linux/drivers/infiniband/ulp/iser/iser_verbs.c
<<
>>
Prefs
   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
  47static 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
  53static 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                 event->device->name, 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 */
  68static 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, ib_dev->name,
  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
 112cq_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);
 120pd_err:
 121        kfree(device->comps);
 122comps_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 */
 131static 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 */
 155int 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
 208err_fmr:
 209        kfree(page_vec);
 210err_frpl:
 211        kfree(desc);
 212
 213        return ret;
 214}
 215
 216/**
 217 * iser_free_fmr_pool - releases the FMR pool and page vec
 218 */
 219void 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
 236static int
 237iser_alloc_reg_res(struct iser_device *device,
 238                   struct ib_pd *pd,
 239                   struct iser_reg_resources *res,
 240                   unsigned int size)
 241{
 242        struct ib_device *ib_dev = device->ib_device;
 243        enum ib_mr_type mr_type;
 244        int ret;
 245
 246        if (ib_dev->attrs.device_cap_flags & IB_DEVICE_SG_GAPS_REG)
 247                mr_type = IB_MR_TYPE_SG_GAPS;
 248        else
 249                mr_type = IB_MR_TYPE_MEM_REG;
 250
 251        res->mr = ib_alloc_mr(pd, mr_type, size);
 252        if (IS_ERR(res->mr)) {
 253                ret = PTR_ERR(res->mr);
 254                iser_err("Failed to allocate ib_fast_reg_mr err=%d\n", ret);
 255                return ret;
 256        }
 257        res->mr_valid = 0;
 258
 259        return 0;
 260}
 261
 262static void
 263iser_free_reg_res(struct iser_reg_resources *rsc)
 264{
 265        ib_dereg_mr(rsc->mr);
 266}
 267
 268static int
 269iser_alloc_pi_ctx(struct iser_device *device,
 270                  struct ib_pd *pd,
 271                  struct iser_fr_desc *desc,
 272                  unsigned int size)
 273{
 274        struct iser_pi_context *pi_ctx = NULL;
 275        int ret;
 276
 277        desc->pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
 278        if (!desc->pi_ctx)
 279                return -ENOMEM;
 280
 281        pi_ctx = desc->pi_ctx;
 282
 283        ret = iser_alloc_reg_res(device, pd, &pi_ctx->rsc, size);
 284        if (ret) {
 285                iser_err("failed to allocate reg_resources\n");
 286                goto alloc_reg_res_err;
 287        }
 288
 289        pi_ctx->sig_mr = ib_alloc_mr(pd, IB_MR_TYPE_SIGNATURE, 2);
 290        if (IS_ERR(pi_ctx->sig_mr)) {
 291                ret = PTR_ERR(pi_ctx->sig_mr);
 292                goto sig_mr_failure;
 293        }
 294        pi_ctx->sig_mr_valid = 0;
 295        desc->pi_ctx->sig_protected = 0;
 296
 297        return 0;
 298
 299sig_mr_failure:
 300        iser_free_reg_res(&pi_ctx->rsc);
 301alloc_reg_res_err:
 302        kfree(desc->pi_ctx);
 303
 304        return ret;
 305}
 306
 307static void
 308iser_free_pi_ctx(struct iser_pi_context *pi_ctx)
 309{
 310        iser_free_reg_res(&pi_ctx->rsc);
 311        ib_dereg_mr(pi_ctx->sig_mr);
 312        kfree(pi_ctx);
 313}
 314
 315static struct iser_fr_desc *
 316iser_create_fastreg_desc(struct iser_device *device,
 317                         struct ib_pd *pd,
 318                         bool pi_enable,
 319                         unsigned int size)
 320{
 321        struct iser_fr_desc *desc;
 322        int ret;
 323
 324        desc = kzalloc(sizeof(*desc), GFP_KERNEL);
 325        if (!desc)
 326                return ERR_PTR(-ENOMEM);
 327
 328        ret = iser_alloc_reg_res(device, pd, &desc->rsc, size);
 329        if (ret)
 330                goto reg_res_alloc_failure;
 331
 332        if (pi_enable) {
 333                ret = iser_alloc_pi_ctx(device, pd, desc, size);
 334                if (ret)
 335                        goto pi_ctx_alloc_failure;
 336        }
 337
 338        return desc;
 339
 340pi_ctx_alloc_failure:
 341        iser_free_reg_res(&desc->rsc);
 342reg_res_alloc_failure:
 343        kfree(desc);
 344
 345        return ERR_PTR(ret);
 346}
 347
 348/**
 349 * iser_alloc_fastreg_pool - Creates pool of fast_reg descriptors
 350 * for fast registration work requests.
 351 * returns 0 on success, or errno code on failure
 352 */
 353int iser_alloc_fastreg_pool(struct ib_conn *ib_conn,
 354                            unsigned cmds_max,
 355                            unsigned int size)
 356{
 357        struct iser_device *device = ib_conn->device;
 358        struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
 359        struct iser_fr_desc *desc;
 360        int i, ret;
 361
 362        INIT_LIST_HEAD(&fr_pool->list);
 363        INIT_LIST_HEAD(&fr_pool->all_list);
 364        spin_lock_init(&fr_pool->lock);
 365        fr_pool->size = 0;
 366        for (i = 0; i < cmds_max; i++) {
 367                desc = iser_create_fastreg_desc(device, device->pd,
 368                                                ib_conn->pi_support, size);
 369                if (IS_ERR(desc)) {
 370                        ret = PTR_ERR(desc);
 371                        goto err;
 372                }
 373
 374                list_add_tail(&desc->list, &fr_pool->list);
 375                list_add_tail(&desc->all_list, &fr_pool->all_list);
 376                fr_pool->size++;
 377        }
 378
 379        return 0;
 380
 381err:
 382        iser_free_fastreg_pool(ib_conn);
 383        return ret;
 384}
 385
 386/**
 387 * iser_free_fastreg_pool - releases the pool of fast_reg descriptors
 388 */
 389void iser_free_fastreg_pool(struct ib_conn *ib_conn)
 390{
 391        struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
 392        struct iser_fr_desc *desc, *tmp;
 393        int i = 0;
 394
 395        if (list_empty(&fr_pool->all_list))
 396                return;
 397
 398        iser_info("freeing conn %p fr pool\n", ib_conn);
 399
 400        list_for_each_entry_safe(desc, tmp, &fr_pool->all_list, all_list) {
 401                list_del(&desc->all_list);
 402                iser_free_reg_res(&desc->rsc);
 403                if (desc->pi_ctx)
 404                        iser_free_pi_ctx(desc->pi_ctx);
 405                kfree(desc);
 406                ++i;
 407        }
 408
 409        if (i < fr_pool->size)
 410                iser_warn("pool still has %d regions registered\n",
 411                          fr_pool->size - i);
 412}
 413
 414/**
 415 * iser_create_ib_conn_res - Queue-Pair (QP)
 416 *
 417 * returns 0 on success, -1 on failure
 418 */
 419static int iser_create_ib_conn_res(struct ib_conn *ib_conn)
 420{
 421        struct iser_conn *iser_conn = to_iser_conn(ib_conn);
 422        struct iser_device      *device;
 423        struct ib_device        *ib_dev;
 424        struct ib_qp_init_attr  init_attr;
 425        int                     ret = -ENOMEM;
 426        int index, min_index = 0;
 427
 428        BUG_ON(ib_conn->device == NULL);
 429
 430        device = ib_conn->device;
 431        ib_dev = device->ib_device;
 432
 433        memset(&init_attr, 0, sizeof init_attr);
 434
 435        mutex_lock(&ig.connlist_mutex);
 436        /* select the CQ with the minimal number of usages */
 437        for (index = 0; index < device->comps_used; index++) {
 438                if (device->comps[index].active_qps <
 439                    device->comps[min_index].active_qps)
 440                        min_index = index;
 441        }
 442        ib_conn->comp = &device->comps[min_index];
 443        ib_conn->comp->active_qps++;
 444        mutex_unlock(&ig.connlist_mutex);
 445        iser_info("cq index %d used for ib_conn %p\n", min_index, ib_conn);
 446
 447        init_attr.event_handler = iser_qp_event_callback;
 448        init_attr.qp_context    = (void *)ib_conn;
 449        init_attr.send_cq       = ib_conn->comp->cq;
 450        init_attr.recv_cq       = ib_conn->comp->cq;
 451        init_attr.cap.max_recv_wr  = ISER_QP_MAX_RECV_DTOS;
 452        init_attr.cap.max_send_sge = 2;
 453        init_attr.cap.max_recv_sge = 1;
 454        init_attr.sq_sig_type   = IB_SIGNAL_REQ_WR;
 455        init_attr.qp_type       = IB_QPT_RC;
 456        if (ib_conn->pi_support) {
 457                init_attr.cap.max_send_wr = ISER_QP_SIG_MAX_REQ_DTOS + 1;
 458                init_attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
 459                iser_conn->max_cmds =
 460                        ISER_GET_MAX_XMIT_CMDS(ISER_QP_SIG_MAX_REQ_DTOS);
 461        } else {
 462                if (ib_dev->attrs.max_qp_wr > ISER_QP_MAX_REQ_DTOS) {
 463                        init_attr.cap.max_send_wr  = ISER_QP_MAX_REQ_DTOS + 1;
 464                        iser_conn->max_cmds =
 465                                ISER_GET_MAX_XMIT_CMDS(ISER_QP_MAX_REQ_DTOS);
 466                } else {
 467                        init_attr.cap.max_send_wr = ib_dev->attrs.max_qp_wr;
 468                        iser_conn->max_cmds =
 469                                ISER_GET_MAX_XMIT_CMDS(ib_dev->attrs.max_qp_wr);
 470                        iser_dbg("device %s supports max_send_wr %d\n",
 471                                 device->ib_device->name, ib_dev->attrs.max_qp_wr);
 472                }
 473        }
 474
 475        ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
 476        if (ret)
 477                goto out_err;
 478
 479        ib_conn->qp = ib_conn->cma_id->qp;
 480        iser_info("setting conn %p cma_id %p qp %p\n",
 481                  ib_conn, ib_conn->cma_id,
 482                  ib_conn->cma_id->qp);
 483        return ret;
 484
 485out_err:
 486        mutex_lock(&ig.connlist_mutex);
 487        ib_conn->comp->active_qps--;
 488        mutex_unlock(&ig.connlist_mutex);
 489        iser_err("unable to alloc mem or create resource, err %d\n", ret);
 490
 491        return ret;
 492}
 493
 494/**
 495 * based on the resolved device node GUID see if there already allocated
 496 * device for this device. If there's no such, create one.
 497 */
 498static
 499struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
 500{
 501        struct iser_device *device;
 502
 503        mutex_lock(&ig.device_list_mutex);
 504
 505        list_for_each_entry(device, &ig.device_list, ig_list)
 506                /* find if there's a match using the node GUID */
 507                if (device->ib_device->node_guid == cma_id->device->node_guid)
 508                        goto inc_refcnt;
 509
 510        device = kzalloc(sizeof *device, GFP_KERNEL);
 511        if (device == NULL)
 512                goto out;
 513
 514        /* assign this device to the device */
 515        device->ib_device = cma_id->device;
 516        /* init the device and link it into ig device list */
 517        if (iser_create_device_ib_res(device)) {
 518                kfree(device);
 519                device = NULL;
 520                goto out;
 521        }
 522        list_add(&device->ig_list, &ig.device_list);
 523
 524inc_refcnt:
 525        device->refcount++;
 526out:
 527        mutex_unlock(&ig.device_list_mutex);
 528        return device;
 529}
 530
 531/* if there's no demand for this device, release it */
 532static void iser_device_try_release(struct iser_device *device)
 533{
 534        mutex_lock(&ig.device_list_mutex);
 535        device->refcount--;
 536        iser_info("device %p refcount %d\n", device, device->refcount);
 537        if (!device->refcount) {
 538                iser_free_device_ib_res(device);
 539                list_del(&device->ig_list);
 540                kfree(device);
 541        }
 542        mutex_unlock(&ig.device_list_mutex);
 543}
 544
 545/**
 546 * Called with state mutex held
 547 **/
 548static int iser_conn_state_comp_exch(struct iser_conn *iser_conn,
 549                                     enum iser_conn_state comp,
 550                                     enum iser_conn_state exch)
 551{
 552        int ret;
 553
 554        ret = (iser_conn->state == comp);
 555        if (ret)
 556                iser_conn->state = exch;
 557
 558        return ret;
 559}
 560
 561void iser_release_work(struct work_struct *work)
 562{
 563        struct iser_conn *iser_conn;
 564
 565        iser_conn = container_of(work, struct iser_conn, release_work);
 566
 567        /* Wait for conn_stop to complete */
 568        wait_for_completion(&iser_conn->stop_completion);
 569        /* Wait for IB resouces cleanup to complete */
 570        wait_for_completion(&iser_conn->ib_completion);
 571
 572        mutex_lock(&iser_conn->state_mutex);
 573        iser_conn->state = ISER_CONN_DOWN;
 574        mutex_unlock(&iser_conn->state_mutex);
 575
 576        iser_conn_release(iser_conn);
 577}
 578
 579/**
 580 * iser_free_ib_conn_res - release IB related resources
 581 * @iser_conn: iser connection struct
 582 * @destroy: indicator if we need to try to release the
 583 *     iser device and memory regoins pool (only iscsi
 584 *     shutdown and DEVICE_REMOVAL will use this).
 585 *
 586 * This routine is called with the iser state mutex held
 587 * so the cm_id removal is out of here. It is Safe to
 588 * be invoked multiple times.
 589 */
 590static void iser_free_ib_conn_res(struct iser_conn *iser_conn,
 591                                  bool destroy)
 592{
 593        struct ib_conn *ib_conn = &iser_conn->ib_conn;
 594        struct iser_device *device = ib_conn->device;
 595
 596        iser_info("freeing conn %p cma_id %p qp %p\n",
 597                  iser_conn, ib_conn->cma_id, ib_conn->qp);
 598
 599        if (ib_conn->qp != NULL) {
 600                mutex_lock(&ig.connlist_mutex);
 601                ib_conn->comp->active_qps--;
 602                mutex_unlock(&ig.connlist_mutex);
 603                rdma_destroy_qp(ib_conn->cma_id);
 604                ib_conn->qp = NULL;
 605        }
 606
 607        if (destroy) {
 608                if (iser_conn->rx_descs)
 609                        iser_free_rx_descriptors(iser_conn);
 610
 611                if (device != NULL) {
 612                        iser_device_try_release(device);
 613                        ib_conn->device = NULL;
 614                }
 615        }
 616}
 617
 618/**
 619 * Frees all conn objects and deallocs conn descriptor
 620 */
 621void iser_conn_release(struct iser_conn *iser_conn)
 622{
 623        struct ib_conn *ib_conn = &iser_conn->ib_conn;
 624
 625        mutex_lock(&ig.connlist_mutex);
 626        list_del(&iser_conn->conn_list);
 627        mutex_unlock(&ig.connlist_mutex);
 628
 629        mutex_lock(&iser_conn->state_mutex);
 630        /* In case we endup here without ep_disconnect being invoked. */
 631        if (iser_conn->state != ISER_CONN_DOWN) {
 632                iser_warn("iser conn %p state %d, expected state down.\n",
 633                          iser_conn, iser_conn->state);
 634                iscsi_destroy_endpoint(iser_conn->ep);
 635                iser_conn->state = ISER_CONN_DOWN;
 636        }
 637        /*
 638         * In case we never got to bind stage, we still need to
 639         * release IB resources (which is safe to call more than once).
 640         */
 641        iser_free_ib_conn_res(iser_conn, true);
 642        mutex_unlock(&iser_conn->state_mutex);
 643
 644        if (ib_conn->cma_id != NULL) {
 645                rdma_destroy_id(ib_conn->cma_id);
 646                ib_conn->cma_id = NULL;
 647        }
 648
 649        kfree(iser_conn);
 650}
 651
 652/**
 653 * triggers start of the disconnect procedures and wait for them to be done
 654 * Called with state mutex held
 655 */
 656int iser_conn_terminate(struct iser_conn *iser_conn)
 657{
 658        struct ib_conn *ib_conn = &iser_conn->ib_conn;
 659        int err = 0;
 660
 661        /* terminate the iser conn only if the conn state is UP */
 662        if (!iser_conn_state_comp_exch(iser_conn, ISER_CONN_UP,
 663                                       ISER_CONN_TERMINATING))
 664                return 0;
 665
 666        iser_info("iser_conn %p state %d\n", iser_conn, iser_conn->state);
 667
 668        /* suspend queuing of new iscsi commands */
 669        if (iser_conn->iscsi_conn)
 670                iscsi_suspend_queue(iser_conn->iscsi_conn);
 671
 672        /*
 673         * In case we didn't already clean up the cma_id (peer initiated
 674         * a disconnection), we need to Cause the CMA to change the QP
 675         * state to ERROR.
 676         */
 677        if (ib_conn->cma_id) {
 678                err = rdma_disconnect(ib_conn->cma_id);
 679                if (err)
 680                        iser_err("Failed to disconnect, conn: 0x%p err %d\n",
 681                                 iser_conn, err);
 682
 683                /* block until all flush errors are consumed */
 684                ib_drain_sq(ib_conn->qp);
 685        }
 686
 687        return 1;
 688}
 689
 690/**
 691 * Called with state mutex held
 692 **/
 693static void iser_connect_error(struct rdma_cm_id *cma_id)
 694{
 695        struct iser_conn *iser_conn;
 696
 697        iser_conn = (struct iser_conn *)cma_id->context;
 698        iser_conn->state = ISER_CONN_TERMINATING;
 699}
 700
 701static void
 702iser_calc_scsi_params(struct iser_conn *iser_conn,
 703                      unsigned int max_sectors)
 704{
 705        struct iser_device *device = iser_conn->ib_conn.device;
 706        struct ib_device_attr *attr = &device->ib_device->attrs;
 707        unsigned short sg_tablesize, sup_sg_tablesize;
 708        unsigned short reserved_mr_pages;
 709
 710        /*
 711         * FRs without SG_GAPS or FMRs can only map up to a (device) page per
 712         * entry, but if the first entry is misaligned we'll end up using two
 713         * entries (head and tail) for a single page worth data, so one
 714         * additional entry is required.
 715         */
 716        if ((attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) &&
 717            (attr->device_cap_flags & IB_DEVICE_SG_GAPS_REG))
 718                reserved_mr_pages = 0;
 719        else
 720                reserved_mr_pages = 1;
 721
 722        sg_tablesize = DIV_ROUND_UP(max_sectors * 512, SIZE_4K);
 723        if (attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS)
 724                sup_sg_tablesize =
 725                        min_t(
 726                         uint, ISCSI_ISER_MAX_SG_TABLESIZE,
 727                         attr->max_fast_reg_page_list_len - reserved_mr_pages);
 728        else
 729                sup_sg_tablesize = ISCSI_ISER_MAX_SG_TABLESIZE;
 730
 731        iser_conn->scsi_sg_tablesize = min(sg_tablesize, sup_sg_tablesize);
 732        iser_conn->pages_per_mr =
 733                iser_conn->scsi_sg_tablesize + reserved_mr_pages;
 734}
 735
 736/**
 737 * Called with state mutex held
 738 **/
 739static void iser_addr_handler(struct rdma_cm_id *cma_id)
 740{
 741        struct iser_device *device;
 742        struct iser_conn   *iser_conn;
 743        struct ib_conn   *ib_conn;
 744        int    ret;
 745
 746        iser_conn = (struct iser_conn *)cma_id->context;
 747        if (iser_conn->state != ISER_CONN_PENDING)
 748                /* bailout */
 749                return;
 750
 751        ib_conn = &iser_conn->ib_conn;
 752        device = iser_device_find_by_ib_device(cma_id);
 753        if (!device) {
 754                iser_err("device lookup/creation failed\n");
 755                iser_connect_error(cma_id);
 756                return;
 757        }
 758
 759        ib_conn->device = device;
 760
 761        /* connection T10-PI support */
 762        if (iser_pi_enable) {
 763                if (!(device->ib_device->attrs.device_cap_flags &
 764                      IB_DEVICE_SIGNATURE_HANDOVER)) {
 765                        iser_warn("T10-PI requested but not supported on %s, "
 766                                  "continue without T10-PI\n",
 767                                  ib_conn->device->ib_device->name);
 768                        ib_conn->pi_support = false;
 769                } else {
 770                        ib_conn->pi_support = true;
 771                }
 772        }
 773
 774        iser_calc_scsi_params(iser_conn, iser_max_sectors);
 775
 776        ret = rdma_resolve_route(cma_id, 1000);
 777        if (ret) {
 778                iser_err("resolve route failed: %d\n", ret);
 779                iser_connect_error(cma_id);
 780                return;
 781        }
 782}
 783
 784/**
 785 * Called with state mutex held
 786 **/
 787static void iser_route_handler(struct rdma_cm_id *cma_id)
 788{
 789        struct rdma_conn_param conn_param;
 790        int    ret;
 791        struct iser_cm_hdr req_hdr;
 792        struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
 793        struct ib_conn *ib_conn = &iser_conn->ib_conn;
 794        struct iser_device *device = ib_conn->device;
 795
 796        if (iser_conn->state != ISER_CONN_PENDING)
 797                /* bailout */
 798                return;
 799
 800        ret = iser_create_ib_conn_res(ib_conn);
 801        if (ret)
 802                goto failure;
 803
 804        memset(&conn_param, 0, sizeof conn_param);
 805        conn_param.responder_resources = device->ib_device->attrs.max_qp_rd_atom;
 806        conn_param.initiator_depth     = 1;
 807        conn_param.retry_count         = 7;
 808        conn_param.rnr_retry_count     = 6;
 809
 810        memset(&req_hdr, 0, sizeof(req_hdr));
 811        req_hdr.flags = ISER_ZBVA_NOT_SUP;
 812        if (!device->remote_inv_sup)
 813                req_hdr.flags |= ISER_SEND_W_INV_NOT_SUP;
 814        conn_param.private_data = (void *)&req_hdr;
 815        conn_param.private_data_len = sizeof(struct iser_cm_hdr);
 816
 817        ret = rdma_connect(cma_id, &conn_param);
 818        if (ret) {
 819                iser_err("failure connecting: %d\n", ret);
 820                goto failure;
 821        }
 822
 823        return;
 824failure:
 825        iser_connect_error(cma_id);
 826}
 827
 828static void iser_connected_handler(struct rdma_cm_id *cma_id,
 829                                   const void *private_data)
 830{
 831        struct iser_conn *iser_conn;
 832        struct ib_qp_attr attr;
 833        struct ib_qp_init_attr init_attr;
 834
 835        iser_conn = (struct iser_conn *)cma_id->context;
 836        if (iser_conn->state != ISER_CONN_PENDING)
 837                /* bailout */
 838                return;
 839
 840        (void)ib_query_qp(cma_id->qp, &attr, ~0, &init_attr);
 841        iser_info("remote qpn:%x my qpn:%x\n", attr.dest_qp_num, cma_id->qp->qp_num);
 842
 843        if (private_data) {
 844                u8 flags = *(u8 *)private_data;
 845
 846                iser_conn->snd_w_inv = !(flags & ISER_SEND_W_INV_NOT_SUP);
 847        }
 848
 849        iser_info("conn %p: negotiated %s invalidation\n",
 850                  iser_conn, iser_conn->snd_w_inv ? "remote" : "local");
 851
 852        iser_conn->state = ISER_CONN_UP;
 853        complete(&iser_conn->up_completion);
 854}
 855
 856static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
 857{
 858        struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
 859
 860        if (iser_conn_terminate(iser_conn)) {
 861                if (iser_conn->iscsi_conn)
 862                        iscsi_conn_failure(iser_conn->iscsi_conn,
 863                                           ISCSI_ERR_CONN_FAILED);
 864                else
 865                        iser_err("iscsi_iser connection isn't bound\n");
 866        }
 867}
 868
 869static void iser_cleanup_handler(struct rdma_cm_id *cma_id,
 870                                 bool destroy)
 871{
 872        struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
 873
 874        /*
 875         * We are not guaranteed that we visited disconnected_handler
 876         * by now, call it here to be safe that we handle CM drep
 877         * and flush errors.
 878         */
 879        iser_disconnected_handler(cma_id);
 880        iser_free_ib_conn_res(iser_conn, destroy);
 881        complete(&iser_conn->ib_completion);
 882};
 883
 884static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 885{
 886        struct iser_conn *iser_conn;
 887        int ret = 0;
 888
 889        iser_conn = (struct iser_conn *)cma_id->context;
 890        iser_info("%s (%d): status %d conn %p id %p\n",
 891                  rdma_event_msg(event->event), event->event,
 892                  event->status, cma_id->context, cma_id);
 893
 894        mutex_lock(&iser_conn->state_mutex);
 895        switch (event->event) {
 896        case RDMA_CM_EVENT_ADDR_RESOLVED:
 897                iser_addr_handler(cma_id);
 898                break;
 899        case RDMA_CM_EVENT_ROUTE_RESOLVED:
 900                iser_route_handler(cma_id);
 901                break;
 902        case RDMA_CM_EVENT_ESTABLISHED:
 903                iser_connected_handler(cma_id, event->param.conn.private_data);
 904                break;
 905        case RDMA_CM_EVENT_REJECTED:
 906                iser_info("Connection rejected: %s\n",
 907                         rdma_reject_msg(cma_id, event->status));
 908                /* FALLTHROUGH */
 909        case RDMA_CM_EVENT_ADDR_ERROR:
 910        case RDMA_CM_EVENT_ROUTE_ERROR:
 911        case RDMA_CM_EVENT_CONNECT_ERROR:
 912        case RDMA_CM_EVENT_UNREACHABLE:
 913                iser_connect_error(cma_id);
 914                break;
 915        case RDMA_CM_EVENT_DISCONNECTED:
 916        case RDMA_CM_EVENT_ADDR_CHANGE:
 917        case RDMA_CM_EVENT_TIMEWAIT_EXIT:
 918                iser_cleanup_handler(cma_id, false);
 919                break;
 920        case RDMA_CM_EVENT_DEVICE_REMOVAL:
 921                /*
 922                 * we *must* destroy the device as we cannot rely
 923                 * on iscsid to be around to initiate error handling.
 924                 * also if we are not in state DOWN implicitly destroy
 925                 * the cma_id.
 926                 */
 927                iser_cleanup_handler(cma_id, true);
 928                if (iser_conn->state != ISER_CONN_DOWN) {
 929                        iser_conn->ib_conn.cma_id = NULL;
 930                        ret = 1;
 931                }
 932                break;
 933        default:
 934                iser_err("Unexpected RDMA CM event: %s (%d)\n",
 935                         rdma_event_msg(event->event), event->event);
 936                break;
 937        }
 938        mutex_unlock(&iser_conn->state_mutex);
 939
 940        return ret;
 941}
 942
 943void iser_conn_init(struct iser_conn *iser_conn)
 944{
 945        struct ib_conn *ib_conn = &iser_conn->ib_conn;
 946
 947        iser_conn->state = ISER_CONN_INIT;
 948        init_completion(&iser_conn->stop_completion);
 949        init_completion(&iser_conn->ib_completion);
 950        init_completion(&iser_conn->up_completion);
 951        INIT_LIST_HEAD(&iser_conn->conn_list);
 952        mutex_init(&iser_conn->state_mutex);
 953
 954        ib_conn->post_recv_buf_count = 0;
 955        ib_conn->reg_cqe.done = iser_reg_comp;
 956}
 957
 958 /**
 959 * starts the process of connecting to the target
 960 * sleeps until the connection is established or rejected
 961 */
 962int iser_connect(struct iser_conn   *iser_conn,
 963                 struct sockaddr    *src_addr,
 964                 struct sockaddr    *dst_addr,
 965                 int                 non_blocking)
 966{
 967        struct ib_conn *ib_conn = &iser_conn->ib_conn;
 968        int err = 0;
 969
 970        mutex_lock(&iser_conn->state_mutex);
 971
 972        sprintf(iser_conn->name, "%pISp", dst_addr);
 973
 974        iser_info("connecting to: %s\n", iser_conn->name);
 975
 976        /* the device is known only --after-- address resolution */
 977        ib_conn->device = NULL;
 978
 979        iser_conn->state = ISER_CONN_PENDING;
 980
 981        ib_conn->cma_id = rdma_create_id(&init_net, iser_cma_handler,
 982                                         (void *)iser_conn,
 983                                         RDMA_PS_TCP, IB_QPT_RC);
 984        if (IS_ERR(ib_conn->cma_id)) {
 985                err = PTR_ERR(ib_conn->cma_id);
 986                iser_err("rdma_create_id failed: %d\n", err);
 987                goto id_failure;
 988        }
 989
 990        err = rdma_resolve_addr(ib_conn->cma_id, src_addr, dst_addr, 1000);
 991        if (err) {
 992                iser_err("rdma_resolve_addr failed: %d\n", err);
 993                goto addr_failure;
 994        }
 995
 996        if (!non_blocking) {
 997                wait_for_completion_interruptible(&iser_conn->up_completion);
 998
 999                if (iser_conn->state != ISER_CONN_UP) {
1000                        err =  -EIO;
1001                        goto connect_failure;
1002                }
1003        }
1004        mutex_unlock(&iser_conn->state_mutex);
1005
1006        mutex_lock(&ig.connlist_mutex);
1007        list_add(&iser_conn->conn_list, &ig.connlist);
1008        mutex_unlock(&ig.connlist_mutex);
1009        return 0;
1010
1011id_failure:
1012        ib_conn->cma_id = NULL;
1013addr_failure:
1014        iser_conn->state = ISER_CONN_DOWN;
1015connect_failure:
1016        mutex_unlock(&iser_conn->state_mutex);
1017        iser_conn_release(iser_conn);
1018        return err;
1019}
1020
1021int iser_post_recvl(struct iser_conn *iser_conn)
1022{
1023        struct ib_conn *ib_conn = &iser_conn->ib_conn;
1024        struct iser_login_desc *desc = &iser_conn->login_desc;
1025        struct ib_recv_wr wr;
1026        int ib_ret;
1027
1028        desc->sge.addr = desc->rsp_dma;
1029        desc->sge.length = ISER_RX_LOGIN_SIZE;
1030        desc->sge.lkey = ib_conn->device->pd->local_dma_lkey;
1031
1032        desc->cqe.done = iser_login_rsp;
1033        wr.wr_cqe = &desc->cqe;
1034        wr.sg_list = &desc->sge;
1035        wr.num_sge = 1;
1036        wr.next = NULL;
1037
1038        ib_conn->post_recv_buf_count++;
1039        ib_ret = ib_post_recv(ib_conn->qp, &wr, NULL);
1040        if (ib_ret) {
1041                iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1042                ib_conn->post_recv_buf_count--;
1043        }
1044
1045        return ib_ret;
1046}
1047
1048int iser_post_recvm(struct iser_conn *iser_conn, int count)
1049{
1050        struct ib_conn *ib_conn = &iser_conn->ib_conn;
1051        unsigned int my_rx_head = iser_conn->rx_desc_head;
1052        struct iser_rx_desc *rx_desc;
1053        struct ib_recv_wr *wr;
1054        int i, ib_ret;
1055
1056        for (wr = ib_conn->rx_wr, i = 0; i < count; i++, wr++) {
1057                rx_desc = &iser_conn->rx_descs[my_rx_head];
1058                rx_desc->cqe.done = iser_task_rsp;
1059                wr->wr_cqe = &rx_desc->cqe;
1060                wr->sg_list = &rx_desc->rx_sg;
1061                wr->num_sge = 1;
1062                wr->next = wr + 1;
1063                my_rx_head = (my_rx_head + 1) & iser_conn->qp_max_recv_dtos_mask;
1064        }
1065
1066        wr--;
1067        wr->next = NULL; /* mark end of work requests list */
1068
1069        ib_conn->post_recv_buf_count += count;
1070        ib_ret = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, NULL);
1071        if (ib_ret) {
1072                iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1073                ib_conn->post_recv_buf_count -= count;
1074        } else
1075                iser_conn->rx_desc_head = my_rx_head;
1076
1077        return ib_ret;
1078}
1079
1080
1081/**
1082 * iser_start_send - Initiate a Send DTO operation
1083 *
1084 * returns 0 on success, -1 on failure
1085 */
1086int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc,
1087                   bool signal)
1088{
1089        struct ib_send_wr *wr = iser_tx_next_wr(tx_desc);
1090        int ib_ret;
1091
1092        ib_dma_sync_single_for_device(ib_conn->device->ib_device,
1093                                      tx_desc->dma_addr, ISER_HEADERS_LEN,
1094                                      DMA_TO_DEVICE);
1095
1096        wr->next = NULL;
1097        wr->wr_cqe = &tx_desc->cqe;
1098        wr->sg_list = tx_desc->tx_sg;
1099        wr->num_sge = tx_desc->num_sge;
1100        wr->opcode = IB_WR_SEND;
1101        wr->send_flags = signal ? IB_SEND_SIGNALED : 0;
1102
1103        ib_ret = ib_post_send(ib_conn->qp, &tx_desc->wrs[0].send, NULL);
1104        if (ib_ret)
1105                iser_err("ib_post_send failed, ret:%d opcode:%d\n",
1106                         ib_ret, wr->opcode);
1107
1108        return ib_ret;
1109}
1110
1111u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
1112                             enum iser_data_dir cmd_dir, sector_t *sector)
1113{
1114        struct iser_mem_reg *reg = &iser_task->rdma_reg[cmd_dir];
1115        struct iser_fr_desc *desc = reg->mem_h;
1116        unsigned long sector_size = iser_task->sc->device->sector_size;
1117        struct ib_mr_status mr_status;
1118        int ret;
1119
1120        if (desc && desc->pi_ctx->sig_protected) {
1121                desc->pi_ctx->sig_protected = 0;
1122                ret = ib_check_mr_status(desc->pi_ctx->sig_mr,
1123                                         IB_MR_CHECK_SIG_STATUS, &mr_status);
1124                if (ret) {
1125                        pr_err("ib_check_mr_status failed, ret %d\n", ret);
1126                        goto err;
1127                }
1128
1129                if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1130                        sector_t sector_off = mr_status.sig_err.sig_err_offset;
1131
1132                        sector_div(sector_off, sector_size + 8);
1133                        *sector = scsi_get_lba(iser_task->sc) + sector_off;
1134
1135                        pr_err("PI error found type %d at sector %llx "
1136                               "expected %x vs actual %x\n",
1137                               mr_status.sig_err.err_type,
1138                               (unsigned long long)*sector,
1139                               mr_status.sig_err.expected,
1140                               mr_status.sig_err.actual);
1141
1142                        switch (mr_status.sig_err.err_type) {
1143                        case IB_SIG_BAD_GUARD:
1144                                return 0x1;
1145                        case IB_SIG_BAD_REFTAG:
1146                                return 0x3;
1147                        case IB_SIG_BAD_APPTAG:
1148                                return 0x2;
1149                        }
1150                }
1151        }
1152
1153        return 0;
1154err:
1155        /* Not alot we can do here, return ambiguous guard error */
1156        return 0x1;
1157}
1158
1159void iser_err_comp(struct ib_wc *wc, const char *type)
1160{
1161        if (wc->status != IB_WC_WR_FLUSH_ERR) {
1162                struct iser_conn *iser_conn = to_iser_conn(wc->qp->qp_context);
1163
1164                iser_err("%s failure: %s (%d) vend_err %#x\n", type,
1165                         ib_wc_status_msg(wc->status), wc->status,
1166                         wc->vendor_err);
1167
1168                if (iser_conn->iscsi_conn)
1169                        iscsi_conn_failure(iser_conn->iscsi_conn,
1170                                           ISCSI_ERR_CONN_FAILED);
1171        } else {
1172                iser_dbg("%s failure: %s (%d)\n", type,
1173                         ib_wc_status_msg(wc->status), wc->status);
1174        }
1175}
1176