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                  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  */
  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, 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         if (ib_register_event_handler(&device->event_handler))
 110                 goto cq_err;
 111
 112         return 0;
 113
 114 cq_err:
 115         for (i = 0; i < device->comps_used; i++) {
 116                 struct iser_comp *comp = &device->comps[i];
 117
 118                 if (comp->cq)
 119                         ib_free_cq(comp->cq);
 120         }
 121         ib_dealloc_pd(device->pd);
 122 pd_err:
 123         kfree(device->comps);
 124 comps_err:
 125         iser_err("failed to allocate an IB resource\n");
 126         return -1;
 127 }
 128
 129 /**
 130  * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
 131  * CQ and PD created with the device associated with the adapator.
 132  */
 133 static void iser_free_device_ib_res(struct iser_device *device)
 134 {
 135         int i;
 136
 137         for (i = 0; i < device->comps_used; i++) {
 138                 struct iser_comp *comp = &device->comps[i];
 139
 140                 ib_free_cq(comp->cq);
 141                 comp->cq = NULL;
 142         }
 143
 144         (void)ib_unregister_event_handler(&device->event_handler);
 145         ib_dealloc_pd(device->pd);
 146
 147         kfree(device->comps);
 148         device->comps = NULL;
 149         device->pd = NULL;
 150 }
 151
 152 /**
 153  * iser_alloc_fmr_pool - Creates FMR pool and page_vector
 154  *
 155  * returns 0 on success, or errno code on failure
 156  */
 157 int iser_alloc_fmr_pool(struct ib_conn *ib_conn,
 158                         unsigned cmds_max,
 159                         unsigned int size)
 160 {
 161         struct iser_device *device = ib_conn->device;
 162         struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
 163         struct iser_page_vec *page_vec;
 164         struct iser_fr_desc *desc;
 165         struct ib_fmr_pool *fmr_pool;
 166         struct ib_fmr_pool_param params;
 167         int ret;
 168
 169         INIT_LIST_HEAD(&fr_pool->list);
 170         spin_lock_init(&fr_pool->lock);
 171
 172         desc = kzalloc(sizeof(*desc), GFP_KERNEL);
 173         if (!desc)
 174                 return -ENOMEM;
 175
 176         page_vec = kmalloc(sizeof(*page_vec) + (sizeof(u64) * size),
 177                            GFP_KERNEL);
 178         if (!page_vec) {
 179                 ret = -ENOMEM;
 180                 goto err_frpl;
 181         }
 182
 183         page_vec->pages = (u64 *)(page_vec + 1);
 184
 185         params.page_shift        = SHIFT_4K;
 186         params.max_pages_per_fmr = size;
 187         /* make the pool size twice the max number of SCSI commands *
 188          * the ML is expected to queue, watermark for unmap at 50%  */
 189         params.pool_size         = cmds_max * 2;
 190         params.dirty_watermark   = cmds_max;
 191         params.cache             = 0;
 192         params.flush_function    = NULL;
 193         params.access            = (IB_ACCESS_LOCAL_WRITE  |
 194                                     IB_ACCESS_REMOTE_WRITE |
 195                                     IB_ACCESS_REMOTE_READ);
 196
 197         fmr_pool = ib_create_fmr_pool(device->pd, &params);
 198         if (IS_ERR(fmr_pool)) {
 199                 ret = PTR_ERR(fmr_pool);
 200                 iser_err("FMR allocation failed, err %d\n", ret);
 201                 goto err_fmr;
 202         }
 203
 204         desc->rsc.page_vec = page_vec;
 205         desc->rsc.fmr_pool = fmr_pool;
 206         list_add(&desc->list, &fr_pool->list);
 207
 208         return 0;
 209
 210 err_fmr:
 211         kfree(page_vec);
 212 err_frpl:
 213         kfree(desc);
 214
 215         return ret;
 216 }
 217
 218 /**
 219  * iser_free_fmr_pool - releases the FMR pool and page vec
 220  */
 221 void iser_free_fmr_pool(struct ib_conn *ib_conn)
 222 {
 223         struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
 224         struct iser_fr_desc *desc;
 225
 226         desc = list_first_entry(&fr_pool->list,
 227                                 struct iser_fr_desc, list);
 228         list_del(&desc->list);
 229
 230         iser_info("freeing conn %p fmr pool %p\n",
 231                   ib_conn, desc->rsc.fmr_pool);
 232
 233         ib_destroy_fmr_pool(desc->rsc.fmr_pool);
 234         kfree(desc->rsc.page_vec);
 235         kfree(desc);
 236 }
 237
 238 static int
 239 iser_alloc_reg_res(struct iser_device *device,
 240                    struct ib_pd *pd,
 241                    struct iser_reg_resources *res,
 242                    unsigned int size)
 243 {
 244         struct ib_device *ib_dev = device->ib_device;
 245         enum ib_mr_type mr_type;
 246         int ret;
 247
 248         if (ib_dev->attrs.device_cap_flags & IB_DEVICE_SG_GAPS_REG)
 249                 mr_type = IB_MR_TYPE_SG_GAPS;
 250         else
 251                 mr_type = IB_MR_TYPE_MEM_REG;
 252
 253         res->mr = ib_alloc_mr(pd, mr_type, size);
 254         if (IS_ERR(res->mr)) {
 255                 ret = PTR_ERR(res->mr);
 256                 iser_err("Failed to allocate ib_fast_reg_mr err=%d\n", ret);
 257                 return ret;
 258         }
 259         res->mr_valid = 0;
 260
 261         return 0;
 262 }
 263
 264 static void
 265 iser_free_reg_res(struct iser_reg_resources *rsc)
 266 {
 267         ib_dereg_mr(rsc->mr);
 268 }
 269
 270 static int
 271 iser_alloc_pi_ctx(struct iser_device *device,
 272                   struct ib_pd *pd,
 273                   struct iser_fr_desc *desc,
 274                   unsigned int size)
 275 {
 276         struct iser_pi_context *pi_ctx = NULL;
 277         int ret;
 278
 279         desc->pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
 280         if (!desc->pi_ctx)
 281                 return -ENOMEM;
 282
 283         pi_ctx = desc->pi_ctx;
 284
 285         ret = iser_alloc_reg_res(device, pd, &pi_ctx->rsc, size);
 286         if (ret) {
 287                 iser_err("failed to allocate reg_resources\n");
 288                 goto alloc_reg_res_err;
 289         }
 290
 291         pi_ctx->sig_mr = ib_alloc_mr(pd, IB_MR_TYPE_SIGNATURE, 2);
 292         if (IS_ERR(pi_ctx->sig_mr)) {
 293                 ret = PTR_ERR(pi_ctx->sig_mr);
 294                 goto sig_mr_failure;
 295         }
 296         pi_ctx->sig_mr_valid = 0;
 297         desc->pi_ctx->sig_protected = 0;
 298
 299         return 0;
 300
 301 sig_mr_failure:
 302         iser_free_reg_res(&pi_ctx->rsc);
 303 alloc_reg_res_err:
 304         kfree(desc->pi_ctx);
 305
 306         return ret;
 307 }
 308
 309 static void
 310 iser_free_pi_ctx(struct iser_pi_context *pi_ctx)
 311 {
 312         iser_free_reg_res(&pi_ctx->rsc);
 313         ib_dereg_mr(pi_ctx->sig_mr);
 314         kfree(pi_ctx);
 315 }
 316
 317 static struct iser_fr_desc *
 318 iser_create_fastreg_desc(struct iser_device *device,
 319                          struct ib_pd *pd,
 320                          bool pi_enable,
 321                          unsigned int size)
 322 {
 323         struct iser_fr_desc *desc;
 324         int ret;
 325
 326         desc = kzalloc(sizeof(*desc), GFP_KERNEL);
 327         if (!desc)
 328                 return ERR_PTR(-ENOMEM);
 329
 330         ret = iser_alloc_reg_res(device, pd, &desc->rsc, size);
 331         if (ret)
 332                 goto reg_res_alloc_failure;
 333
 334         if (pi_enable) {
 335                 ret = iser_alloc_pi_ctx(device, pd, desc, size);
 336                 if (ret)
 337                         goto pi_ctx_alloc_failure;
 338         }
 339
 340         return desc;
 341
 342 pi_ctx_alloc_failure:
 343         iser_free_reg_res(&desc->rsc);
 344 reg_res_alloc_failure:
 345         kfree(desc);
 346
 347         return ERR_PTR(ret);
 348 }
 349
 350 /**
 351  * iser_alloc_fastreg_pool - Creates pool of fast_reg descriptors
 352  * for fast registration work requests.
 353  * returns 0 on success, or errno code on failure
 354  */
 355 int iser_alloc_fastreg_pool(struct ib_conn *ib_conn,
 356                             unsigned cmds_max,
 357                             unsigned int size)
 358 {
 359         struct iser_device *device = ib_conn->device;
 360         struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
 361         struct iser_fr_desc *desc;
 362         int i, ret;
 363
 364         INIT_LIST_HEAD(&fr_pool->list);
 365         INIT_LIST_HEAD(&fr_pool->all_list);
 366         spin_lock_init(&fr_pool->lock);
 367         fr_pool->size = 0;
 368         for (i = 0; i < cmds_max; i++) {
 369                 desc = iser_create_fastreg_desc(device, device->pd,
 370                                                 ib_conn->pi_support, size);
 371                 if (IS_ERR(desc)) {
 372                         ret = PTR_ERR(desc);
 373                         goto err;
 374                 }
 375
 376                 list_add_tail(&desc->list, &fr_pool->list);
 377                 list_add_tail(&desc->all_list, &fr_pool->all_list);
 378                 fr_pool->size++;
 379         }
 380
 381         return 0;
 382
 383 err:
 384         iser_free_fastreg_pool(ib_conn);
 385         return ret;
 386 }
 387
 388 /**
 389  * iser_free_fastreg_pool - releases the pool of fast_reg descriptors
 390  */
 391 void iser_free_fastreg_pool(struct ib_conn *ib_conn)
 392 {
 393         struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
 394         struct iser_fr_desc *desc, *tmp;
 395         int i = 0;
 396
 397         if (list_empty(&fr_pool->all_list))
 398                 return;
 399
 400         iser_info("freeing conn %p fr pool\n", ib_conn);
 401
 402         list_for_each_entry_safe(desc, tmp, &fr_pool->all_list, all_list) {
 403                 list_del(&desc->all_list);
 404                 iser_free_reg_res(&desc->rsc);
 405                 if (desc->pi_ctx)
 406                         iser_free_pi_ctx(desc->pi_ctx);
 407                 kfree(desc);
 408                 ++i;
 409         }
 410
 411         if (i < fr_pool->size)
 412                 iser_warn("pool still has %d regions registered\n",
 413                           fr_pool->size - i);
 414 }
 415
 416 /**
 417  * iser_create_ib_conn_res - Queue-Pair (QP)
 418  *
 419  * returns 0 on success, -1 on failure
 420  */
 421 static int iser_create_ib_conn_res(struct ib_conn *ib_conn)
 422 {
 423         struct iser_conn *iser_conn = to_iser_conn(ib_conn);
 424         struct iser_device      *device;
 425         struct ib_device        *ib_dev;
 426         struct ib_qp_init_attr  init_attr;
 427         int                     ret = -ENOMEM;
 428         int index, min_index = 0;
 429
 430         BUG_ON(ib_conn->device == NULL);
 431
 432         device = ib_conn->device;
 433         ib_dev = device->ib_device;
 434
 435         memset(&init_attr, 0, sizeof init_attr);
 436
 437         mutex_lock(&ig.connlist_mutex);
 438         /* select the CQ with the minimal number of usages */
 439         for (index = 0; index < device->comps_used; index++) {
 440                 if (device->comps[index].active_qps <
 441                     device->comps[min_index].active_qps)
 442                         min_index = index;
 443         }
 444         ib_conn->comp = &device->comps[min_index];
 445         ib_conn->comp->active_qps++;
 446         mutex_unlock(&ig.connlist_mutex);
 447         iser_info("cq index %d used for ib_conn %p\n", min_index, ib_conn);
 448
 449         init_attr.event_handler = iser_qp_event_callback;
 450         init_attr.qp_context    = (void *)ib_conn;
 451         init_attr.send_cq       = ib_conn->comp->cq;
 452         init_attr.recv_cq       = ib_conn->comp->cq;
 453         init_attr.cap.max_recv_wr  = ISER_QP_MAX_RECV_DTOS;
 454         init_attr.cap.max_send_sge = 2;
 455         init_attr.cap.max_recv_sge = 1;
 456         init_attr.sq_sig_type   = IB_SIGNAL_REQ_WR;
 457         init_attr.qp_type       = IB_QPT_RC;
 458         if (ib_conn->pi_support) {
 459                 init_attr.cap.max_send_wr = ISER_QP_SIG_MAX_REQ_DTOS + 1;
 460                 init_attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
 461                 iser_conn->max_cmds =
 462                         ISER_GET_MAX_XMIT_CMDS(ISER_QP_SIG_MAX_REQ_DTOS);
 463         } else {
 464                 if (ib_dev->attrs.max_qp_wr > ISER_QP_MAX_REQ_DTOS) {
 465                         init_attr.cap.max_send_wr  = ISER_QP_MAX_REQ_DTOS + 1;
 466                         iser_conn->max_cmds =
 467                                 ISER_GET_MAX_XMIT_CMDS(ISER_QP_MAX_REQ_DTOS);
 468                 } else {
 469                         init_attr.cap.max_send_wr = ib_dev->attrs.max_qp_wr;
 470                         iser_conn->max_cmds =
 471                                 ISER_GET_MAX_XMIT_CMDS(ib_dev->attrs.max_qp_wr);
 472                         iser_dbg("device %s supports max_send_wr %d\n",
 473                                  device->ib_device->name, ib_dev->attrs.max_qp_wr);
 474                 }
 475         }
 476
 477         ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
 478         if (ret)
 479                 goto out_err;
 480
 481         ib_conn->qp = ib_conn->cma_id->qp;
 482         iser_info("setting conn %p cma_id %p qp %p\n",
 483                   ib_conn, ib_conn->cma_id,
 484                   ib_conn->cma_id->qp);
 485         return ret;
 486
 487 out_err:
 488         mutex_lock(&ig.connlist_mutex);
 489         ib_conn->comp->active_qps--;
 490         mutex_unlock(&ig.connlist_mutex);
 491         iser_err("unable to alloc mem or create resource, err %d\n", ret);
 492
 493         return ret;
 494 }
 495
 496 /**
 497  * based on the resolved device node GUID see if there already allocated
 498  * device for this device. If there's no such, create one.
 499  */
 500 static
 501 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
 502 {
 503         struct iser_device *device;
 504
 505         mutex_lock(&ig.device_list_mutex);
 506
 507         list_for_each_entry(device, &ig.device_list, ig_list)
 508                 /* find if there's a match using the node GUID */
 509                 if (device->ib_device->node_guid == cma_id->device->node_guid)
 510                         goto inc_refcnt;
 511
 512         device = kzalloc(sizeof *device, GFP_KERNEL);
 513         if (device == NULL)
 514                 goto out;
 515
 516         /* assign this device to the device */
 517         device->ib_device = cma_id->device;
 518         /* init the device and link it into ig device list */
 519         if (iser_create_device_ib_res(device)) {
 520                 kfree(device);
 521                 device = NULL;
 522                 goto out;
 523         }
 524         list_add(&device->ig_list, &ig.device_list);
 525
 526 inc_refcnt:
 527         device->refcount++;
 528 out:
 529         mutex_unlock(&ig.device_list_mutex);
 530         return device;
 531 }
 532
 533 /* if there's no demand for this device, release it */
 534 static void iser_device_try_release(struct iser_device *device)
 535 {
 536         mutex_lock(&ig.device_list_mutex);
 537         device->refcount--;
 538         iser_info("device %p refcount %d\n", device, device->refcount);
 539         if (!device->refcount) {
 540                 iser_free_device_ib_res(device);
 541                 list_del(&device->ig_list);
 542                 kfree(device);
 543         }
 544         mutex_unlock(&ig.device_list_mutex);
 545 }
 546
 547 /**
 548  * Called with state mutex held
 549  **/
 550 static int iser_conn_state_comp_exch(struct iser_conn *iser_conn,
 551                                      enum iser_conn_state comp,
 552                                      enum iser_conn_state exch)
 553 {
 554         int ret;
 555
 556         ret = (iser_conn->state == comp);
 557         if (ret)
 558                 iser_conn->state = exch;
 559
 560         return ret;
 561 }
 562
 563 void iser_release_work(struct work_struct *work)
 564 {
 565         struct iser_conn *iser_conn;
 566
 567         iser_conn = container_of(work, struct iser_conn, release_work);
 568
 569         /* Wait for conn_stop to complete */
 570         wait_for_completion(&iser_conn->stop_completion);
 571         /* Wait for IB resouces cleanup to complete */
 572         wait_for_completion(&iser_conn->ib_completion);
 573
 574         mutex_lock(&iser_conn->state_mutex);
 575         iser_conn->state = ISER_CONN_DOWN;
 576         mutex_unlock(&iser_conn->state_mutex);
 577
 578         iser_conn_release(iser_conn);
 579 }
 580
 581 /**
 582  * iser_free_ib_conn_res - release IB related resources
 583  * @iser_conn: iser connection struct
 584  * @destroy: indicator if we need to try to release the
 585  *     iser device and memory regoins pool (only iscsi
 586  *     shutdown and DEVICE_REMOVAL will use this).
 587  *
 588  * This routine is called with the iser state mutex held
 589  * so the cm_id removal is out of here. It is Safe to
 590  * be invoked multiple times.
 591  */
 592 static void iser_free_ib_conn_res(struct iser_conn *iser_conn,
 593                                   bool destroy)
 594 {
 595         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 596         struct iser_device *device = ib_conn->device;
 597
 598         iser_info("freeing conn %p cma_id %p qp %p\n",
 599                   iser_conn, ib_conn->cma_id, ib_conn->qp);
 600
 601         if (ib_conn->qp != NULL) {
 602                 mutex_lock(&ig.connlist_mutex);
 603                 ib_conn->comp->active_qps--;
 604                 mutex_unlock(&ig.connlist_mutex);
 605                 rdma_destroy_qp(ib_conn->cma_id);
 606                 ib_conn->qp = NULL;
 607         }
 608
 609         if (destroy) {
 610                 if (iser_conn->rx_descs)
 611                         iser_free_rx_descriptors(iser_conn);
 612
 613                 if (device != NULL) {
 614                         iser_device_try_release(device);
 615                         ib_conn->device = NULL;
 616                 }
 617         }
 618 }
 619
 620 /**
 621  * Frees all conn objects and deallocs conn descriptor
 622  */
 623 void iser_conn_release(struct iser_conn *iser_conn)
 624 {
 625         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 626
 627         mutex_lock(&ig.connlist_mutex);
 628         list_del(&iser_conn->conn_list);
 629         mutex_unlock(&ig.connlist_mutex);
 630
 631         mutex_lock(&iser_conn->state_mutex);
 632         /* In case we endup here without ep_disconnect being invoked. */
 633         if (iser_conn->state != ISER_CONN_DOWN) {
 634                 iser_warn("iser conn %p state %d, expected state down.\n",
 635                           iser_conn, iser_conn->state);
 636                 iscsi_destroy_endpoint(iser_conn->ep);
 637                 iser_conn->state = ISER_CONN_DOWN;
 638         }
 639         /*
 640          * In case we never got to bind stage, we still need to
 641          * release IB resources (which is safe to call more than once).
 642          */
 643         iser_free_ib_conn_res(iser_conn, true);
 644         mutex_unlock(&iser_conn->state_mutex);
 645
 646         if (ib_conn->cma_id != NULL) {
 647                 rdma_destroy_id(ib_conn->cma_id);
 648                 ib_conn->cma_id = NULL;
 649         }
 650
 651         kfree(iser_conn);
 652 }
 653
 654 /**
 655  * triggers start of the disconnect procedures and wait for them to be done
 656  * Called with state mutex held
 657  */
 658 int iser_conn_terminate(struct iser_conn *iser_conn)
 659 {
 660         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 661         int err = 0;
 662
 663         /* terminate the iser conn only if the conn state is UP */
 664         if (!iser_conn_state_comp_exch(iser_conn, ISER_CONN_UP,
 665                                        ISER_CONN_TERMINATING))
 666                 return 0;
 667
 668         iser_info("iser_conn %p state %d\n", iser_conn, iser_conn->state);
 669
 670         /* suspend queuing of new iscsi commands */
 671         if (iser_conn->iscsi_conn)
 672                 iscsi_suspend_queue(iser_conn->iscsi_conn);
 673
 674         /*
 675          * In case we didn't already clean up the cma_id (peer initiated
 676          * a disconnection), we need to Cause the CMA to change the QP
 677          * state to ERROR.
 678          */
 679         if (ib_conn->cma_id) {
 680                 err = rdma_disconnect(ib_conn->cma_id);
 681                 if (err)
 682                         iser_err("Failed to disconnect, conn: 0x%p err %d\n",
 683                                  iser_conn, err);
 684
 685                 /* block until all flush errors are consumed */
 686                 ib_drain_sq(ib_conn->qp);
 687         }
 688
 689         return 1;
 690 }
 691
 692 /**
 693  * Called with state mutex held
 694  **/
 695 static void iser_connect_error(struct rdma_cm_id *cma_id)
 696 {
 697         struct iser_conn *iser_conn;
 698
 699         iser_conn = (struct iser_conn *)cma_id->context;
 700         iser_conn->state = ISER_CONN_TERMINATING;
 701 }
 702
 703 static void
 704 iser_calc_scsi_params(struct iser_conn *iser_conn,
 705                       unsigned int max_sectors)
 706 {
 707         struct iser_device *device = iser_conn->ib_conn.device;
 708         unsigned short sg_tablesize, sup_sg_tablesize;
 709
 710         sg_tablesize = DIV_ROUND_UP(max_sectors * 512, SIZE_4K);
 711         if (device->ib_device->attrs.device_cap_flags &
 712                         IB_DEVICE_MEM_MGT_EXTENSIONS)
 713                 sup_sg_tablesize =
 714                         min_t(
 715                          uint, ISCSI_ISER_MAX_SG_TABLESIZE,
 716                          device->ib_device->attrs.max_fast_reg_page_list_len);
 717         else
 718                 sup_sg_tablesize = ISCSI_ISER_MAX_SG_TABLESIZE;
 719
 720         iser_conn->scsi_sg_tablesize = min(sg_tablesize, sup_sg_tablesize);
 721 }
 722
 723 /**
 724  * Called with state mutex held
 725  **/
 726 static void iser_addr_handler(struct rdma_cm_id *cma_id)
 727 {
 728         struct iser_device *device;
 729         struct iser_conn   *iser_conn;
 730         struct ib_conn   *ib_conn;
 731         int    ret;
 732
 733         iser_conn = (struct iser_conn *)cma_id->context;
 734         if (iser_conn->state != ISER_CONN_PENDING)
 735                 /* bailout */
 736                 return;
 737
 738         ib_conn = &iser_conn->ib_conn;
 739         device = iser_device_find_by_ib_device(cma_id);
 740         if (!device) {
 741                 iser_err("device lookup/creation failed\n");
 742                 iser_connect_error(cma_id);
 743                 return;
 744         }
 745
 746         ib_conn->device = device;
 747
 748         /* connection T10-PI support */
 749         if (iser_pi_enable) {
 750                 if (!(device->ib_device->attrs.device_cap_flags &
 751                       IB_DEVICE_SIGNATURE_HANDOVER)) {
 752                         iser_warn("T10-PI requested but not supported on %s, "
 753                                   "continue without T10-PI\n",
 754                                   ib_conn->device->ib_device->name);
 755                         ib_conn->pi_support = false;
 756                 } else {
 757                         ib_conn->pi_support = true;
 758                 }
 759         }
 760
 761         iser_calc_scsi_params(iser_conn, iser_max_sectors);
 762
 763         ret = rdma_resolve_route(cma_id, 1000);
 764         if (ret) {
 765                 iser_err("resolve route failed: %d\n", ret);
 766                 iser_connect_error(cma_id);
 767                 return;
 768         }
 769 }
 770
 771 /**
 772  * Called with state mutex held
 773  **/
 774 static void iser_route_handler(struct rdma_cm_id *cma_id)
 775 {
 776         struct rdma_conn_param conn_param;
 777         int    ret;
 778         struct iser_cm_hdr req_hdr;
 779         struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
 780         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 781         struct iser_device *device = ib_conn->device;
 782
 783         if (iser_conn->state != ISER_CONN_PENDING)
 784                 /* bailout */
 785                 return;
 786
 787         ret = iser_create_ib_conn_res(ib_conn);
 788         if (ret)
 789                 goto failure;
 790
 791         memset(&conn_param, 0, sizeof conn_param);
 792         conn_param.responder_resources = device->ib_device->attrs.max_qp_rd_atom;
 793         conn_param.initiator_depth     = 1;
 794         conn_param.retry_count         = 7;
 795         conn_param.rnr_retry_count     = 6;
 796
 797         memset(&req_hdr, 0, sizeof(req_hdr));
 798         req_hdr.flags = ISER_ZBVA_NOT_SUP;
 799         if (!device->remote_inv_sup)
 800                 req_hdr.flags |= ISER_SEND_W_INV_NOT_SUP;
 801         conn_param.private_data = (void *)&req_hdr;
 802         conn_param.private_data_len = sizeof(struct iser_cm_hdr);
 803
 804         ret = rdma_connect(cma_id, &conn_param);
 805         if (ret) {
 806                 iser_err("failure connecting: %d\n", ret);
 807                 goto failure;
 808         }
 809
 810         return;
 811 failure:
 812         iser_connect_error(cma_id);
 813 }
 814
 815 static void iser_connected_handler(struct rdma_cm_id *cma_id,
 816                                    const void *private_data)
 817 {
 818         struct iser_conn *iser_conn;
 819         struct ib_qp_attr attr;
 820         struct ib_qp_init_attr init_attr;
 821
 822         iser_conn = (struct iser_conn *)cma_id->context;
 823         if (iser_conn->state != ISER_CONN_PENDING)
 824                 /* bailout */
 825                 return;
 826
 827         (void)ib_query_qp(cma_id->qp, &attr, ~0, &init_attr);
 828         iser_info("remote qpn:%x my qpn:%x\n", attr.dest_qp_num, cma_id->qp->qp_num);
 829
 830         if (private_data) {
 831                 u8 flags = *(u8 *)private_data;
 832
 833                 iser_conn->snd_w_inv = !(flags & ISER_SEND_W_INV_NOT_SUP);
 834         }
 835
 836         iser_info("conn %p: negotiated %s invalidation\n",
 837                   iser_conn, iser_conn->snd_w_inv ? "remote" : "local");
 838
 839         iser_conn->state = ISER_CONN_UP;
 840         complete(&iser_conn->up_completion);
 841 }
 842
 843 static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
 844 {
 845         struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
 846
 847         if (iser_conn_terminate(iser_conn)) {
 848                 if (iser_conn->iscsi_conn)
 849                         iscsi_conn_failure(iser_conn->iscsi_conn,
 850                                            ISCSI_ERR_CONN_FAILED);
 851                 else
 852                         iser_err("iscsi_iser connection isn't bound\n");
 853         }
 854 }
 855
 856 static void iser_cleanup_handler(struct rdma_cm_id *cma_id,
 857                                  bool destroy)
 858 {
 859         struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
 860
 861         /*
 862          * We are not guaranteed that we visited disconnected_handler
 863          * by now, call it here to be safe that we handle CM drep
 864          * and flush errors.
 865          */
 866         iser_disconnected_handler(cma_id);
 867         iser_free_ib_conn_res(iser_conn, destroy);
 868         complete(&iser_conn->ib_completion);
 869 };
 870
 871 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 872 {
 873         struct iser_conn *iser_conn;
 874         int ret = 0;
 875
 876         iser_conn = (struct iser_conn *)cma_id->context;
 877         iser_info("%s (%d): status %d conn %p id %p\n",
 878                   rdma_event_msg(event->event), event->event,
 879                   event->status, cma_id->context, cma_id);
 880
 881         mutex_lock(&iser_conn->state_mutex);
 882         switch (event->event) {
 883         case RDMA_CM_EVENT_ADDR_RESOLVED:
 884                 iser_addr_handler(cma_id);
 885                 break;
 886         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 887                 iser_route_handler(cma_id);
 888                 break;
 889         case RDMA_CM_EVENT_ESTABLISHED:
 890                 iser_connected_handler(cma_id, event->param.conn.private_data);
 891                 break;
 892         case RDMA_CM_EVENT_REJECTED:
 893                 iser_info("Connection rejected: %s\n",
 894                          rdma_reject_msg(cma_id, event->status));
 895                 /* FALLTHROUGH */
 896         case RDMA_CM_EVENT_ADDR_ERROR:
 897         case RDMA_CM_EVENT_ROUTE_ERROR:
 898         case RDMA_CM_EVENT_CONNECT_ERROR:
 899         case RDMA_CM_EVENT_UNREACHABLE:
 900                 iser_connect_error(cma_id);
 901                 break;
 902         case RDMA_CM_EVENT_DISCONNECTED:
 903         case RDMA_CM_EVENT_ADDR_CHANGE:
 904         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
 905                 iser_cleanup_handler(cma_id, false);
 906                 break;
 907         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 908                 /*
 909                  * we *must* destroy the device as we cannot rely
 910                  * on iscsid to be around to initiate error handling.
 911                  * also if we are not in state DOWN implicitly destroy
 912                  * the cma_id.
 913                  */
 914                 iser_cleanup_handler(cma_id, true);
 915                 if (iser_conn->state != ISER_CONN_DOWN) {
 916                         iser_conn->ib_conn.cma_id = NULL;
 917                         ret = 1;
 918                 }
 919                 break;
 920         default:
 921                 iser_err("Unexpected RDMA CM event: %s (%d)\n",
 922                          rdma_event_msg(event->event), event->event);
 923                 break;
 924         }
 925         mutex_unlock(&iser_conn->state_mutex);
 926
 927         return ret;
 928 }
 929
 930 void iser_conn_init(struct iser_conn *iser_conn)
 931 {
 932         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 933
 934         iser_conn->state = ISER_CONN_INIT;
 935         init_completion(&iser_conn->stop_completion);
 936         init_completion(&iser_conn->ib_completion);
 937         init_completion(&iser_conn->up_completion);
 938         INIT_LIST_HEAD(&iser_conn->conn_list);
 939         mutex_init(&iser_conn->state_mutex);
 940
 941         ib_conn->post_recv_buf_count = 0;
 942         ib_conn->reg_cqe.done = iser_reg_comp;
 943 }
 944
 945  /**
 946  * starts the process of connecting to the target
 947  * sleeps until the connection is established or rejected
 948  */
 949 int iser_connect(struct iser_conn   *iser_conn,
 950                  struct sockaddr    *src_addr,
 951                  struct sockaddr    *dst_addr,
 952                  int                 non_blocking)
 953 {
 954         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 955         int err = 0;
 956
 957         mutex_lock(&iser_conn->state_mutex);
 958
 959         sprintf(iser_conn->name, "%pISp", dst_addr);
 960
 961         iser_info("connecting to: %s\n", iser_conn->name);
 962
 963         /* the device is known only --after-- address resolution */
 964         ib_conn->device = NULL;
 965
 966         iser_conn->state = ISER_CONN_PENDING;
 967
 968         ib_conn->cma_id = rdma_create_id(&init_net, iser_cma_handler,
 969                                          (void *)iser_conn,
 970                                          RDMA_PS_TCP, IB_QPT_RC);
 971         if (IS_ERR(ib_conn->cma_id)) {
 972                 err = PTR_ERR(ib_conn->cma_id);
 973                 iser_err("rdma_create_id failed: %d\n", err);
 974                 goto id_failure;
 975         }
 976
 977         err = rdma_resolve_addr(ib_conn->cma_id, src_addr, dst_addr, 1000);
 978         if (err) {
 979                 iser_err("rdma_resolve_addr failed: %d\n", err);
 980                 goto addr_failure;
 981         }
 982
 983         if (!non_blocking) {
 984                 wait_for_completion_interruptible(&iser_conn->up_completion);
 985
 986                 if (iser_conn->state != ISER_CONN_UP) {
 987                         err =  -EIO;
 988                         goto connect_failure;
 989                 }
 990         }
 991         mutex_unlock(&iser_conn->state_mutex);
 992
 993         mutex_lock(&ig.connlist_mutex);
 994         list_add(&iser_conn->conn_list, &ig.connlist);
 995         mutex_unlock(&ig.connlist_mutex);
 996         return 0;
 997
 998 id_failure:
 999         ib_conn->cma_id = NULL;
1000 addr_failure:
1001         iser_conn->state = ISER_CONN_DOWN;
1002 connect_failure:
1003         mutex_unlock(&iser_conn->state_mutex);
1004         iser_conn_release(iser_conn);
1005         return err;
1006 }
1007
1008 int iser_post_recvl(struct iser_conn *iser_conn)
1009 {
1010         struct ib_conn *ib_conn = &iser_conn->ib_conn;
1011         struct iser_login_desc *desc = &iser_conn->login_desc;
1012         struct ib_recv_wr wr, *wr_failed;
1013         int ib_ret;
1014
1015         desc->sge.addr = desc->rsp_dma;
1016         desc->sge.length = ISER_RX_LOGIN_SIZE;
1017         desc->sge.lkey = ib_conn->device->pd->local_dma_lkey;
1018
1019         desc->cqe.done = iser_login_rsp;
1020         wr.wr_cqe = &desc->cqe;
1021         wr.sg_list = &desc->sge;
1022         wr.num_sge = 1;
1023         wr.next = NULL;
1024
1025         ib_conn->post_recv_buf_count++;
1026         ib_ret = ib_post_recv(ib_conn->qp, &wr, &wr_failed);
1027         if (ib_ret) {
1028                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1029                 ib_conn->post_recv_buf_count--;
1030         }
1031
1032         return ib_ret;
1033 }
1034
1035 int iser_post_recvm(struct iser_conn *iser_conn, int count)
1036 {
1037         struct ib_conn *ib_conn = &iser_conn->ib_conn;
1038         unsigned int my_rx_head = iser_conn->rx_desc_head;
1039         struct iser_rx_desc *rx_desc;
1040         struct ib_recv_wr *wr, *wr_failed;
1041         int i, ib_ret;
1042
1043         for (wr = ib_conn->rx_wr, i = 0; i < count; i++, wr++) {
1044                 rx_desc = &iser_conn->rx_descs[my_rx_head];
1045                 rx_desc->cqe.done = iser_task_rsp;
1046                 wr->wr_cqe = &rx_desc->cqe;
1047                 wr->sg_list = &rx_desc->rx_sg;
1048                 wr->num_sge = 1;
1049                 wr->next = wr + 1;
1050                 my_rx_head = (my_rx_head + 1) & iser_conn->qp_max_recv_dtos_mask;
1051         }
1052
1053         wr--;
1054         wr->next = NULL; /* mark end of work requests list */
1055
1056         ib_conn->post_recv_buf_count += count;
1057         ib_ret = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &wr_failed);
1058         if (ib_ret) {
1059                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1060                 ib_conn->post_recv_buf_count -= count;
1061         } else
1062                 iser_conn->rx_desc_head = my_rx_head;
1063
1064         return ib_ret;
1065 }
1066
1067
1068 /**
1069  * iser_start_send - Initiate a Send DTO operation
1070  *
1071  * returns 0 on success, -1 on failure
1072  */
1073 int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc,
1074                    bool signal)
1075 {
1076         struct ib_send_wr *bad_wr, *wr = iser_tx_next_wr(tx_desc);
1077         int ib_ret;
1078
1079         ib_dma_sync_single_for_device(ib_conn->device->ib_device,
1080                                       tx_desc->dma_addr, ISER_HEADERS_LEN,
1081                                       DMA_TO_DEVICE);
1082
1083         wr->next = NULL;
1084         wr->wr_cqe = &tx_desc->cqe;
1085         wr->sg_list = tx_desc->tx_sg;
1086         wr->num_sge = tx_desc->num_sge;
1087         wr->opcode = IB_WR_SEND;
1088         wr->send_flags = signal ? IB_SEND_SIGNALED : 0;
1089
1090         ib_ret = ib_post_send(ib_conn->qp, &tx_desc->wrs[0].send, &bad_wr);
1091         if (ib_ret)
1092                 iser_err("ib_post_send failed, ret:%d opcode:%d\n",
1093                          ib_ret, bad_wr->opcode);
1094
1095         return ib_ret;
1096 }
1097
1098 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
1099                              enum iser_data_dir cmd_dir, sector_t *sector)
1100 {
1101         struct iser_mem_reg *reg = &iser_task->rdma_reg[cmd_dir];
1102         struct iser_fr_desc *desc = reg->mem_h;
1103         unsigned long sector_size = iser_task->sc->device->sector_size;
1104         struct ib_mr_status mr_status;
1105         int ret;
1106
1107         if (desc && desc->pi_ctx->sig_protected) {
1108                 desc->pi_ctx->sig_protected = 0;
1109                 ret = ib_check_mr_status(desc->pi_ctx->sig_mr,
1110                                          IB_MR_CHECK_SIG_STATUS, &mr_status);
1111                 if (ret) {
1112                         pr_err("ib_check_mr_status failed, ret %d\n", ret);
1113                         goto err;
1114                 }
1115
1116                 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1117                         sector_t sector_off = mr_status.sig_err.sig_err_offset;
1118
1119                         sector_div(sector_off, sector_size + 8);
1120                         *sector = scsi_get_lba(iser_task->sc) + sector_off;
1121
1122                         pr_err("PI error found type %d at sector %llx "
1123                                "expected %x vs actual %x\n",
1124                                mr_status.sig_err.err_type,
1125                                (unsigned long long)*sector,
1126                                mr_status.sig_err.expected,
1127                                mr_status.sig_err.actual);
1128
1129                         switch (mr_status.sig_err.err_type) {
1130                         case IB_SIG_BAD_GUARD:
1131                                 return 0x1;
1132                         case IB_SIG_BAD_REFTAG:
1133                                 return 0x3;
1134                         case IB_SIG_BAD_APPTAG:
1135                                 return 0x2;
1136                         }
1137                 }
1138         }
1139
1140         return 0;
1141 err:
1142         /* Not alot we can do here, return ambiguous guard error */
1143         return 0x1;
1144 }
1145
1146 void iser_err_comp(struct ib_wc *wc, const char *type)
1147 {
1148         if (wc->status != IB_WC_WR_FLUSH_ERR) {
1149                 struct iser_conn *iser_conn = to_iser_conn(wc->qp->qp_context);
1150
1151                 iser_err("%s failure: %s (%d) vend_err %x\n", type,
1152                          ib_wc_status_msg(wc->status), wc->status,
1153                          wc->vendor_err);
1154
1155                 if (iser_conn->iscsi_conn)
1156                         iscsi_conn_failure(iser_conn->iscsi_conn,
1157                                            ISCSI_ERR_CONN_FAILED);
1158         } else {
1159                 iser_dbg("%s failure: %s (%d)\n", type,
1160                          ib_wc_status_msg(wc->status), wc->status);
1161         }
1162 }