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         spin_lock_init(&fr_pool->lock);
 366         fr_pool->size = 0;
 367         for (i = 0; i < cmds_max; i++) {
 368                 desc = iser_create_fastreg_desc(device, device->pd,
 369                                                 ib_conn->pi_support, size);
 370                 if (IS_ERR(desc)) {
 371                         ret = PTR_ERR(desc);
 372                         goto err;
 373                 }
 374
 375                 list_add_tail(&desc->list, &fr_pool->list);
 376                 fr_pool->size++;
 377         }
 378
 379         return 0;
 380
 381 err:
 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  */
 389 void 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->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->list, list) {
 401                 list_del(&desc->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  */
 419 static 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
 485 out_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  */
 498 static
 499 struct 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
 524 inc_refcnt:
 525         device->refcount++;
 526 out:
 527         mutex_unlock(&ig.device_list_mutex);
 528         return device;
 529 }
 530
 531 /* if there's no demand for this device, release it */
 532 static 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  **/
 548 static 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
 561 void 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  */
 590 static 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  */
 621 void 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  */
 656 int 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  **/
 693 static 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
 701 static void
 702 iser_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         unsigned short sg_tablesize, sup_sg_tablesize;
 707
 708         sg_tablesize = DIV_ROUND_UP(max_sectors * 512, SIZE_4K);
 709         sup_sg_tablesize = min_t(unsigned, ISCSI_ISER_MAX_SG_TABLESIZE,
 710                                  device->ib_device->attrs.max_fast_reg_page_list_len);
 711
 712         iser_conn->scsi_sg_tablesize = min(sg_tablesize, sup_sg_tablesize);
 713 }
 714
 715 /**
 716  * Called with state mutex held
 717  **/
 718 static void iser_addr_handler(struct rdma_cm_id *cma_id)
 719 {
 720         struct iser_device *device;
 721         struct iser_conn   *iser_conn;
 722         struct ib_conn   *ib_conn;
 723         int    ret;
 724
 725         iser_conn = (struct iser_conn *)cma_id->context;
 726         if (iser_conn->state != ISER_CONN_PENDING)
 727                 /* bailout */
 728                 return;
 729
 730         ib_conn = &iser_conn->ib_conn;
 731         device = iser_device_find_by_ib_device(cma_id);
 732         if (!device) {
 733                 iser_err("device lookup/creation failed\n");
 734                 iser_connect_error(cma_id);
 735                 return;
 736         }
 737
 738         ib_conn->device = device;
 739
 740         /* connection T10-PI support */
 741         if (iser_pi_enable) {
 742                 if (!(device->ib_device->attrs.device_cap_flags &
 743                       IB_DEVICE_SIGNATURE_HANDOVER)) {
 744                         iser_warn("T10-PI requested but not supported on %s, "
 745                                   "continue without T10-PI\n",
 746                                   ib_conn->device->ib_device->name);
 747                         ib_conn->pi_support = false;
 748                 } else {
 749                         ib_conn->pi_support = true;
 750                 }
 751         }
 752
 753         iser_calc_scsi_params(iser_conn, iser_max_sectors);
 754
 755         ret = rdma_resolve_route(cma_id, 1000);
 756         if (ret) {
 757                 iser_err("resolve route failed: %d\n", ret);
 758                 iser_connect_error(cma_id);
 759                 return;
 760         }
 761 }
 762
 763 /**
 764  * Called with state mutex held
 765  **/
 766 static void iser_route_handler(struct rdma_cm_id *cma_id)
 767 {
 768         struct rdma_conn_param conn_param;
 769         int    ret;
 770         struct iser_cm_hdr req_hdr;
 771         struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
 772         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 773         struct iser_device *device = ib_conn->device;
 774
 775         if (iser_conn->state != ISER_CONN_PENDING)
 776                 /* bailout */
 777                 return;
 778
 779         ret = iser_create_ib_conn_res(ib_conn);
 780         if (ret)
 781                 goto failure;
 782
 783         memset(&conn_param, 0, sizeof conn_param);
 784         conn_param.responder_resources = device->ib_device->attrs.max_qp_rd_atom;
 785         conn_param.initiator_depth     = 1;
 786         conn_param.retry_count         = 7;
 787         conn_param.rnr_retry_count     = 6;
 788
 789         memset(&req_hdr, 0, sizeof(req_hdr));
 790         req_hdr.flags = ISER_ZBVA_NOT_SUP;
 791         if (!device->remote_inv_sup)
 792                 req_hdr.flags |= ISER_SEND_W_INV_NOT_SUP;
 793         conn_param.private_data = (void *)&req_hdr;
 794         conn_param.private_data_len = sizeof(struct iser_cm_hdr);
 795
 796         ret = rdma_connect(cma_id, &conn_param);
 797         if (ret) {
 798                 iser_err("failure connecting: %d\n", ret);
 799                 goto failure;
 800         }
 801
 802         return;
 803 failure:
 804         iser_connect_error(cma_id);
 805 }
 806
 807 static void iser_connected_handler(struct rdma_cm_id *cma_id,
 808                                    const void *private_data)
 809 {
 810         struct iser_conn *iser_conn;
 811         struct ib_qp_attr attr;
 812         struct ib_qp_init_attr init_attr;
 813
 814         iser_conn = (struct iser_conn *)cma_id->context;
 815         if (iser_conn->state != ISER_CONN_PENDING)
 816                 /* bailout */
 817                 return;
 818
 819         (void)ib_query_qp(cma_id->qp, &attr, ~0, &init_attr);
 820         iser_info("remote qpn:%x my qpn:%x\n", attr.dest_qp_num, cma_id->qp->qp_num);
 821
 822         if (private_data) {
 823                 u8 flags = *(u8 *)private_data;
 824
 825                 iser_conn->snd_w_inv = !(flags & ISER_SEND_W_INV_NOT_SUP);
 826         }
 827
 828         iser_info("conn %p: negotiated %s invalidation\n",
 829                   iser_conn, iser_conn->snd_w_inv ? "remote" : "local");
 830
 831         iser_conn->state = ISER_CONN_UP;
 832         complete(&iser_conn->up_completion);
 833 }
 834
 835 static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
 836 {
 837         struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
 838
 839         if (iser_conn_terminate(iser_conn)) {
 840                 if (iser_conn->iscsi_conn)
 841                         iscsi_conn_failure(iser_conn->iscsi_conn,
 842                                            ISCSI_ERR_CONN_FAILED);
 843                 else
 844                         iser_err("iscsi_iser connection isn't bound\n");
 845         }
 846 }
 847
 848 static void iser_cleanup_handler(struct rdma_cm_id *cma_id,
 849                                  bool destroy)
 850 {
 851         struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
 852
 853         /*
 854          * We are not guaranteed that we visited disconnected_handler
 855          * by now, call it here to be safe that we handle CM drep
 856          * and flush errors.
 857          */
 858         iser_disconnected_handler(cma_id);
 859         iser_free_ib_conn_res(iser_conn, destroy);
 860         complete(&iser_conn->ib_completion);
 861 };
 862
 863 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
 864 {
 865         struct iser_conn *iser_conn;
 866         int ret = 0;
 867
 868         iser_conn = (struct iser_conn *)cma_id->context;
 869         iser_info("%s (%d): status %d conn %p id %p\n",
 870                   rdma_event_msg(event->event), event->event,
 871                   event->status, cma_id->context, cma_id);
 872
 873         mutex_lock(&iser_conn->state_mutex);
 874         switch (event->event) {
 875         case RDMA_CM_EVENT_ADDR_RESOLVED:
 876                 iser_addr_handler(cma_id);
 877                 break;
 878         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 879                 iser_route_handler(cma_id);
 880                 break;
 881         case RDMA_CM_EVENT_ESTABLISHED:
 882                 iser_connected_handler(cma_id, event->param.conn.private_data);
 883                 break;
 884         case RDMA_CM_EVENT_REJECTED:
 885                 iser_info("Connection rejected: %s\n",
 886                          rdma_reject_msg(cma_id, event->status));
 887                 /* FALLTHROUGH */
 888         case RDMA_CM_EVENT_ADDR_ERROR:
 889         case RDMA_CM_EVENT_ROUTE_ERROR:
 890         case RDMA_CM_EVENT_CONNECT_ERROR:
 891         case RDMA_CM_EVENT_UNREACHABLE:
 892                 iser_connect_error(cma_id);
 893                 break;
 894         case RDMA_CM_EVENT_DISCONNECTED:
 895         case RDMA_CM_EVENT_ADDR_CHANGE:
 896         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
 897                 iser_cleanup_handler(cma_id, false);
 898                 break;
 899         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 900                 /*
 901                  * we *must* destroy the device as we cannot rely
 902                  * on iscsid to be around to initiate error handling.
 903                  * also if we are not in state DOWN implicitly destroy
 904                  * the cma_id.
 905                  */
 906                 iser_cleanup_handler(cma_id, true);
 907                 if (iser_conn->state != ISER_CONN_DOWN) {
 908                         iser_conn->ib_conn.cma_id = NULL;
 909                         ret = 1;
 910                 }
 911                 break;
 912         default:
 913                 iser_err("Unexpected RDMA CM event: %s (%d)\n",
 914                          rdma_event_msg(event->event), event->event);
 915                 break;
 916         }
 917         mutex_unlock(&iser_conn->state_mutex);
 918
 919         return ret;
 920 }
 921
 922 void iser_conn_init(struct iser_conn *iser_conn)
 923 {
 924         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 925
 926         iser_conn->state = ISER_CONN_INIT;
 927         init_completion(&iser_conn->stop_completion);
 928         init_completion(&iser_conn->ib_completion);
 929         init_completion(&iser_conn->up_completion);
 930         INIT_LIST_HEAD(&iser_conn->conn_list);
 931         mutex_init(&iser_conn->state_mutex);
 932
 933         ib_conn->post_recv_buf_count = 0;
 934         ib_conn->reg_cqe.done = iser_reg_comp;
 935 }
 936
 937  /**
 938  * starts the process of connecting to the target
 939  * sleeps until the connection is established or rejected
 940  */
 941 int iser_connect(struct iser_conn   *iser_conn,
 942                  struct sockaddr    *src_addr,
 943                  struct sockaddr    *dst_addr,
 944                  int                 non_blocking)
 945 {
 946         struct ib_conn *ib_conn = &iser_conn->ib_conn;
 947         int err = 0;
 948
 949         mutex_lock(&iser_conn->state_mutex);
 950
 951         sprintf(iser_conn->name, "%pISp", dst_addr);
 952
 953         iser_info("connecting to: %s\n", iser_conn->name);
 954
 955         /* the device is known only --after-- address resolution */
 956         ib_conn->device = NULL;
 957
 958         iser_conn->state = ISER_CONN_PENDING;
 959
 960         ib_conn->cma_id = rdma_create_id(&init_net, iser_cma_handler,
 961                                          (void *)iser_conn,
 962                                          RDMA_PS_TCP, IB_QPT_RC);
 963         if (IS_ERR(ib_conn->cma_id)) {
 964                 err = PTR_ERR(ib_conn->cma_id);
 965                 iser_err("rdma_create_id failed: %d\n", err);
 966                 goto id_failure;
 967         }
 968
 969         err = rdma_resolve_addr(ib_conn->cma_id, src_addr, dst_addr, 1000);
 970         if (err) {
 971                 iser_err("rdma_resolve_addr failed: %d\n", err);
 972                 goto addr_failure;
 973         }
 974
 975         if (!non_blocking) {
 976                 wait_for_completion_interruptible(&iser_conn->up_completion);
 977
 978                 if (iser_conn->state != ISER_CONN_UP) {
 979                         err =  -EIO;
 980                         goto connect_failure;
 981                 }
 982         }
 983         mutex_unlock(&iser_conn->state_mutex);
 984
 985         mutex_lock(&ig.connlist_mutex);
 986         list_add(&iser_conn->conn_list, &ig.connlist);
 987         mutex_unlock(&ig.connlist_mutex);
 988         return 0;
 989
 990 id_failure:
 991         ib_conn->cma_id = NULL;
 992 addr_failure:
 993         iser_conn->state = ISER_CONN_DOWN;
 994 connect_failure:
 995         mutex_unlock(&iser_conn->state_mutex);
 996         iser_conn_release(iser_conn);
 997         return err;
 998 }
 999
1000 int iser_post_recvl(struct iser_conn *iser_conn)
1001 {
1002         struct ib_conn *ib_conn = &iser_conn->ib_conn;
1003         struct iser_login_desc *desc = &iser_conn->login_desc;
1004         struct ib_recv_wr wr, *wr_failed;
1005         int ib_ret;
1006
1007         desc->sge.addr = desc->rsp_dma;
1008         desc->sge.length = ISER_RX_LOGIN_SIZE;
1009         desc->sge.lkey = ib_conn->device->pd->local_dma_lkey;
1010
1011         desc->cqe.done = iser_login_rsp;
1012         wr.wr_cqe = &desc->cqe;
1013         wr.sg_list = &desc->sge;
1014         wr.num_sge = 1;
1015         wr.next = NULL;
1016
1017         ib_conn->post_recv_buf_count++;
1018         ib_ret = ib_post_recv(ib_conn->qp, &wr, &wr_failed);
1019         if (ib_ret) {
1020                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1021                 ib_conn->post_recv_buf_count--;
1022         }
1023
1024         return ib_ret;
1025 }
1026
1027 int iser_post_recvm(struct iser_conn *iser_conn, int count)
1028 {
1029         struct ib_conn *ib_conn = &iser_conn->ib_conn;
1030         unsigned int my_rx_head = iser_conn->rx_desc_head;
1031         struct iser_rx_desc *rx_desc;
1032         struct ib_recv_wr *wr, *wr_failed;
1033         int i, ib_ret;
1034
1035         for (wr = ib_conn->rx_wr, i = 0; i < count; i++, wr++) {
1036                 rx_desc = &iser_conn->rx_descs[my_rx_head];
1037                 rx_desc->cqe.done = iser_task_rsp;
1038                 wr->wr_cqe = &rx_desc->cqe;
1039                 wr->sg_list = &rx_desc->rx_sg;
1040                 wr->num_sge = 1;
1041                 wr->next = wr + 1;
1042                 my_rx_head = (my_rx_head + 1) & iser_conn->qp_max_recv_dtos_mask;
1043         }
1044
1045         wr--;
1046         wr->next = NULL; /* mark end of work requests list */
1047
1048         ib_conn->post_recv_buf_count += count;
1049         ib_ret = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &wr_failed);
1050         if (ib_ret) {
1051                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1052                 ib_conn->post_recv_buf_count -= count;
1053         } else
1054                 iser_conn->rx_desc_head = my_rx_head;
1055
1056         return ib_ret;
1057 }
1058
1059
1060 /**
1061  * iser_start_send - Initiate a Send DTO operation
1062  *
1063  * returns 0 on success, -1 on failure
1064  */
1065 int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc,
1066                    bool signal)
1067 {
1068         struct ib_send_wr *bad_wr, *wr = iser_tx_next_wr(tx_desc);
1069         int ib_ret;
1070
1071         ib_dma_sync_single_for_device(ib_conn->device->ib_device,
1072                                       tx_desc->dma_addr, ISER_HEADERS_LEN,
1073                                       DMA_TO_DEVICE);
1074
1075         wr->next = NULL;
1076         wr->wr_cqe = &tx_desc->cqe;
1077         wr->sg_list = tx_desc->tx_sg;
1078         wr->num_sge = tx_desc->num_sge;
1079         wr->opcode = IB_WR_SEND;
1080         wr->send_flags = signal ? IB_SEND_SIGNALED : 0;
1081
1082         ib_ret = ib_post_send(ib_conn->qp, &tx_desc->wrs[0].send, &bad_wr);
1083         if (ib_ret)
1084                 iser_err("ib_post_send failed, ret:%d opcode:%d\n",
1085                          ib_ret, bad_wr->opcode);
1086
1087         return ib_ret;
1088 }
1089
1090 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
1091                              enum iser_data_dir cmd_dir, sector_t *sector)
1092 {
1093         struct iser_mem_reg *reg = &iser_task->rdma_reg[cmd_dir];
1094         struct iser_fr_desc *desc = reg->mem_h;
1095         unsigned long sector_size = iser_task->sc->device->sector_size;
1096         struct ib_mr_status mr_status;
1097         int ret;
1098
1099         if (desc && desc->pi_ctx->sig_protected) {
1100                 desc->pi_ctx->sig_protected = 0;
1101                 ret = ib_check_mr_status(desc->pi_ctx->sig_mr,
1102                                          IB_MR_CHECK_SIG_STATUS, &mr_status);
1103                 if (ret) {
1104                         pr_err("ib_check_mr_status failed, ret %d\n", ret);
1105                         goto err;
1106                 }
1107
1108                 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1109                         sector_t sector_off = mr_status.sig_err.sig_err_offset;
1110
1111                         sector_div(sector_off, sector_size + 8);
1112                         *sector = scsi_get_lba(iser_task->sc) + sector_off;
1113
1114                         pr_err("PI error found type %d at sector %llx "
1115                                "expected %x vs actual %x\n",
1116                                mr_status.sig_err.err_type,
1117                                (unsigned long long)*sector,
1118                                mr_status.sig_err.expected,
1119                                mr_status.sig_err.actual);
1120
1121                         switch (mr_status.sig_err.err_type) {
1122                         case IB_SIG_BAD_GUARD:
1123                                 return 0x1;
1124                         case IB_SIG_BAD_REFTAG:
1125                                 return 0x3;
1126                         case IB_SIG_BAD_APPTAG:
1127                                 return 0x2;
1128                         }
1129                 }
1130         }
1131
1132         return 0;
1133 err:
1134         /* Not alot we can do here, return ambiguous guard error */
1135         return 0x1;
1136 }
1137
1138 void iser_err_comp(struct ib_wc *wc, const char *type)
1139 {
1140         if (wc->status != IB_WC_WR_FLUSH_ERR) {
1141                 struct iser_conn *iser_conn = to_iser_conn(wc->qp->qp_context);
1142
1143                 iser_err("%s failure: %s (%d) vend_err %x\n", type,
1144                          ib_wc_status_msg(wc->status), wc->status,
1145                          wc->vendor_err);
1146
1147                 if (iser_conn->iscsi_conn)
1148                         iscsi_conn_failure(iser_conn->iscsi_conn,
1149                                            ISCSI_ERR_CONN_FAILED);
1150         } else {
1151                 iser_dbg("%s failure: %s (%d)\n", type,
1152                          ib_wc_status_msg(wc->status), wc->status);
1153         }
1154 }