drivers/infiniband/ulp/isert/ib_isert.c

   1 /*******************************************************************************
   2  * This file contains iSCSI extentions for RDMA (iSER) Verbs
   3  *
   4  * (c) Copyright 2013 Datera, Inc.
   5  *
   6  * Nicholas A. Bellinger <nab@linux-iscsi.org>
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License as published by
  10  * the Free Software Foundation; either version 2 of the License, or
  11  * (at your option) any later version.
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  ****************************************************************************/
  18
  19 #include <linux/string.h>
  20 #include <linux/module.h>
  21 #include <linux/scatterlist.h>
  22 #include <linux/socket.h>
  23 #include <linux/in.h>
  24 #include <linux/in6.h>
  25 #include <rdma/ib_verbs.h>
  26 #include <rdma/rdma_cm.h>
  27 #include <target/target_core_base.h>
  28 #include <target/target_core_fabric.h>
  29 #include <target/iscsi/iscsi_transport.h>
  30 #include <linux/semaphore.h>
  31
  32 #include "ib_isert.h"
  33
  34 #define ISERT_MAX_CONN          8
  35 #define ISER_MAX_RX_CQ_LEN      (ISERT_QP_MAX_RECV_DTOS * ISERT_MAX_CONN)
  36 #define ISER_MAX_TX_CQ_LEN \
  37         ((ISERT_QP_MAX_REQ_DTOS + ISCSI_DEF_XMIT_CMDS_MAX) * ISERT_MAX_CONN)
  38 #define ISER_MAX_CQ_LEN         (ISER_MAX_RX_CQ_LEN + ISER_MAX_TX_CQ_LEN + \
  39                                  ISERT_MAX_CONN)
  40
  41 static int isert_debug_level;
  42 module_param_named(debug_level, isert_debug_level, int, 0644);
  43 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0 (default:0)");
  44
  45 static DEFINE_MUTEX(device_list_mutex);
  46 static LIST_HEAD(device_list);
  47 static struct workqueue_struct *isert_comp_wq;
  48 static struct workqueue_struct *isert_release_wq;
  49
  50 static int
  51 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd);
  52 static int
  53 isert_login_post_recv(struct isert_conn *isert_conn);
  54 static int
  55 isert_rdma_accept(struct isert_conn *isert_conn);
  56 struct rdma_cm_id *isert_setup_id(struct isert_np *isert_np);
  57
  58 static void isert_release_work(struct work_struct *work);
  59 static void isert_recv_done(struct ib_cq *cq, struct ib_wc *wc);
  60 static void isert_send_done(struct ib_cq *cq, struct ib_wc *wc);
  61 static void isert_login_recv_done(struct ib_cq *cq, struct ib_wc *wc);
  62 static void isert_login_send_done(struct ib_cq *cq, struct ib_wc *wc);
  63
  64 static inline bool
  65 isert_prot_cmd(struct isert_conn *conn, struct se_cmd *cmd)
  66 {
  67         return (conn->pi_support &&
  68                 cmd->prot_op != TARGET_PROT_NORMAL);
  69 }
  70
  71
  72 static void
  73 isert_qp_event_callback(struct ib_event *e, void *context)
  74 {
  75         struct isert_conn *isert_conn = context;
  76
  77         isert_err("%s (%d): conn %p\n",
  78                   ib_event_msg(e->event), e->event, isert_conn);
  79
  80         switch (e->event) {
  81         case IB_EVENT_COMM_EST:
  82                 rdma_notify(isert_conn->cm_id, IB_EVENT_COMM_EST);
  83                 break;
  84         case IB_EVENT_QP_LAST_WQE_REACHED:
  85                 isert_warn("Reached TX IB_EVENT_QP_LAST_WQE_REACHED\n");
  86                 break;
  87         default:
  88                 break;
  89         }
  90 }
  91
  92 static struct isert_comp *
  93 isert_comp_get(struct isert_conn *isert_conn)
  94 {
  95         struct isert_device *device = isert_conn->device;
  96         struct isert_comp *comp;
  97         int i, min = 0;
  98
  99         mutex_lock(&device_list_mutex);
 100         for (i = 0; i < device->comps_used; i++)
 101                 if (device->comps[i].active_qps <
 102                     device->comps[min].active_qps)
 103                         min = i;
 104         comp = &device->comps[min];
 105         comp->active_qps++;
 106         mutex_unlock(&device_list_mutex);
 107
 108         isert_info("conn %p, using comp %p min_index: %d\n",
 109                    isert_conn, comp, min);
 110
 111         return comp;
 112 }
 113
 114 static void
 115 isert_comp_put(struct isert_comp *comp)
 116 {
 117         mutex_lock(&device_list_mutex);
 118         comp->active_qps--;
 119         mutex_unlock(&device_list_mutex);
 120 }
 121
 122 static struct ib_qp *
 123 isert_create_qp(struct isert_conn *isert_conn,
 124                 struct isert_comp *comp,
 125                 struct rdma_cm_id *cma_id)
 126 {
 127         struct isert_device *device = isert_conn->device;
 128         struct ib_qp_init_attr attr;
 129         int ret;
 130
 131         memset(&attr, 0, sizeof(struct ib_qp_init_attr));
 132         attr.event_handler = isert_qp_event_callback;
 133         attr.qp_context = isert_conn;
 134         attr.send_cq = comp->cq;
 135         attr.recv_cq = comp->cq;
 136         attr.cap.max_send_wr = ISERT_QP_MAX_REQ_DTOS + 1;
 137         attr.cap.max_recv_wr = ISERT_QP_MAX_RECV_DTOS + 1;
 138         attr.cap.max_rdma_ctxs = ISCSI_DEF_XMIT_CMDS_MAX;
 139         attr.cap.max_send_sge = device->ib_device->attrs.max_sge;
 140         attr.cap.max_recv_sge = 1;
 141         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 142         attr.qp_type = IB_QPT_RC;
 143         if (device->pi_capable)
 144                 attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
 145
 146         ret = rdma_create_qp(cma_id, device->pd, &attr);
 147         if (ret) {
 148                 isert_err("rdma_create_qp failed for cma_id %d\n", ret);
 149                 return ERR_PTR(ret);
 150         }
 151
 152         return cma_id->qp;
 153 }
 154
 155 static int
 156 isert_conn_setup_qp(struct isert_conn *isert_conn, struct rdma_cm_id *cma_id)
 157 {
 158         struct isert_comp *comp;
 159         int ret;
 160
 161         comp = isert_comp_get(isert_conn);
 162         isert_conn->qp = isert_create_qp(isert_conn, comp, cma_id);
 163         if (IS_ERR(isert_conn->qp)) {
 164                 ret = PTR_ERR(isert_conn->qp);
 165                 goto err;
 166         }
 167
 168         return 0;
 169 err:
 170         isert_comp_put(comp);
 171         return ret;
 172 }
 173
 174 static int
 175 isert_alloc_rx_descriptors(struct isert_conn *isert_conn)
 176 {
 177         struct isert_device *device = isert_conn->device;
 178         struct ib_device *ib_dev = device->ib_device;
 179         struct iser_rx_desc *rx_desc;
 180         struct ib_sge *rx_sg;
 181         u64 dma_addr;
 182         int i, j;
 183
 184         isert_conn->rx_descs = kzalloc(ISERT_QP_MAX_RECV_DTOS *
 185                                 sizeof(struct iser_rx_desc), GFP_KERNEL);
 186         if (!isert_conn->rx_descs)
 187                 return -ENOMEM;
 188
 189         rx_desc = isert_conn->rx_descs;
 190
 191         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 192                 dma_addr = ib_dma_map_single(ib_dev, (void *)rx_desc,
 193                                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 194                 if (ib_dma_mapping_error(ib_dev, dma_addr))
 195                         goto dma_map_fail;
 196
 197                 rx_desc->dma_addr = dma_addr;
 198
 199                 rx_sg = &rx_desc->rx_sg;
 200                 rx_sg->addr = rx_desc->dma_addr;
 201                 rx_sg->length = ISER_RX_PAYLOAD_SIZE;
 202                 rx_sg->lkey = device->pd->local_dma_lkey;
 203                 rx_desc->rx_cqe.done = isert_recv_done;
 204         }
 205
 206         return 0;
 207
 208 dma_map_fail:
 209         rx_desc = isert_conn->rx_descs;
 210         for (j = 0; j < i; j++, rx_desc++) {
 211                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 212                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 213         }
 214         kfree(isert_conn->rx_descs);
 215         isert_conn->rx_descs = NULL;
 216         isert_err("conn %p failed to allocate rx descriptors\n", isert_conn);
 217         return -ENOMEM;
 218 }
 219
 220 static void
 221 isert_free_rx_descriptors(struct isert_conn *isert_conn)
 222 {
 223         struct ib_device *ib_dev = isert_conn->device->ib_device;
 224         struct iser_rx_desc *rx_desc;
 225         int i;
 226
 227         if (!isert_conn->rx_descs)
 228                 return;
 229
 230         rx_desc = isert_conn->rx_descs;
 231         for (i = 0; i < ISERT_QP_MAX_RECV_DTOS; i++, rx_desc++)  {
 232                 ib_dma_unmap_single(ib_dev, rx_desc->dma_addr,
 233                                     ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 234         }
 235
 236         kfree(isert_conn->rx_descs);
 237         isert_conn->rx_descs = NULL;
 238 }
 239
 240 static void
 241 isert_free_comps(struct isert_device *device)
 242 {
 243         int i;
 244
 245         for (i = 0; i < device->comps_used; i++) {
 246                 struct isert_comp *comp = &device->comps[i];
 247
 248                 if (comp->cq)
 249                         ib_free_cq(comp->cq);
 250         }
 251         kfree(device->comps);
 252 }
 253
 254 static int
 255 isert_alloc_comps(struct isert_device *device)
 256 {
 257         int i, max_cqe, ret = 0;
 258
 259         device->comps_used = min(ISERT_MAX_CQ, min_t(int, num_online_cpus(),
 260                                  device->ib_device->num_comp_vectors));
 261
 262         isert_info("Using %d CQs, %s supports %d vectors support "
 263                    "pi_capable %d\n",
 264                    device->comps_used, device->ib_device->name,
 265                    device->ib_device->num_comp_vectors,
 266                    device->pi_capable);
 267
 268         device->comps = kcalloc(device->comps_used, sizeof(struct isert_comp),
 269                                 GFP_KERNEL);
 270         if (!device->comps)
 271                 return -ENOMEM;
 272
 273         max_cqe = min(ISER_MAX_CQ_LEN, device->ib_device->attrs.max_cqe);
 274
 275         for (i = 0; i < device->comps_used; i++) {
 276                 struct isert_comp *comp = &device->comps[i];
 277
 278                 comp->device = device;
 279                 comp->cq = ib_alloc_cq(device->ib_device, comp, max_cqe, i,
 280                                 IB_POLL_WORKQUEUE);
 281                 if (IS_ERR(comp->cq)) {
 282                         isert_err("Unable to allocate cq\n");
 283                         ret = PTR_ERR(comp->cq);
 284                         comp->cq = NULL;
 285                         goto out_cq;
 286                 }
 287         }
 288
 289         return 0;
 290 out_cq:
 291         isert_free_comps(device);
 292         return ret;
 293 }
 294
 295 static int
 296 isert_create_device_ib_res(struct isert_device *device)
 297 {
 298         struct ib_device *ib_dev = device->ib_device;
 299         int ret;
 300
 301         isert_dbg("devattr->max_sge: %d\n", ib_dev->attrs.max_sge);
 302         isert_dbg("devattr->max_sge_rd: %d\n", ib_dev->attrs.max_sge_rd);
 303
 304         ret = isert_alloc_comps(device);
 305         if (ret)
 306                 goto out;
 307
 308         device->pd = ib_alloc_pd(ib_dev, 0);
 309         if (IS_ERR(device->pd)) {
 310                 ret = PTR_ERR(device->pd);
 311                 isert_err("failed to allocate pd, device %p, ret=%d\n",
 312                           device, ret);
 313                 goto out_cq;
 314         }
 315
 316         /* Check signature cap */
 317         device->pi_capable = ib_dev->attrs.device_cap_flags &
 318                              IB_DEVICE_SIGNATURE_HANDOVER ? true : false;
 319
 320         return 0;
 321
 322 out_cq:
 323         isert_free_comps(device);
 324 out:
 325         if (ret > 0)
 326                 ret = -EINVAL;
 327         return ret;
 328 }
 329
 330 static void
 331 isert_free_device_ib_res(struct isert_device *device)
 332 {
 333         isert_info("device %p\n", device);
 334
 335         ib_dealloc_pd(device->pd);
 336         isert_free_comps(device);
 337 }
 338
 339 static void
 340 isert_device_put(struct isert_device *device)
 341 {
 342         mutex_lock(&device_list_mutex);
 343         device->refcount--;
 344         isert_info("device %p refcount %d\n", device, device->refcount);
 345         if (!device->refcount) {
 346                 isert_free_device_ib_res(device);
 347                 list_del(&device->dev_node);
 348                 kfree(device);
 349         }
 350         mutex_unlock(&device_list_mutex);
 351 }
 352
 353 static struct isert_device *
 354 isert_device_get(struct rdma_cm_id *cma_id)
 355 {
 356         struct isert_device *device;
 357         int ret;
 358
 359         mutex_lock(&device_list_mutex);
 360         list_for_each_entry(device, &device_list, dev_node) {
 361                 if (device->ib_device->node_guid == cma_id->device->node_guid) {
 362                         device->refcount++;
 363                         isert_info("Found iser device %p refcount %d\n",
 364                                    device, device->refcount);
 365                         mutex_unlock(&device_list_mutex);
 366                         return device;
 367                 }
 368         }
 369
 370         device = kzalloc(sizeof(struct isert_device), GFP_KERNEL);
 371         if (!device) {
 372                 mutex_unlock(&device_list_mutex);
 373                 return ERR_PTR(-ENOMEM);
 374         }
 375
 376         INIT_LIST_HEAD(&device->dev_node);
 377
 378         device->ib_device = cma_id->device;
 379         ret = isert_create_device_ib_res(device);
 380         if (ret) {
 381                 kfree(device);
 382                 mutex_unlock(&device_list_mutex);
 383                 return ERR_PTR(ret);
 384         }
 385
 386         device->refcount++;
 387         list_add_tail(&device->dev_node, &device_list);
 388         isert_info("Created a new iser device %p refcount %d\n",
 389                    device, device->refcount);
 390         mutex_unlock(&device_list_mutex);
 391
 392         return device;
 393 }
 394
 395 static void
 396 isert_init_conn(struct isert_conn *isert_conn)
 397 {
 398         isert_conn->state = ISER_CONN_INIT;
 399         INIT_LIST_HEAD(&isert_conn->node);
 400         init_completion(&isert_conn->login_comp);
 401         init_completion(&isert_conn->login_req_comp);
 402         init_waitqueue_head(&isert_conn->rem_wait);
 403         kref_init(&isert_conn->kref);
 404         mutex_init(&isert_conn->mutex);
 405         INIT_WORK(&isert_conn->release_work, isert_release_work);
 406 }
 407
 408 static void
 409 isert_free_login_buf(struct isert_conn *isert_conn)
 410 {
 411         struct ib_device *ib_dev = isert_conn->device->ib_device;
 412
 413         ib_dma_unmap_single(ib_dev, isert_conn->login_rsp_dma,
 414                             ISER_RX_PAYLOAD_SIZE, DMA_TO_DEVICE);
 415         kfree(isert_conn->login_rsp_buf);
 416
 417         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 418                             ISER_RX_PAYLOAD_SIZE,
 419                             DMA_FROM_DEVICE);
 420         kfree(isert_conn->login_req_buf);
 421 }
 422
 423 static int
 424 isert_alloc_login_buf(struct isert_conn *isert_conn,
 425                       struct ib_device *ib_dev)
 426 {
 427         int ret;
 428
 429         isert_conn->login_req_buf = kzalloc(sizeof(*isert_conn->login_req_buf),
 430                         GFP_KERNEL);
 431         if (!isert_conn->login_req_buf)
 432                 return -ENOMEM;
 433
 434         isert_conn->login_req_dma = ib_dma_map_single(ib_dev,
 435                                 isert_conn->login_req_buf,
 436                                 ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 437         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_req_dma);
 438         if (ret) {
 439                 isert_err("login_req_dma mapping error: %d\n", ret);
 440                 isert_conn->login_req_dma = 0;
 441                 goto out_free_login_req_buf;
 442         }
 443
 444         isert_conn->login_rsp_buf = kzalloc(ISER_RX_PAYLOAD_SIZE, GFP_KERNEL);
 445         if (!isert_conn->login_rsp_buf) {
 446                 ret = -ENOMEM;
 447                 goto out_unmap_login_req_buf;
 448         }
 449
 450         isert_conn->login_rsp_dma = ib_dma_map_single(ib_dev,
 451                                         isert_conn->login_rsp_buf,
 452                                         ISER_RX_PAYLOAD_SIZE, DMA_TO_DEVICE);
 453         ret = ib_dma_mapping_error(ib_dev, isert_conn->login_rsp_dma);
 454         if (ret) {
 455                 isert_err("login_rsp_dma mapping error: %d\n", ret);
 456                 isert_conn->login_rsp_dma = 0;
 457                 goto out_free_login_rsp_buf;
 458         }
 459
 460         return 0;
 461
 462 out_free_login_rsp_buf:
 463         kfree(isert_conn->login_rsp_buf);
 464 out_unmap_login_req_buf:
 465         ib_dma_unmap_single(ib_dev, isert_conn->login_req_dma,
 466                             ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
 467 out_free_login_req_buf:
 468         kfree(isert_conn->login_req_buf);
 469         return ret;
 470 }
 471
 472 static void
 473 isert_set_nego_params(struct isert_conn *isert_conn,
 474                       struct rdma_conn_param *param)
 475 {
 476         struct ib_device_attr *attr = &isert_conn->device->ib_device->attrs;
 477
 478         /* Set max inflight RDMA READ requests */
 479         isert_conn->initiator_depth = min_t(u8, param->initiator_depth,
 480                                 attr->max_qp_init_rd_atom);
 481         isert_dbg("Using initiator_depth: %u\n", isert_conn->initiator_depth);
 482
 483         if (param->private_data) {
 484                 u8 flags = *(u8 *)param->private_data;
 485
 486                 /*
 487                  * use remote invalidation if the both initiator
 488                  * and the HCA support it
 489                  */
 490                 isert_conn->snd_w_inv = !(flags & ISER_SEND_W_INV_NOT_SUP) &&
 491                                           (attr->device_cap_flags &
 492                                            IB_DEVICE_MEM_MGT_EXTENSIONS);
 493                 if (isert_conn->snd_w_inv)
 494                         isert_info("Using remote invalidation\n");
 495         }
 496 }
 497
 498 static int
 499 isert_connect_request(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 500 {
 501         struct isert_np *isert_np = cma_id->context;
 502         struct iscsi_np *np = isert_np->np;
 503         struct isert_conn *isert_conn;
 504         struct isert_device *device;
 505         int ret = 0;
 506
 507         spin_lock_bh(&np->np_thread_lock);
 508         if (!np->enabled) {
 509                 spin_unlock_bh(&np->np_thread_lock);
 510                 isert_dbg("iscsi_np is not enabled, reject connect request\n");
 511                 return rdma_reject(cma_id, NULL, 0);
 512         }
 513         spin_unlock_bh(&np->np_thread_lock);
 514
 515         isert_dbg("cma_id: %p, portal: %p\n",
 516                  cma_id, cma_id->context);
 517
 518         isert_conn = kzalloc(sizeof(struct isert_conn), GFP_KERNEL);
 519         if (!isert_conn)
 520                 return -ENOMEM;
 521
 522         isert_init_conn(isert_conn);
 523         isert_conn->cm_id = cma_id;
 524
 525         ret = isert_alloc_login_buf(isert_conn, cma_id->device);
 526         if (ret)
 527                 goto out;
 528
 529         device = isert_device_get(cma_id);
 530         if (IS_ERR(device)) {
 531                 ret = PTR_ERR(device);
 532                 goto out_rsp_dma_map;
 533         }
 534         isert_conn->device = device;
 535
 536         isert_set_nego_params(isert_conn, &event->param.conn);
 537
 538         ret = isert_conn_setup_qp(isert_conn, cma_id);
 539         if (ret)
 540                 goto out_conn_dev;
 541
 542         ret = isert_login_post_recv(isert_conn);
 543         if (ret)
 544                 goto out_conn_dev;
 545
 546         ret = isert_rdma_accept(isert_conn);
 547         if (ret)
 548                 goto out_conn_dev;
 549
 550         mutex_lock(&isert_np->mutex);
 551         list_add_tail(&isert_conn->node, &isert_np->accepted);
 552         mutex_unlock(&isert_np->mutex);
 553
 554         return 0;
 555
 556 out_conn_dev:
 557         isert_device_put(device);
 558 out_rsp_dma_map:
 559         isert_free_login_buf(isert_conn);
 560 out:
 561         kfree(isert_conn);
 562         rdma_reject(cma_id, NULL, 0);
 563         return ret;
 564 }
 565
 566 static void
 567 isert_connect_release(struct isert_conn *isert_conn)
 568 {
 569         struct isert_device *device = isert_conn->device;
 570
 571         isert_dbg("conn %p\n", isert_conn);
 572
 573         BUG_ON(!device);
 574
 575         isert_free_rx_descriptors(isert_conn);
 576         if (isert_conn->cm_id &&
 577             !isert_conn->dev_removed)
 578                 rdma_destroy_id(isert_conn->cm_id);
 579
 580         if (isert_conn->qp) {
 581                 struct isert_comp *comp = isert_conn->qp->recv_cq->cq_context;
 582
 583                 isert_comp_put(comp);
 584                 ib_destroy_qp(isert_conn->qp);
 585         }
 586
 587         if (isert_conn->login_req_buf)
 588                 isert_free_login_buf(isert_conn);
 589
 590         isert_device_put(device);
 591
 592         if (isert_conn->dev_removed)
 593                 wake_up_interruptible(&isert_conn->rem_wait);
 594         else
 595                 kfree(isert_conn);
 596 }
 597
 598 static void
 599 isert_connected_handler(struct rdma_cm_id *cma_id)
 600 {
 601         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 602         struct isert_np *isert_np = cma_id->context;
 603
 604         isert_info("conn %p\n", isert_conn);
 605
 606         mutex_lock(&isert_conn->mutex);
 607         isert_conn->state = ISER_CONN_UP;
 608         kref_get(&isert_conn->kref);
 609         mutex_unlock(&isert_conn->mutex);
 610
 611         mutex_lock(&isert_np->mutex);
 612         list_move_tail(&isert_conn->node, &isert_np->pending);
 613         mutex_unlock(&isert_np->mutex);
 614
 615         isert_info("np %p: Allow accept_np to continue\n", isert_np);
 616         up(&isert_np->sem);
 617 }
 618
 619 static void
 620 isert_release_kref(struct kref *kref)
 621 {
 622         struct isert_conn *isert_conn = container_of(kref,
 623                                 struct isert_conn, kref);
 624
 625         isert_info("conn %p final kref %s/%d\n", isert_conn, current->comm,
 626                    current->pid);
 627
 628         isert_connect_release(isert_conn);
 629 }
 630
 631 static void
 632 isert_put_conn(struct isert_conn *isert_conn)
 633 {
 634         kref_put(&isert_conn->kref, isert_release_kref);
 635 }
 636
 637 static void
 638 isert_handle_unbound_conn(struct isert_conn *isert_conn)
 639 {
 640         struct isert_np *isert_np = isert_conn->cm_id->context;
 641
 642         mutex_lock(&isert_np->mutex);
 643         if (!list_empty(&isert_conn->node)) {
 644                 /*
 645                  * This means iscsi doesn't know this connection
 646                  * so schedule a cleanup ourselves
 647                  */
 648                 list_del_init(&isert_conn->node);
 649                 isert_put_conn(isert_conn);
 650                 queue_work(isert_release_wq, &isert_conn->release_work);
 651         }
 652         mutex_unlock(&isert_np->mutex);
 653 }
 654
 655 /**
 656  * isert_conn_terminate() - Initiate connection termination
 657  * @isert_conn: isert connection struct
 658  *
 659  * Notes:
 660  * In case the connection state is BOUND, move state
 661  * to TEMINATING and start teardown sequence (rdma_disconnect).
 662  * In case the connection state is UP, complete flush as well.
 663  *
 664  * This routine must be called with mutex held. Thus it is
 665  * safe to call multiple times.
 666  */
 667 static void
 668 isert_conn_terminate(struct isert_conn *isert_conn)
 669 {
 670         int err;
 671
 672         if (isert_conn->state >= ISER_CONN_TERMINATING)
 673                 return;
 674
 675         isert_info("Terminating conn %p state %d\n",
 676                    isert_conn, isert_conn->state);
 677         isert_conn->state = ISER_CONN_TERMINATING;
 678         err = rdma_disconnect(isert_conn->cm_id);
 679         if (err)
 680                 isert_warn("Failed rdma_disconnect isert_conn %p\n",
 681                            isert_conn);
 682 }
 683
 684 static int
 685 isert_np_cma_handler(struct isert_np *isert_np,
 686                      enum rdma_cm_event_type event)
 687 {
 688         isert_dbg("%s (%d): isert np %p\n",
 689                   rdma_event_msg(event), event, isert_np);
 690
 691         switch (event) {
 692         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 693                 isert_np->cm_id = NULL;
 694                 break;
 695         case RDMA_CM_EVENT_ADDR_CHANGE:
 696                 isert_np->cm_id = isert_setup_id(isert_np);
 697                 if (IS_ERR(isert_np->cm_id)) {
 698                         isert_err("isert np %p setup id failed: %ld\n",
 699                                   isert_np, PTR_ERR(isert_np->cm_id));
 700                         isert_np->cm_id = NULL;
 701                 }
 702                 break;
 703         default:
 704                 isert_err("isert np %p Unexpected event %d\n",
 705                           isert_np, event);
 706         }
 707
 708         return -1;
 709 }
 710
 711 static int
 712 isert_disconnected_handler(struct rdma_cm_id *cma_id,
 713                            enum rdma_cm_event_type event)
 714 {
 715         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 716
 717         mutex_lock(&isert_conn->mutex);
 718         switch (isert_conn->state) {
 719         case ISER_CONN_TERMINATING:
 720                 break;
 721         case ISER_CONN_UP:
 722                 isert_conn_terminate(isert_conn);
 723                 ib_drain_qp(isert_conn->qp);
 724                 isert_handle_unbound_conn(isert_conn);
 725                 break;
 726         case ISER_CONN_BOUND:
 727         case ISER_CONN_FULL_FEATURE: /* FALLTHRU */
 728                 iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
 729                 break;
 730         default:
 731                 isert_warn("conn %p terminating in state %d\n",
 732                            isert_conn, isert_conn->state);
 733         }
 734         mutex_unlock(&isert_conn->mutex);
 735
 736         return 0;
 737 }
 738
 739 static int
 740 isert_connect_error(struct rdma_cm_id *cma_id)
 741 {
 742         struct isert_conn *isert_conn = cma_id->qp->qp_context;
 743
 744         ib_drain_qp(isert_conn->qp);
 745         list_del_init(&isert_conn->node);
 746         isert_conn->cm_id = NULL;
 747         isert_put_conn(isert_conn);
 748
 749         return -1;
 750 }
 751
 752 static int
 753 isert_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 754 {
 755         struct isert_np *isert_np = cma_id->context;
 756         struct isert_conn *isert_conn;
 757         int ret = 0;
 758
 759         isert_info("%s (%d): status %d id %p np %p\n",
 760                    rdma_event_msg(event->event), event->event,
 761                    event->status, cma_id, cma_id->context);
 762
 763         if (isert_np->cm_id == cma_id)
 764                 return isert_np_cma_handler(cma_id->context, event->event);
 765
 766         switch (event->event) {
 767         case RDMA_CM_EVENT_CONNECT_REQUEST:
 768                 ret = isert_connect_request(cma_id, event);
 769                 if (ret)
 770                         isert_err("failed handle connect request %d\n", ret);
 771                 break;
 772         case RDMA_CM_EVENT_ESTABLISHED:
 773                 isert_connected_handler(cma_id);
 774                 break;
 775         case RDMA_CM_EVENT_ADDR_CHANGE:    /* FALLTHRU */
 776         case RDMA_CM_EVENT_DISCONNECTED:   /* FALLTHRU */
 777         case RDMA_CM_EVENT_TIMEWAIT_EXIT:  /* FALLTHRU */
 778                 ret = isert_disconnected_handler(cma_id, event->event);
 779                 break;
 780         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 781                 isert_conn = cma_id->qp->qp_context;
 782                 isert_conn->dev_removed = true;
 783                 isert_disconnected_handler(cma_id, event->event);
 784                 wait_event_interruptible(isert_conn->rem_wait,
 785                                          isert_conn->state == ISER_CONN_DOWN);
 786                 kfree(isert_conn);
 787                 /*
 788                  * return non-zero from the callback to destroy
 789                  * the rdma cm id
 790                  */
 791                 return 1;
 792         case RDMA_CM_EVENT_REJECTED:       /* FALLTHRU */
 793                 isert_info("Connection rejected: %s\n",
 794                            rdma_reject_msg(cma_id, event->status));
 795         case RDMA_CM_EVENT_UNREACHABLE:    /* FALLTHRU */
 796         case RDMA_CM_EVENT_CONNECT_ERROR:
 797                 ret = isert_connect_error(cma_id);
 798                 break;
 799         default:
 800                 isert_err("Unhandled RDMA CMA event: %d\n", event->event);
 801                 break;
 802         }
 803
 804         return ret;
 805 }
 806
 807 static int
 808 isert_post_recvm(struct isert_conn *isert_conn, u32 count)
 809 {
 810         struct ib_recv_wr *rx_wr, *rx_wr_failed;
 811         int i, ret;
 812         struct iser_rx_desc *rx_desc;
 813
 814         for (rx_wr = isert_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
 815                 rx_desc = &isert_conn->rx_descs[i];
 816
 817                 rx_wr->wr_cqe = &rx_desc->rx_cqe;
 818                 rx_wr->sg_list = &rx_desc->rx_sg;
 819                 rx_wr->num_sge = 1;
 820                 rx_wr->next = rx_wr + 1;
 821                 rx_desc->in_use = false;
 822         }
 823         rx_wr--;
 824         rx_wr->next = NULL; /* mark end of work requests list */
 825
 826         ret = ib_post_recv(isert_conn->qp, isert_conn->rx_wr,
 827                            &rx_wr_failed);
 828         if (ret)
 829                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
 830
 831         return ret;
 832 }
 833
 834 static int
 835 isert_post_recv(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc)
 836 {
 837         struct ib_recv_wr *rx_wr_failed, rx_wr;
 838         int ret;
 839
 840         if (!rx_desc->in_use) {
 841                 /*
 842                  * if the descriptor is not in-use we already reposted it
 843                  * for recv, so just silently return
 844                  */
 845                 return 0;
 846         }
 847
 848         rx_desc->in_use = false;
 849         rx_wr.wr_cqe = &rx_desc->rx_cqe;
 850         rx_wr.sg_list = &rx_desc->rx_sg;
 851         rx_wr.num_sge = 1;
 852         rx_wr.next = NULL;
 853
 854         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_failed);
 855         if (ret)
 856                 isert_err("ib_post_recv() failed with ret: %d\n", ret);
 857
 858         return ret;
 859 }
 860
 861 static int
 862 isert_login_post_send(struct isert_conn *isert_conn, struct iser_tx_desc *tx_desc)
 863 {
 864         struct ib_device *ib_dev = isert_conn->cm_id->device;
 865         struct ib_send_wr send_wr, *send_wr_failed;
 866         int ret;
 867
 868         ib_dma_sync_single_for_device(ib_dev, tx_desc->dma_addr,
 869                                       ISER_HEADERS_LEN, DMA_TO_DEVICE);
 870
 871         tx_desc->tx_cqe.done = isert_login_send_done;
 872
 873         send_wr.next    = NULL;
 874         send_wr.wr_cqe  = &tx_desc->tx_cqe;
 875         send_wr.sg_list = tx_desc->tx_sg;
 876         send_wr.num_sge = tx_desc->num_sge;
 877         send_wr.opcode  = IB_WR_SEND;
 878         send_wr.send_flags = IB_SEND_SIGNALED;
 879
 880         ret = ib_post_send(isert_conn->qp, &send_wr, &send_wr_failed);
 881         if (ret)
 882                 isert_err("ib_post_send() failed, ret: %d\n", ret);
 883
 884         return ret;
 885 }
 886
 887 static void
 888 __isert_create_send_desc(struct isert_device *device,
 889                          struct iser_tx_desc *tx_desc)
 890 {
 891
 892         memset(&tx_desc->iser_header, 0, sizeof(struct iser_ctrl));
 893         tx_desc->iser_header.flags = ISCSI_CTRL;
 894
 895         tx_desc->num_sge = 1;
 896
 897         if (tx_desc->tx_sg[0].lkey != device->pd->local_dma_lkey) {
 898                 tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
 899                 isert_dbg("tx_desc %p lkey mismatch, fixing\n", tx_desc);
 900         }
 901 }
 902
 903 static void
 904 isert_create_send_desc(struct isert_conn *isert_conn,
 905                        struct isert_cmd *isert_cmd,
 906                        struct iser_tx_desc *tx_desc)
 907 {
 908         struct isert_device *device = isert_conn->device;
 909         struct ib_device *ib_dev = device->ib_device;
 910
 911         ib_dma_sync_single_for_cpu(ib_dev, tx_desc->dma_addr,
 912                                    ISER_HEADERS_LEN, DMA_TO_DEVICE);
 913
 914         __isert_create_send_desc(device, tx_desc);
 915 }
 916
 917 static int
 918 isert_init_tx_hdrs(struct isert_conn *isert_conn,
 919                    struct iser_tx_desc *tx_desc)
 920 {
 921         struct isert_device *device = isert_conn->device;
 922         struct ib_device *ib_dev = device->ib_device;
 923         u64 dma_addr;
 924
 925         dma_addr = ib_dma_map_single(ib_dev, (void *)tx_desc,
 926                         ISER_HEADERS_LEN, DMA_TO_DEVICE);
 927         if (ib_dma_mapping_error(ib_dev, dma_addr)) {
 928                 isert_err("ib_dma_mapping_error() failed\n");
 929                 return -ENOMEM;
 930         }
 931
 932         tx_desc->dma_addr = dma_addr;
 933         tx_desc->tx_sg[0].addr  = tx_desc->dma_addr;
 934         tx_desc->tx_sg[0].length = ISER_HEADERS_LEN;
 935         tx_desc->tx_sg[0].lkey = device->pd->local_dma_lkey;
 936
 937         isert_dbg("Setup tx_sg[0].addr: 0x%llx length: %u lkey: 0x%x\n",
 938                   tx_desc->tx_sg[0].addr, tx_desc->tx_sg[0].length,
 939                   tx_desc->tx_sg[0].lkey);
 940
 941         return 0;
 942 }
 943
 944 static void
 945 isert_init_send_wr(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
 946                    struct ib_send_wr *send_wr)
 947 {
 948         struct iser_tx_desc *tx_desc = &isert_cmd->tx_desc;
 949
 950         tx_desc->tx_cqe.done = isert_send_done;
 951         send_wr->wr_cqe = &tx_desc->tx_cqe;
 952
 953         if (isert_conn->snd_w_inv && isert_cmd->inv_rkey) {
 954                 send_wr->opcode  = IB_WR_SEND_WITH_INV;
 955                 send_wr->ex.invalidate_rkey = isert_cmd->inv_rkey;
 956         } else {
 957                 send_wr->opcode = IB_WR_SEND;
 958         }
 959
 960         send_wr->sg_list = &tx_desc->tx_sg[0];
 961         send_wr->num_sge = isert_cmd->tx_desc.num_sge;
 962         send_wr->send_flags = IB_SEND_SIGNALED;
 963 }
 964
 965 static int
 966 isert_login_post_recv(struct isert_conn *isert_conn)
 967 {
 968         struct ib_recv_wr rx_wr, *rx_wr_fail;
 969         struct ib_sge sge;
 970         int ret;
 971
 972         memset(&sge, 0, sizeof(struct ib_sge));
 973         sge.addr = isert_conn->login_req_dma;
 974         sge.length = ISER_RX_PAYLOAD_SIZE;
 975         sge.lkey = isert_conn->device->pd->local_dma_lkey;
 976
 977         isert_dbg("Setup sge: addr: %llx length: %d 0x%08x\n",
 978                 sge.addr, sge.length, sge.lkey);
 979
 980         isert_conn->login_req_buf->rx_cqe.done = isert_login_recv_done;
 981
 982         memset(&rx_wr, 0, sizeof(struct ib_recv_wr));
 983         rx_wr.wr_cqe = &isert_conn->login_req_buf->rx_cqe;
 984         rx_wr.sg_list = &sge;
 985         rx_wr.num_sge = 1;
 986
 987         ret = ib_post_recv(isert_conn->qp, &rx_wr, &rx_wr_fail);
 988         if (ret)
 989                 isert_err("ib_post_recv() failed: %d\n", ret);
 990
 991         return ret;
 992 }
 993
 994 static int
 995 isert_put_login_tx(struct iscsi_conn *conn, struct iscsi_login *login,
 996                    u32 length)
 997 {
 998         struct isert_conn *isert_conn = conn->context;
 999         struct isert_device *device = isert_conn->device;
1000         struct ib_device *ib_dev = device->ib_device;
1001         struct iser_tx_desc *tx_desc = &isert_conn->login_tx_desc;
1002         int ret;
1003
1004         __isert_create_send_desc(device, tx_desc);
1005
1006         memcpy(&tx_desc->iscsi_header, &login->rsp[0],
1007                sizeof(struct iscsi_hdr));
1008
1009         isert_init_tx_hdrs(isert_conn, tx_desc);
1010
1011         if (length > 0) {
1012                 struct ib_sge *tx_dsg = &tx_desc->tx_sg[1];
1013
1014                 ib_dma_sync_single_for_cpu(ib_dev, isert_conn->login_rsp_dma,
1015                                            length, DMA_TO_DEVICE);
1016
1017                 memcpy(isert_conn->login_rsp_buf, login->rsp_buf, length);
1018
1019                 ib_dma_sync_single_for_device(ib_dev, isert_conn->login_rsp_dma,
1020                                               length, DMA_TO_DEVICE);
1021
1022                 tx_dsg->addr    = isert_conn->login_rsp_dma;
1023                 tx_dsg->length  = length;
1024                 tx_dsg->lkey    = isert_conn->device->pd->local_dma_lkey;
1025                 tx_desc->num_sge = 2;
1026         }
1027         if (!login->login_failed) {
1028                 if (login->login_complete) {
1029                         ret = isert_alloc_rx_descriptors(isert_conn);
1030                         if (ret)
1031                                 return ret;
1032
1033                         ret = isert_post_recvm(isert_conn,
1034                                                ISERT_QP_MAX_RECV_DTOS);
1035                         if (ret)
1036                                 return ret;
1037
1038                         /* Now we are in FULL_FEATURE phase */
1039                         mutex_lock(&isert_conn->mutex);
1040                         isert_conn->state = ISER_CONN_FULL_FEATURE;
1041                         mutex_unlock(&isert_conn->mutex);
1042                         goto post_send;
1043                 }
1044
1045                 ret = isert_login_post_recv(isert_conn);
1046                 if (ret)
1047                         return ret;
1048         }
1049 post_send:
1050         ret = isert_login_post_send(isert_conn, tx_desc);
1051         if (ret)
1052                 return ret;
1053
1054         return 0;
1055 }
1056
1057 static void
1058 isert_rx_login_req(struct isert_conn *isert_conn)
1059 {
1060         struct iser_rx_desc *rx_desc = isert_conn->login_req_buf;
1061         int rx_buflen = isert_conn->login_req_len;
1062         struct iscsi_conn *conn = isert_conn->conn;
1063         struct iscsi_login *login = conn->conn_login;
1064         int size;
1065
1066         isert_info("conn %p\n", isert_conn);
1067
1068         WARN_ON_ONCE(!login);
1069
1070         if (login->first_request) {
1071                 struct iscsi_login_req *login_req =
1072                         (struct iscsi_login_req *)&rx_desc->iscsi_header;
1073                 /*
1074                  * Setup the initial iscsi_login values from the leading
1075                  * login request PDU.
1076                  */
1077                 login->leading_connection = (!login_req->tsih) ? 1 : 0;
1078                 login->current_stage =
1079                         (login_req->flags & ISCSI_FLAG_LOGIN_CURRENT_STAGE_MASK)
1080                          >> 2;
1081                 login->version_min      = login_req->min_version;
1082                 login->version_max      = login_req->max_version;
1083                 memcpy(login->isid, login_req->isid, 6);
1084                 login->cmd_sn           = be32_to_cpu(login_req->cmdsn);
1085                 login->init_task_tag    = login_req->itt;
1086                 login->initial_exp_statsn = be32_to_cpu(login_req->exp_statsn);
1087                 login->cid              = be16_to_cpu(login_req->cid);
1088                 login->tsih             = be16_to_cpu(login_req->tsih);
1089         }
1090
1091         memcpy(&login->req[0], (void *)&rx_desc->iscsi_header, ISCSI_HDR_LEN);
1092
1093         size = min(rx_buflen, MAX_KEY_VALUE_PAIRS);
1094         isert_dbg("Using login payload size: %d, rx_buflen: %d "
1095                   "MAX_KEY_VALUE_PAIRS: %d\n", size, rx_buflen,
1096                   MAX_KEY_VALUE_PAIRS);
1097         memcpy(login->req_buf, &rx_desc->data[0], size);
1098
1099         if (login->first_request) {
1100                 complete(&isert_conn->login_comp);
1101                 return;
1102         }
1103         schedule_delayed_work(&conn->login_work, 0);
1104 }
1105
1106 static struct iscsi_cmd
1107 *isert_allocate_cmd(struct iscsi_conn *conn, struct iser_rx_desc *rx_desc)
1108 {
1109         struct isert_conn *isert_conn = conn->context;
1110         struct isert_cmd *isert_cmd;
1111         struct iscsi_cmd *cmd;
1112
1113         cmd = iscsit_allocate_cmd(conn, TASK_INTERRUPTIBLE);
1114         if (!cmd) {
1115                 isert_err("Unable to allocate iscsi_cmd + isert_cmd\n");
1116                 return NULL;
1117         }
1118         isert_cmd = iscsit_priv_cmd(cmd);
1119         isert_cmd->conn = isert_conn;
1120         isert_cmd->iscsi_cmd = cmd;
1121         isert_cmd->rx_desc = rx_desc;
1122
1123         return cmd;
1124 }
1125
1126 static int
1127 isert_handle_scsi_cmd(struct isert_conn *isert_conn,
1128                       struct isert_cmd *isert_cmd, struct iscsi_cmd *cmd,
1129                       struct iser_rx_desc *rx_desc, unsigned char *buf)
1130 {
1131         struct iscsi_conn *conn = isert_conn->conn;
1132         struct iscsi_scsi_req *hdr = (struct iscsi_scsi_req *)buf;
1133         int imm_data, imm_data_len, unsol_data, sg_nents, rc;
1134         bool dump_payload = false;
1135         unsigned int data_len;
1136
1137         rc = iscsit_setup_scsi_cmd(conn, cmd, buf);
1138         if (rc < 0)
1139                 return rc;
1140
1141         imm_data = cmd->immediate_data;
1142         imm_data_len = cmd->first_burst_len;
1143         unsol_data = cmd->unsolicited_data;
1144         data_len = cmd->se_cmd.data_length;
1145
1146         if (imm_data && imm_data_len == data_len)
1147                 cmd->se_cmd.se_cmd_flags |= SCF_PASSTHROUGH_SG_TO_MEM_NOALLOC;
1148         rc = iscsit_process_scsi_cmd(conn, cmd, hdr);
1149         if (rc < 0) {
1150                 return 0;
1151         } else if (rc > 0) {
1152                 dump_payload = true;
1153                 goto sequence_cmd;
1154         }
1155
1156         if (!imm_data)
1157                 return 0;
1158
1159         if (imm_data_len != data_len) {
1160                 sg_nents = max(1UL, DIV_ROUND_UP(imm_data_len, PAGE_SIZE));
1161                 sg_copy_from_buffer(cmd->se_cmd.t_data_sg, sg_nents,
1162                                     &rx_desc->data[0], imm_data_len);
1163                 isert_dbg("Copy Immediate sg_nents: %u imm_data_len: %d\n",
1164                           sg_nents, imm_data_len);
1165         } else {
1166                 sg_init_table(&isert_cmd->sg, 1);
1167                 cmd->se_cmd.t_data_sg = &isert_cmd->sg;
1168                 cmd->se_cmd.t_data_nents = 1;
1169                 sg_set_buf(&isert_cmd->sg, &rx_desc->data[0], imm_data_len);
1170                 isert_dbg("Transfer Immediate imm_data_len: %d\n",
1171                           imm_data_len);
1172         }
1173
1174         cmd->write_data_done += imm_data_len;
1175
1176         if (cmd->write_data_done == cmd->se_cmd.data_length) {
1177                 spin_lock_bh(&cmd->istate_lock);
1178                 cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1179                 cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1180                 spin_unlock_bh(&cmd->istate_lock);
1181         }
1182
1183 sequence_cmd:
1184         rc = iscsit_sequence_cmd(conn, cmd, buf, hdr->cmdsn);
1185
1186         if (!rc && dump_payload == false && unsol_data)
1187                 iscsit_set_unsoliticed_dataout(cmd);
1188         else if (dump_payload && imm_data)
1189                 target_put_sess_cmd(&cmd->se_cmd);
1190
1191         return 0;
1192 }
1193
1194 static int
1195 isert_handle_iscsi_dataout(struct isert_conn *isert_conn,
1196                            struct iser_rx_desc *rx_desc, unsigned char *buf)
1197 {
1198         struct scatterlist *sg_start;
1199         struct iscsi_conn *conn = isert_conn->conn;
1200         struct iscsi_cmd *cmd = NULL;
1201         struct iscsi_data *hdr = (struct iscsi_data *)buf;
1202         u32 unsol_data_len = ntoh24(hdr->dlength);
1203         int rc, sg_nents, sg_off, page_off;
1204
1205         rc = iscsit_check_dataout_hdr(conn, buf, &cmd);
1206         if (rc < 0)
1207                 return rc;
1208         else if (!cmd)
1209                 return 0;
1210         /*
1211          * FIXME: Unexpected unsolicited_data out
1212          */
1213         if (!cmd->unsolicited_data) {
1214                 isert_err("Received unexpected solicited data payload\n");
1215                 dump_stack();
1216                 return -1;
1217         }
1218
1219         isert_dbg("Unsolicited DataOut unsol_data_len: %u, "
1220                   "write_data_done: %u, data_length: %u\n",
1221                   unsol_data_len,  cmd->write_data_done,
1222                   cmd->se_cmd.data_length);
1223
1224         sg_off = cmd->write_data_done / PAGE_SIZE;
1225         sg_start = &cmd->se_cmd.t_data_sg[sg_off];
1226         sg_nents = max(1UL, DIV_ROUND_UP(unsol_data_len, PAGE_SIZE));
1227         page_off = cmd->write_data_done % PAGE_SIZE;
1228         /*
1229          * FIXME: Non page-aligned unsolicited_data out
1230          */
1231         if (page_off) {
1232                 isert_err("unexpected non-page aligned data payload\n");
1233                 dump_stack();
1234                 return -1;
1235         }
1236         isert_dbg("Copying DataOut: sg_start: %p, sg_off: %u "
1237                   "sg_nents: %u from %p %u\n", sg_start, sg_off,
1238                   sg_nents, &rx_desc->data[0], unsol_data_len);
1239
1240         sg_copy_from_buffer(sg_start, sg_nents, &rx_desc->data[0],
1241                             unsol_data_len);
1242
1243         rc = iscsit_check_dataout_payload(cmd, hdr, false);
1244         if (rc < 0)
1245                 return rc;
1246
1247         /*
1248          * multiple data-outs on the same command can arrive -
1249          * so post the buffer before hand
1250          */
1251         rc = isert_post_recv(isert_conn, rx_desc);
1252         if (rc) {
1253                 isert_err("ib_post_recv failed with %d\n", rc);
1254                 return rc;
1255         }
1256         return 0;
1257 }
1258
1259 static int
1260 isert_handle_nop_out(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1261                      struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1262                      unsigned char *buf)
1263 {
1264         struct iscsi_conn *conn = isert_conn->conn;
1265         struct iscsi_nopout *hdr = (struct iscsi_nopout *)buf;
1266         int rc;
1267
1268         rc = iscsit_setup_nop_out(conn, cmd, hdr);
1269         if (rc < 0)
1270                 return rc;
1271         /*
1272          * FIXME: Add support for NOPOUT payload using unsolicited RDMA payload
1273          */
1274
1275         return iscsit_process_nop_out(conn, cmd, hdr);
1276 }
1277
1278 static int
1279 isert_handle_text_cmd(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd,
1280                       struct iscsi_cmd *cmd, struct iser_rx_desc *rx_desc,
1281                       struct iscsi_text *hdr)
1282 {
1283         struct iscsi_conn *conn = isert_conn->conn;
1284         u32 payload_length = ntoh24(hdr->dlength);
1285         int rc;
1286         unsigned char *text_in = NULL;
1287
1288         rc = iscsit_setup_text_cmd(conn, cmd, hdr);
1289         if (rc < 0)
1290                 return rc;
1291
1292         if (payload_length) {
1293                 text_in = kzalloc(payload_length, GFP_KERNEL);
1294                 if (!text_in)
1295                         return -ENOMEM;
1296         }
1297         cmd->text_in_ptr = text_in;
1298
1299         memcpy(cmd->text_in_ptr, &rx_desc->data[0], payload_length);
1300
1301         return iscsit_process_text_cmd(conn, cmd, hdr);
1302 }
1303
1304 static int
1305 isert_rx_opcode(struct isert_conn *isert_conn, struct iser_rx_desc *rx_desc,
1306                 uint32_t read_stag, uint64_t read_va,
1307                 uint32_t write_stag, uint64_t write_va)
1308 {
1309         struct iscsi_hdr *hdr = &rx_desc->iscsi_header;
1310         struct iscsi_conn *conn = isert_conn->conn;
1311         struct iscsi_cmd *cmd;
1312         struct isert_cmd *isert_cmd;
1313         int ret = -EINVAL;
1314         u8 opcode = (hdr->opcode & ISCSI_OPCODE_MASK);
1315
1316         if (conn->sess->sess_ops->SessionType &&
1317            (!(opcode & ISCSI_OP_TEXT) || !(opcode & ISCSI_OP_LOGOUT))) {
1318                 isert_err("Got illegal opcode: 0x%02x in SessionType=Discovery,"
1319                           " ignoring\n", opcode);
1320                 return 0;
1321         }
1322
1323         switch (opcode) {
1324         case ISCSI_OP_SCSI_CMD:
1325                 cmd = isert_allocate_cmd(conn, rx_desc);
1326                 if (!cmd)
1327                         break;
1328
1329                 isert_cmd = iscsit_priv_cmd(cmd);
1330                 isert_cmd->read_stag = read_stag;
1331                 isert_cmd->read_va = read_va;
1332                 isert_cmd->write_stag = write_stag;
1333                 isert_cmd->write_va = write_va;
1334                 isert_cmd->inv_rkey = read_stag ? read_stag : write_stag;
1335
1336                 ret = isert_handle_scsi_cmd(isert_conn, isert_cmd, cmd,
1337                                         rx_desc, (unsigned char *)hdr);
1338                 break;
1339         case ISCSI_OP_NOOP_OUT:
1340                 cmd = isert_allocate_cmd(conn, rx_desc);
1341                 if (!cmd)
1342                         break;
1343
1344                 isert_cmd = iscsit_priv_cmd(cmd);
1345                 ret = isert_handle_nop_out(isert_conn, isert_cmd, cmd,
1346                                            rx_desc, (unsigned char *)hdr);
1347                 break;
1348         case ISCSI_OP_SCSI_DATA_OUT:
1349                 ret = isert_handle_iscsi_dataout(isert_conn, rx_desc,
1350                                                 (unsigned char *)hdr);
1351                 break;
1352         case ISCSI_OP_SCSI_TMFUNC:
1353                 cmd = isert_allocate_cmd(conn, rx_desc);
1354                 if (!cmd)
1355                         break;
1356
1357                 ret = iscsit_handle_task_mgt_cmd(conn, cmd,
1358                                                 (unsigned char *)hdr);
1359                 break;
1360         case ISCSI_OP_LOGOUT:
1361                 cmd = isert_allocate_cmd(conn, rx_desc);
1362                 if (!cmd)
1363                         break;
1364
1365                 ret = iscsit_handle_logout_cmd(conn, cmd, (unsigned char *)hdr);
1366                 break;
1367         case ISCSI_OP_TEXT:
1368                 if (be32_to_cpu(hdr->ttt) != 0xFFFFFFFF)
1369                         cmd = iscsit_find_cmd_from_itt(conn, hdr->itt);
1370                 else
1371                         cmd = isert_allocate_cmd(conn, rx_desc);
1372
1373                 if (!cmd)
1374                         break;
1375
1376                 isert_cmd = iscsit_priv_cmd(cmd);
1377                 ret = isert_handle_text_cmd(isert_conn, isert_cmd, cmd,
1378                                             rx_desc, (struct iscsi_text *)hdr);
1379                 break;
1380         default:
1381                 isert_err("Got unknown iSCSI OpCode: 0x%02x\n", opcode);
1382                 dump_stack();
1383                 break;
1384         }
1385
1386         return ret;
1387 }
1388
1389 static void
1390 isert_print_wc(struct ib_wc *wc, const char *type)
1391 {
1392         if (wc->status != IB_WC_WR_FLUSH_ERR)
1393                 isert_err("%s failure: %s (%d) vend_err %x\n", type,
1394                           ib_wc_status_msg(wc->status), wc->status,
1395                           wc->vendor_err);
1396         else
1397                 isert_dbg("%s failure: %s (%d)\n", type,
1398                           ib_wc_status_msg(wc->status), wc->status);
1399 }
1400
1401 static void
1402 isert_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1403 {
1404         struct isert_conn *isert_conn = wc->qp->qp_context;
1405         struct ib_device *ib_dev = isert_conn->cm_id->device;
1406         struct iser_rx_desc *rx_desc = cqe_to_rx_desc(wc->wr_cqe);
1407         struct iscsi_hdr *hdr = &rx_desc->iscsi_header;
1408         struct iser_ctrl *iser_ctrl = &rx_desc->iser_header;
1409         uint64_t read_va = 0, write_va = 0;
1410         uint32_t read_stag = 0, write_stag = 0;
1411
1412         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1413                 isert_print_wc(wc, "recv");
1414                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1415                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1416                 return;
1417         }
1418
1419         rx_desc->in_use = true;
1420
1421         ib_dma_sync_single_for_cpu(ib_dev, rx_desc->dma_addr,
1422                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1423
1424         isert_dbg("DMA: 0x%llx, iSCSI opcode: 0x%02x, ITT: 0x%08x, flags: 0x%02x dlen: %d\n",
1425                  rx_desc->dma_addr, hdr->opcode, hdr->itt, hdr->flags,
1426                  (int)(wc->byte_len - ISER_HEADERS_LEN));
1427
1428         switch (iser_ctrl->flags & 0xF0) {
1429         case ISCSI_CTRL:
1430                 if (iser_ctrl->flags & ISER_RSV) {
1431                         read_stag = be32_to_cpu(iser_ctrl->read_stag);
1432                         read_va = be64_to_cpu(iser_ctrl->read_va);
1433                         isert_dbg("ISER_RSV: read_stag: 0x%x read_va: 0x%llx\n",
1434                                   read_stag, (unsigned long long)read_va);
1435                 }
1436                 if (iser_ctrl->flags & ISER_WSV) {
1437                         write_stag = be32_to_cpu(iser_ctrl->write_stag);
1438                         write_va = be64_to_cpu(iser_ctrl->write_va);
1439                         isert_dbg("ISER_WSV: write_stag: 0x%x write_va: 0x%llx\n",
1440                                   write_stag, (unsigned long long)write_va);
1441                 }
1442
1443                 isert_dbg("ISER ISCSI_CTRL PDU\n");
1444                 break;
1445         case ISER_HELLO:
1446                 isert_err("iSER Hello message\n");
1447                 break;
1448         default:
1449                 isert_warn("Unknown iSER hdr flags: 0x%02x\n", iser_ctrl->flags);
1450                 break;
1451         }
1452
1453         isert_rx_opcode(isert_conn, rx_desc,
1454                         read_stag, read_va, write_stag, write_va);
1455
1456         ib_dma_sync_single_for_device(ib_dev, rx_desc->dma_addr,
1457                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1458 }
1459
1460 static void
1461 isert_login_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1462 {
1463         struct isert_conn *isert_conn = wc->qp->qp_context;
1464         struct ib_device *ib_dev = isert_conn->device->ib_device;
1465
1466         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1467                 isert_print_wc(wc, "login recv");
1468                 return;
1469         }
1470
1471         ib_dma_sync_single_for_cpu(ib_dev, isert_conn->login_req_dma,
1472                         ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1473
1474         isert_conn->login_req_len = wc->byte_len - ISER_HEADERS_LEN;
1475
1476         if (isert_conn->conn) {
1477                 struct iscsi_login *login = isert_conn->conn->conn_login;
1478
1479                 if (login && !login->first_request)
1480                         isert_rx_login_req(isert_conn);
1481         }
1482
1483         mutex_lock(&isert_conn->mutex);
1484         complete(&isert_conn->login_req_comp);
1485         mutex_unlock(&isert_conn->mutex);
1486
1487         ib_dma_sync_single_for_device(ib_dev, isert_conn->login_req_dma,
1488                                 ISER_RX_PAYLOAD_SIZE, DMA_FROM_DEVICE);
1489 }
1490
1491 static void
1492 isert_rdma_rw_ctx_destroy(struct isert_cmd *cmd, struct isert_conn *conn)
1493 {
1494         struct se_cmd *se_cmd = &cmd->iscsi_cmd->se_cmd;
1495         enum dma_data_direction dir = target_reverse_dma_direction(se_cmd);
1496
1497         if (!cmd->rw.nr_ops)
1498                 return;
1499
1500         if (isert_prot_cmd(conn, se_cmd)) {
1501                 rdma_rw_ctx_destroy_signature(&cmd->rw, conn->qp,
1502                                 conn->cm_id->port_num, se_cmd->t_data_sg,
1503                                 se_cmd->t_data_nents, se_cmd->t_prot_sg,
1504                                 se_cmd->t_prot_nents, dir);
1505         } else {
1506                 rdma_rw_ctx_destroy(&cmd->rw, conn->qp, conn->cm_id->port_num,
1507                                 se_cmd->t_data_sg, se_cmd->t_data_nents, dir);
1508         }
1509
1510         cmd->rw.nr_ops = 0;
1511 }
1512
1513 static void
1514 isert_put_cmd(struct isert_cmd *isert_cmd, bool comp_err)
1515 {
1516         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1517         struct isert_conn *isert_conn = isert_cmd->conn;
1518         struct iscsi_conn *conn = isert_conn->conn;
1519         struct iscsi_text_rsp *hdr;
1520
1521         isert_dbg("Cmd %p\n", isert_cmd);
1522
1523         switch (cmd->iscsi_opcode) {
1524         case ISCSI_OP_SCSI_CMD:
1525                 spin_lock_bh(&conn->cmd_lock);
1526                 if (!list_empty(&cmd->i_conn_node))
1527                         list_del_init(&cmd->i_conn_node);
1528                 spin_unlock_bh(&conn->cmd_lock);
1529
1530                 if (cmd->data_direction == DMA_TO_DEVICE) {
1531                         iscsit_stop_dataout_timer(cmd);
1532                         /*
1533                          * Check for special case during comp_err where
1534                          * WRITE_PENDING has been handed off from core,
1535                          * but requires an extra target_put_sess_cmd()
1536                          * before transport_generic_free_cmd() below.
1537                          */
1538                         if (comp_err &&
1539                             cmd->se_cmd.t_state == TRANSPORT_WRITE_PENDING) {
1540                                 struct se_cmd *se_cmd = &cmd->se_cmd;
1541
1542                                 target_put_sess_cmd(se_cmd);
1543                         }
1544                 }
1545
1546                 isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1547                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1548                 break;
1549         case ISCSI_OP_SCSI_TMFUNC:
1550                 spin_lock_bh(&conn->cmd_lock);
1551                 if (!list_empty(&cmd->i_conn_node))
1552                         list_del_init(&cmd->i_conn_node);
1553                 spin_unlock_bh(&conn->cmd_lock);
1554
1555                 transport_generic_free_cmd(&cmd->se_cmd, 0);
1556                 break;
1557         case ISCSI_OP_REJECT:
1558         case ISCSI_OP_NOOP_OUT:
1559         case ISCSI_OP_TEXT:
1560                 hdr = (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
1561                 /* If the continue bit is on, keep the command alive */
1562                 if (hdr->flags & ISCSI_FLAG_TEXT_CONTINUE)
1563                         break;
1564
1565                 spin_lock_bh(&conn->cmd_lock);
1566                 if (!list_empty(&cmd->i_conn_node))
1567                         list_del_init(&cmd->i_conn_node);
1568                 spin_unlock_bh(&conn->cmd_lock);
1569
1570                 /*
1571                  * Handle special case for REJECT when iscsi_add_reject*() has
1572                  * overwritten the original iscsi_opcode assignment, and the
1573                  * associated cmd->se_cmd needs to be released.
1574                  */
1575                 if (cmd->se_cmd.se_tfo != NULL) {
1576                         isert_dbg("Calling transport_generic_free_cmd for 0x%02x\n",
1577                                  cmd->iscsi_opcode);
1578                         transport_generic_free_cmd(&cmd->se_cmd, 0);
1579                         break;
1580                 }
1581                 /*
1582                  * Fall-through
1583                  */
1584         default:
1585                 iscsit_release_cmd(cmd);
1586                 break;
1587         }
1588 }
1589
1590 static void
1591 isert_unmap_tx_desc(struct iser_tx_desc *tx_desc, struct ib_device *ib_dev)
1592 {
1593         if (tx_desc->dma_addr != 0) {
1594                 isert_dbg("unmap single for tx_desc->dma_addr\n");
1595                 ib_dma_unmap_single(ib_dev, tx_desc->dma_addr,
1596                                     ISER_HEADERS_LEN, DMA_TO_DEVICE);
1597                 tx_desc->dma_addr = 0;
1598         }
1599 }
1600
1601 static void
1602 isert_completion_put(struct iser_tx_desc *tx_desc, struct isert_cmd *isert_cmd,
1603                      struct ib_device *ib_dev, bool comp_err)
1604 {
1605         if (isert_cmd->pdu_buf_dma != 0) {
1606                 isert_dbg("unmap single for isert_cmd->pdu_buf_dma\n");
1607                 ib_dma_unmap_single(ib_dev, isert_cmd->pdu_buf_dma,
1608                                     isert_cmd->pdu_buf_len, DMA_TO_DEVICE);
1609                 isert_cmd->pdu_buf_dma = 0;
1610         }
1611
1612         isert_unmap_tx_desc(tx_desc, ib_dev);
1613         isert_put_cmd(isert_cmd, comp_err);
1614 }
1615
1616 static int
1617 isert_check_pi_status(struct se_cmd *se_cmd, struct ib_mr *sig_mr)
1618 {
1619         struct ib_mr_status mr_status;
1620         int ret;
1621
1622         ret = ib_check_mr_status(sig_mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
1623         if (ret) {
1624                 isert_err("ib_check_mr_status failed, ret %d\n", ret);
1625                 goto fail_mr_status;
1626         }
1627
1628         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1629                 u64 sec_offset_err;
1630                 u32 block_size = se_cmd->se_dev->dev_attrib.block_size + 8;
1631
1632                 switch (mr_status.sig_err.err_type) {
1633                 case IB_SIG_BAD_GUARD:
1634                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_GUARD_CHECK_FAILED;
1635                         break;
1636                 case IB_SIG_BAD_REFTAG:
1637                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_REF_TAG_CHECK_FAILED;
1638                         break;
1639                 case IB_SIG_BAD_APPTAG:
1640                         se_cmd->pi_err = TCM_LOGICAL_BLOCK_APP_TAG_CHECK_FAILED;
1641                         break;
1642                 }
1643                 sec_offset_err = mr_status.sig_err.sig_err_offset;
1644                 do_div(sec_offset_err, block_size);
1645                 se_cmd->bad_sector = sec_offset_err + se_cmd->t_task_lba;
1646
1647                 isert_err("PI error found type %d at sector 0x%llx "
1648                           "expected 0x%x vs actual 0x%x\n",
1649                           mr_status.sig_err.err_type,
1650                           (unsigned long long)se_cmd->bad_sector,
1651                           mr_status.sig_err.expected,
1652                           mr_status.sig_err.actual);
1653                 ret = 1;
1654         }
1655
1656 fail_mr_status:
1657         return ret;
1658 }
1659
1660 static void
1661 isert_rdma_write_done(struct ib_cq *cq, struct ib_wc *wc)
1662 {
1663         struct isert_conn *isert_conn = wc->qp->qp_context;
1664         struct isert_device *device = isert_conn->device;
1665         struct iser_tx_desc *desc = cqe_to_tx_desc(wc->wr_cqe);
1666         struct isert_cmd *isert_cmd = tx_desc_to_cmd(desc);
1667         struct se_cmd *cmd = &isert_cmd->iscsi_cmd->se_cmd;
1668         int ret = 0;
1669
1670         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1671                 isert_print_wc(wc, "rdma write");
1672                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1673                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1674                 isert_completion_put(desc, isert_cmd, device->ib_device, true);
1675                 return;
1676         }
1677
1678         isert_dbg("Cmd %p\n", isert_cmd);
1679
1680         ret = isert_check_pi_status(cmd, isert_cmd->rw.sig->sig_mr);
1681         isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1682
1683         if (ret) {
1684                 /*
1685                  * transport_generic_request_failure() expects to have
1686                  * plus two references to handle queue-full, so re-add
1687                  * one here as target-core will have already dropped
1688                  * it after the first isert_put_datain() callback.
1689                  */
1690                 kref_get(&cmd->cmd_kref);
1691                 transport_generic_request_failure(cmd, cmd->pi_err);
1692         } else {
1693                 /*
1694                  * XXX: isert_put_response() failure is not retried.
1695                  */
1696                 ret = isert_put_response(isert_conn->conn, isert_cmd->iscsi_cmd);
1697                 if (ret)
1698                         pr_warn_ratelimited("isert_put_response() ret: %d\n", ret);
1699         }
1700 }
1701
1702 static void
1703 isert_rdma_read_done(struct ib_cq *cq, struct ib_wc *wc)
1704 {
1705         struct isert_conn *isert_conn = wc->qp->qp_context;
1706         struct isert_device *device = isert_conn->device;
1707         struct iser_tx_desc *desc = cqe_to_tx_desc(wc->wr_cqe);
1708         struct isert_cmd *isert_cmd = tx_desc_to_cmd(desc);
1709         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1710         struct se_cmd *se_cmd = &cmd->se_cmd;
1711         int ret = 0;
1712
1713         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1714                 isert_print_wc(wc, "rdma read");
1715                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1716                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1717                 isert_completion_put(desc, isert_cmd, device->ib_device, true);
1718                 return;
1719         }
1720
1721         isert_dbg("Cmd %p\n", isert_cmd);
1722
1723         iscsit_stop_dataout_timer(cmd);
1724
1725         if (isert_prot_cmd(isert_conn, se_cmd))
1726                 ret = isert_check_pi_status(se_cmd, isert_cmd->rw.sig->sig_mr);
1727         isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1728         cmd->write_data_done = 0;
1729
1730         isert_dbg("Cmd: %p RDMA_READ comp calling execute_cmd\n", isert_cmd);
1731         spin_lock_bh(&cmd->istate_lock);
1732         cmd->cmd_flags |= ICF_GOT_LAST_DATAOUT;
1733         cmd->i_state = ISTATE_RECEIVED_LAST_DATAOUT;
1734         spin_unlock_bh(&cmd->istate_lock);
1735
1736         /*
1737          * transport_generic_request_failure() will drop the extra
1738          * se_cmd->cmd_kref reference after T10-PI error, and handle
1739          * any non-zero ->queue_status() callback error retries.
1740          */
1741         if (ret)
1742                 transport_generic_request_failure(se_cmd, se_cmd->pi_err);
1743         else
1744                 target_execute_cmd(se_cmd);
1745 }
1746
1747 static void
1748 isert_do_control_comp(struct work_struct *work)
1749 {
1750         struct isert_cmd *isert_cmd = container_of(work,
1751                         struct isert_cmd, comp_work);
1752         struct isert_conn *isert_conn = isert_cmd->conn;
1753         struct ib_device *ib_dev = isert_conn->cm_id->device;
1754         struct iscsi_cmd *cmd = isert_cmd->iscsi_cmd;
1755
1756         isert_dbg("Cmd %p i_state %d\n", isert_cmd, cmd->i_state);
1757
1758         switch (cmd->i_state) {
1759         case ISTATE_SEND_TASKMGTRSP:
1760                 iscsit_tmr_post_handler(cmd, cmd->conn);
1761         case ISTATE_SEND_REJECT:   /* FALLTHRU */
1762         case ISTATE_SEND_TEXTRSP:  /* FALLTHRU */
1763                 cmd->i_state = ISTATE_SENT_STATUS;
1764                 isert_completion_put(&isert_cmd->tx_desc, isert_cmd,
1765                                      ib_dev, false);
1766                 break;
1767         case ISTATE_SEND_LOGOUTRSP:
1768                 iscsit_logout_post_handler(cmd, cmd->conn);
1769                 break;
1770         default:
1771                 isert_err("Unknown i_state %d\n", cmd->i_state);
1772                 dump_stack();
1773                 break;
1774         }
1775 }
1776
1777 static void
1778 isert_login_send_done(struct ib_cq *cq, struct ib_wc *wc)
1779 {
1780         struct isert_conn *isert_conn = wc->qp->qp_context;
1781         struct ib_device *ib_dev = isert_conn->cm_id->device;
1782         struct iser_tx_desc *tx_desc = cqe_to_tx_desc(wc->wr_cqe);
1783
1784         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1785                 isert_print_wc(wc, "login send");
1786                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1787                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1788         }
1789
1790         isert_unmap_tx_desc(tx_desc, ib_dev);
1791 }
1792
1793 static void
1794 isert_send_done(struct ib_cq *cq, struct ib_wc *wc)
1795 {
1796         struct isert_conn *isert_conn = wc->qp->qp_context;
1797         struct ib_device *ib_dev = isert_conn->cm_id->device;
1798         struct iser_tx_desc *tx_desc = cqe_to_tx_desc(wc->wr_cqe);
1799         struct isert_cmd *isert_cmd = tx_desc_to_cmd(tx_desc);
1800
1801         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1802                 isert_print_wc(wc, "send");
1803                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1804                         iscsit_cause_connection_reinstatement(isert_conn->conn, 0);
1805                 isert_completion_put(tx_desc, isert_cmd, ib_dev, true);
1806                 return;
1807         }
1808
1809         isert_dbg("Cmd %p\n", isert_cmd);
1810
1811         switch (isert_cmd->iscsi_cmd->i_state) {
1812         case ISTATE_SEND_TASKMGTRSP:
1813         case ISTATE_SEND_LOGOUTRSP:
1814         case ISTATE_SEND_REJECT:
1815         case ISTATE_SEND_TEXTRSP:
1816                 isert_unmap_tx_desc(tx_desc, ib_dev);
1817
1818                 INIT_WORK(&isert_cmd->comp_work, isert_do_control_comp);
1819                 queue_work(isert_comp_wq, &isert_cmd->comp_work);
1820                 return;
1821         default:
1822                 isert_cmd->iscsi_cmd->i_state = ISTATE_SENT_STATUS;
1823                 isert_completion_put(tx_desc, isert_cmd, ib_dev, false);
1824                 break;
1825         }
1826 }
1827
1828 static int
1829 isert_post_response(struct isert_conn *isert_conn, struct isert_cmd *isert_cmd)
1830 {
1831         struct ib_send_wr *wr_failed;
1832         int ret;
1833
1834         ret = isert_post_recv(isert_conn, isert_cmd->rx_desc);
1835         if (ret) {
1836                 isert_err("ib_post_recv failed with %d\n", ret);
1837                 return ret;
1838         }
1839
1840         ret = ib_post_send(isert_conn->qp, &isert_cmd->tx_desc.send_wr,
1841                            &wr_failed);
1842         if (ret) {
1843                 isert_err("ib_post_send failed with %d\n", ret);
1844                 return ret;
1845         }
1846         return ret;
1847 }
1848
1849 static int
1850 isert_put_response(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
1851 {
1852         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1853         struct isert_conn *isert_conn = conn->context;
1854         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1855         struct iscsi_scsi_rsp *hdr = (struct iscsi_scsi_rsp *)
1856                                 &isert_cmd->tx_desc.iscsi_header;
1857
1858         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1859         iscsit_build_rsp_pdu(cmd, conn, true, hdr);
1860         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1861         /*
1862          * Attach SENSE DATA payload to iSCSI Response PDU
1863          */
1864         if (cmd->se_cmd.sense_buffer &&
1865             ((cmd->se_cmd.se_cmd_flags & SCF_TRANSPORT_TASK_SENSE) ||
1866             (cmd->se_cmd.se_cmd_flags & SCF_EMULATED_TASK_SENSE))) {
1867                 struct isert_device *device = isert_conn->device;
1868                 struct ib_device *ib_dev = device->ib_device;
1869                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
1870                 u32 padding, pdu_len;
1871
1872                 put_unaligned_be16(cmd->se_cmd.scsi_sense_length,
1873                                    cmd->sense_buffer);
1874                 cmd->se_cmd.scsi_sense_length += sizeof(__be16);
1875
1876                 padding = -(cmd->se_cmd.scsi_sense_length) & 3;
1877                 hton24(hdr->dlength, (u32)cmd->se_cmd.scsi_sense_length);
1878                 pdu_len = cmd->se_cmd.scsi_sense_length + padding;
1879
1880                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
1881                                 (void *)cmd->sense_buffer, pdu_len,
1882                                 DMA_TO_DEVICE);
1883                 if (ib_dma_mapping_error(ib_dev, isert_cmd->pdu_buf_dma))
1884                         return -ENOMEM;
1885
1886                 isert_cmd->pdu_buf_len = pdu_len;
1887                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
1888                 tx_dsg->length  = pdu_len;
1889                 tx_dsg->lkey    = device->pd->local_dma_lkey;
1890                 isert_cmd->tx_desc.num_sge = 2;
1891         }
1892
1893         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1894
1895         isert_dbg("Posting SCSI Response\n");
1896
1897         return isert_post_response(isert_conn, isert_cmd);
1898 }
1899
1900 static void
1901 isert_aborted_task(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
1902 {
1903         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1904         struct isert_conn *isert_conn = conn->context;
1905
1906         spin_lock_bh(&conn->cmd_lock);
1907         if (!list_empty(&cmd->i_conn_node))
1908                 list_del_init(&cmd->i_conn_node);
1909         spin_unlock_bh(&conn->cmd_lock);
1910
1911         if (cmd->data_direction == DMA_TO_DEVICE)
1912                 iscsit_stop_dataout_timer(cmd);
1913         isert_rdma_rw_ctx_destroy(isert_cmd, isert_conn);
1914 }
1915
1916 static enum target_prot_op
1917 isert_get_sup_prot_ops(struct iscsi_conn *conn)
1918 {
1919         struct isert_conn *isert_conn = conn->context;
1920         struct isert_device *device = isert_conn->device;
1921
1922         if (conn->tpg->tpg_attrib.t10_pi) {
1923                 if (device->pi_capable) {
1924                         isert_info("conn %p PI offload enabled\n", isert_conn);
1925                         isert_conn->pi_support = true;
1926                         return TARGET_PROT_ALL;
1927                 }
1928         }
1929
1930         isert_info("conn %p PI offload disabled\n", isert_conn);
1931         isert_conn->pi_support = false;
1932
1933         return TARGET_PROT_NORMAL;
1934 }
1935
1936 static int
1937 isert_put_nopin(struct iscsi_cmd *cmd, struct iscsi_conn *conn,
1938                 bool nopout_response)
1939 {
1940         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1941         struct isert_conn *isert_conn = conn->context;
1942         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1943
1944         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1945         iscsit_build_nopin_rsp(cmd, conn, (struct iscsi_nopin *)
1946                                &isert_cmd->tx_desc.iscsi_header,
1947                                nopout_response);
1948         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1949         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1950
1951         isert_dbg("conn %p Posting NOPIN Response\n", isert_conn);
1952
1953         return isert_post_response(isert_conn, isert_cmd);
1954 }
1955
1956 static int
1957 isert_put_logout_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
1958 {
1959         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1960         struct isert_conn *isert_conn = conn->context;
1961         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1962
1963         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1964         iscsit_build_logout_rsp(cmd, conn, (struct iscsi_logout_rsp *)
1965                                 &isert_cmd->tx_desc.iscsi_header);
1966         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1967         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1968
1969         isert_dbg("conn %p Posting Logout Response\n", isert_conn);
1970
1971         return isert_post_response(isert_conn, isert_cmd);
1972 }
1973
1974 static int
1975 isert_put_tm_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
1976 {
1977         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1978         struct isert_conn *isert_conn = conn->context;
1979         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1980
1981         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
1982         iscsit_build_task_mgt_rsp(cmd, conn, (struct iscsi_tm_rsp *)
1983                                   &isert_cmd->tx_desc.iscsi_header);
1984         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
1985         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
1986
1987         isert_dbg("conn %p Posting Task Management Response\n", isert_conn);
1988
1989         return isert_post_response(isert_conn, isert_cmd);
1990 }
1991
1992 static int
1993 isert_put_reject(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
1994 {
1995         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
1996         struct isert_conn *isert_conn = conn->context;
1997         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
1998         struct isert_device *device = isert_conn->device;
1999         struct ib_device *ib_dev = device->ib_device;
2000         struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2001         struct iscsi_reject *hdr =
2002                 (struct iscsi_reject *)&isert_cmd->tx_desc.iscsi_header;
2003
2004         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2005         iscsit_build_reject(cmd, conn, hdr);
2006         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2007
2008         hton24(hdr->dlength, ISCSI_HDR_LEN);
2009         isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2010                         (void *)cmd->buf_ptr, ISCSI_HDR_LEN,
2011                         DMA_TO_DEVICE);
2012         if (ib_dma_mapping_error(ib_dev, isert_cmd->pdu_buf_dma))
2013                 return -ENOMEM;
2014         isert_cmd->pdu_buf_len = ISCSI_HDR_LEN;
2015         tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2016         tx_dsg->length  = ISCSI_HDR_LEN;
2017         tx_dsg->lkey    = device->pd->local_dma_lkey;
2018         isert_cmd->tx_desc.num_sge = 2;
2019
2020         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2021
2022         isert_dbg("conn %p Posting Reject\n", isert_conn);
2023
2024         return isert_post_response(isert_conn, isert_cmd);
2025 }
2026
2027 static int
2028 isert_put_text_rsp(struct iscsi_cmd *cmd, struct iscsi_conn *conn)
2029 {
2030         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2031         struct isert_conn *isert_conn = conn->context;
2032         struct ib_send_wr *send_wr = &isert_cmd->tx_desc.send_wr;
2033         struct iscsi_text_rsp *hdr =
2034                 (struct iscsi_text_rsp *)&isert_cmd->tx_desc.iscsi_header;
2035         u32 txt_rsp_len;
2036         int rc;
2037
2038         isert_create_send_desc(isert_conn, isert_cmd, &isert_cmd->tx_desc);
2039         rc = iscsit_build_text_rsp(cmd, conn, hdr, ISCSI_INFINIBAND);
2040         if (rc < 0)
2041                 return rc;
2042
2043         txt_rsp_len = rc;
2044         isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2045
2046         if (txt_rsp_len) {
2047                 struct isert_device *device = isert_conn->device;
2048                 struct ib_device *ib_dev = device->ib_device;
2049                 struct ib_sge *tx_dsg = &isert_cmd->tx_desc.tx_sg[1];
2050                 void *txt_rsp_buf = cmd->buf_ptr;
2051
2052                 isert_cmd->pdu_buf_dma = ib_dma_map_single(ib_dev,
2053                                 txt_rsp_buf, txt_rsp_len, DMA_TO_DEVICE);
2054                 if (ib_dma_mapping_error(ib_dev, isert_cmd->pdu_buf_dma))
2055                         return -ENOMEM;
2056
2057                 isert_cmd->pdu_buf_len = txt_rsp_len;
2058                 tx_dsg->addr    = isert_cmd->pdu_buf_dma;
2059                 tx_dsg->length  = txt_rsp_len;
2060                 tx_dsg->lkey    = device->pd->local_dma_lkey;
2061                 isert_cmd->tx_desc.num_sge = 2;
2062         }
2063         isert_init_send_wr(isert_conn, isert_cmd, send_wr);
2064
2065         isert_dbg("conn %p Text Response\n", isert_conn);
2066
2067         return isert_post_response(isert_conn, isert_cmd);
2068 }
2069
2070 static inline void
2071 isert_set_dif_domain(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs,
2072                      struct ib_sig_domain *domain)
2073 {
2074         domain->sig_type = IB_SIG_TYPE_T10_DIF;
2075         domain->sig.dif.bg_type = IB_T10DIF_CRC;
2076         domain->sig.dif.pi_interval = se_cmd->se_dev->dev_attrib.block_size;
2077         domain->sig.dif.ref_tag = se_cmd->reftag_seed;
2078         /*
2079          * At the moment we hard code those, but if in the future
2080          * the target core would like to use it, we will take it
2081          * from se_cmd.
2082          */
2083         domain->sig.dif.apptag_check_mask = 0xffff;
2084         domain->sig.dif.app_escape = true;
2085         domain->sig.dif.ref_escape = true;
2086         if (se_cmd->prot_type == TARGET_DIF_TYPE1_PROT ||
2087             se_cmd->prot_type == TARGET_DIF_TYPE2_PROT)
2088                 domain->sig.dif.ref_remap = true;
2089 };
2090
2091 static int
2092 isert_set_sig_attrs(struct se_cmd *se_cmd, struct ib_sig_attrs *sig_attrs)
2093 {
2094         memset(sig_attrs, 0, sizeof(*sig_attrs));
2095
2096         switch (se_cmd->prot_op) {
2097         case TARGET_PROT_DIN_INSERT:
2098         case TARGET_PROT_DOUT_STRIP:
2099                 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
2100                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2101                 break;
2102         case TARGET_PROT_DOUT_INSERT:
2103         case TARGET_PROT_DIN_STRIP:
2104                 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
2105                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2106                 break;
2107         case TARGET_PROT_DIN_PASS:
2108         case TARGET_PROT_DOUT_PASS:
2109                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->wire);
2110                 isert_set_dif_domain(se_cmd, sig_attrs, &sig_attrs->mem);
2111                 break;
2112         default:
2113                 isert_err("Unsupported PI operation %d\n", se_cmd->prot_op);
2114                 return -EINVAL;
2115         }
2116
2117         sig_attrs->check_mask =
2118                (se_cmd->prot_checks & TARGET_DIF_CHECK_GUARD  ? 0xc0 : 0) |
2119                (se_cmd->prot_checks & TARGET_DIF_CHECK_APPTAG ? 0x30 : 0) |
2120                (se_cmd->prot_checks & TARGET_DIF_CHECK_REFTAG ? 0x0f : 0);
2121         return 0;
2122 }
2123
2124 static int
2125 isert_rdma_rw_ctx_post(struct isert_cmd *cmd, struct isert_conn *conn,
2126                 struct ib_cqe *cqe, struct ib_send_wr *chain_wr)
2127 {
2128         struct se_cmd *se_cmd = &cmd->iscsi_cmd->se_cmd;
2129         enum dma_data_direction dir = target_reverse_dma_direction(se_cmd);
2130         u8 port_num = conn->cm_id->port_num;
2131         u64 addr;
2132         u32 rkey, offset;
2133         int ret;
2134
2135         if (cmd->ctx_init_done)
2136                 goto rdma_ctx_post;
2137
2138         if (dir == DMA_FROM_DEVICE) {
2139                 addr = cmd->write_va;
2140                 rkey = cmd->write_stag;
2141                 offset = cmd->iscsi_cmd->write_data_done;
2142         } else {
2143                 addr = cmd->read_va;
2144                 rkey = cmd->read_stag;
2145                 offset = 0;
2146         }
2147
2148         if (isert_prot_cmd(conn, se_cmd)) {
2149                 struct ib_sig_attrs sig_attrs;
2150
2151                 ret = isert_set_sig_attrs(se_cmd, &sig_attrs);
2152                 if (ret)
2153                         return ret;
2154
2155                 WARN_ON_ONCE(offset);
2156                 ret = rdma_rw_ctx_signature_init(&cmd->rw, conn->qp, port_num,
2157                                 se_cmd->t_data_sg, se_cmd->t_data_nents,
2158                                 se_cmd->t_prot_sg, se_cmd->t_prot_nents,
2159                                 &sig_attrs, addr, rkey, dir);
2160         } else {
2161                 ret = rdma_rw_ctx_init(&cmd->rw, conn->qp, port_num,
2162                                 se_cmd->t_data_sg, se_cmd->t_data_nents,
2163                                 offset, addr, rkey, dir);
2164         }
2165
2166         if (ret < 0) {
2167                 isert_err("Cmd: %p failed to prepare RDMA res\n", cmd);
2168                 return ret;
2169         }
2170
2171         cmd->ctx_init_done = true;
2172
2173 rdma_ctx_post:
2174         ret = rdma_rw_ctx_post(&cmd->rw, conn->qp, port_num, cqe, chain_wr);
2175         if (ret < 0)
2176                 isert_err("Cmd: %p failed to post RDMA res\n", cmd);
2177         return ret;
2178 }
2179
2180 static int
2181 isert_put_datain(struct iscsi_conn *conn, struct iscsi_cmd *cmd)
2182 {
2183         struct se_cmd *se_cmd = &cmd->se_cmd;
2184         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2185         struct isert_conn *isert_conn = conn->context;
2186         struct ib_cqe *cqe = NULL;
2187         struct ib_send_wr *chain_wr = NULL;
2188         int rc;
2189
2190         isert_dbg("Cmd: %p RDMA_WRITE data_length: %u\n",
2191                  isert_cmd, se_cmd->data_length);
2192
2193         if (isert_prot_cmd(isert_conn, se_cmd)) {
2194                 isert_cmd->tx_desc.tx_cqe.done = isert_rdma_write_done;
2195                 cqe = &isert_cmd->tx_desc.tx_cqe;
2196         } else {
2197                 /*
2198                  * Build isert_conn->tx_desc for iSCSI response PDU and attach
2199                  */
2200                 isert_create_send_desc(isert_conn, isert_cmd,
2201                                        &isert_cmd->tx_desc);
2202                 iscsit_build_rsp_pdu(cmd, conn, true, (struct iscsi_scsi_rsp *)
2203                                      &isert_cmd->tx_desc.iscsi_header);
2204                 isert_init_tx_hdrs(isert_conn, &isert_cmd->tx_desc);
2205                 isert_init_send_wr(isert_conn, isert_cmd,
2206                                    &isert_cmd->tx_desc.send_wr);
2207
2208                 rc = isert_post_recv(isert_conn, isert_cmd->rx_desc);
2209                 if (rc) {
2210                         isert_err("ib_post_recv failed with %d\n", rc);
2211                         return rc;
2212                 }
2213
2214                 chain_wr = &isert_cmd->tx_desc.send_wr;
2215         }
2216
2217         rc = isert_rdma_rw_ctx_post(isert_cmd, isert_conn, cqe, chain_wr);
2218         isert_dbg("Cmd: %p posted RDMA_WRITE for iSER Data READ rc: %d\n",
2219                   isert_cmd, rc);
2220         return rc;
2221 }
2222
2223 static int
2224 isert_get_dataout(struct iscsi_conn *conn, struct iscsi_cmd *cmd, bool recovery)
2225 {
2226         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2227         int ret;
2228
2229         isert_dbg("Cmd: %p RDMA_READ data_length: %u write_data_done: %u\n",
2230                  isert_cmd, cmd->se_cmd.data_length, cmd->write_data_done);
2231
2232         isert_cmd->tx_desc.tx_cqe.done = isert_rdma_read_done;
2233         ret = isert_rdma_rw_ctx_post(isert_cmd, conn->context,
2234                                      &isert_cmd->tx_desc.tx_cqe, NULL);
2235
2236         isert_dbg("Cmd: %p posted RDMA_READ memory for ISER Data WRITE rc: %d\n",
2237                  isert_cmd, ret);
2238         return ret;
2239 }
2240
2241 static int
2242 isert_immediate_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2243 {
2244         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2245         int ret = 0;
2246
2247         switch (state) {
2248         case ISTATE_REMOVE:
2249                 spin_lock_bh(&conn->cmd_lock);
2250                 list_del_init(&cmd->i_conn_node);
2251                 spin_unlock_bh(&conn->cmd_lock);
2252                 isert_put_cmd(isert_cmd, true);
2253                 break;
2254         case ISTATE_SEND_NOPIN_WANT_RESPONSE:
2255                 ret = isert_put_nopin(cmd, conn, false);
2256                 break;
2257         default:
2258                 isert_err("Unknown immediate state: 0x%02x\n", state);
2259                 ret = -EINVAL;
2260                 break;
2261         }
2262
2263         return ret;
2264 }
2265
2266 static int
2267 isert_response_queue(struct iscsi_conn *conn, struct iscsi_cmd *cmd, int state)
2268 {
2269         struct isert_conn *isert_conn = conn->context;
2270         int ret;
2271
2272         switch (state) {
2273         case ISTATE_SEND_LOGOUTRSP:
2274                 ret = isert_put_logout_rsp(cmd, conn);
2275                 if (!ret)
2276                         isert_conn->logout_posted = true;
2277                 break;
2278         case ISTATE_SEND_NOPIN:
2279                 ret = isert_put_nopin(cmd, conn, true);
2280                 break;
2281         case ISTATE_SEND_TASKMGTRSP:
2282                 ret = isert_put_tm_rsp(cmd, conn);
2283                 break;
2284         case ISTATE_SEND_REJECT:
2285                 ret = isert_put_reject(cmd, conn);
2286                 break;
2287         case ISTATE_SEND_TEXTRSP:
2288                 ret = isert_put_text_rsp(cmd, conn);
2289                 break;
2290         case ISTATE_SEND_STATUS:
2291                 /*
2292                  * Special case for sending non GOOD SCSI status from TX thread
2293                  * context during pre se_cmd excecution failure.
2294                  */
2295                 ret = isert_put_response(conn, cmd);
2296                 break;
2297         default:
2298                 isert_err("Unknown response state: 0x%02x\n", state);
2299                 ret = -EINVAL;
2300                 break;
2301         }
2302
2303         return ret;
2304 }
2305
2306 struct rdma_cm_id *
2307 isert_setup_id(struct isert_np *isert_np)
2308 {
2309         struct iscsi_np *np = isert_np->np;
2310         struct rdma_cm_id *id;
2311         struct sockaddr *sa;
2312         int ret;
2313
2314         sa = (struct sockaddr *)&np->np_sockaddr;
2315         isert_dbg("ksockaddr: %p, sa: %p\n", &np->np_sockaddr, sa);
2316
2317         id = rdma_create_id(&init_net, isert_cma_handler, isert_np,
2318                             RDMA_PS_TCP, IB_QPT_RC);
2319         if (IS_ERR(id)) {
2320                 isert_err("rdma_create_id() failed: %ld\n", PTR_ERR(id));
2321                 ret = PTR_ERR(id);
2322                 goto out;
2323         }
2324         isert_dbg("id %p context %p\n", id, id->context);
2325
2326         ret = rdma_bind_addr(id, sa);
2327         if (ret) {
2328                 isert_err("rdma_bind_addr() failed: %d\n", ret);
2329                 goto out_id;
2330         }
2331
2332         ret = rdma_listen(id, 0);
2333         if (ret) {
2334                 isert_err("rdma_listen() failed: %d\n", ret);
2335                 goto out_id;
2336         }
2337
2338         return id;
2339 out_id:
2340         rdma_destroy_id(id);
2341 out:
2342         return ERR_PTR(ret);
2343 }
2344
2345 static int
2346 isert_setup_np(struct iscsi_np *np,
2347                struct sockaddr_storage *ksockaddr)
2348 {
2349         struct isert_np *isert_np;
2350         struct rdma_cm_id *isert_lid;
2351         int ret;
2352
2353         isert_np = kzalloc(sizeof(struct isert_np), GFP_KERNEL);
2354         if (!isert_np)
2355                 return -ENOMEM;
2356
2357         sema_init(&isert_np->sem, 0);
2358         mutex_init(&isert_np->mutex);
2359         INIT_LIST_HEAD(&isert_np->accepted);
2360         INIT_LIST_HEAD(&isert_np->pending);
2361         isert_np->np = np;
2362
2363         /*
2364          * Setup the np->np_sockaddr from the passed sockaddr setup
2365          * in iscsi_target_configfs.c code..
2366          */
2367         memcpy(&np->np_sockaddr, ksockaddr,
2368                sizeof(struct sockaddr_storage));
2369
2370         isert_lid = isert_setup_id(isert_np);
2371         if (IS_ERR(isert_lid)) {
2372                 ret = PTR_ERR(isert_lid);
2373                 goto out;
2374         }
2375
2376         isert_np->cm_id = isert_lid;
2377         np->np_context = isert_np;
2378
2379         return 0;
2380
2381 out:
2382         kfree(isert_np);
2383
2384         return ret;
2385 }
2386
2387 static int
2388 isert_rdma_accept(struct isert_conn *isert_conn)
2389 {
2390         struct rdma_cm_id *cm_id = isert_conn->cm_id;
2391         struct rdma_conn_param cp;
2392         int ret;
2393         struct iser_cm_hdr rsp_hdr;
2394
2395         memset(&cp, 0, sizeof(struct rdma_conn_param));
2396         cp.initiator_depth = isert_conn->initiator_depth;
2397         cp.retry_count = 7;
2398         cp.rnr_retry_count = 7;
2399
2400         memset(&rsp_hdr, 0, sizeof(rsp_hdr));
2401         rsp_hdr.flags = ISERT_ZBVA_NOT_USED;
2402         if (!isert_conn->snd_w_inv)
2403                 rsp_hdr.flags = rsp_hdr.flags | ISERT_SEND_W_INV_NOT_USED;
2404         cp.private_data = (void *)&rsp_hdr;
2405         cp.private_data_len = sizeof(rsp_hdr);
2406
2407         ret = rdma_accept(cm_id, &cp);
2408         if (ret) {
2409                 isert_err("rdma_accept() failed with: %d\n", ret);
2410                 return ret;
2411         }
2412
2413         return 0;
2414 }
2415
2416 static int
2417 isert_get_login_rx(struct iscsi_conn *conn, struct iscsi_login *login)
2418 {
2419         struct isert_conn *isert_conn = conn->context;
2420         int ret;
2421
2422         isert_info("before login_req comp conn: %p\n", isert_conn);
2423         ret = wait_for_completion_interruptible(&isert_conn->login_req_comp);
2424         if (ret) {
2425                 isert_err("isert_conn %p interrupted before got login req\n",
2426                           isert_conn);
2427                 return ret;
2428         }
2429         reinit_completion(&isert_conn->login_req_comp);
2430
2431         /*
2432          * For login requests after the first PDU, isert_rx_login_req() will
2433          * kick schedule_delayed_work(&conn->login_work) as the packet is
2434          * received, which turns this callback from iscsi_target_do_login_rx()
2435          * into a NOP.
2436          */
2437         if (!login->first_request)
2438                 return 0;
2439
2440         isert_rx_login_req(isert_conn);
2441
2442         isert_info("before login_comp conn: %p\n", conn);
2443         ret = wait_for_completion_interruptible(&isert_conn->login_comp);
2444         if (ret)
2445                 return ret;
2446
2447         isert_info("processing login->req: %p\n", login->req);
2448
2449         return 0;
2450 }
2451
2452 static void
2453 isert_set_conn_info(struct iscsi_np *np, struct iscsi_conn *conn,
2454                     struct isert_conn *isert_conn)
2455 {
2456         struct rdma_cm_id *cm_id = isert_conn->cm_id;
2457         struct rdma_route *cm_route = &cm_id->route;
2458
2459         conn->login_family = np->np_sockaddr.ss_family;
2460
2461         conn->login_sockaddr = cm_route->addr.dst_addr;
2462         conn->local_sockaddr = cm_route->addr.src_addr;
2463 }
2464
2465 static int
2466 isert_accept_np(struct iscsi_np *np, struct iscsi_conn *conn)
2467 {
2468         struct isert_np *isert_np = np->np_context;
2469         struct isert_conn *isert_conn;
2470         int ret;
2471
2472 accept_wait:
2473         ret = down_interruptible(&isert_np->sem);
2474         if (ret)
2475                 return -ENODEV;
2476
2477         spin_lock_bh(&np->np_thread_lock);
2478         if (np->np_thread_state >= ISCSI_NP_THREAD_RESET) {
2479                 spin_unlock_bh(&np->np_thread_lock);
2480                 isert_dbg("np_thread_state %d\n",
2481                          np->np_thread_state);
2482                 /**
2483                  * No point in stalling here when np_thread
2484                  * is in state RESET/SHUTDOWN/EXIT - bail
2485                  **/
2486                 return -ENODEV;
2487         }
2488         spin_unlock_bh(&np->np_thread_lock);
2489
2490         mutex_lock(&isert_np->mutex);
2491         if (list_empty(&isert_np->pending)) {
2492                 mutex_unlock(&isert_np->mutex);
2493                 goto accept_wait;
2494         }
2495         isert_conn = list_first_entry(&isert_np->pending,
2496                         struct isert_conn, node);
2497         list_del_init(&isert_conn->node);
2498         mutex_unlock(&isert_np->mutex);
2499
2500         conn->context = isert_conn;
2501         isert_conn->conn = conn;
2502         isert_conn->state = ISER_CONN_BOUND;
2503
2504         isert_set_conn_info(np, conn, isert_conn);
2505
2506         isert_dbg("Processing isert_conn: %p\n", isert_conn);
2507
2508         return 0;
2509 }
2510
2511 static void
2512 isert_free_np(struct iscsi_np *np)
2513 {
2514         struct isert_np *isert_np = np->np_context;
2515         struct isert_conn *isert_conn, *n;
2516
2517         if (isert_np->cm_id)
2518                 rdma_destroy_id(isert_np->cm_id);
2519
2520         /*
2521          * FIXME: At this point we don't have a good way to insure
2522          * that at this point we don't have hanging connections that
2523          * completed RDMA establishment but didn't start iscsi login
2524          * process. So work-around this by cleaning up what ever piled
2525          * up in accepted and pending lists.
2526          */
2527         mutex_lock(&isert_np->mutex);
2528         if (!list_empty(&isert_np->pending)) {
2529                 isert_info("Still have isert pending connections\n");
2530                 list_for_each_entry_safe(isert_conn, n,
2531                                          &isert_np->pending,
2532                                          node) {
2533                         isert_info("cleaning isert_conn %p state (%d)\n",
2534                                    isert_conn, isert_conn->state);
2535                         isert_connect_release(isert_conn);
2536                 }
2537         }
2538
2539         if (!list_empty(&isert_np->accepted)) {
2540                 isert_info("Still have isert accepted connections\n");
2541                 list_for_each_entry_safe(isert_conn, n,
2542                                          &isert_np->accepted,
2543                                          node) {
2544                         isert_info("cleaning isert_conn %p state (%d)\n",
2545                                    isert_conn, isert_conn->state);
2546                         isert_connect_release(isert_conn);
2547                 }
2548         }
2549         mutex_unlock(&isert_np->mutex);
2550
2551         np->np_context = NULL;
2552         kfree(isert_np);
2553 }
2554
2555 static void isert_release_work(struct work_struct *work)
2556 {
2557         struct isert_conn *isert_conn = container_of(work,
2558                                                      struct isert_conn,
2559                                                      release_work);
2560
2561         isert_info("Starting release conn %p\n", isert_conn);
2562
2563         mutex_lock(&isert_conn->mutex);
2564         isert_conn->state = ISER_CONN_DOWN;
2565         mutex_unlock(&isert_conn->mutex);
2566
2567         isert_info("Destroying conn %p\n", isert_conn);
2568         isert_put_conn(isert_conn);
2569 }
2570
2571 static void
2572 isert_wait4logout(struct isert_conn *isert_conn)
2573 {
2574         struct iscsi_conn *conn = isert_conn->conn;
2575
2576         isert_info("conn %p\n", isert_conn);
2577
2578         if (isert_conn->logout_posted) {
2579                 isert_info("conn %p wait for conn_logout_comp\n", isert_conn);
2580                 wait_for_completion_timeout(&conn->conn_logout_comp,
2581                                             SECONDS_FOR_LOGOUT_COMP * HZ);
2582         }
2583 }
2584
2585 static void
2586 isert_wait4cmds(struct iscsi_conn *conn)
2587 {
2588         isert_info("iscsi_conn %p\n", conn);
2589
2590         if (conn->sess) {
2591                 target_sess_cmd_list_set_waiting(conn->sess->se_sess);
2592                 target_wait_for_sess_cmds(conn->sess->se_sess);
2593         }
2594 }
2595
2596 /**
2597  * isert_put_unsol_pending_cmds() - Drop commands waiting for
2598  *     unsolicitate dataout
2599  * @conn:    iscsi connection
2600  *
2601  * We might still have commands that are waiting for unsolicited
2602  * dataouts messages. We must put the extra reference on those
2603  * before blocking on the target_wait_for_session_cmds
2604  */
2605 static void
2606 isert_put_unsol_pending_cmds(struct iscsi_conn *conn)
2607 {
2608         struct iscsi_cmd *cmd, *tmp;
2609         static LIST_HEAD(drop_cmd_list);
2610
2611         spin_lock_bh(&conn->cmd_lock);
2612         list_for_each_entry_safe(cmd, tmp, &conn->conn_cmd_list, i_conn_node) {
2613                 if ((cmd->cmd_flags & ICF_NON_IMMEDIATE_UNSOLICITED_DATA) &&
2614                     (cmd->write_data_done < conn->sess->sess_ops->FirstBurstLength) &&
2615                     (cmd->write_data_done < cmd->se_cmd.data_length))
2616                         list_move_tail(&cmd->i_conn_node, &drop_cmd_list);
2617         }
2618         spin_unlock_bh(&conn->cmd_lock);
2619
2620         list_for_each_entry_safe(cmd, tmp, &drop_cmd_list, i_conn_node) {
2621                 list_del_init(&cmd->i_conn_node);
2622                 if (cmd->i_state != ISTATE_REMOVE) {
2623                         struct isert_cmd *isert_cmd = iscsit_priv_cmd(cmd);
2624
2625                         isert_info("conn %p dropping cmd %p\n", conn, cmd);
2626                         isert_put_cmd(isert_cmd, true);
2627                 }
2628         }
2629 }
2630
2631 static void isert_wait_conn(struct iscsi_conn *conn)
2632 {
2633         struct isert_conn *isert_conn = conn->context;
2634
2635         isert_info("Starting conn %p\n", isert_conn);
2636
2637         mutex_lock(&isert_conn->mutex);
2638         isert_conn_terminate(isert_conn);
2639         mutex_unlock(&isert_conn->mutex);
2640
2641         ib_drain_qp(isert_conn->qp);
2642         isert_put_unsol_pending_cmds(conn);
2643         isert_wait4cmds(conn);
2644         isert_wait4logout(isert_conn);
2645
2646         queue_work(isert_release_wq, &isert_conn->release_work);
2647 }
2648
2649 static void isert_free_conn(struct iscsi_conn *conn)
2650 {
2651         struct isert_conn *isert_conn = conn->context;
2652
2653         ib_drain_qp(isert_conn->qp);
2654         isert_put_conn(isert_conn);
2655 }
2656
2657 static void isert_get_rx_pdu(struct iscsi_conn *conn)
2658 {
2659         struct completion comp;
2660
2661         init_completion(&comp);
2662
2663         wait_for_completion_interruptible(&comp);
2664 }
2665
2666 static struct iscsit_transport iser_target_transport = {
2667         .name                   = "IB/iSER",
2668         .transport_type         = ISCSI_INFINIBAND,
2669         .rdma_shutdown          = true,
2670         .priv_size              = sizeof(struct isert_cmd),
2671         .owner                  = THIS_MODULE,
2672         .iscsit_setup_np        = isert_setup_np,
2673         .iscsit_accept_np       = isert_accept_np,
2674         .iscsit_free_np         = isert_free_np,
2675         .iscsit_wait_conn       = isert_wait_conn,
2676         .iscsit_free_conn       = isert_free_conn,
2677         .iscsit_get_login_rx    = isert_get_login_rx,
2678         .iscsit_put_login_tx    = isert_put_login_tx,
2679         .iscsit_immediate_queue = isert_immediate_queue,
2680         .iscsit_response_queue  = isert_response_queue,
2681         .iscsit_get_dataout     = isert_get_dataout,
2682         .iscsit_queue_data_in   = isert_put_datain,
2683         .iscsit_queue_status    = isert_put_response,
2684         .iscsit_aborted_task    = isert_aborted_task,
2685         .iscsit_get_rx_pdu      = isert_get_rx_pdu,
2686         .iscsit_get_sup_prot_ops = isert_get_sup_prot_ops,
2687 };
2688
2689 static int __init isert_init(void)
2690 {
2691         int ret;
2692
2693         isert_comp_wq = alloc_workqueue("isert_comp_wq",
2694                                         WQ_UNBOUND | WQ_HIGHPRI, 0);
2695         if (!isert_comp_wq) {
2696                 isert_err("Unable to allocate isert_comp_wq\n");
2697                 return -ENOMEM;
2698         }
2699
2700         isert_release_wq = alloc_workqueue("isert_release_wq", WQ_UNBOUND,
2701                                         WQ_UNBOUND_MAX_ACTIVE);
2702         if (!isert_release_wq) {
2703                 isert_err("Unable to allocate isert_release_wq\n");
2704                 ret = -ENOMEM;
2705                 goto destroy_comp_wq;
2706         }
2707
2708         iscsit_register_transport(&iser_target_transport);
2709         isert_info("iSER_TARGET[0] - Loaded iser_target_transport\n");
2710
2711         return 0;
2712
2713 destroy_comp_wq:
2714         destroy_workqueue(isert_comp_wq);
2715
2716         return ret;
2717 }
2718
2719 static void __exit isert_exit(void)
2720 {
2721         flush_scheduled_work();
2722         destroy_workqueue(isert_release_wq);
2723         destroy_workqueue(isert_comp_wq);
2724         iscsit_unregister_transport(&iser_target_transport);
2725         isert_info("iSER_TARGET[0] - Released iser_target_transport\n");
2726 }
2727
2728 MODULE_DESCRIPTION("iSER-Target for mainline target infrastructure");
2729 MODULE_AUTHOR("nab@Linux-iSCSI.org");
2730 MODULE_LICENSE("GPL");
2731
2732 module_init(isert_init);
2733 module_exit(isert_exit);