drivers/infiniband/ulp/iser/iser_verbs.c

   1 /*
   2  * Copyright (c) 2004, 2005, 2006 Voltaire, Inc. All rights reserved.
   3  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
   4  * Copyright (c) 2013-2014 Mellanox Technologies. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34 #include <linux/kernel.h>
  35 #include <linux/module.h>
  36 #include <linux/slab.h>
  37 #include <linux/delay.h>
  38
  39 #include "iscsi_iser.h"
  40
  41 #define ISCSI_ISER_MAX_CONN     8
  42 #define ISER_MAX_RX_LEN         (ISER_QP_MAX_RECV_DTOS * ISCSI_ISER_MAX_CONN)
  43 #define ISER_MAX_TX_LEN         (ISER_QP_MAX_REQ_DTOS  * ISCSI_ISER_MAX_CONN)
  44 #define ISER_MAX_CQ_LEN         (ISER_MAX_RX_LEN + ISER_MAX_TX_LEN + \
  45                                  ISCSI_ISER_MAX_CONN)
  46
  47 static void iser_qp_event_callback(struct ib_event *cause, void *context)
  48 {
  49         iser_err("qp event %s (%d)\n",
  50                  ib_event_msg(cause->event), cause->event);
  51 }
  52
  53 static void iser_event_handler(struct ib_event_handler *handler,
  54                                 struct ib_event *event)
  55 {
  56         iser_err("async event %s (%d) on device %s port %d\n",
  57                  ib_event_msg(event->event), event->event,
  58                 dev_name(&event->device->dev), event->element.port_num);
  59 }
  60
  61 /**
  62  * iser_create_device_ib_res - creates Protection Domain (PD), Completion
  63  * Queue (CQ), DMA Memory Region (DMA MR) with the device associated with
  64  * the adapator.
  65  *
  66  * returns 0 on success, -1 on failure
  67  */
  68 static int iser_create_device_ib_res(struct iser_device *device)
  69 {
  70         struct ib_device *ib_dev = device->ib_device;
  71         int ret, i, max_cqe;
  72
  73         ret = iser_assign_reg_ops(device);
  74         if (ret)
  75                 return ret;
  76
  77         device->comps_used = min_t(int, num_online_cpus(),
  78                                  ib_dev->num_comp_vectors);
  79
  80         device->comps = kcalloc(device->comps_used, sizeof(*device->comps),
  81                                 GFP_KERNEL);
  82         if (!device->comps)
  83                 goto comps_err;
  84
  85         max_cqe = min(ISER_MAX_CQ_LEN, ib_dev->attrs.max_cqe);
  86
  87         iser_info("using %d CQs, device %s supports %d vectors max_cqe %d\n",
  88                   device->comps_used, dev_name(&ib_dev->dev),
  89                   ib_dev->num_comp_vectors, max_cqe);
  90
  91         device->pd = ib_alloc_pd(ib_dev,
  92                 iser_always_reg ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
  93         if (IS_ERR(device->pd))
  94                 goto pd_err;
  95
  96         for (i = 0; i < device->comps_used; i++) {
  97                 struct iser_comp *comp = &device->comps[i];
  98
  99                 comp->cq = ib_alloc_cq(ib_dev, comp, max_cqe, i,
 100                                        IB_POLL_SOFTIRQ);
 101                 if (IS_ERR(comp->cq)) {
 102                         comp->cq = NULL;
 103                         goto cq_err;
 104                 }
 105         }
 106
 107         INIT_IB_EVENT_HANDLER(&device->event_handler, ib_dev,
 108                               iser_event_handler);
 109         ib_register_event_handler(&device->event_handler);
 110         return 0;
 111
 112 cq_err:
 113         for (i = 0; i < device->comps_used; i++) {
 114                 struct iser_comp *comp = &device->comps[i];
 115
 116                 if (comp->cq)
 117                         ib_free_cq(comp->cq);
 118         }
 119         ib_dealloc_pd(device->pd);
 120 pd_err:
 121         kfree(device->comps);
 122 comps_err:
 123         iser_err("failed to allocate an IB resource\n");
 124         return -1;
 125 }
 126
 127 /**
 128  * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
 129  * CQ and PD created with the device associated with the adapator.
 130  */
 131 static void iser_free_device_ib_res(struct iser_device *device)
 132 {
 133         int i;
 134
 135         for (i = 0; i < device->comps_used; i++) {
 136                 struct iser_comp *comp = &device->comps[i];
 137
 138                 ib_free_cq(comp->cq);
 139                 comp->cq = NULL;
 140         }
 141
 142         ib_unregister_event_handler(&device->event_handler);
 143         ib_dealloc_pd(device->pd);
 144
 145         kfree(device->comps);
 146         device->comps = NULL;
 147         device->pd = NULL;
 148 }
 149
 150 /**
 151  * iser_alloc_fmr_pool - Creates FMR pool and page_vector
 152  *
 153  * returns 0 on success, or errno code on failure
 154  */
 155 int iser_alloc_fmr_pool(struct ib_conn *ib_conn,
 156                         unsigned cmds_max,
 157                         unsigned int size)
 158 {
 159         struct iser_device *device = ib_conn->device;
 160         struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
 161         struct iser_page_vec *page_vec;
 162         struct iser_fr_desc *desc;
 163         struct ib_fmr_pool *fmr_pool;
 164         struct ib_fmr_pool_param params;
 165         int ret;
 166
 167         INIT_LIST_HEAD(&fr_pool->list);
 168         spin_lock_init(&fr_pool->lock);
 169
 170         desc = kzalloc(sizeof(*desc), GFP_KERNEL);
 171         if (!desc)
 172                 return -ENOMEM;
 173
 174         page_vec = kmalloc(sizeof(*page_vec) + (sizeof(u64) * size),
 175                            GFP_KERNEL);
 176         if (!page_vec) {
 177                 ret = -ENOMEM;
 178                 goto err_frpl;
 179         }
 180
 181         page_vec->pages = (u64 *)(page_vec + 1);
 182
 183         params.page_shift        = SHIFT_4K;
 184         params.max_pages_per_fmr = size;
 185         /* make the pool size twice the max number of SCSI commands *
 186          * the ML is expected to queue, watermark for unmap at 50%  */
 187         params.pool_size         = cmds_max * 2;
 188         params.dirty_watermark   = cmds_max;
 189         params.cache             = 0;
 190         params.flush_function    = NULL;
 191         params.access            = (IB_ACCESS_LOCAL_WRITE  |
 192                                     IB_ACCESS_REMOTE_WRITE |
 193                                     IB_ACCESS_REMOTE_READ);
 194
 195         fmr_pool = ib_create_fmr_pool(device->pd, &params);
 196         if (IS_ERR(fmr_pool)) {
 197                 ret = PTR_ERR(fmr_pool);
 198                 iser_err("FMR allocation failed, err %d\n", ret);
 199                 goto err_fmr;
 200         }
 201
 202         desc->rsc.page_vec = page_vec;
 203         desc->rsc.fmr_pool = fmr_pool;
 204         list_add(&desc->list, &fr_pool->list);
 205
 206         return 0;
 207
 208 err_fmr:
 209         kfree(page_vec);
 210 err_frpl:
 211         kfree(desc);
 212
 213         return ret;
 214 }
 215
 216 /**
 217  * iser_free_fmr_pool - releases the FMR pool and page vec
 218  */
 219 void iser_free_fmr_pool(struct ib_conn *ib_conn)
 220 {
 221         struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
 222         struct iser_fr_desc *desc;
 223
 224         desc = list_first_entry(&fr_pool->list,
 225                                 struct iser_fr_desc, list);
 226         list_del(&desc->list);
 227
 228         iser_info("freeing conn %p fmr pool %p\n",
 229                   ib_conn, desc->rsc.fmr_pool);
 230
 231         ib_destroy_fmr_pool(desc->rsc.fmr_pool);
 232         kfree(desc->rsc.page_vec);
 233         kfree(desc);
 234 }
 235
 236 static struct iser_fr_desc *
 237 iser_create_fastreg_desc(struct iser_device *device,
 238                          struct ib_pd *pd,
 239                          bool pi_enable,
 240                          unsigned int size)
 241 {
 242         struct iser_fr_desc *desc;
 243         struct ib_device *ib_dev = device->ib_device;
 244         enum ib_mr_type mr_type;
 245         int ret;
 246
 247         desc = kzalloc(sizeof(*desc), GFP_KERNEL);
 248         if (!desc)
 249                 return ERR_PTR(-ENOMEM);
 250
 251         if (ib_dev->attrs.device_cap_flags & IB_DEVICE_SG_GAPS_REG)
 252                 mr_type = IB_MR_TYPE_SG_GAPS;
 253         else
 254                 mr_type = IB_MR_TYPE_MEM_REG;
 255
 256         desc->rsc.mr = ib_alloc_mr(pd, mr_type, size);
 257         if (IS_ERR(desc->rsc.mr)) {
 258                 ret = PTR_ERR(desc->rsc.mr);
 259                 iser_err("Failed to allocate ib_fast_reg_mr err=%d\n", ret);
 260                 goto err_alloc_mr;
 261         }
 262
 263         if (pi_enable) {
 264                 desc->rsc.sig_mr = ib_alloc_mr_integrity(pd, size, size);
 265                 if (IS_ERR(desc->rsc.sig_mr)) {
 266                         ret = PTR_ERR(desc->rsc.sig_mr);
 267                         iser_err("Failed to allocate sig_mr err=%d\n", ret);
 268                         goto err_alloc_mr_integrity;
 269                 }
 270         }
 271         desc->rsc.mr_valid = 0;
 272
 273         return desc;
 274
 275 err_alloc_mr_integrity:
 276         ib_dereg_mr(desc->rsc.mr);
 277 err_alloc_mr:
 278         kfree(desc);
 279
 280         return ERR_PTR(ret);
 281 }
 282
 283 static void iser_destroy_fastreg_desc(struct iser_fr_desc *desc)
 284 {
 285         struct iser_reg_resources *res = &desc->rsc;
 286
 287         ib_dereg_mr(res->mr);
 288         if (res->sig_mr) {
 289                 ib_dereg_mr(res->sig_mr);
 290                 res->sig_mr = NULL;
 291         }
 292         kfree(desc);
 293 }
 294
 295 /**
 296  * iser_alloc_fastreg_pool - Creates pool of fast_reg descriptors
 297  * for fast registration work requests.
 298  * returns 0 on success, or errno code on failure
 299  */
 300 int iser_alloc_fastreg_pool(struct ib_conn *ib_conn,
 301                             unsigned cmds_max,
 302                             unsigned int size)
 303 {
 304         struct iser_device *device = ib_conn->device;
 305         struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
 306         struct iser_fr_desc *desc;
 307         int i, ret;
 308
 309         INIT_LIST_HEAD(&fr_pool->list);
 310         INIT_LIST_HEAD(&fr_pool->all_list);
 311         spin_lock_init(&fr_pool->lock);
 312         fr_pool->size = 0;
 313         for (i = 0; i < cmds_max; i++) {
 314                 desc = iser_create_fastreg_desc(device, device->pd,
 315                                                 ib_conn->pi_support, size);
 316                 if (IS_ERR(desc)) {
 317                         ret = PTR_ERR(desc);
 318                         goto err;
 319                 }
 320
 321                 list_add_tail(&desc->list, &fr_pool->list);
 322                 list_add_tail(&desc->all_list, &fr_pool->all_list);
 323                 fr_pool->size++;
 324         }
 325
 326         return 0;
 327
 328 err:
 329         iser_free_fastreg_pool(ib_conn);
 330         return ret;
 331 }
 332
 333 /**
 334  * iser_free_fastreg_pool - releases the pool of fast_reg descriptors
 335  */
 336 void iser_free_fastreg_pool(struct ib_conn *ib_conn)
 337 {
 338         struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
 339         struct iser_fr_desc *desc, *tmp;
 340         int i = 0;
 341
 342         if (list_empty(&fr_pool->all_list))
 343                 return;
 344
 345         iser_info("freeing conn %p fr pool\n", ib_conn);
 346
 347         list_for_each_entry_safe(desc, tmp, &fr_pool->all_list, all_list) {
 348                 list_del(&desc->all_list);
 349                 iser_destroy_fastreg_desc(desc);
 350                 ++i;
 351         }
 352
 353         if (i < fr_pool->size)
 354                 iser_warn("pool still has %d regions registered\n",
 355                           fr_pool->size - i);
 356 }
 357
 358 /**
 359  * iser_create_ib_conn_res - Queue-Pair (QP)
 360  *
 361  * returns 0 on success, -1 on failure
 362  */
 363 static int iser_create_ib_conn_res(struct ib_conn *ib_conn)
 364 {
 365         struct iser_conn *iser_conn = to_iser_conn(ib_conn);
 366         struct iser_device      *device;
 367         struct ib_device        *ib_dev;
 368         struct ib_qp_init_attr  init_attr;
 369         int                     ret = -ENOMEM;
 370         int index, min_index = 0;
 371
 372         BUG_ON(ib_conn->device == NULL);
 373
 374         device = ib_conn->device;
 375         ib_dev = device->ib_device;
 376
 377         memset(&init_attr, 0, sizeof init_attr);
 378
 379         mutex_lock(&ig.connlist_mutex);
 380         /* select the CQ with the minimal number of usages */
 381         for (index = 0; index < device->comps_used; index++) {
 382                 if (device->comps[index].active_qps <
 383                     device->comps[min_index].active_qps)
 384                         min_index = index;
 385         }
 386         ib_conn->comp = &device->comps[min_index];
 387         ib_conn->comp->active_qps++;
 388         mutex_unlock(&ig.connlist_mutex);
 389         iser_info("cq index %d used for ib_conn %p\n", min_index, ib_conn);
 390
 391         init_attr.event_handler = iser_qp_event_callback;
 392         init_attr.qp_context    = (void *)ib_conn;
 393         init_attr.send_cq       = ib_conn->comp->cq;
 394         init_attr.recv_cq       = ib_conn->comp->cq;
 395         init_attr.cap.max_recv_wr  = ISER_QP_MAX_RECV_DTOS;
 396         init_attr.cap.max_send_sge = 2;
 397         init_attr.cap.max_recv_sge = 1;
 398         init_attr.sq_sig_type   = IB_SIGNAL_REQ_WR;
 399         init_attr.qp_type       = IB_QPT_RC;
 400         if (ib_conn->pi_support) {
 401                 init_attr.cap.max_send_wr = ISER_QP_SIG_MAX_REQ_DTOS + 1;
 402                 init_attr.create_flags |= IB_QP_CREATE_INTEGRITY_EN;
 403                 iser_conn->max_cmds =
 404                         ISER_GET_MAX_XMIT_CMDS(ISER_QP_SIG_MAX_REQ_DTOS);
 405         } else {
 406                 if (ib_dev->attrs.max_qp_wr > ISER_QP_MAX_REQ_DTOS) {
 407                         init_attr.cap.max_send_wr  = ISER_QP_MAX_REQ_DTOS + 1;
 408                         iser_conn->max_cmds =
 409                                 ISER_GET_MAX_XMIT_CMDS(ISER_QP_MAX_REQ_DTOS);
 410                 } else {
 411                         init_attr.cap.max_send_wr = ib_dev->attrs.max_qp_wr;
 412                         iser_conn->max_cmds =
 413                                 ISER_GET_MAX_XMIT_CMDS(ib_dev->attrs.max_qp_wr);
 414                         iser_dbg("device %s supports max_send_wr %d\n",
 415                                  dev_name(&device->ib_device->dev),
 416                                  ib_dev->attrs.max_qp_wr);
 417                 }
 418         }
 419
 420         ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
 421         if (ret)
 422                 goto out_err;
 423
 424         ib_conn->qp = ib_conn->cma_id->qp;
 425         iser_info("setting conn %p cma_id %p qp %p\n",
 426                   ib_conn, ib_conn->cma_id,
 427                   ib_conn->cma_id->qp);
 428         return ret;
 429
 430 out_err:
 431         mutex_lock(&ig.connlist_mutex);
 432         ib_conn->comp->active_qps--;
 433         mutex_unlock(&ig.connlist_mutex);
 434         iser_err("unable to alloc mem or create resource, err %d\n", ret);
 435
 436         return ret;
 437 }
 438
 439 /**
 440  * based on the resolved device node GUID see if there already allocated
 441  * device for this device. If there's no such, create one.
 442  */
 443 static
 444 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
 445 {
 446         struct iser_device *device;
 447
 448         mutex_lock(&ig.device_list_mutex);
 449
 450         list_for_each_entry(device, &ig.device_list, ig_list)
 451                 /* find if there's a match using the node GUID */
 452                 if (device->ib_device->node_guid == cma_id->device->node_guid)
 453                         goto inc_refcnt;
 454
 455         device = kzalloc(sizeof *device, GFP_KERNEL);
 456         if (device == NULL)
 457                 goto out;
 458
 459         /* assign this device to the device */
 460         device->ib_device = cma_id->device;
 461         /* init the device and link it into ig device list */
 462         if (iser_create_device_ib_res(device)) {
 463                 kfree(device);
 464                 device = NULL;
 465                 goto out;
 466         }
 467         list_add(&device->ig_list, &ig.device_list);
 468
 469 inc_refcnt:
 470         device->refcount++;
 471 out:
 472         mutex_unlock(&ig.device_list_mutex);
 473         return device;
 474 }
 475
 476 /* if there's no demand for this device, release it */
 477 static void iser_device_try_release(struct iser_device *device)
 478 {
 479         mutex_lock(&ig.device_list_mutex);
 480         device->refcount--;
 481         iser_info("device %p refcount %d\n", device, device->refcount);
 482         if (!device->refcount) {
 483                 iser_free_device_ib_res(device);
 484                 list_del(&device->ig_list);
 485                 kfree(device);
 486         }
 487         mutex_unlock(&ig.device_list_mutex);
 488 }
 489
 490 /**
 491  * Called with state mutex held
 492  **/
 493 static int iser_conn_state_comp_exch(struct iser_conn *iser_conn,
 494                                      enum iser_conn_state comp,
 495                                      enum iser_conn_state exch)
 496 {
 497         int ret;
 498
 499         ret = (iser_conn->state == comp);
 500         if (ret)
 501                 iser_conn->state = exch;
 502
 503         return ret;
 504 }
 505
 506 void iser_release_work(struct work_struct *work)
 507 {
 508         struct iser_conn *iser_conn;
 509
 510         iser_conn = container_of(work, struct iser_conn, release_work);
 511
 512         /* Wait for conn_stop to complete */
 513         wait_for_completion(&iser_conn->stop_completion);
 514         /* Wait for IB resouces cleanup to complete */
 515         wait_for_completion(&iser_conn->ib_completion);
 516
 517         mutex_lock(&iser_conn->state_mutex);
 518         iser_conn->state = ISER_CONN_DOWN;
 519         mutex_unlock(&iser_conn->state_mutex);
 520
 521         iser_conn_release(iser_conn);
 522 }
 523
 524 /**
 525  * iser_free_ib_conn_res - release IB related resources
 526  * @iser_conn: iser connection struct
 527  * @destroy: indicator if we need to try to release the
 528  *     iser device and memory regoins pool (only iscsi
 529  *     shutdown and DEVICE_REMOVAL will use this).
 530  *
 531  * This routine is called with the iser state mutex held
 532  * so the cm_id removal is out of here. It is Safe to
 533  * be invoked multiple times.
 534  */
 535 static void iser_free_ib_conn_res(struct iser_conn *iser_conn,
 536                                   bool destroy)
 537 {
 538         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 539         struct iser_device *device = ib_conn->device;
 540
 541         iser_info("freeing conn %p cma_id %p qp %p\n",
 542                   iser_conn, ib_conn->cma_id, ib_conn->qp);
 543
 544         if (ib_conn->qp != NULL) {
 545                 mutex_lock(&ig.connlist_mutex);
 546                 ib_conn->comp->active_qps--;
 547                 mutex_unlock(&ig.connlist_mutex);
 548                 rdma_destroy_qp(ib_conn->cma_id);
 549                 ib_conn->qp = NULL;
 550         }
 551
 552         if (destroy) {
 553                 if (iser_conn->rx_descs)
 554                         iser_free_rx_descriptors(iser_conn);
 555
 556                 if (device != NULL) {
 557                         iser_device_try_release(device);
 558                         ib_conn->device = NULL;
 559                 }
 560         }
 561 }
 562
 563 /**
 564  * Frees all conn objects and deallocs conn descriptor
 565  */
 566 void iser_conn_release(struct iser_conn *iser_conn)
 567 {
 568         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 569
 570         mutex_lock(&ig.connlist_mutex);
 571         list_del(&iser_conn->conn_list);
 572         mutex_unlock(&ig.connlist_mutex);
 573
 574         mutex_lock(&iser_conn->state_mutex);
 575         /* In case we endup here without ep_disconnect being invoked. */
 576         if (iser_conn->state != ISER_CONN_DOWN) {
 577                 iser_warn("iser conn %p state %d, expected state down.\n",
 578                           iser_conn, iser_conn->state);
 579                 iscsi_destroy_endpoint(iser_conn->ep);
 580                 iser_conn->state = ISER_CONN_DOWN;
 581         }
 582         /*
 583          * In case we never got to bind stage, we still need to
 584          * release IB resources (which is safe to call more than once).
 585          */
 586         iser_free_ib_conn_res(iser_conn, true);
 587         mutex_unlock(&iser_conn->state_mutex);
 588
 589         if (ib_conn->cma_id != NULL) {
 590                 rdma_destroy_id(ib_conn->cma_id);
 591                 ib_conn->cma_id = NULL;
 592         }
 593
 594         kfree(iser_conn);
 595 }
 596
 597 /**
 598  * triggers start of the disconnect procedures and wait for them to be done
 599  * Called with state mutex held
 600  */
 601 int iser_conn_terminate(struct iser_conn *iser_conn)
 602 {
 603         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 604         int err = 0;
 605
 606         /* terminate the iser conn only if the conn state is UP */
 607         if (!iser_conn_state_comp_exch(iser_conn, ISER_CONN_UP,
 608                                        ISER_CONN_TERMINATING))
 609                 return 0;
 610
 611         iser_info("iser_conn %p state %d\n", iser_conn, iser_conn->state);
 612
 613         /* suspend queuing of new iscsi commands */
 614         if (iser_conn->iscsi_conn)
 615                 iscsi_suspend_queue(iser_conn->iscsi_conn);
 616
 617         /*
 618          * In case we didn't already clean up the cma_id (peer initiated
 619          * a disconnection), we need to Cause the CMA to change the QP
 620          * state to ERROR.
 621          */
 622         if (ib_conn->cma_id) {
 623                 err = rdma_disconnect(ib_conn->cma_id);
 624                 if (err)
 625                         iser_err("Failed to disconnect, conn: 0x%p err %d\n",
 626                                  iser_conn, err);
 627
 628                 /* block until all flush errors are consumed */
 629                 ib_drain_sq(ib_conn->qp);
 630         }
 631
 632         return 1;
 633 }
 634
 635 /**
 636  * Called with state mutex held
 637  **/
 638 static void iser_connect_error(struct rdma_cm_id *cma_id)
 639 {
 640         struct iser_conn *iser_conn;
 641
 642         iser_conn = (struct iser_conn *)cma_id->context;
 643         iser_conn->state = ISER_CONN_TERMINATING;
 644 }
 645
 646 static void
 647 iser_calc_scsi_params(struct iser_conn *iser_conn,
 648                       unsigned int max_sectors)
 649 {
 650         struct iser_device *device = iser_conn->ib_conn.device;
 651         struct ib_device_attr *attr = &device->ib_device->attrs;
 652         unsigned short sg_tablesize, sup_sg_tablesize;
 653         unsigned short reserved_mr_pages;
 654         u32 max_num_sg;
 655
 656         /*
 657          * FRs without SG_GAPS or FMRs can only map up to a (device) page per
 658          * entry, but if the first entry is misaligned we'll end up using two
 659          * entries (head and tail) for a single page worth data, so one
 660          * additional entry is required.
 661          */
 662         if ((attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) &&
 663             (attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG))
 664                 reserved_mr_pages = 0;
 665         else
 666                 reserved_mr_pages = 1;
 667
 668         if (iser_conn->ib_conn.pi_support)
 669                 max_num_sg = attr->max_pi_fast_reg_page_list_len;
 670         else
 671                 max_num_sg = attr->max_fast_reg_page_list_len;
 672
 673         sg_tablesize = DIV_ROUND_UP(max_sectors * 512, SIZE_4K);
 674         if (attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS)
 675                 sup_sg_tablesize =
 676                         min_t(
 677                          uint, ISCSI_ISER_MAX_SG_TABLESIZE,
 678                          max_num_sg - reserved_mr_pages);
 679         else
 680                 sup_sg_tablesize = ISCSI_ISER_MAX_SG_TABLESIZE;
 681
 682         iser_conn->scsi_sg_tablesize = min(sg_tablesize, sup_sg_tablesize);
 683         iser_conn->pages_per_mr =
 684                 iser_conn->scsi_sg_tablesize + reserved_mr_pages;
 685 }
 686
 687 /**
 688  * Called with state mutex held
 689  **/
 690 static void iser_addr_handler(struct rdma_cm_id *cma_id)
 691 {
 692         struct iser_device *device;
 693         struct iser_conn   *iser_conn;
 694         struct ib_conn   *ib_conn;
 695         int    ret;
 696
 697         iser_conn = (struct iser_conn *)cma_id->context;
 698         if (iser_conn->state != ISER_CONN_PENDING)
 699                 /* bailout */
 700                 return;
 701
 702         ib_conn = &iser_conn->ib_conn;
 703         device = iser_device_find_by_ib_device(cma_id);
 704         if (!device) {
 705                 iser_err("device lookup/creation failed\n");
 706                 iser_connect_error(cma_id);
 707                 return;
 708         }
 709
 710         ib_conn->device = device;
 711
 712         /* connection T10-PI support */
 713         if (iser_pi_enable) {
 714                 if (!(device->ib_device->attrs.device_cap_flags &
 715                       IB_DEVICE_INTEGRITY_HANDOVER)) {
 716                         iser_warn("T10-PI requested but not supported on %s, "
 717                                   "continue without T10-PI\n",
 718                                   dev_name(&ib_conn->device->ib_device->dev));
 719                         ib_conn->pi_support = false;
 720                 } else {
 721                         ib_conn->pi_support = true;
 722                 }
 723         }
 724
 725         iser_calc_scsi_params(iser_conn, iser_max_sectors);
 726
 727         ret = rdma_resolve_route(cma_id, 1000);
 728         if (ret) {
 729                 iser_err("resolve route failed: %d\n", ret);
 730                 iser_connect_error(cma_id);
 731                 return;
 732         }
 733 }
 734
 735 /**
 736  * Called with state mutex held
 737  **/
 738 static void iser_route_handler(struct rdma_cm_id *cma_id)
 739 {
 740         struct rdma_conn_param conn_param;
 741         int    ret;
 742         struct iser_cm_hdr req_hdr;
 743         struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
 744         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 745         struct iser_device *device = ib_conn->device;
 746
 747         if (iser_conn->state != ISER_CONN_PENDING)
 748                 /* bailout */
 749                 return;
 750
 751         ret = iser_create_ib_conn_res(ib_conn);
 752         if (ret)
 753                 goto failure;
 754
 755         memset(&conn_param, 0, sizeof conn_param);
 756         conn_param.responder_resources = device->ib_device->attrs.max_qp_rd_atom;
 757         conn_param.initiator_depth     = 1;
 758         conn_param.retry_count         = 7;
 759         conn_param.rnr_retry_count     = 6;
 760
 761         memset(&req_hdr, 0, sizeof(req_hdr));
 762         req_hdr.flags = ISER_ZBVA_NOT_SUP;
 763         if (!device->remote_inv_sup)
 764                 req_hdr.flags |= ISER_SEND_W_INV_NOT_SUP;
 765         conn_param.private_data = (void *)&req_hdr;
 766         conn_param.private_data_len = sizeof(struct iser_cm_hdr);
 767
 768         ret = rdma_connect(cma_id, &conn_param);
 769         if (ret) {
 770                 iser_err("failure connecting: %d\n", ret);
 771                 goto failure;
 772         }
 773
 774         return;
 775 failure:
 776         iser_connect_error(cma_id);
 777 }
 778
 779 static void iser_connected_handler(struct rdma_cm_id *cma_id,
 780                                    const void *private_data)
 781 {
 782         struct iser_conn *iser_conn;
 783         struct ib_qp_attr attr;
 784         struct ib_qp_init_attr init_attr;
 785
 786         iser_conn = (struct iser_conn *)cma_id->context;
 787         if (iser_conn->state != ISER_CONN_PENDING)
 788                 /* bailout */
 789                 return;
 790
 791         (void)ib_query_qp(cma_id->qp, &attr, ~0, &init_attr);
 792         iser_info("remote qpn:%x my qpn:%x\n", attr.dest_qp_num, cma_id->qp->qp_num);
 793
 794         if (private_data) {
 795                 u8 flags = *(u8 *)private_data;
 796
 797                 iser_conn->snd_w_inv = !(flags & ISER_SEND_W_INV_NOT_SUP);
 798         }
 799
 800         iser_info("conn %p: negotiated %s invalidation\n",
 801                   iser_conn, iser_conn->snd_w_inv ? "remote" : "local");
 802
 803         iser_conn->state = ISER_CONN_UP;
 804         complete(&iser_conn->up_completion);
 805 }
 806
 807 static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
 808 {
 809         struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
 810
 811         if (iser_conn_terminate(iser_conn)) {
 812                 if (iser_conn->iscsi_conn)
 813                         iscsi_conn_failure(iser_conn->iscsi_conn,
 814                                            ISCSI_ERR_CONN_FAILED);
 815                 else
 816                         iser_err("iscsi_iser connection isn't bound\n");
 817         }
 818 }
 819
 820 static void iser_cleanup_handler(struct rdma_cm_id *cma_id,
 821                                  bool destroy)
 822 {
 823         struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
 824
 825         /*
 826          * We are not guaranteed that we visited disconnected_handler
 827          * by now, call it here to be safe that we handle CM drep
 828          * and flush errors.
 829          */
 830         iser_disconnected_handler(cma_id);
 831         iser_free_ib_conn_res(iser_conn, destroy);
 832         complete(&iser_conn->ib_completion);
 833 };
 834
 835 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 836 {
 837         struct iser_conn *iser_conn;
 838         int ret = 0;
 839
 840         iser_conn = (struct iser_conn *)cma_id->context;
 841         iser_info("%s (%d): status %d conn %p id %p\n",
 842                   rdma_event_msg(event->event), event->event,
 843                   event->status, cma_id->context, cma_id);
 844
 845         mutex_lock(&iser_conn->state_mutex);
 846         switch (event->event) {
 847         case RDMA_CM_EVENT_ADDR_RESOLVED:
 848                 iser_addr_handler(cma_id);
 849                 break;
 850         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 851                 iser_route_handler(cma_id);
 852                 break;
 853         case RDMA_CM_EVENT_ESTABLISHED:
 854                 iser_connected_handler(cma_id, event->param.conn.private_data);
 855                 break;
 856         case RDMA_CM_EVENT_REJECTED:
 857                 iser_info("Connection rejected: %s\n",
 858                          rdma_reject_msg(cma_id, event->status));
 859                 /* FALLTHROUGH */
 860         case RDMA_CM_EVENT_ADDR_ERROR:
 861         case RDMA_CM_EVENT_ROUTE_ERROR:
 862         case RDMA_CM_EVENT_CONNECT_ERROR:
 863         case RDMA_CM_EVENT_UNREACHABLE:
 864                 iser_connect_error(cma_id);
 865                 break;
 866         case RDMA_CM_EVENT_DISCONNECTED:
 867         case RDMA_CM_EVENT_ADDR_CHANGE:
 868         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
 869                 iser_cleanup_handler(cma_id, false);
 870                 break;
 871         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 872                 /*
 873                  * we *must* destroy the device as we cannot rely
 874                  * on iscsid to be around to initiate error handling.
 875                  * also if we are not in state DOWN implicitly destroy
 876                  * the cma_id.
 877                  */
 878                 iser_cleanup_handler(cma_id, true);
 879                 if (iser_conn->state != ISER_CONN_DOWN) {
 880                         iser_conn->ib_conn.cma_id = NULL;
 881                         ret = 1;
 882                 }
 883                 break;
 884         default:
 885                 iser_err("Unexpected RDMA CM event: %s (%d)\n",
 886                          rdma_event_msg(event->event), event->event);
 887                 break;
 888         }
 889         mutex_unlock(&iser_conn->state_mutex);
 890
 891         return ret;
 892 }
 893
 894 void iser_conn_init(struct iser_conn *iser_conn)
 895 {
 896         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 897
 898         iser_conn->state = ISER_CONN_INIT;
 899         init_completion(&iser_conn->stop_completion);
 900         init_completion(&iser_conn->ib_completion);
 901         init_completion(&iser_conn->up_completion);
 902         INIT_LIST_HEAD(&iser_conn->conn_list);
 903         mutex_init(&iser_conn->state_mutex);
 904
 905         ib_conn->post_recv_buf_count = 0;
 906         ib_conn->reg_cqe.done = iser_reg_comp;
 907 }
 908
 909  /**
 910  * starts the process of connecting to the target
 911  * sleeps until the connection is established or rejected
 912  */
 913 int iser_connect(struct iser_conn   *iser_conn,
 914                  struct sockaddr    *src_addr,
 915                  struct sockaddr    *dst_addr,
 916                  int                 non_blocking)
 917 {
 918         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 919         int err = 0;
 920
 921         mutex_lock(&iser_conn->state_mutex);
 922
 923         sprintf(iser_conn->name, "%pISp", dst_addr);
 924
 925         iser_info("connecting to: %s\n", iser_conn->name);
 926
 927         /* the device is known only --after-- address resolution */
 928         ib_conn->device = NULL;
 929
 930         iser_conn->state = ISER_CONN_PENDING;
 931
 932         ib_conn->cma_id = rdma_create_id(&init_net, iser_cma_handler,
 933                                          (void *)iser_conn,
 934                                          RDMA_PS_TCP, IB_QPT_RC);
 935         if (IS_ERR(ib_conn->cma_id)) {
 936                 err = PTR_ERR(ib_conn->cma_id);
 937                 iser_err("rdma_create_id failed: %d\n", err);
 938                 goto id_failure;
 939         }
 940
 941         err = rdma_resolve_addr(ib_conn->cma_id, src_addr, dst_addr, 1000);
 942         if (err) {
 943                 iser_err("rdma_resolve_addr failed: %d\n", err);
 944                 goto addr_failure;
 945         }
 946
 947         if (!non_blocking) {
 948                 wait_for_completion_interruptible(&iser_conn->up_completion);
 949
 950                 if (iser_conn->state != ISER_CONN_UP) {
 951                         err =  -EIO;
 952                         goto connect_failure;
 953                 }
 954         }
 955         mutex_unlock(&iser_conn->state_mutex);
 956
 957         mutex_lock(&ig.connlist_mutex);
 958         list_add(&iser_conn->conn_list, &ig.connlist);
 959         mutex_unlock(&ig.connlist_mutex);
 960         return 0;
 961
 962 id_failure:
 963         ib_conn->cma_id = NULL;
 964 addr_failure:
 965         iser_conn->state = ISER_CONN_DOWN;
 966 connect_failure:
 967         mutex_unlock(&iser_conn->state_mutex);
 968         iser_conn_release(iser_conn);
 969         return err;
 970 }
 971
 972 int iser_post_recvl(struct iser_conn *iser_conn)
 973 {
 974         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 975         struct iser_login_desc *desc = &iser_conn->login_desc;
 976         struct ib_recv_wr wr;
 977         int ib_ret;
 978
 979         desc->sge.addr = desc->rsp_dma;
 980         desc->sge.length = ISER_RX_LOGIN_SIZE;
 981         desc->sge.lkey = ib_conn->device->pd->local_dma_lkey;
 982
 983         desc->cqe.done = iser_login_rsp;
 984         wr.wr_cqe = &desc->cqe;
 985         wr.sg_list = &desc->sge;
 986         wr.num_sge = 1;
 987         wr.next = NULL;
 988
 989         ib_conn->post_recv_buf_count++;
 990         ib_ret = ib_post_recv(ib_conn->qp, &wr, NULL);
 991         if (ib_ret) {
 992                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 993                 ib_conn->post_recv_buf_count--;
 994         }
 995
 996         return ib_ret;
 997 }
 998
 999 int iser_post_recvm(struct iser_conn *iser_conn, int count)
1000 {
1001         struct ib_conn *ib_conn = &iser_conn->ib_conn;
1002         unsigned int my_rx_head = iser_conn->rx_desc_head;
1003         struct iser_rx_desc *rx_desc;
1004         struct ib_recv_wr *wr;
1005         int i, ib_ret;
1006
1007         for (wr = ib_conn->rx_wr, i = 0; i < count; i++, wr++) {
1008                 rx_desc = &iser_conn->rx_descs[my_rx_head];
1009                 rx_desc->cqe.done = iser_task_rsp;
1010                 wr->wr_cqe = &rx_desc->cqe;
1011                 wr->sg_list = &rx_desc->rx_sg;
1012                 wr->num_sge = 1;
1013                 wr->next = wr + 1;
1014                 my_rx_head = (my_rx_head + 1) & iser_conn->qp_max_recv_dtos_mask;
1015         }
1016
1017         wr--;
1018         wr->next = NULL; /* mark end of work requests list */
1019
1020         ib_conn->post_recv_buf_count += count;
1021         ib_ret = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, NULL);
1022         if (ib_ret) {
1023                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1024                 ib_conn->post_recv_buf_count -= count;
1025         } else
1026                 iser_conn->rx_desc_head = my_rx_head;
1027
1028         return ib_ret;
1029 }
1030
1031
1032 /**
1033  * iser_start_send - Initiate a Send DTO operation
1034  *
1035  * returns 0 on success, -1 on failure
1036  */
1037 int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc,
1038                    bool signal)
1039 {
1040         struct ib_send_wr *wr = &tx_desc->send_wr;
1041         struct ib_send_wr *first_wr;
1042         int ib_ret;
1043
1044         ib_dma_sync_single_for_device(ib_conn->device->ib_device,
1045                                       tx_desc->dma_addr, ISER_HEADERS_LEN,
1046                                       DMA_TO_DEVICE);
1047
1048         wr->next = NULL;
1049         wr->wr_cqe = &tx_desc->cqe;
1050         wr->sg_list = tx_desc->tx_sg;
1051         wr->num_sge = tx_desc->num_sge;
1052         wr->opcode = IB_WR_SEND;
1053         wr->send_flags = signal ? IB_SEND_SIGNALED : 0;
1054
1055         if (tx_desc->inv_wr.next)
1056                 first_wr = &tx_desc->inv_wr;
1057         else if (tx_desc->reg_wr.wr.next)
1058                 first_wr = &tx_desc->reg_wr.wr;
1059         else
1060                 first_wr = wr;
1061
1062         ib_ret = ib_post_send(ib_conn->qp, first_wr, NULL);
1063         if (ib_ret)
1064                 iser_err("ib_post_send failed, ret:%d opcode:%d\n",
1065                          ib_ret, wr->opcode);
1066
1067         return ib_ret;
1068 }
1069
1070 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
1071                              enum iser_data_dir cmd_dir, sector_t *sector)
1072 {
1073         struct iser_mem_reg *reg = &iser_task->rdma_reg[cmd_dir];
1074         struct iser_fr_desc *desc = reg->mem_h;
1075         unsigned long sector_size = iser_task->sc->device->sector_size;
1076         struct ib_mr_status mr_status;
1077         int ret;
1078
1079         if (desc && desc->sig_protected) {
1080                 desc->sig_protected = 0;
1081                 ret = ib_check_mr_status(desc->rsc.sig_mr,
1082                                          IB_MR_CHECK_SIG_STATUS, &mr_status);
1083                 if (ret) {
1084                         pr_err("ib_check_mr_status failed, ret %d\n", ret);
1085                         /* Not a lot we can do, return ambiguous guard error */
1086                         *sector = 0;
1087                         return 0x1;
1088                 }
1089
1090                 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1091                         sector_t sector_off = mr_status.sig_err.sig_err_offset;
1092
1093                         sector_div(sector_off, sector_size + 8);
1094                         *sector = scsi_get_lba(iser_task->sc) + sector_off;
1095
1096                         pr_err("PI error found type %d at sector %llx "
1097                                "expected %x vs actual %x\n",
1098                                mr_status.sig_err.err_type,
1099                                (unsigned long long)*sector,
1100                                mr_status.sig_err.expected,
1101                                mr_status.sig_err.actual);
1102
1103                         switch (mr_status.sig_err.err_type) {
1104                         case IB_SIG_BAD_GUARD:
1105                                 return 0x1;
1106                         case IB_SIG_BAD_REFTAG:
1107                                 return 0x3;
1108                         case IB_SIG_BAD_APPTAG:
1109                                 return 0x2;
1110                         }
1111                 }
1112         }
1113
1114         return 0;
1115 }
1116
1117 void iser_err_comp(struct ib_wc *wc, const char *type)
1118 {
1119         if (wc->status != IB_WC_WR_FLUSH_ERR) {
1120                 struct iser_conn *iser_conn = to_iser_conn(wc->qp->qp_context);
1121
1122                 iser_err("%s failure: %s (%d) vend_err %#x\n", type,
1123                          ib_wc_status_msg(wc->status), wc->status,
1124                          wc->vendor_err);
1125
1126                 if (iser_conn->iscsi_conn)
1127                         iscsi_conn_failure(iser_conn->iscsi_conn,
1128                                            ISCSI_ERR_CONN_FAILED);
1129         } else {
1130                 iser_dbg("%s failure: %s (%d)\n", type,
1131                          ib_wc_status_msg(wc->status), wc->status);
1132         }
1133 }