdu_len;
1893                 tx_dsg->lkey    = device->pd->local_dma_lkey;
1894                 isert_cmd->tx_desc.num_sge = 2;
1895         }
1896
1897         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1898
1899         isert_dbg("Posting SCSI Response\n");
1900
1901         return isert_post_response(isert_conn, isert_cmd);
1902 }
1903
1904 static void
1905 isert_aborted_task(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
1906 {
1907         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1908         struct isert_conn *isert_conn = conn->context;
1909
1910         spin_lock_bh(&conn->cmd_lock);
1911         if (!list_empty(&cmd->i_conn_node))
1912                 list_del_init(&cmd->i_conn_node);
1913         spin_unlock_bh(&conn->cmd_lock);
1914
1915         if (cmd->data_direction == DMA_TO_DEVICE)
1916                 iscsit_stop_dataout_timer(cmd);
1917         isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1918 }
1919
1920 static enum target_prot_op
1921 isert_get_sup_prot_ops(struct iscsi_conn *conn)
1922 {
1923         struct isert_conn *isert_conn = conn->context;
1924         struct isert_device *device = isert_conn->device;
1925
1926         if (conn->tpg->tpg_attrib.t10_pi) {
1927                 if (device->pi_capable) {
1928                         isert_info("conn %p PI offload enabled\n", isert_conn);
1929                         isert_conn->pi_support = true;
1930                         return TARGET_PROT_ALL;
1931                 }
1932         }
1933
1934         isert_info("conn %p PI offload disabled\n", isert_conn);
1935         isert_conn->pi_support = false;
1936
1937         return TARGET_PROT_NORMAL;
1938 }
1939
1940 static int
1941 isert_put_nopin(struct iscsi_cmd *cmd, struct iscsi_conn *conn,
1942                 bool nopout_response)
1943 {
1944         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1945         struct isert_conn *isert_conn = conn->context;
1946         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1947
1948         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1949         iscsit_build_nopin_rsp(cmd, conn, (struct iscsi_nopin *)
1950                                &isert_cmd->tx_desc.iscsi_header,
1951                                nopout_response);
1952         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1953         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1954
1955         isert_dbg("conn %p Posting NOPIN Response\n", isert_conn);
1956
1957         return isert_post_response(isert_conn, isert_cmd);
1958 }
1959
1960 static int
1961 isert_put_logout_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
1962 {
1963         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1964         struct isert_conn *isert_conn = conn->context;
1965         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1966
1967         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1968         iscsit_build_logout_rsp(cmd, conn, (struct iscsi_logout_rsp *)
1969                                 &isert_cmd->tx_desc.iscsi_header);
1970         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1971         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1972
1973         isert_dbg("conn %p Posting Logout Response\n", isert_conn);
1974
1975         return isert_post_response(isert_conn, isert_cmd);
1976 }
1977
1978 static int
1979 isert_put_tm_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
1980 {
1981         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1982         struct isert_conn *isert_conn = conn->context;
1983         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1984
1985         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1986         iscsit_build_task_mgt_rsp(cmd, conn, (struct iscsi_tm_rsp *)
1987                                   &isert_cmd->tx_desc.iscsi_header);
1988         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1989         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1990
1991         isert_dbg("conn %p Posting Task Management Response\n", isert_conn);
1992
1993         return isert_post_response(isert_conn, isert_cmd);
1994 }
1995
1996 static int
1997 isert_put_reject(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
1998 {
1999         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2000         struct isert_conn *isert_conn = conn->context;
2001         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2002         struct isert_device *device = isert_conn->device;
2003         struct ib_device *ib_dev = device->ib_device;
2004         struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2005         struct iscsi_reject *hdr =
2006                 (struct iscsi_reject *)&isert_cmd->tx_desc.iscsi_header;
2007
2008         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2009         iscsit_build_reject(cmd, conn, hdr);
2010         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2011
2012         hton24(hdr->dlength, ISCSI_HDR_LEN);
2013         isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2014                         (void *)cmd->buf_ptr, ISCSI_HDR_LEN,
2015                         DMA_TO_DEVICE);
2016         if (ib_dma_mapping_error(ib_dev, isert_cmd->pdu_buf_dma))
2017                 return -ENOMEM;
2018         isert_cmd->pdu_buf_len = ISCSI_HDR_LEN;
2019         tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2020         tx_dsg->length  = ISCSI_HDR_LEN;
2021         tx_dsg->lkey    = device->pd->local_dma_lkey;
2022         isert_cmd->tx_desc.num_sge = 2;
2023
2024         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2025
2026         isert_dbg("conn %p Posting Reject\n", isert_conn);
2027
2028         return isert_post_response(isert_conn, isert_cmd);
2029 }
2030
2031 static int
2032 isert_put_text_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2033 {
2034         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2035         struct isert_conn *isert_conn = conn->context;
2036         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2037         struct iscsi_text_rsp *hdr =
2038                 (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
2039         u32 txt_rsp_len;
2040         int rc;
2041
2042         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2043         rc = iscsit_build_text_rsp(cmd, conn, hdr, ISCSI_INFINIBAND);
2044         if (rc < 0)
2045                 return rc;
2046
2047         txt_rsp_len = rc;
2048         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2049
2050         if (txt_rsp_len) {
2051                 struct isert_device *device = isert_conn->device;
2052                 struct ib_device *ib_dev = device->ib_device;
2053                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2054                 void *txt_rsp_buf = cmd->buf_ptr;
2055
2056                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2057                                 txt_rsp_buf, txt_rsp_len, DMA_TO_DEVICE);
2058                 if (ib_dma_mapping_error(ib_dev, isert_cmd->pdu_buf_dma))
2059                         return -ENOMEM;
2060
2061                 isert_cmd->pdu_buf_len = txt_rsp_len;
2062                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2063                 tx_dsg->length  = txt_rsp_len;
2064                 tx_dsg->lkey    = device->pd->local_dma_lkey;
2065                 isert_cmd->tx_desc.num_sge = 2;
2066         }
2067         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2068
2069         isert_dbg("conn %p Text Response\n", isert_conn);
2070
2071         return isert_post_response(isert_conn, isert_cmd);
2072 }
2073
2074 static inline void
2075 isert_set_dif_domain(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs,
2076                      struct ib_sig_domain *domain)
2077 {
2078         domain->sig_type = IB_SIG_TYPE_T10_DIF;
2079         domain->sig.dif.bg_type = IB_T10DIF_CRC;
2080         domain->sig.dif.pi_interval = se_cmd->se_dev->dev_attrib.block_size;
2081         domain->sig.dif.ref_tag = se_cmd->reftag_seed;
2082         /*
2083          * At the moment we hard code those, but if in the future
2084          * the target core would like to use it, we will take it
2085          * from se_cmd.
2086          */
2087         domain->sig.dif.apptag_check_mask = 0xffff;
2088         domain->sig.dif.app_escape = true;
2089         domain->sig.dif.ref_escape = true;
2090         if (se_cmd->prot_type == TARGET_DIF_TYPE1_PROT ||
2091             se_cmd->prot_type == TARGET_DIF_TYPE2_PROT)
2092                 domain->sig.dif.ref_remap = true;
2093 };
2094
2095 static int
2096 isert_set_sig_attrs(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs)
2097 {
2098         memset(sig_attrs, 0, sizeof(*sig_attrs));
2099
2100         switch (se_cmd->prot_op) {
2101         case TARGET_PROT_DIN_INSERT:
2102         case TARGET_PROT_DOUT_STRIP:
2103                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
2104                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2105                 break;
2106         case TARGET_PROT_DOUT_INSERT:
2107         case TARGET_PROT_DIN_STRIP:
2108                 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
2109                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2110                 break;
2111         case TARGET_PROT_DIN_PASS:
2112         case TARGET_PROT_DOUT_PASS:
2113                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2114                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2115                 break;
2116         default:
2117                 isert_err("Unsupported PI operation %d\n", se_cmd->prot_op);
2118                 return -EINVAL;
2119         }
2120
2121         sig_attrs->check_mask =
2122                (se_cmd->prot_checks & TARGET_DIF_CHECK_GUARD  ? 0xc0 : 0) |
2123                (se_cmd->prot_checks & TARGET_DIF_CHECK_APPTAG ? 0x30 : 0) |
2124                (se_cmd->prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x0f : 0);
2125         return 0;
2126 }
2127
2128 static int
2129 isert_rdma_rw_ctx_post(struct isert_cmd *cmd, struct isert_conn *conn,
2130                 struct ib_cqe *cqe, struct ib_send_wr *chain_wr)
2131 {
2132         struct se_cmd *se_cmd = &cmd->iscsi_cmd->se_cmd;
2133         enum dma_data_direction dir = target_reverse_dma_direction(se_cmd);
2134         u8 port_num = conn->cm_id->port_num;
2135         u64 addr;
2136         u32 rkey, offset;
2137         int ret;
2138
2139         if (cmd->ctx_init_done)
2140                 goto rdma_ctx_post;
2141
2142         if (dir == DMA_FROM_DEVICE) {
2143                 addr = cmd->write_va;
2144                 rkey = cmd->write_stag;
2145                 offset = cmd->iscsi_cmd->write_data_done;
2146         } else {
2147                 addr = cmd->read_va;
2148                 rkey = cmd->read_stag;
2149                 offset = 0;
2150         }
2151
2152         if (isert_prot_cmd(conn, se_cmd)) {
2153                 struct ib_sig_attrs sig_attrs;
2154
2155                 ret = isert_set_sig_attrs(se_cmd, &sig_attrs);
2156                 if (ret)
2157                         return ret;
2158
2159                 WARN_ON_ONCE(offset);
2160                 ret = rdma_rw_ctx_signature_init(&cmd->rw, conn->qp, port_num,
2161                                 se_cmd->t_data_sg, se_cmd->t_data_nents,
2162                                 se_cmd->t_prot_sg, se_cmd->t_prot_nents,
2163                                 &sig_attrs, addr, rkey, dir);
2164         } else {
2165                 ret = rdma_rw_ctx_init(&cmd->rw, conn->qp, port_num,
2166                                 se_cmd->t_data_sg, se_cmd->t_data_nents,
2167                                 offset, addr, rkey, dir);
2168         }
2169
2170         if (ret < 0) {
2171                 isert_err("Cmd: %p failed to prepare RDMA res\n", cmd);
2172                 return ret;
2173         }
2174
2175         cmd->ctx_init_done = true;
2176
2177 rdma_ctx_post:
2178         ret = rdma_rw_ctx_post(&cmd->rw, conn->qp, port_num, cqe, chain_wr);
2179         if (ret < 0)
2180                 isert_err("Cmd: %p failed to post RDMA res\n", cmd);
2181         return ret;
2182 }
2183
2184 static int
2185 isert_put_datain(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2186 {
2187         struct se_cmd *se_cmd = &cmd->se_cmd;
2188         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2189         struct isert_conn *isert_conn = conn->context;
2190         struct ib_cqe *cqe = NULL;
2191         struct ib_send_wr *chain_wr = NULL;
2192         int rc;
2193
2194         isert_dbg("Cmd: %p RDMA_WRITE data_length: %u\n",
2195                  isert_cmd, se_cmd->data_length);
2196
2197         if (isert_prot_cmd(isert_conn, se_cmd)) {
2198                 isert_cmd->tx_desc.tx_cqe.done = isert_rdma_write_done;
2199                 cqe = &isert_cmd->tx_desc.tx_cqe;
2200         } else {
2201                 /*
2202                  * Build isert_conn->tx_desc for iSCSI response PDU and attach
2203                  */
2204                 isert_create_send_desc(isert_conn, isert_cmd,
2205                                        &isert_cmd->tx_desc);
2206                 iscsit_build_rsp_pdu(cmd, conn, true, (struct iscsi_scsi_rsp *)
2207                                      &isert_cmd->tx_desc.iscsi_header);
2208                 isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2209                 isert_init_send_wr(isert_conn, isert_cmd,
2210                                    &isert_cmd->tx_desc.send_wr);
2211
2212                 rc = isert_post_recv(isert_conn, isert_cmd->rx_desc);
2213                 if (rc) {
2214                         isert_err("ib_post_recv failed with %d\n", rc);
2215                         return rc;
2216                 }
2217
2218                 chain_wr = &isert_cmd->tx_desc.send_wr;
2219         }
2220
2221         rc = isert_rdma_rw_ctx_post(isert_cmd, isert_conn, cqe, chain_wr);
2222         isert_dbg("Cmd: %p posted RDMA_WRITE for iSER Data READ rc: %d\n",
2223                   isert_cmd, rc);
2224         return rc;
2225 }
2226
2227 static int
2228 isert_get_dataout(struct iscsi_conn *conn, struct iscsi_cmd *cmd, bool recovery)
2229 {
2230         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2231         int ret;
2232
2233         isert_dbg("Cmd: %p RDMA_READ data_length: %u write_data_done: %u\n",
2234                  isert_cmd, cmd->se_cmd.data_length, cmd->write_data_done);
2235
2236         isert_cmd->tx_desc.tx_cqe.done = isert_rdma_read_done;
2237         ret = isert_rdma_rw_ctx_post(isert_cmd, conn->context,
2238                                      &isert_cmd->tx_desc.tx_cqe, NULL);
2239
2240         isert_dbg("Cmd: %p posted RDMA_READ memory for ISER Data WRITE rc: %d\n",
2241                  isert_cmd, ret);
2242         return ret;
2243 }
2244
2245 static int
2246 isert_immediate_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2247 {
2248         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2249         int ret = 0;
2250
2251         switch (state) {
2252         case ISTATE_REMOVE:
2253                 spin_lock_bh(&conn->cmd_lock);
2254                 list_del_init(&cmd->i_conn_node);
2255                 spin_unlock_bh(&conn->cmd_lock);
2256                 isert_put_cmd(isert_cmd, true);
2257                 break;
2258         case ISTATE_SEND_NOPIN_WANT_RESPONSE:
2259                 ret = isert_put_nopin(cmd, conn, false);
2260                 break;
2261         default:
2262                 isert_err("Unknown immediate state: 0x%02x\n", state);
2263                 ret = -EINVAL;
2264                 break;
2265         }
2266
2267         return ret;
2268 }
2269
2270 static int
2271 isert_response_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2272 {
2273         struct isert_conn *isert_conn = conn->context;
2274         int ret;
2275
2276         switch (state) {
2277         case ISTATE_SEND_LOGOUTRSP:
2278                 ret = isert_put_logout_rsp(cmd, conn);
2279                 if (!ret)
2280                         isert_conn->logout_posted = true;
2281                 break;
2282         case ISTATE_SEND_NOPIN:
2283                 ret = isert_put_nopin(cmd, conn, true);
2284                 break;
2285         case ISTATE_SEND_TASKMGTRSP:
2286                 ret = isert_put_tm_rsp(cmd, conn);
2287                 break;
2288         case ISTATE_SEND_REJECT:
2289                 ret = isert_put_reject(cmd, conn);
2290                 break;
2291         case ISTATE_SEND_TEXTRSP:
2292                 ret = isert_put_text_rsp(cmd, conn);
2293                 break;
2294         case ISTATE_SEND_STATUS:
2295                 /*
2296                  * Special case for sending non GOOD SCSI status from TX thread
2297                  * context during pre se_cmd excecution failure.
2298                  */
2299                 ret = isert_put_response(conn, cmd);
2300                 break;
2301         default:
2302                 isert_err("Unknown response state: 0x%02x\n", state);
2303                 ret = -EINVAL;
2304                 break;
2305         }
2306
2307         return ret;
2308 }
2309
2310 struct rdma_cm_id *
2311 isert_setup_id(struct isert_np *isert_np)
2312 {
2313         struct iscsi_np *np = isert_np->np;
2314         struct rdma_cm_id *id;
2315         struct sockaddr *sa;
2316         int ret;
2317
2318         sa = (struct sockaddr *)&np->np_sockaddr;
2319         isert_dbg("ksockaddr: %p, sa: %p\n", &np->np_sockaddr, sa);
2320
2321         id = rdma_create_id(&init_net, isert_cma_handler, isert_np,
2322                             RDMA_PS_TCP, IB_QPT_RC);
2323         if (IS_ERR(id)) {
2324                 isert_err("rdma_create_id() failed: %ld\n", PTR_ERR(id));
2325                 ret = PTR_ERR(id);
2326                 goto out;
2327         }
2328         isert_dbg("id %p context %p\n", id, id->context);
2329
2330         ret = rdma_bind_addr(id, sa);
2331         if (ret) {
2332                 isert_err("rdma_bind_addr() failed: %d\n", ret);
2333                 goto out_id;
2334         }
2335
2336         ret = rdma_listen(id, 0);
2337         if (ret) {
2338                 isert_err("rdma_listen() failed: %d\n", ret);
2339                 goto out_id;
2340         }
2341
2342         return id;
2343 out_id:
2344         rdma_destroy_id(id);
2345 out:
2346         return ERR_PTR(ret);
2347 }
2348
2349 static int
2350 isert_setup_np(struct iscsi_np *np,
2351                struct sockaddr_storage *ksockaddr)
2352 {
2353         struct isert_np *isert_np;
2354         struct rdma_cm_id *isert_lid;
2355         int ret;
2356
2357         isert_np = kzalloc(sizeof(struct isert_np), GFP_KERNEL);
2358         if (!isert_np)
2359                 return -ENOMEM;
2360
2361         sema_init(&isert_np->sem, 0);
2362         mutex_init(&isert_np->mutex);
2363         INIT_LIST_HEAD(&isert_np->accepted);
2364         INIT_LIST_HEAD(&isert_np->pending);
2365         isert_np->np = np;
2366
2367         /*
2368          * Setup the np->np_sockaddr from the passed sockaddr setup
2369          * in iscsi_target_configfs.c code..
2370          */
2371         memcpy(&np->np_sockaddr, ksockaddr,
2372                sizeof(struct sockaddr_storage));
2373
2374         isert_lid = isert_setup_id(isert_np);
2375         if (IS_ERR(isert_lid)) {
2376                 ret = PTR_ERR(isert_lid);
2377                 goto out;
2378         }
2379
2380         isert_np->cm_id = isert_lid;
2381         np->np_context = isert_np;
2382
2383         return 0;
2384
2385 out:
2386         kfree(isert_np);
2387
2388         return ret;
2389 }
2390
2391 static int
2392 isert_rdma_accept(struct isert_conn *isert_conn)
2393 {
2394         struct rdma_cm_id *cm_id = isert_conn->cm_id;
2395         struct rdma_conn_param cp;
2396         int ret;
2397         struct iser_cm_hdr rsp_hdr;
2398
2399         memset(&cp, 0, sizeof(struct rdma_conn_param));
2400         cp.initiator_depth = isert_conn->initiator_depth;
2401         cp.retry_count = 7;
2402         cp.rnr_retry_count = 7;
2403
2404         memset(&rsp_hdr, 0, sizeof(rsp_hdr));
2405         rsp_hdr.flags = ISERT_ZBVA_NOT_USED;
2406         if (!isert_conn->snd_w_inv)
2407                 rsp_hdr.flags = rsp_hdr.flags | ISERT_SEND_W_INV_NOT_USED;
2408         cp.private_data = (void *)&rsp_hdr;
2409         cp.private_data_len = sizeof(rsp_hdr);
2410
2411         ret = rdma_accept(cm_id, &cp);
2412         if (ret) {
2413                 isert_err("rdma_accept() failed with: %d\n", ret);
2414                 return ret;
2415         }
2416
2417         return 0;
2418 }
2419
2420 static int
2421 isert_get_login_rx(struct iscsi_conn *conn, struct iscsi_login *login)
2422 {
2423         struct isert_conn *isert_conn = conn->context;
2424         int ret;
2425
2426         isert_info("before login_req comp conn: %p\n", isert_conn);
2427         ret = wait_for_completion_interruptible(&isert_conn->login_req_comp);
2428         if (ret) {
2429                 isert_err("isert_conn %p interrupted before got login req\n",
2430                           isert_conn);
2431                 return ret;
2432         }
2433         reinit_completion(&isert_conn->login_req_comp);
2434
2435         /*
2436          * For login requests after the first PDU, isert_rx_login_req() will
2437          * kick schedule_delayed_work(&conn->login_work) as the packet is
2438          * received, which turns this callback from iscsi_target_do_login_rx()
2439          * into a NOP.
2440          */
2441         if (!login->first_request)
2442                 return 0;
2443
2444         isert_rx_login_req(isert_conn);
2445
2446         isert_info("before login_comp conn: %p\n", conn);
2447         ret = wait_for_completion_interruptible(&isert_conn->login_comp);
2448         if (ret)
2449                 return ret;
2450
2451         isert_info("processing login->req: %p\n", login->req);
2452
2453         return 0;
2454 }
2455
2456 static void
2457 isert_set_conn_info(struct iscsi_np *np, struct iscsi_conn *conn,
2458                     struct isert_conn *isert_conn)
2459 {
2460         struct rdma_cm_id *cm_id = isert_conn->cm_id;
2461         struct rdma_route *cm_route = &cm_id->route;
2462
2463         conn->login_family = np->np_sockaddr.ss_family;
2464
2465         conn->login_sockaddr = cm_route->addr.dst_addr;
2466         conn->local_sockaddr = cm_route->addr.src_addr;
2467 }
2468
2469 static int
2470 isert_accept_np(struct iscsi_np *np, struct iscsi_conn *conn)
2471 {
2472         struct isert_np *isert_np = np->np_context;
2473         struct isert_conn *isert_conn;
2474         int ret;
2475
2476 accept_wait:
2477         ret = down_interruptible(&isert_np->sem);
2478         if (ret)
2479                 return -ENODEV;
2480
2481         spin_lock_bh(&np->np_thread_lock);
2482         if (np->np_thread_state >= ISCSI_NP_THREAD_RESET) {
2483                 spin_unlock_bh(&np->np_thread_lock);
2484                 isert_dbg("np_thread_state %d\n",
2485                          np->np_thread_state);
2486                 /**
2487                  * No point in stalling here when np_thread
2488                  * is in state RESET/SHUTDOWN/EXIT - bail
2489                  **/
2490                 return -ENODEV;
2491         }
2492         spin_unlock_bh(&np->np_thread_lock);
2493
2494         mutex_lock(&isert_np->mutex);
2495         if (list_empty(&isert_np->pending)) {
2496                 mutex_unlock(&isert_np->mutex);
2497                 goto accept_wait;
2498         }
2499         isert_conn = list_first_entry(&isert_np->pending,
2500                         struct isert_conn, node);
2501         list_del_init(&isert_conn->node);
2502         mutex_unlock(&isert_np->mutex);
2503
2504         conn->context = isert_conn;
2505         isert_conn->conn = conn;
2506         isert_conn->state = ISER_CONN_BOUND;
2507
2508         isert_set_conn_info(np, conn, isert_conn);
2509
2510         isert_dbg("Processing isert_conn: %p\n", isert_conn);
2511
2512         return 0;
2513 }
2514
2515 static void
2516 isert_free_np(struct iscsi_np *np)
2517 {
2518         struct isert_np *isert_np = np->np_context;
2519         struct isert_conn *isert_conn, *n;
2520         LIST_HEAD(drop_conn_list);
2521
2522         if (isert_np->cm_id)
2523                 rdma_destroy_id(isert_np->cm_id);
2524
2525         /*
2526          * FIXME: At this point we don't have a good way to insure
2527          * that at this point we don't have hanging connections that
2528          * completed RDMA establishment but didn't start iscsi login
2529          * process. So work-around this by cleaning up what ever piled
2530          * up in accepted and pending lists.
2531          */
2532         mutex_lock(&isert_np->mutex);
2533         if (!list_empty(&isert_np->pending)) {
2534                 isert_info("Still have isert pending connections\n");
2535                 list_for_each_entry_safe(isert_conn, n,
2536                                          &isert_np->pending,
2537                                          node) {
2538                         isert_info("cleaning isert_conn %p state (%d)\n",
2539                                    isert_conn, isert_conn->state);
2540                         list_move_tail(&isert_conn->node, &drop_conn_list);
2541                 }
2542         }
2543
2544         if (!list_empty(&isert_np->accepted)) {
2545                 isert_info("Still have isert accepted connections\n");
2546                 list_for_each_entry_safe(isert_conn, n,
2547                                          &isert_np->accepted,
2548                                          node) {
2549                         isert_info("cleaning isert_conn %p state (%d)\n",
2550                                    isert_conn, isert_conn->state);
2551                         list_move_tail(&isert_conn->node, &drop_conn_list);
2552                 }
2553         }
2554         mutex_unlock(&isert_np->mutex);
2555
2556         list_for_each_entry_safe(isert_conn, n, &drop_conn_list, node) {
2557                 list_del_init(&isert_conn->node);
2558                 isert_connect_release(isert_conn);
2559         }
2560
2561         np->np_context = NULL;
2562         kfree(isert_np);
2563 }
2564
2565 static void isert_release_work(struct work_struct *work)
2566 {
2567         struct isert_conn *isert_conn = container_of(work,
2568                                                      struct isert_conn,
2569                                                      release_work);
2570
2571         isert_info("Starting release conn %p\n", isert_conn);
2572
2573         mutex_lock(&isert_conn->mutex);
2574         isert_conn->state = ISER_CONN_DOWN;
2575         mutex_unlock(&isert_conn->mutex);
2576
2577         isert_info("Destroying conn %p\n", isert_conn);
2578         isert_put_conn(isert_conn);
2579 }
2580
2581 static void
2582 isert_wait4logout(struct isert_conn *isert_conn)
2583 {
2584         struct iscsi_conn *conn = isert_conn->conn;
2585
2586         isert_info("conn %p\n", isert_conn);
2587
2588         if (isert_conn->logout_posted) {
2589                 isert_info("conn %p wait for conn_logout_comp\n", isert_conn);
2590                 wait_for_completion_timeout(&conn->conn_logout_comp,
2591                                             SECONDS_FOR_LOGOUT_COMP * HZ);
2592         }
2593 }
2594
2595 static void
2596 isert_wait4cmds(struct iscsi_conn *conn)
2597 {
2598         isert_info("iscsi_conn %p\n", conn);
2599
2600         if (conn->sess) {
2601                 target_sess_cmd_list_set_waiting(conn->sess->se_sess);
2602                 target_wait_for_sess_cmds(conn->sess->se_sess);
2603         }
2604 }
2605
2606 /**
2607  * isert_put_unsol_pending_cmds() - Drop commands waiting for
2608  *     unsolicitate dataout
2609  * @conn:    iscsi connection
2610  *
2611  * We might still have commands that are waiting for unsolicited
2612  * dataouts messages. We must put the extra reference on those
2613  * before blocking on the target_wait_for_session_cmds
2614  */
2615 static void
2616 isert_put_unsol_pending_cmds(struct iscsi_conn *conn)
2617 {
2618         struct iscsi_cmd *cmd, *tmp;
2619         static LIST_HEAD(drop_cmd_list);
2620
2621         spin_lock_bh(&conn->cmd_lock);
2622         list_for_each_entry_safe(cmd, tmp, &conn->conn_cmd_list, i_conn_node) {
2623                 if ((cmd->cmd_flags & ICF_NON_IMMEDIATE_UNSOLICITED_DATA) &&
2624                     (cmd->write_data_done < conn->sess->sess_ops->FirstBurstLength) &&
2625                     (cmd->write_data_done < cmd->se_cmd.data_length))
2626                         list_move_tail(&cmd->i_conn_node, &drop_cmd_list);
2627         }
2628         spin_unlock_bh(&conn->cmd_lock);
2629
2630         list_for_each_entry_safe(cmd, tmp, &drop_cmd_list, i_conn_node) {
2631                 list_del_init(&cmd->i_conn_node);
2632                 if (cmd->i_state != ISTATE_REMOVE) {
2633                         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2634
2635                         isert_info("conn %p dropping cmd %p\n", conn, cmd);
2636                         isert_put_cmd(isert_cmd, true);
2637                 }
2638         }
2639 }
2640
2641 static void isert_wait_conn(struct iscsi_conn *conn)
2642 {
2643         struct isert_conn *isert_conn = conn->context;
2644
2645         isert_info("Starting conn %p\n", isert_conn);
2646
2647         mutex_lock(&isert_conn->mutex);
2648         isert_conn_terminate(isert_conn);
2649         mutex_unlock(&isert_conn->mutex);
2650
2651         ib_drain_qp(isert_conn->qp);
2652         isert_put_unsol_pending_cmds(conn);
2653         isert_wait4cmds(conn);
2654         isert_wait4logout(isert_conn);
2655
2656         queue_work(isert_release_wq, &isert_conn->release_work);
2657 }
2658
2659 static void isert_free_conn(struct iscsi_conn *conn)
2660 {
2661         struct isert_conn *isert_conn = conn->context;
2662
2663         ib_drain_qp(isert_conn->qp);
2664         isert_put_conn(isert_conn);
2665 }
2666
2667 static void isert_get_rx_pdu(struct iscsi_conn *conn)
2668 {
2669         struct completion comp;
2670
2671         init_completion(&comp);
2672
2673         wait_for_completion_interruptible(&comp);
2674 }
2675
2676 static struct iscsit_transport iser_target_transport = {
2677         .name                   = "IB/iSER",
2678         .transport_type         = ISCSI_INFINIBAND,
2679         .rdma_shutdown          = true,
2680         .priv_size              = sizeof(struct isert_cmd),
2681         .owner                  = THIS_MODULE,
2682         .iscsit_setup_np        = isert_setup_np,
2683         .iscsit_accept_np       = isert_accept_np,
2684         .iscsit_free_np         = isert_free_np,
2685         .iscsit_wait_conn       = isert_wait_conn,
2686         .iscsit_free_conn       = isert_free_conn,
2687         .iscsit_get_login_rx    = isert_get_login_rx,
2688         .iscsit_put_login_tx    = isert_put_login_tx,
2689         .iscsit_immediate_queue = isert_immediate_queue,
2690         .iscsit_response_queue  = isert_response_queue,
2691         .iscsit_get_dataout     = isert_get_dataout,
2692         .iscsit_queue_data_in   = isert_put_datain,
2693         .iscsit_queue_status    = isert_put_response,
2694         .iscsit_aborted_task    = isert_aborted_task,
2695         .iscsit_get_rx_pdu      = isert_get_rx_pdu,
2696         .iscsit_get_sup_prot_ops = isert_get_sup_prot_ops,
2697 };
2698
2699 static int __init isert_init(void)
2700 {
2701         int ret;
2702
2703         isert_comp_wq = alloc_workqueue("isert_comp_wq",
2704                                         WQ_UNBOUND | WQ_HIGHPRI, 0);
2705         if (!isert_comp_wq) {
2706                 isert_err("Unable to allocate isert_comp_wq\n");
2707                 return -ENOMEM;
2708         }
2709
2710         isert_release_wq = alloc_workqueue("isert_release_wq", WQ_UNBOUND,
2711                                         WQ_UNBOUND_MAX_ACTIVE);
2712         if (!isert_release_wq) {
2713                 isert_err("Unable to allocate isert_release_wq\n");
2714                 ret = -ENOMEM;
2715                 goto destroy_comp_wq;
2716         }
2717
2718         iscsit_register_transport(&iser_target_transport);
2719         isert_info("iSER_TARGET[0] - Loaded iser_target_transport\n");
2720
2721         return 0;
2722
2723 destroy_comp_wq:
2724         destroy_workqueue(isert_comp_wq);
2725
2726         return ret;
2727 }
2728
2729 static void __exit isert_exit(void)
2730 {
2731         flush_scheduled_work();
2732         destroy_workqueue(isert_release_wq);
2733         destroy_workqueue(isert_comp_wq);
2734         iscsit_unregister_transport(&iser_target_transport);
2735         isert_info("iSER_TARGET[0] - Released iser_target_transport\n");
2736 }
2737
2738 MODULE_DESCRIPTION("iSER-Target for mainline target infrastructure");
2739 MODULE_AUTHOR("nab@Linux-iSCSI.org");
2740 MODULE_LICENSE("GPL");
2741
2742 module_init(isert_init);
2743 module_exit(isert_exit);