drivers/infiniband/hw/ocrdma/ocrdma_verbs.c

   1 /*******************************************************************
   2  * This file is part of the Emulex RoCE Device Driver for          *
   3  * RoCE (RDMA over Converged Ethernet) adapters.                   *
   4  * Copyright (C) 2008-2012 Emulex. All rights reserved.            *
   5  * EMULEX and SLI are trademarks of Emulex.                        *
   6  * www.emulex.com                                                  *
   7  *                                                                 *
   8  * This program is free software; you can redistribute it and/or   *
   9  * modify it under the terms of version 2 of the GNU General       *
  10  * Public License as published by the Free Software Foundation.    *
  11  * This program is distributed in the hope that it will be useful. *
  12  * ALL EXPRESS OR IMPLIED CONDITIONS, REPRESENTATIONS AND          *
  13  * WARRANTIES, INCLUDING ANY IMPLIED WARRANTY OF MERCHANTABILITY,  *
  14  * FITNESS FOR A PARTICULAR PURPOSE, OR NON-INFRINGEMENT, ARE      *
  15  * DISCLAIMED, EXCEPT TO THE EXTENT THAT SUCH DISCLAIMERS ARE HELD *
  16  * TO BE LEGALLY INVALID.  See the GNU General Public License for  *
  17  * more details, a copy of which can be found in the file COPYING  *
  18  * included with this package.                                     *
  19  *
  20  * Contact Information:
  21  * linux-drivers@emulex.com
  22  *
  23  * Emulex
  24  * 3333 Susan Street
  25  * Costa Mesa, CA 92626
  26  *******************************************************************/
  27
  28 #include <linux/dma-mapping.h>
  29 #include <rdma/ib_verbs.h>
  30 #include <rdma/ib_user_verbs.h>
  31 #include <rdma/iw_cm.h>
  32 #include <rdma/ib_umem.h>
  33 #include <rdma/ib_addr.h>
  34
  35 #include "ocrdma.h"
  36 #include "ocrdma_hw.h"
  37 #include "ocrdma_verbs.h"
  38 #include "ocrdma_abi.h"
  39
  40 int ocrdma_query_pkey(struct ib_device *ibdev, u8 port, u16 index, u16 *pkey)
  41 {
  42         if (index > 1)
  43                 return -EINVAL;
  44
  45         *pkey = 0xffff;
  46         return 0;
  47 }
  48
  49 int ocrdma_query_gid(struct ib_device *ibdev, u8 port,
  50                      int index, union ib_gid *sgid)
  51 {
  52         struct ocrdma_dev *dev;
  53
  54         dev = get_ocrdma_dev(ibdev);
  55         memset(sgid, 0, sizeof(*sgid));
  56         if (index >= OCRDMA_MAX_SGID)
  57                 return -EINVAL;
  58
  59         memcpy(sgid, &dev->sgid_tbl[index], sizeof(*sgid));
  60
  61         return 0;
  62 }
  63
  64 int ocrdma_query_device(struct ib_device *ibdev, struct ib_device_attr *attr,
  65                         struct ib_udata *uhw)
  66 {
  67         struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
  68
  69         if (uhw->inlen || uhw->outlen)
  70                 return -EINVAL;
  71
  72         memset(attr, 0, sizeof *attr);
  73         memcpy(&attr->fw_ver, &dev->attr.fw_ver[0],
  74                min(sizeof(dev->attr.fw_ver), sizeof(attr->fw_ver)));
  75         ocrdma_get_guid(dev, (u8 *)&attr->sys_image_guid);
  76         attr->max_mr_size = dev->attr.max_mr_size;
  77         attr->page_size_cap = 0xffff000;
  78         attr->vendor_id = dev->nic_info.pdev->vendor;
  79         attr->vendor_part_id = dev->nic_info.pdev->device;
  80         attr->hw_ver = dev->asic_id;
  81         attr->max_qp = dev->attr.max_qp;
  82         attr->max_ah = OCRDMA_MAX_AH;
  83         attr->max_qp_wr = dev->attr.max_wqe;
  84
  85         attr->device_cap_flags = IB_DEVICE_CURR_QP_STATE_MOD |
  86                                         IB_DEVICE_RC_RNR_NAK_GEN |
  87                                         IB_DEVICE_SHUTDOWN_PORT |
  88                                         IB_DEVICE_SYS_IMAGE_GUID |
  89                                         IB_DEVICE_LOCAL_DMA_LKEY |
  90                                         IB_DEVICE_MEM_MGT_EXTENSIONS;
  91         attr->max_sge = min(dev->attr.max_send_sge, dev->attr.max_srq_sge);
  92         attr->max_sge_rd = 0;
  93         attr->max_cq = dev->attr.max_cq;
  94         attr->max_cqe = dev->attr.max_cqe;
  95         attr->max_mr = dev->attr.max_mr;
  96         attr->max_mw = dev->attr.max_mw;
  97         attr->max_pd = dev->attr.max_pd;
  98         attr->atomic_cap = 0;
  99         attr->max_fmr = 0;
 100         attr->max_map_per_fmr = 0;
 101         attr->max_qp_rd_atom =
 102             min(dev->attr.max_ord_per_qp, dev->attr.max_ird_per_qp);
 103         attr->max_qp_init_rd_atom = dev->attr.max_ord_per_qp;
 104         attr->max_srq = dev->attr.max_srq;
 105         attr->max_srq_sge = dev->attr.max_srq_sge;
 106         attr->max_srq_wr = dev->attr.max_rqe;
 107         attr->local_ca_ack_delay = dev->attr.local_ca_ack_delay;
 108         attr->max_fast_reg_page_list_len = dev->attr.max_pages_per_frmr;
 109         attr->max_pkeys = 1;
 110         return 0;
 111 }
 112
 113 static inline void get_link_speed_and_width(struct ocrdma_dev *dev,
 114                                             u8 *ib_speed, u8 *ib_width)
 115 {
 116         int status;
 117         u8 speed;
 118
 119         status = ocrdma_mbx_get_link_speed(dev, &speed);
 120         if (status)
 121                 speed = OCRDMA_PHYS_LINK_SPEED_ZERO;
 122
 123         switch (speed) {
 124         case OCRDMA_PHYS_LINK_SPEED_1GBPS:
 125                 *ib_speed = IB_SPEED_SDR;
 126                 *ib_width = IB_WIDTH_1X;
 127                 break;
 128
 129         case OCRDMA_PHYS_LINK_SPEED_10GBPS:
 130                 *ib_speed = IB_SPEED_QDR;
 131                 *ib_width = IB_WIDTH_1X;
 132                 break;
 133
 134         case OCRDMA_PHYS_LINK_SPEED_20GBPS:
 135                 *ib_speed = IB_SPEED_DDR;
 136                 *ib_width = IB_WIDTH_4X;
 137                 break;
 138
 139         case OCRDMA_PHYS_LINK_SPEED_40GBPS:
 140                 *ib_speed = IB_SPEED_QDR;
 141                 *ib_width = IB_WIDTH_4X;
 142                 break;
 143
 144         default:
 145                 /* Unsupported */
 146                 *ib_speed = IB_SPEED_SDR;
 147                 *ib_width = IB_WIDTH_1X;
 148         }
 149 }
 150
 151 int ocrdma_query_port(struct ib_device *ibdev,
 152                       u8 port, struct ib_port_attr *props)
 153 {
 154         enum ib_port_state port_state;
 155         struct ocrdma_dev *dev;
 156         struct net_device *netdev;
 157
 158         dev = get_ocrdma_dev(ibdev);
 159         if (port > 1) {
 160                 pr_err("%s(%d) invalid_port=0x%x\n", __func__,
 161                        dev->id, port);
 162                 return -EINVAL;
 163         }
 164         netdev = dev->nic_info.netdev;
 165         if (netif_running(netdev) && netif_oper_up(netdev)) {
 166                 port_state = IB_PORT_ACTIVE;
 167                 props->phys_state = 5;
 168         } else {
 169                 port_state = IB_PORT_DOWN;
 170                 props->phys_state = 3;
 171         }
 172         props->max_mtu = IB_MTU_4096;
 173         props->active_mtu = iboe_get_mtu(netdev->mtu);
 174         props->lid = 0;
 175         props->lmc = 0;
 176         props->sm_lid = 0;
 177         props->sm_sl = 0;
 178         props->state = port_state;
 179         props->port_cap_flags =
 180             IB_PORT_CM_SUP |
 181             IB_PORT_REINIT_SUP |
 182             IB_PORT_DEVICE_MGMT_SUP | IB_PORT_VENDOR_CLASS_SUP | IB_PORT_IP_BASED_GIDS;
 183         props->gid_tbl_len = OCRDMA_MAX_SGID;
 184         props->pkey_tbl_len = 1;
 185         props->bad_pkey_cntr = 0;
 186         props->qkey_viol_cntr = 0;
 187         get_link_speed_and_width(dev, &props->active_speed,
 188                                  &props->active_width);
 189         props->max_msg_sz = 0x80000000;
 190         props->max_vl_num = 4;
 191         return 0;
 192 }
 193
 194 int ocrdma_modify_port(struct ib_device *ibdev, u8 port, int mask,
 195                        struct ib_port_modify *props)
 196 {
 197         struct ocrdma_dev *dev;
 198
 199         dev = get_ocrdma_dev(ibdev);
 200         if (port > 1) {
 201                 pr_err("%s(%d) invalid_port=0x%x\n", __func__, dev->id, port);
 202                 return -EINVAL;
 203         }
 204         return 0;
 205 }
 206
 207 static int ocrdma_add_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
 208                            unsigned long len)
 209 {
 210         struct ocrdma_mm *mm;
 211
 212         mm = kzalloc(sizeof(*mm), GFP_KERNEL);
 213         if (mm == NULL)
 214                 return -ENOMEM;
 215         mm->key.phy_addr = phy_addr;
 216         mm->key.len = len;
 217         INIT_LIST_HEAD(&mm->entry);
 218
 219         mutex_lock(&uctx->mm_list_lock);
 220         list_add_tail(&mm->entry, &uctx->mm_head);
 221         mutex_unlock(&uctx->mm_list_lock);
 222         return 0;
 223 }
 224
 225 static void ocrdma_del_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
 226                             unsigned long len)
 227 {
 228         struct ocrdma_mm *mm, *tmp;
 229
 230         mutex_lock(&uctx->mm_list_lock);
 231         list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
 232                 if (len != mm->key.len && phy_addr != mm->key.phy_addr)
 233                         continue;
 234
 235                 list_del(&mm->entry);
 236                 kfree(mm);
 237                 break;
 238         }
 239         mutex_unlock(&uctx->mm_list_lock);
 240 }
 241
 242 static bool ocrdma_search_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
 243                               unsigned long len)
 244 {
 245         bool found = false;
 246         struct ocrdma_mm *mm;
 247
 248         mutex_lock(&uctx->mm_list_lock);
 249         list_for_each_entry(mm, &uctx->mm_head, entry) {
 250                 if (len != mm->key.len && phy_addr != mm->key.phy_addr)
 251                         continue;
 252
 253                 found = true;
 254                 break;
 255         }
 256         mutex_unlock(&uctx->mm_list_lock);
 257         return found;
 258 }
 259
 260
 261 static u16 _ocrdma_pd_mgr_get_bitmap(struct ocrdma_dev *dev, bool dpp_pool)
 262 {
 263         u16 pd_bitmap_idx = 0;
 264         const unsigned long *pd_bitmap;
 265
 266         if (dpp_pool) {
 267                 pd_bitmap = dev->pd_mgr->pd_dpp_bitmap;
 268                 pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
 269                                                     dev->pd_mgr->max_dpp_pd);
 270                 __set_bit(pd_bitmap_idx, dev->pd_mgr->pd_dpp_bitmap);
 271                 dev->pd_mgr->pd_dpp_count++;
 272                 if (dev->pd_mgr->pd_dpp_count > dev->pd_mgr->pd_dpp_thrsh)
 273                         dev->pd_mgr->pd_dpp_thrsh = dev->pd_mgr->pd_dpp_count;
 274         } else {
 275                 pd_bitmap = dev->pd_mgr->pd_norm_bitmap;
 276                 pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
 277                                                     dev->pd_mgr->max_normal_pd);
 278                 __set_bit(pd_bitmap_idx, dev->pd_mgr->pd_norm_bitmap);
 279                 dev->pd_mgr->pd_norm_count++;
 280                 if (dev->pd_mgr->pd_norm_count > dev->pd_mgr->pd_norm_thrsh)
 281                         dev->pd_mgr->pd_norm_thrsh = dev->pd_mgr->pd_norm_count;
 282         }
 283         return pd_bitmap_idx;
 284 }
 285
 286 static int _ocrdma_pd_mgr_put_bitmap(struct ocrdma_dev *dev, u16 pd_id,
 287                                         bool dpp_pool)
 288 {
 289         u16 pd_count;
 290         u16 pd_bit_index;
 291
 292         pd_count = dpp_pool ? dev->pd_mgr->pd_dpp_count :
 293                               dev->pd_mgr->pd_norm_count;
 294         if (pd_count == 0)
 295                 return -EINVAL;
 296
 297         if (dpp_pool) {
 298                 pd_bit_index = pd_id - dev->pd_mgr->pd_dpp_start;
 299                 if (pd_bit_index >= dev->pd_mgr->max_dpp_pd) {
 300                         return -EINVAL;
 301                 } else {
 302                         __clear_bit(pd_bit_index, dev->pd_mgr->pd_dpp_bitmap);
 303                         dev->pd_mgr->pd_dpp_count--;
 304                 }
 305         } else {
 306                 pd_bit_index = pd_id - dev->pd_mgr->pd_norm_start;
 307                 if (pd_bit_index >= dev->pd_mgr->max_normal_pd) {
 308                         return -EINVAL;
 309                 } else {
 310                         __clear_bit(pd_bit_index, dev->pd_mgr->pd_norm_bitmap);
 311                         dev->pd_mgr->pd_norm_count--;
 312                 }
 313         }
 314
 315         return 0;
 316 }
 317
 318 static u8 ocrdma_put_pd_num(struct ocrdma_dev *dev, u16 pd_id,
 319                                    bool dpp_pool)
 320 {
 321         int status;
 322
 323         mutex_lock(&dev->dev_lock);
 324         status = _ocrdma_pd_mgr_put_bitmap(dev, pd_id, dpp_pool);
 325         mutex_unlock(&dev->dev_lock);
 326         return status;
 327 }
 328
 329 static int ocrdma_get_pd_num(struct ocrdma_dev *dev, struct ocrdma_pd *pd)
 330 {
 331         u16 pd_idx = 0;
 332         int status = 0;
 333
 334         mutex_lock(&dev->dev_lock);
 335         if (pd->dpp_enabled) {
 336                 /* try allocating DPP PD, if not available then normal PD */
 337                 if (dev->pd_mgr->pd_dpp_count < dev->pd_mgr->max_dpp_pd) {
 338                         pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, true);
 339                         pd->id = dev->pd_mgr->pd_dpp_start + pd_idx;
 340                         pd->dpp_page = dev->pd_mgr->dpp_page_index + pd_idx;
 341                 } else if (dev->pd_mgr->pd_norm_count <
 342                            dev->pd_mgr->max_normal_pd) {
 343                         pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
 344                         pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
 345                         pd->dpp_enabled = false;
 346                 } else {
 347                         status = -EINVAL;
 348                 }
 349         } else {
 350                 if (dev->pd_mgr->pd_norm_count < dev->pd_mgr->max_normal_pd) {
 351                         pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
 352                         pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
 353                 } else {
 354                         status = -EINVAL;
 355                 }
 356         }
 357         mutex_unlock(&dev->dev_lock);
 358         return status;
 359 }
 360
 361 static struct ocrdma_pd *_ocrdma_alloc_pd(struct ocrdma_dev *dev,
 362                                           struct ocrdma_ucontext *uctx,
 363                                           struct ib_udata *udata)
 364 {
 365         struct ocrdma_pd *pd = NULL;
 366         int status = 0;
 367
 368         pd = kzalloc(sizeof(*pd), GFP_KERNEL);
 369         if (!pd)
 370                 return ERR_PTR(-ENOMEM);
 371
 372         if (udata && uctx && dev->attr.max_dpp_pds) {
 373                 pd->dpp_enabled =
 374                         ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R;
 375                 pd->num_dpp_qp =
 376                         pd->dpp_enabled ? (dev->nic_info.db_page_size /
 377                                            dev->attr.wqe_size) : 0;
 378         }
 379
 380         if (dev->pd_mgr->pd_prealloc_valid) {
 381                 status = ocrdma_get_pd_num(dev, pd);
 382                 if (status == 0) {
 383                         return pd;
 384                 } else {
 385                         kfree(pd);
 386                         return ERR_PTR(status);
 387                 }
 388         }
 389
 390 retry:
 391         status = ocrdma_mbx_alloc_pd(dev, pd);
 392         if (status) {
 393                 if (pd->dpp_enabled) {
 394                         pd->dpp_enabled = false;
 395                         pd->num_dpp_qp = 0;
 396                         goto retry;
 397                 } else {
 398                         kfree(pd);
 399                         return ERR_PTR(status);
 400                 }
 401         }
 402
 403         return pd;
 404 }
 405
 406 static inline int is_ucontext_pd(struct ocrdma_ucontext *uctx,
 407                                  struct ocrdma_pd *pd)
 408 {
 409         return (uctx->cntxt_pd == pd ? true : false);
 410 }
 411
 412 static int _ocrdma_dealloc_pd(struct ocrdma_dev *dev,
 413                               struct ocrdma_pd *pd)
 414 {
 415         int status = 0;
 416
 417         if (dev->pd_mgr->pd_prealloc_valid)
 418                 status = ocrdma_put_pd_num(dev, pd->id, pd->dpp_enabled);
 419         else
 420                 status = ocrdma_mbx_dealloc_pd(dev, pd);
 421
 422         kfree(pd);
 423         return status;
 424 }
 425
 426 static int ocrdma_alloc_ucontext_pd(struct ocrdma_dev *dev,
 427                                     struct ocrdma_ucontext *uctx,
 428                                     struct ib_udata *udata)
 429 {
 430         int status = 0;
 431
 432         uctx->cntxt_pd = _ocrdma_alloc_pd(dev, uctx, udata);
 433         if (IS_ERR(uctx->cntxt_pd)) {
 434                 status = PTR_ERR(uctx->cntxt_pd);
 435                 uctx->cntxt_pd = NULL;
 436                 goto err;
 437         }
 438
 439         uctx->cntxt_pd->uctx = uctx;
 440         uctx->cntxt_pd->ibpd.device = &dev->ibdev;
 441 err:
 442         return status;
 443 }
 444
 445 static int ocrdma_dealloc_ucontext_pd(struct ocrdma_ucontext *uctx)
 446 {
 447         struct ocrdma_pd *pd = uctx->cntxt_pd;
 448         struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
 449
 450         if (uctx->pd_in_use) {
 451                 pr_err("%s(%d) Freeing in use pdid=0x%x.\n",
 452                        __func__, dev->id, pd->id);
 453         }
 454         uctx->cntxt_pd = NULL;
 455         (void)_ocrdma_dealloc_pd(dev, pd);
 456         return 0;
 457 }
 458
 459 static struct ocrdma_pd *ocrdma_get_ucontext_pd(struct ocrdma_ucontext *uctx)
 460 {
 461         struct ocrdma_pd *pd = NULL;
 462
 463         mutex_lock(&uctx->mm_list_lock);
 464         if (!uctx->pd_in_use) {
 465                 uctx->pd_in_use = true;
 466                 pd = uctx->cntxt_pd;
 467         }
 468         mutex_unlock(&uctx->mm_list_lock);
 469
 470         return pd;
 471 }
 472
 473 static void ocrdma_release_ucontext_pd(struct ocrdma_ucontext *uctx)
 474 {
 475         mutex_lock(&uctx->mm_list_lock);
 476         uctx->pd_in_use = false;
 477         mutex_unlock(&uctx->mm_list_lock);
 478 }
 479
 480 struct ib_ucontext *ocrdma_alloc_ucontext(struct ib_device *ibdev,
 481                                           struct ib_udata *udata)
 482 {
 483         int status;
 484         struct ocrdma_ucontext *ctx;
 485         struct ocrdma_alloc_ucontext_resp resp;
 486         struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
 487         struct pci_dev *pdev = dev->nic_info.pdev;
 488         u32 map_len = roundup(sizeof(u32) * 2048, PAGE_SIZE);
 489
 490         if (!udata)
 491                 return ERR_PTR(-EFAULT);
 492         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
 493         if (!ctx)
 494                 return ERR_PTR(-ENOMEM);
 495         INIT_LIST_HEAD(&ctx->mm_head);
 496         mutex_init(&ctx->mm_list_lock);
 497
 498         ctx->ah_tbl.va = dma_alloc_coherent(&pdev->dev, map_len,
 499                                             &ctx->ah_tbl.pa, GFP_KERNEL);
 500         if (!ctx->ah_tbl.va) {
 501                 kfree(ctx);
 502                 return ERR_PTR(-ENOMEM);
 503         }
 504         memset(ctx->ah_tbl.va, 0, map_len);
 505         ctx->ah_tbl.len = map_len;
 506
 507         memset(&resp, 0, sizeof(resp));
 508         resp.ah_tbl_len = ctx->ah_tbl.len;
 509         resp.ah_tbl_page = virt_to_phys(ctx->ah_tbl.va);
 510
 511         status = ocrdma_add_mmap(ctx, resp.ah_tbl_page, resp.ah_tbl_len);
 512         if (status)
 513                 goto map_err;
 514
 515         status = ocrdma_alloc_ucontext_pd(dev, ctx, udata);
 516         if (status)
 517                 goto pd_err;
 518
 519         resp.dev_id = dev->id;
 520         resp.max_inline_data = dev->attr.max_inline_data;
 521         resp.wqe_size = dev->attr.wqe_size;
 522         resp.rqe_size = dev->attr.rqe_size;
 523         resp.dpp_wqe_size = dev->attr.wqe_size;
 524
 525         memcpy(resp.fw_ver, dev->attr.fw_ver, sizeof(resp.fw_ver));
 526         status = ib_copy_to_udata(udata, &resp, sizeof(resp));
 527         if (status)
 528                 goto cpy_err;
 529         return &ctx->ibucontext;
 530
 531 cpy_err:
 532 pd_err:
 533         ocrdma_del_mmap(ctx, ctx->ah_tbl.pa, ctx->ah_tbl.len);
 534 map_err:
 535         dma_free_coherent(&pdev->dev, ctx->ah_tbl.len, ctx->ah_tbl.va,
 536                           ctx->ah_tbl.pa);
 537         kfree(ctx);
 538         return ERR_PTR(status);
 539 }
 540
 541 int ocrdma_dealloc_ucontext(struct ib_ucontext *ibctx)
 542 {
 543         int status = 0;
 544         struct ocrdma_mm *mm, *tmp;
 545         struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ibctx);
 546         struct ocrdma_dev *dev = get_ocrdma_dev(ibctx->device);
 547         struct pci_dev *pdev = dev->nic_info.pdev;
 548
 549         status = ocrdma_dealloc_ucontext_pd(uctx);
 550
 551         ocrdma_del_mmap(uctx, uctx->ah_tbl.pa, uctx->ah_tbl.len);
 552         dma_free_coherent(&pdev->dev, uctx->ah_tbl.len, uctx->ah_tbl.va,
 553                           uctx->ah_tbl.pa);
 554
 555         list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
 556                 list_del(&mm->entry);
 557                 kfree(mm);
 558         }
 559         kfree(uctx);
 560         return status;
 561 }
 562
 563 int ocrdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
 564 {
 565         struct ocrdma_ucontext *ucontext = get_ocrdma_ucontext(context);
 566         struct ocrdma_dev *dev = get_ocrdma_dev(context->device);
 567         unsigned long vm_page = vma->vm_pgoff << PAGE_SHIFT;
 568         u64 unmapped_db = (u64) dev->nic_info.unmapped_db;
 569         unsigned long len = (vma->vm_end - vma->vm_start);
 570         int status = 0;
 571         bool found;
 572
 573         if (vma->vm_start & (PAGE_SIZE - 1))
 574                 return -EINVAL;
 575         found = ocrdma_search_mmap(ucontext, vma->vm_pgoff << PAGE_SHIFT, len);
 576         if (!found)
 577                 return -EINVAL;
 578
 579         if ((vm_page >= unmapped_db) && (vm_page <= (unmapped_db +
 580                 dev->nic_info.db_total_size)) &&
 581                 (len <= dev->nic_info.db_page_size)) {
 582                 if (vma->vm_flags & VM_READ)
 583                         return -EPERM;
 584
 585                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 586                 status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
 587                                             len, vma->vm_page_prot);
 588         } else if (dev->nic_info.dpp_unmapped_len &&
 589                 (vm_page >= (u64) dev->nic_info.dpp_unmapped_addr) &&
 590                 (vm_page <= (u64) (dev->nic_info.dpp_unmapped_addr +
 591                         dev->nic_info.dpp_unmapped_len)) &&
 592                 (len <= dev->nic_info.dpp_unmapped_len)) {
 593                 if (vma->vm_flags & VM_READ)
 594                         return -EPERM;
 595
 596                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 597                 status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
 598                                             len, vma->vm_page_prot);
 599         } else {
 600                 status = remap_pfn_range(vma, vma->vm_start,
 601                                          vma->vm_pgoff, len, vma->vm_page_prot);
 602         }
 603         return status;
 604 }
 605
 606 static int ocrdma_copy_pd_uresp(struct ocrdma_dev *dev, struct ocrdma_pd *pd,
 607                                 struct ib_ucontext *ib_ctx,
 608                                 struct ib_udata *udata)
 609 {
 610         int status;
 611         u64 db_page_addr;
 612         u64 dpp_page_addr = 0;
 613         u32 db_page_size;
 614         struct ocrdma_alloc_pd_uresp rsp;
 615         struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
 616
 617         memset(&rsp, 0, sizeof(rsp));
 618         rsp.id = pd->id;
 619         rsp.dpp_enabled = pd->dpp_enabled;
 620         db_page_addr = ocrdma_get_db_addr(dev, pd->id);
 621         db_page_size = dev->nic_info.db_page_size;
 622
 623         status = ocrdma_add_mmap(uctx, db_page_addr, db_page_size);
 624         if (status)
 625                 return status;
 626
 627         if (pd->dpp_enabled) {
 628                 dpp_page_addr = dev->nic_info.dpp_unmapped_addr +
 629                                 (pd->id * PAGE_SIZE);
 630                 status = ocrdma_add_mmap(uctx, dpp_page_addr,
 631                                  PAGE_SIZE);
 632                 if (status)
 633                         goto dpp_map_err;
 634                 rsp.dpp_page_addr_hi = upper_32_bits(dpp_page_addr);
 635                 rsp.dpp_page_addr_lo = dpp_page_addr;
 636         }
 637
 638         status = ib_copy_to_udata(udata, &rsp, sizeof(rsp));
 639         if (status)
 640                 goto ucopy_err;
 641
 642         pd->uctx = uctx;
 643         return 0;
 644
 645 ucopy_err:
 646         if (pd->dpp_enabled)
 647                 ocrdma_del_mmap(pd->uctx, dpp_page_addr, PAGE_SIZE);
 648 dpp_map_err:
 649         ocrdma_del_mmap(pd->uctx, db_page_addr, db_page_size);
 650         return status;
 651 }
 652
 653 struct ib_pd *ocrdma_alloc_pd(struct ib_device *ibdev,
 654                               struct ib_ucontext *context,
 655                               struct ib_udata *udata)
 656 {
 657         struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
 658         struct ocrdma_pd *pd;
 659         struct ocrdma_ucontext *uctx = NULL;
 660         int status;
 661         u8 is_uctx_pd = false;
 662
 663         if (udata && context) {
 664                 uctx = get_ocrdma_ucontext(context);
 665                 pd = ocrdma_get_ucontext_pd(uctx);
 666                 if (pd) {
 667                         is_uctx_pd = true;
 668                         goto pd_mapping;
 669                 }
 670         }
 671
 672         pd = _ocrdma_alloc_pd(dev, uctx, udata);
 673         if (IS_ERR(pd)) {
 674                 status = PTR_ERR(pd);
 675                 goto exit;
 676         }
 677
 678 pd_mapping:
 679         if (udata && context) {
 680                 status = ocrdma_copy_pd_uresp(dev, pd, context, udata);
 681                 if (status)
 682                         goto err;
 683         }
 684         return &pd->ibpd;
 685
 686 err:
 687         if (is_uctx_pd) {
 688                 ocrdma_release_ucontext_pd(uctx);
 689         } else {
 690                 status = _ocrdma_dealloc_pd(dev, pd);
 691         }
 692 exit:
 693         return ERR_PTR(status);
 694 }
 695
 696 int ocrdma_dealloc_pd(struct ib_pd *ibpd)
 697 {
 698         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
 699         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
 700         struct ocrdma_ucontext *uctx = NULL;
 701         int status = 0;
 702         u64 usr_db;
 703
 704         uctx = pd->uctx;
 705         if (uctx) {
 706                 u64 dpp_db = dev->nic_info.dpp_unmapped_addr +
 707                         (pd->id * PAGE_SIZE);
 708                 if (pd->dpp_enabled)
 709                         ocrdma_del_mmap(pd->uctx, dpp_db, PAGE_SIZE);
 710                 usr_db = ocrdma_get_db_addr(dev, pd->id);
 711                 ocrdma_del_mmap(pd->uctx, usr_db, dev->nic_info.db_page_size);
 712
 713                 if (is_ucontext_pd(uctx, pd)) {
 714                         ocrdma_release_ucontext_pd(uctx);
 715                         return status;
 716                 }
 717         }
 718         status = _ocrdma_dealloc_pd(dev, pd);
 719         return status;
 720 }
 721
 722 static int ocrdma_alloc_lkey(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
 723                             u32 pdid, int acc, u32 num_pbls, u32 addr_check)
 724 {
 725         int status;
 726
 727         mr->hwmr.fr_mr = 0;
 728         mr->hwmr.local_rd = 1;
 729         mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
 730         mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
 731         mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
 732         mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
 733         mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
 734         mr->hwmr.num_pbls = num_pbls;
 735
 736         status = ocrdma_mbx_alloc_lkey(dev, &mr->hwmr, pdid, addr_check);
 737         if (status)
 738                 return status;
 739
 740         mr->ibmr.lkey = mr->hwmr.lkey;
 741         if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
 742                 mr->ibmr.rkey = mr->hwmr.lkey;
 743         return 0;
 744 }
 745
 746 struct ib_mr *ocrdma_get_dma_mr(struct ib_pd *ibpd, int acc)
 747 {
 748         int status;
 749         struct ocrdma_mr *mr;
 750         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
 751         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
 752
 753         if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE)) {
 754                 pr_err("%s err, invalid access rights\n", __func__);
 755                 return ERR_PTR(-EINVAL);
 756         }
 757
 758         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
 759         if (!mr)
 760                 return ERR_PTR(-ENOMEM);
 761
 762         status = ocrdma_alloc_lkey(dev, mr, pd->id, acc, 0,
 763                                    OCRDMA_ADDR_CHECK_DISABLE);
 764         if (status) {
 765                 kfree(mr);
 766                 return ERR_PTR(status);
 767         }
 768
 769         return &mr->ibmr;
 770 }
 771
 772 static void ocrdma_free_mr_pbl_tbl(struct ocrdma_dev *dev,
 773                                    struct ocrdma_hw_mr *mr)
 774 {
 775         struct pci_dev *pdev = dev->nic_info.pdev;
 776         int i = 0;
 777
 778         if (mr->pbl_table) {
 779                 for (i = 0; i < mr->num_pbls; i++) {
 780                         if (!mr->pbl_table[i].va)
 781                                 continue;
 782                         dma_free_coherent(&pdev->dev, mr->pbl_size,
 783                                           mr->pbl_table[i].va,
 784                                           mr->pbl_table[i].pa);
 785                 }
 786                 kfree(mr->pbl_table);
 787                 mr->pbl_table = NULL;
 788         }
 789 }
 790
 791 static int ocrdma_get_pbl_info(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
 792                               u32 num_pbes)
 793 {
 794         u32 num_pbls = 0;
 795         u32 idx = 0;
 796         int status = 0;
 797         u32 pbl_size;
 798
 799         do {
 800                 pbl_size = OCRDMA_MIN_HPAGE_SIZE * (1 << idx);
 801                 if (pbl_size > MAX_OCRDMA_PBL_SIZE) {
 802                         status = -EFAULT;
 803                         break;
 804                 }
 805                 num_pbls = roundup(num_pbes, (pbl_size / sizeof(u64)));
 806                 num_pbls = num_pbls / (pbl_size / sizeof(u64));
 807                 idx++;
 808         } while (num_pbls >= dev->attr.max_num_mr_pbl);
 809
 810         mr->hwmr.num_pbes = num_pbes;
 811         mr->hwmr.num_pbls = num_pbls;
 812         mr->hwmr.pbl_size = pbl_size;
 813         return status;
 814 }
 815
 816 static int ocrdma_build_pbl_tbl(struct ocrdma_dev *dev, struct ocrdma_hw_mr *mr)
 817 {
 818         int status = 0;
 819         int i;
 820         u32 dma_len = mr->pbl_size;
 821         struct pci_dev *pdev = dev->nic_info.pdev;
 822         void *va;
 823         dma_addr_t pa;
 824
 825         mr->pbl_table = kzalloc(sizeof(struct ocrdma_pbl) *
 826                                 mr->num_pbls, GFP_KERNEL);
 827
 828         if (!mr->pbl_table)
 829                 return -ENOMEM;
 830
 831         for (i = 0; i < mr->num_pbls; i++) {
 832                 va = dma_alloc_coherent(&pdev->dev, dma_len, &pa, GFP_KERNEL);
 833                 if (!va) {
 834                         ocrdma_free_mr_pbl_tbl(dev, mr);
 835                         status = -ENOMEM;
 836                         break;
 837                 }
 838                 memset(va, 0, dma_len);
 839                 mr->pbl_table[i].va = va;
 840                 mr->pbl_table[i].pa = pa;
 841         }
 842         return status;
 843 }
 844
 845 static void build_user_pbes(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
 846                             u32 num_pbes)
 847 {
 848         struct ocrdma_pbe *pbe;
 849         struct scatterlist *sg;
 850         struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
 851         struct ib_umem *umem = mr->umem;
 852         int shift, pg_cnt, pages, pbe_cnt, entry, total_num_pbes = 0;
 853
 854         if (!mr->hwmr.num_pbes)
 855                 return;
 856
 857         pbe = (struct ocrdma_pbe *)pbl_tbl->va;
 858         pbe_cnt = 0;
 859
 860         shift = ilog2(umem->page_size);
 861
 862         for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
 863                 pages = sg_dma_len(sg) >> shift;
 864                 for (pg_cnt = 0; pg_cnt < pages; pg_cnt++) {
 865                         /* store the page address in pbe */
 866                         pbe->pa_lo =
 867                             cpu_to_le32(sg_dma_address
 868                                         (sg) +
 869                                         (umem->page_size * pg_cnt));
 870                         pbe->pa_hi =
 871                             cpu_to_le32(upper_32_bits
 872                                         ((sg_dma_address
 873                                           (sg) +
 874                                           umem->page_size * pg_cnt)));
 875                         pbe_cnt += 1;
 876                         total_num_pbes += 1;
 877                         pbe++;
 878
 879                         /* if done building pbes, issue the mbx cmd. */
 880                         if (total_num_pbes == num_pbes)
 881                                 return;
 882
 883                         /* if the given pbl is full storing the pbes,
 884                          * move to next pbl.
 885                          */
 886                         if (pbe_cnt ==
 887                                 (mr->hwmr.pbl_size / sizeof(u64))) {
 888                                 pbl_tbl++;
 889                                 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
 890                                 pbe_cnt = 0;
 891                         }
 892
 893                 }
 894         }
 895 }
 896
 897 struct ib_mr *ocrdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
 898                                  u64 usr_addr, int acc, struct ib_udata *udata)
 899 {
 900         int status = -ENOMEM;
 901         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
 902         struct ocrdma_mr *mr;
 903         struct ocrdma_pd *pd;
 904         u32 num_pbes;
 905
 906         pd = get_ocrdma_pd(ibpd);
 907
 908         if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
 909                 return ERR_PTR(-EINVAL);
 910
 911         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
 912         if (!mr)
 913                 return ERR_PTR(status);
 914         mr->umem = ib_umem_get(ibpd->uobject->context, start, len, acc, 0);
 915         if (IS_ERR(mr->umem)) {
 916                 status = -EFAULT;
 917                 goto umem_err;
 918         }
 919         num_pbes = ib_umem_page_count(mr->umem);
 920         status = ocrdma_get_pbl_info(dev, mr, num_pbes);
 921         if (status)
 922                 goto umem_err;
 923
 924         mr->hwmr.pbe_size = mr->umem->page_size;
 925         mr->hwmr.fbo = ib_umem_offset(mr->umem);
 926         mr->hwmr.va = usr_addr;
 927         mr->hwmr.len = len;
 928         mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
 929         mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
 930         mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
 931         mr->hwmr.local_rd = 1;
 932         mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
 933         status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
 934         if (status)
 935                 goto umem_err;
 936         build_user_pbes(dev, mr, num_pbes);
 937         status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
 938         if (status)
 939                 goto mbx_err;
 940         mr->ibmr.lkey = mr->hwmr.lkey;
 941         if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
 942                 mr->ibmr.rkey = mr->hwmr.lkey;
 943
 944         return &mr->ibmr;
 945
 946 mbx_err:
 947         ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
 948 umem_err:
 949         kfree(mr);
 950         return ERR_PTR(status);
 951 }
 952
 953 int ocrdma_dereg_mr(struct ib_mr *ib_mr)
 954 {
 955         struct ocrdma_mr *mr = get_ocrdma_mr(ib_mr);
 956         struct ocrdma_dev *dev = get_ocrdma_dev(ib_mr->device);
 957
 958         (void) ocrdma_mbx_dealloc_lkey(dev, mr->hwmr.fr_mr, mr->hwmr.lkey);
 959
 960         ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
 961
 962         /* it could be user registered memory. */
 963         if (mr->umem)
 964                 ib_umem_release(mr->umem);
 965         kfree(mr);
 966
 967         /* Don't stop cleanup, in case FW is unresponsive */
 968         if (dev->mqe_ctx.fw_error_state) {
 969                 pr_err("%s(%d) fw not responding.\n",
 970                        __func__, dev->id);
 971         }
 972         return 0;
 973 }
 974
 975 static int ocrdma_copy_cq_uresp(struct ocrdma_dev *dev, struct ocrdma_cq *cq,
 976                                 struct ib_udata *udata,
 977                                 struct ib_ucontext *ib_ctx)
 978 {
 979         int status;
 980         struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
 981         struct ocrdma_create_cq_uresp uresp;
 982
 983         memset(&uresp, 0, sizeof(uresp));
 984         uresp.cq_id = cq->id;
 985         uresp.page_size = PAGE_ALIGN(cq->len);
 986         uresp.num_pages = 1;
 987         uresp.max_hw_cqe = cq->max_hw_cqe;
 988         uresp.page_addr[0] = virt_to_phys(cq->va);
 989         uresp.db_page_addr =  ocrdma_get_db_addr(dev, uctx->cntxt_pd->id);
 990         uresp.db_page_size = dev->nic_info.db_page_size;
 991         uresp.phase_change = cq->phase_change ? 1 : 0;
 992         status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
 993         if (status) {
 994                 pr_err("%s(%d) copy error cqid=0x%x.\n",
 995                        __func__, dev->id, cq->id);
 996                 goto err;
 997         }
 998         status = ocrdma_add_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
 999         if (status)
1000                 goto err;
1001         status = ocrdma_add_mmap(uctx, uresp.page_addr[0], uresp.page_size);
1002         if (status) {
1003                 ocrdma_del_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
1004                 goto err;
1005         }
1006         cq->ucontext = uctx;
1007 err:
1008         return status;
1009 }
1010
1011 struct ib_cq *ocrdma_create_cq(struct ib_device *ibdev,
1012                                const struct ib_cq_init_attr *attr,
1013                                struct ib_ucontext *ib_ctx,
1014                                struct ib_udata *udata)
1015 {
1016         int entries = attr->cqe;
1017         struct ocrdma_cq *cq;
1018         struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
1019         struct ocrdma_ucontext *uctx = NULL;
1020         u16 pd_id = 0;
1021         int status;
1022         struct ocrdma_create_cq_ureq ureq;
1023
1024         if (attr->flags)
1025                 return ERR_PTR(-EINVAL);
1026
1027         if (udata) {
1028                 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1029                         return ERR_PTR(-EFAULT);
1030         } else
1031                 ureq.dpp_cq = 0;
1032         cq = kzalloc(sizeof(*cq), GFP_KERNEL);
1033         if (!cq)
1034                 return ERR_PTR(-ENOMEM);
1035
1036         spin_lock_init(&cq->cq_lock);
1037         spin_lock_init(&cq->comp_handler_lock);
1038         INIT_LIST_HEAD(&cq->sq_head);
1039         INIT_LIST_HEAD(&cq->rq_head);
1040         cq->first_arm = true;
1041
1042         if (ib_ctx) {
1043                 uctx = get_ocrdma_ucontext(ib_ctx);
1044                 pd_id = uctx->cntxt_pd->id;
1045         }
1046
1047         status = ocrdma_mbx_create_cq(dev, cq, entries, ureq.dpp_cq, pd_id);
1048         if (status) {
1049                 kfree(cq);
1050                 return ERR_PTR(status);
1051         }
1052         if (ib_ctx) {
1053                 status = ocrdma_copy_cq_uresp(dev, cq, udata, ib_ctx);
1054                 if (status)
1055                         goto ctx_err;
1056         }
1057         cq->phase = OCRDMA_CQE_VALID;
1058         dev->cq_tbl[cq->id] = cq;
1059         return &cq->ibcq;
1060
1061 ctx_err:
1062         ocrdma_mbx_destroy_cq(dev, cq);
1063         kfree(cq);
1064         return ERR_PTR(status);
1065 }
1066
1067 int ocrdma_resize_cq(struct ib_cq *ibcq, int new_cnt,
1068                      struct ib_udata *udata)
1069 {
1070         int status = 0;
1071         struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1072
1073         if (new_cnt < 1 || new_cnt > cq->max_hw_cqe) {
1074                 status = -EINVAL;
1075                 return status;
1076         }
1077         ibcq->cqe = new_cnt;
1078         return status;
1079 }
1080
1081 static void ocrdma_flush_cq(struct ocrdma_cq *cq)
1082 {
1083         int cqe_cnt;
1084         int valid_count = 0;
1085         unsigned long flags;
1086
1087         struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
1088         struct ocrdma_cqe *cqe = NULL;
1089
1090         cqe = cq->va;
1091         cqe_cnt = cq->cqe_cnt;
1092
1093         /* Last irq might have scheduled a polling thread
1094          * sync-up with it before hard flushing.
1095          */
1096         spin_lock_irqsave(&cq->cq_lock, flags);
1097         while (cqe_cnt) {
1098                 if (is_cqe_valid(cq, cqe))
1099                         valid_count++;
1100                 cqe++;
1101                 cqe_cnt--;
1102         }
1103         ocrdma_ring_cq_db(dev, cq->id, false, false, valid_count);
1104         spin_unlock_irqrestore(&cq->cq_lock, flags);
1105 }
1106
1107 int ocrdma_destroy_cq(struct ib_cq *ibcq)
1108 {
1109         struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1110         struct ocrdma_eq *eq = NULL;
1111         struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
1112         int pdid = 0;
1113         u32 irq, indx;
1114
1115         dev->cq_tbl[cq->id] = NULL;
1116         indx = ocrdma_get_eq_table_index(dev, cq->eqn);
1117         if (indx == -EINVAL)
1118                 BUG();
1119
1120         eq = &dev->eq_tbl[indx];
1121         irq = ocrdma_get_irq(dev, eq);
1122         synchronize_irq(irq);
1123         ocrdma_flush_cq(cq);
1124
1125         (void)ocrdma_mbx_destroy_cq(dev, cq);
1126         if (cq->ucontext) {
1127                 pdid = cq->ucontext->cntxt_pd->id;
1128                 ocrdma_del_mmap(cq->ucontext, (u64) cq->pa,
1129                                 PAGE_ALIGN(cq->len));
1130                 ocrdma_del_mmap(cq->ucontext,
1131                                 ocrdma_get_db_addr(dev, pdid),
1132                                 dev->nic_info.db_page_size);
1133         }
1134
1135         kfree(cq);
1136         return 0;
1137 }
1138
1139 static int ocrdma_add_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1140 {
1141         int status = -EINVAL;
1142
1143         if (qp->id < OCRDMA_MAX_QP && dev->qp_tbl[qp->id] == NULL) {
1144                 dev->qp_tbl[qp->id] = qp;
1145                 status = 0;
1146         }
1147         return status;
1148 }
1149
1150 static void ocrdma_del_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1151 {
1152         dev->qp_tbl[qp->id] = NULL;
1153 }
1154
1155 static int ocrdma_check_qp_params(struct ib_pd *ibpd, struct ocrdma_dev *dev,
1156                                   struct ib_qp_init_attr *attrs)
1157 {
1158         if ((attrs->qp_type != IB_QPT_GSI) &&
1159             (attrs->qp_type != IB_QPT_RC) &&
1160             (attrs->qp_type != IB_QPT_UC) &&
1161             (attrs->qp_type != IB_QPT_UD)) {
1162                 pr_err("%s(%d) unsupported qp type=0x%x requested\n",
1163                        __func__, dev->id, attrs->qp_type);
1164                 return -EINVAL;
1165         }
1166         /* Skip the check for QP1 to support CM size of 128 */
1167         if ((attrs->qp_type != IB_QPT_GSI) &&
1168             (attrs->cap.max_send_wr > dev->attr.max_wqe)) {
1169                 pr_err("%s(%d) unsupported send_wr=0x%x requested\n",
1170                        __func__, dev->id, attrs->cap.max_send_wr);
1171                 pr_err("%s(%d) supported send_wr=0x%x\n",
1172                        __func__, dev->id, dev->attr.max_wqe);
1173                 return -EINVAL;
1174         }
1175         if (!attrs->srq && (attrs->cap.max_recv_wr > dev->attr.max_rqe)) {
1176                 pr_err("%s(%d) unsupported recv_wr=0x%x requested\n",
1177                        __func__, dev->id, attrs->cap.max_recv_wr);
1178                 pr_err("%s(%d) supported recv_wr=0x%x\n",
1179                        __func__, dev->id, dev->attr.max_rqe);
1180                 return -EINVAL;
1181         }
1182         if (attrs->cap.max_inline_data > dev->attr.max_inline_data) {
1183                 pr_err("%s(%d) unsupported inline data size=0x%x requested\n",
1184                        __func__, dev->id, attrs->cap.max_inline_data);
1185                 pr_err("%s(%d) supported inline data size=0x%x\n",
1186                        __func__, dev->id, dev->attr.max_inline_data);
1187                 return -EINVAL;
1188         }
1189         if (attrs->cap.max_send_sge > dev->attr.max_send_sge) {
1190                 pr_err("%s(%d) unsupported send_sge=0x%x requested\n",
1191                        __func__, dev->id, attrs->cap.max_send_sge);
1192                 pr_err("%s(%d) supported send_sge=0x%x\n",
1193                        __func__, dev->id, dev->attr.max_send_sge);
1194                 return -EINVAL;
1195         }
1196         if (attrs->cap.max_recv_sge > dev->attr.max_recv_sge) {
1197                 pr_err("%s(%d) unsupported recv_sge=0x%x requested\n",
1198                        __func__, dev->id, attrs->cap.max_recv_sge);
1199                 pr_err("%s(%d) supported recv_sge=0x%x\n",
1200                        __func__, dev->id, dev->attr.max_recv_sge);
1201                 return -EINVAL;
1202         }
1203         /* unprivileged user space cannot create special QP */
1204         if (ibpd->uobject && attrs->qp_type == IB_QPT_GSI) {
1205                 pr_err
1206                     ("%s(%d) Userspace can't create special QPs of type=0x%x\n",
1207                      __func__, dev->id, attrs->qp_type);
1208                 return -EINVAL;
1209         }
1210         /* allow creating only one GSI type of QP */
1211         if (attrs->qp_type == IB_QPT_GSI && dev->gsi_qp_created) {
1212                 pr_err("%s(%d) GSI special QPs already created.\n",
1213                        __func__, dev->id);
1214                 return -EINVAL;
1215         }
1216         /* verify consumer QPs are not trying to use GSI QP's CQ */
1217         if ((attrs->qp_type != IB_QPT_GSI) && (dev->gsi_qp_created)) {
1218                 if ((dev->gsi_sqcq == get_ocrdma_cq(attrs->send_cq)) ||
1219                         (dev->gsi_rqcq == get_ocrdma_cq(attrs->recv_cq))) {
1220                         pr_err("%s(%d) Consumer QP cannot use GSI CQs.\n",
1221                                 __func__, dev->id);
1222                         return -EINVAL;
1223                 }
1224         }
1225         return 0;
1226 }
1227
1228 static int ocrdma_copy_qp_uresp(struct ocrdma_qp *qp,
1229                                 struct ib_udata *udata, int dpp_offset,
1230                                 int dpp_credit_lmt, int srq)
1231 {
1232         int status = 0;
1233         u64 usr_db;
1234         struct ocrdma_create_qp_uresp uresp;
1235         struct ocrdma_pd *pd = qp->pd;
1236         struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
1237
1238         memset(&uresp, 0, sizeof(uresp));
1239         usr_db = dev->nic_info.unmapped_db +
1240                         (pd->id * dev->nic_info.db_page_size);
1241         uresp.qp_id = qp->id;
1242         uresp.sq_dbid = qp->sq.dbid;
1243         uresp.num_sq_pages = 1;
1244         uresp.sq_page_size = PAGE_ALIGN(qp->sq.len);
1245         uresp.sq_page_addr[0] = virt_to_phys(qp->sq.va);
1246         uresp.num_wqe_allocated = qp->sq.max_cnt;
1247         if (!srq) {
1248                 uresp.rq_dbid = qp->rq.dbid;
1249                 uresp.num_rq_pages = 1;
1250                 uresp.rq_page_size = PAGE_ALIGN(qp->rq.len);
1251                 uresp.rq_page_addr[0] = virt_to_phys(qp->rq.va);
1252                 uresp.num_rqe_allocated = qp->rq.max_cnt;
1253         }
1254         uresp.db_page_addr = usr_db;
1255         uresp.db_page_size = dev->nic_info.db_page_size;
1256         uresp.db_sq_offset = OCRDMA_DB_GEN2_SQ_OFFSET;
1257         uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1258         uresp.db_shift = OCRDMA_DB_RQ_SHIFT;
1259
1260         if (qp->dpp_enabled) {
1261                 uresp.dpp_credit = dpp_credit_lmt;
1262                 uresp.dpp_offset = dpp_offset;
1263         }
1264         status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1265         if (status) {
1266                 pr_err("%s(%d) user copy error.\n", __func__, dev->id);
1267                 goto err;
1268         }
1269         status = ocrdma_add_mmap(pd->uctx, uresp.sq_page_addr[0],
1270                                  uresp.sq_page_size);
1271         if (status)
1272                 goto err;
1273
1274         if (!srq) {
1275                 status = ocrdma_add_mmap(pd->uctx, uresp.rq_page_addr[0],
1276                                          uresp.rq_page_size);
1277                 if (status)
1278                         goto rq_map_err;
1279         }
1280         return status;
1281 rq_map_err:
1282         ocrdma_del_mmap(pd->uctx, uresp.sq_page_addr[0], uresp.sq_page_size);
1283 err:
1284         return status;
1285 }
1286
1287 static void ocrdma_set_qp_db(struct ocrdma_dev *dev, struct ocrdma_qp *qp,
1288                              struct ocrdma_pd *pd)
1289 {
1290         if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1291                 qp->sq_db = dev->nic_info.db +
1292                         (pd->id * dev->nic_info.db_page_size) +
1293                         OCRDMA_DB_GEN2_SQ_OFFSET;
1294                 qp->rq_db = dev->nic_info.db +
1295                         (pd->id * dev->nic_info.db_page_size) +
1296                         OCRDMA_DB_GEN2_RQ_OFFSET;
1297         } else {
1298                 qp->sq_db = dev->nic_info.db +
1299                         (pd->id * dev->nic_info.db_page_size) +
1300                         OCRDMA_DB_SQ_OFFSET;
1301                 qp->rq_db = dev->nic_info.db +
1302                         (pd->id * dev->nic_info.db_page_size) +
1303                         OCRDMA_DB_RQ_OFFSET;
1304         }
1305 }
1306
1307 static int ocrdma_alloc_wr_id_tbl(struct ocrdma_qp *qp)
1308 {
1309         qp->wqe_wr_id_tbl =
1310             kzalloc(sizeof(*(qp->wqe_wr_id_tbl)) * qp->sq.max_cnt,
1311                     GFP_KERNEL);
1312         if (qp->wqe_wr_id_tbl == NULL)
1313                 return -ENOMEM;
1314         qp->rqe_wr_id_tbl =
1315             kzalloc(sizeof(u64) * qp->rq.max_cnt, GFP_KERNEL);
1316         if (qp->rqe_wr_id_tbl == NULL)
1317                 return -ENOMEM;
1318
1319         return 0;
1320 }
1321
1322 static void ocrdma_set_qp_init_params(struct ocrdma_qp *qp,
1323                                       struct ocrdma_pd *pd,
1324                                       struct ib_qp_init_attr *attrs)
1325 {
1326         qp->pd = pd;
1327         spin_lock_init(&qp->q_lock);
1328         INIT_LIST_HEAD(&qp->sq_entry);
1329         INIT_LIST_HEAD(&qp->rq_entry);
1330
1331         qp->qp_type = attrs->qp_type;
1332         qp->cap_flags = OCRDMA_QP_INB_RD | OCRDMA_QP_INB_WR;
1333         qp->max_inline_data = attrs->cap.max_inline_data;
1334         qp->sq.max_sges = attrs->cap.max_send_sge;
1335         qp->rq.max_sges = attrs->cap.max_recv_sge;
1336         qp->state = OCRDMA_QPS_RST;
1337         qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
1338 }
1339
1340 static void ocrdma_store_gsi_qp_cq(struct ocrdma_dev *dev,
1341                                    struct ib_qp_init_attr *attrs)
1342 {
1343         if (attrs->qp_type == IB_QPT_GSI) {
1344                 dev->gsi_qp_created = 1;
1345                 dev->gsi_sqcq = get_ocrdma_cq(attrs->send_cq);
1346                 dev->gsi_rqcq = get_ocrdma_cq(attrs->recv_cq);
1347         }
1348 }
1349
1350 struct ib_qp *ocrdma_create_qp(struct ib_pd *ibpd,
1351                                struct ib_qp_init_attr *attrs,
1352                                struct ib_udata *udata)
1353 {
1354         int status;
1355         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1356         struct ocrdma_qp *qp;
1357         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1358         struct ocrdma_create_qp_ureq ureq;
1359         u16 dpp_credit_lmt, dpp_offset;
1360
1361         status = ocrdma_check_qp_params(ibpd, dev, attrs);
1362         if (status)
1363                 goto gen_err;
1364
1365         memset(&ureq, 0, sizeof(ureq));
1366         if (udata) {
1367                 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1368                         return ERR_PTR(-EFAULT);
1369         }
1370         qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1371         if (!qp) {
1372                 status = -ENOMEM;
1373                 goto gen_err;
1374         }
1375         ocrdma_set_qp_init_params(qp, pd, attrs);
1376         if (udata == NULL)
1377                 qp->cap_flags |= (OCRDMA_QP_MW_BIND | OCRDMA_QP_LKEY0 |
1378                                         OCRDMA_QP_FAST_REG);
1379
1380         mutex_lock(&dev->dev_lock);
1381         status = ocrdma_mbx_create_qp(qp, attrs, ureq.enable_dpp_cq,
1382                                         ureq.dpp_cq_id,
1383                                         &dpp_offset, &dpp_credit_lmt);
1384         if (status)
1385                 goto mbx_err;
1386
1387         /* user space QP's wr_id table are managed in library */
1388         if (udata == NULL) {
1389                 status = ocrdma_alloc_wr_id_tbl(qp);
1390                 if (status)
1391                         goto map_err;
1392         }
1393
1394         status = ocrdma_add_qpn_map(dev, qp);
1395         if (status)
1396                 goto map_err;
1397         ocrdma_set_qp_db(dev, qp, pd);
1398         if (udata) {
1399                 status = ocrdma_copy_qp_uresp(qp, udata, dpp_offset,
1400                                               dpp_credit_lmt,
1401                                               (attrs->srq != NULL));
1402                 if (status)
1403                         goto cpy_err;
1404         }
1405         ocrdma_store_gsi_qp_cq(dev, attrs);
1406         qp->ibqp.qp_num = qp->id;
1407         mutex_unlock(&dev->dev_lock);
1408         return &qp->ibqp;
1409
1410 cpy_err:
1411         ocrdma_del_qpn_map(dev, qp);
1412 map_err:
1413         ocrdma_mbx_destroy_qp(dev, qp);
1414 mbx_err:
1415         mutex_unlock(&dev->dev_lock);
1416         kfree(qp->wqe_wr_id_tbl);
1417         kfree(qp->rqe_wr_id_tbl);
1418         kfree(qp);
1419         pr_err("%s(%d) error=%d\n", __func__, dev->id, status);
1420 gen_err:
1421         return ERR_PTR(status);
1422 }
1423
1424 int _ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1425                       int attr_mask)
1426 {
1427         int status = 0;
1428         struct ocrdma_qp *qp;
1429         struct ocrdma_dev *dev;
1430         enum ib_qp_state old_qps;
1431
1432         qp = get_ocrdma_qp(ibqp);
1433         dev = get_ocrdma_dev(ibqp->device);
1434         if (attr_mask & IB_QP_STATE)
1435                 status = ocrdma_qp_state_change(qp, attr->qp_state, &old_qps);
1436         /* if new and previous states are same hw doesn't need to
1437          * know about it.
1438          */
1439         if (status < 0)
1440                 return status;
1441         status = ocrdma_mbx_modify_qp(dev, qp, attr, attr_mask);
1442
1443         return status;
1444 }
1445
1446 int ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1447                      int attr_mask, struct ib_udata *udata)
1448 {
1449         unsigned long flags;
1450         int status = -EINVAL;
1451         struct ocrdma_qp *qp;
1452         struct ocrdma_dev *dev;
1453         enum ib_qp_state old_qps, new_qps;
1454
1455         qp = get_ocrdma_qp(ibqp);
1456         dev = get_ocrdma_dev(ibqp->device);
1457
1458         /* syncronize with multiple context trying to change, retrive qps */
1459         mutex_lock(&dev->dev_lock);
1460         /* syncronize with wqe, rqe posting and cqe processing contexts */
1461         spin_lock_irqsave(&qp->q_lock, flags);
1462         old_qps = get_ibqp_state(qp->state);
1463         if (attr_mask & IB_QP_STATE)
1464                 new_qps = attr->qp_state;
1465         else
1466                 new_qps = old_qps;
1467         spin_unlock_irqrestore(&qp->q_lock, flags);
1468
1469         if (!ib_modify_qp_is_ok(old_qps, new_qps, ibqp->qp_type, attr_mask,
1470                                 IB_LINK_LAYER_ETHERNET)) {
1471                 pr_err("%s(%d) invalid attribute mask=0x%x specified for\n"
1472                        "qpn=0x%x of type=0x%x old_qps=0x%x, new_qps=0x%x\n",
1473                        __func__, dev->id, attr_mask, qp->id, ibqp->qp_type,
1474                        old_qps, new_qps);
1475                 goto param_err;
1476         }
1477
1478         status = _ocrdma_modify_qp(ibqp, attr, attr_mask);
1479         if (status > 0)
1480                 status = 0;
1481 param_err:
1482         mutex_unlock(&dev->dev_lock);
1483         return status;
1484 }
1485
1486 static enum ib_mtu ocrdma_mtu_int_to_enum(u16 mtu)
1487 {
1488         switch (mtu) {
1489         case 256:
1490                 return IB_MTU_256;
1491         case 512:
1492                 return IB_MTU_512;
1493         case 1024:
1494                 return IB_MTU_1024;
1495         case 2048:
1496                 return IB_MTU_2048;
1497         case 4096:
1498                 return IB_MTU_4096;
1499         default:
1500                 return IB_MTU_1024;
1501         }
1502 }
1503
1504 static int ocrdma_to_ib_qp_acc_flags(int qp_cap_flags)
1505 {
1506         int ib_qp_acc_flags = 0;
1507
1508         if (qp_cap_flags & OCRDMA_QP_INB_WR)
1509                 ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
1510         if (qp_cap_flags & OCRDMA_QP_INB_RD)
1511                 ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
1512         return ib_qp_acc_flags;
1513 }
1514
1515 int ocrdma_query_qp(struct ib_qp *ibqp,
1516                     struct ib_qp_attr *qp_attr,
1517                     int attr_mask, struct ib_qp_init_attr *qp_init_attr)
1518 {
1519         int status;
1520         u32 qp_state;
1521         struct ocrdma_qp_params params;
1522         struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1523         struct ocrdma_dev *dev = get_ocrdma_dev(ibqp->device);
1524
1525         memset(&params, 0, sizeof(params));
1526         mutex_lock(&dev->dev_lock);
1527         status = ocrdma_mbx_query_qp(dev, qp, &params);
1528         mutex_unlock(&dev->dev_lock);
1529         if (status)
1530                 goto mbx_err;
1531         if (qp->qp_type == IB_QPT_UD)
1532                 qp_attr->qkey = params.qkey;
1533         qp_attr->path_mtu =
1534                 ocrdma_mtu_int_to_enum(params.path_mtu_pkey_indx &
1535                                 OCRDMA_QP_PARAMS_PATH_MTU_MASK) >>
1536                                 OCRDMA_QP_PARAMS_PATH_MTU_SHIFT;
1537         qp_attr->path_mig_state = IB_MIG_MIGRATED;
1538         qp_attr->rq_psn = params.hop_lmt_rq_psn & OCRDMA_QP_PARAMS_RQ_PSN_MASK;
1539         qp_attr->sq_psn = params.tclass_sq_psn & OCRDMA_QP_PARAMS_SQ_PSN_MASK;
1540         qp_attr->dest_qp_num =
1541             params.ack_to_rnr_rtc_dest_qpn & OCRDMA_QP_PARAMS_DEST_QPN_MASK;
1542
1543         qp_attr->qp_access_flags = ocrdma_to_ib_qp_acc_flags(qp->cap_flags);
1544         qp_attr->cap.max_send_wr = qp->sq.max_cnt - 1;
1545         qp_attr->cap.max_recv_wr = qp->rq.max_cnt - 1;
1546         qp_attr->cap.max_send_sge = qp->sq.max_sges;
1547         qp_attr->cap.max_recv_sge = qp->rq.max_sges;
1548         qp_attr->cap.max_inline_data = qp->max_inline_data;
1549         qp_init_attr->cap = qp_attr->cap;
1550         memcpy(&qp_attr->ah_attr.grh.dgid, &params.dgid[0],
1551                sizeof(params.dgid));
1552         qp_attr->ah_attr.grh.flow_label = params.rnt_rc_sl_fl &
1553             OCRDMA_QP_PARAMS_FLOW_LABEL_MASK;
1554         qp_attr->ah_attr.grh.sgid_index = qp->sgid_idx;
1555         qp_attr->ah_attr.grh.hop_limit = (params.hop_lmt_rq_psn &
1556                                           OCRDMA_QP_PARAMS_HOP_LMT_MASK) >>
1557                                                 OCRDMA_QP_PARAMS_HOP_LMT_SHIFT;
1558         qp_attr->ah_attr.grh.traffic_class = (params.tclass_sq_psn &
1559                                               OCRDMA_QP_PARAMS_TCLASS_MASK) >>
1560                                                 OCRDMA_QP_PARAMS_TCLASS_SHIFT;
1561
1562         qp_attr->ah_attr.ah_flags = IB_AH_GRH;
1563         qp_attr->ah_attr.port_num = 1;
1564         qp_attr->ah_attr.sl = (params.rnt_rc_sl_fl &
1565                                OCRDMA_QP_PARAMS_SL_MASK) >>
1566                                 OCRDMA_QP_PARAMS_SL_SHIFT;
1567         qp_attr->timeout = (params.ack_to_rnr_rtc_dest_qpn &
1568                             OCRDMA_QP_PARAMS_ACK_TIMEOUT_MASK) >>
1569                                 OCRDMA_QP_PARAMS_ACK_TIMEOUT_SHIFT;
1570         qp_attr->rnr_retry = (params.ack_to_rnr_rtc_dest_qpn &
1571                               OCRDMA_QP_PARAMS_RNR_RETRY_CNT_MASK) >>
1572                                 OCRDMA_QP_PARAMS_RNR_RETRY_CNT_SHIFT;
1573         qp_attr->retry_cnt =
1574             (params.rnt_rc_sl_fl & OCRDMA_QP_PARAMS_RETRY_CNT_MASK) >>
1575                 OCRDMA_QP_PARAMS_RETRY_CNT_SHIFT;
1576         qp_attr->min_rnr_timer = 0;
1577         qp_attr->pkey_index = 0;
1578         qp_attr->port_num = 1;
1579         qp_attr->ah_attr.src_path_bits = 0;
1580         qp_attr->ah_attr.static_rate = 0;
1581         qp_attr->alt_pkey_index = 0;
1582         qp_attr->alt_port_num = 0;
1583         qp_attr->alt_timeout = 0;
1584         memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
1585         qp_state = (params.max_sge_recv_flags & OCRDMA_QP_PARAMS_STATE_MASK) >>
1586                     OCRDMA_QP_PARAMS_STATE_SHIFT;
1587         qp_attr->qp_state = get_ibqp_state(qp_state);
1588         qp_attr->cur_qp_state = qp_attr->qp_state;
1589         qp_attr->sq_draining = (qp_state == OCRDMA_QPS_SQ_DRAINING) ? 1 : 0;
1590         qp_attr->max_dest_rd_atomic =
1591             params.max_ord_ird >> OCRDMA_QP_PARAMS_MAX_ORD_SHIFT;
1592         qp_attr->max_rd_atomic =
1593             params.max_ord_ird & OCRDMA_QP_PARAMS_MAX_IRD_MASK;
1594         qp_attr->en_sqd_async_notify = (params.max_sge_recv_flags &
1595                                 OCRDMA_QP_PARAMS_FLAGS_SQD_ASYNC) ? 1 : 0;
1596         /* Sync driver QP state with FW */
1597         ocrdma_qp_state_change(qp, qp_attr->qp_state, NULL);
1598 mbx_err:
1599         return status;
1600 }
1601
1602 static void ocrdma_srq_toggle_bit(struct ocrdma_srq *srq, unsigned int idx)
1603 {
1604         unsigned int i = idx / 32;
1605         u32 mask = (1U << (idx % 32));
1606
1607         srq->idx_bit_fields[i] ^= mask;
1608 }
1609
1610 static int ocrdma_hwq_free_cnt(struct ocrdma_qp_hwq_info *q)
1611 {
1612         return ((q->max_wqe_idx - q->head) + q->tail) % q->max_cnt;
1613 }
1614
1615 static int is_hw_sq_empty(struct ocrdma_qp *qp)
1616 {
1617         return (qp->sq.tail == qp->sq.head);
1618 }
1619
1620 static int is_hw_rq_empty(struct ocrdma_qp *qp)
1621 {
1622         return (qp->rq.tail == qp->rq.head);
1623 }
1624
1625 static void *ocrdma_hwq_head(struct ocrdma_qp_hwq_info *q)
1626 {
1627         return q->va + (q->head * q->entry_size);
1628 }
1629
1630 static void *ocrdma_hwq_head_from_idx(struct ocrdma_qp_hwq_info *q,
1631                                       u32 idx)
1632 {
1633         return q->va + (idx * q->entry_size);
1634 }
1635
1636 static void ocrdma_hwq_inc_head(struct ocrdma_qp_hwq_info *q)
1637 {
1638         q->head = (q->head + 1) & q->max_wqe_idx;
1639 }
1640
1641 static void ocrdma_hwq_inc_tail(struct ocrdma_qp_hwq_info *q)
1642 {
1643         q->tail = (q->tail + 1) & q->max_wqe_idx;
1644 }
1645
1646 /* discard the cqe for a given QP */
1647 static void ocrdma_discard_cqes(struct ocrdma_qp *qp, struct ocrdma_cq *cq)
1648 {
1649         unsigned long cq_flags;
1650         unsigned long flags;
1651         int discard_cnt = 0;
1652         u32 cur_getp, stop_getp;
1653         struct ocrdma_cqe *cqe;
1654         u32 qpn = 0, wqe_idx = 0;
1655
1656         spin_lock_irqsave(&cq->cq_lock, cq_flags);
1657
1658         /* traverse through the CQEs in the hw CQ,
1659          * find the matching CQE for a given qp,
1660          * mark the matching one discarded by clearing qpn.
1661          * ring the doorbell in the poll_cq() as
1662          * we don't complete out of order cqe.
1663          */
1664
1665         cur_getp = cq->getp;
1666         /* find upto when do we reap the cq. */
1667         stop_getp = cur_getp;
1668         do {
1669                 if (is_hw_sq_empty(qp) && (!qp->srq && is_hw_rq_empty(qp)))
1670                         break;
1671
1672                 cqe = cq->va + cur_getp;
1673                 /* if (a) done reaping whole hw cq, or
1674                  *    (b) qp_xq becomes empty.
1675                  * then exit
1676                  */
1677                 qpn = cqe->cmn.qpn & OCRDMA_CQE_QPN_MASK;
1678                 /* if previously discarded cqe found, skip that too. */
1679                 /* check for matching qp */
1680                 if (qpn == 0 || qpn != qp->id)
1681                         goto skip_cqe;
1682
1683                 if (is_cqe_for_sq(cqe)) {
1684                         ocrdma_hwq_inc_tail(&qp->sq);
1685                 } else {
1686                         if (qp->srq) {
1687                                 wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
1688                                         OCRDMA_CQE_BUFTAG_SHIFT) &
1689                                         qp->srq->rq.max_wqe_idx;
1690                                 if (wqe_idx < 1)
1691                                         BUG();
1692                                 spin_lock_irqsave(&qp->srq->q_lock, flags);
1693                                 ocrdma_hwq_inc_tail(&qp->srq->rq);
1694                                 ocrdma_srq_toggle_bit(qp->srq, wqe_idx - 1);
1695                                 spin_unlock_irqrestore(&qp->srq->q_lock, flags);
1696
1697                         } else {
1698                                 ocrdma_hwq_inc_tail(&qp->rq);
1699                         }
1700                 }
1701                 /* mark cqe discarded so that it is not picked up later
1702                  * in the poll_cq().
1703                  */
1704                 discard_cnt += 1;
1705                 cqe->cmn.qpn = 0;
1706 skip_cqe:
1707                 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
1708         } while (cur_getp != stop_getp);
1709         spin_unlock_irqrestore(&cq->cq_lock, cq_flags);
1710 }
1711
1712 void ocrdma_del_flush_qp(struct ocrdma_qp *qp)
1713 {
1714         int found = false;
1715         unsigned long flags;
1716         struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
1717         /* sync with any active CQ poll */
1718
1719         spin_lock_irqsave(&dev->flush_q_lock, flags);
1720         found = ocrdma_is_qp_in_sq_flushlist(qp->sq_cq, qp);
1721         if (found)
1722                 list_del(&qp->sq_entry);
1723         if (!qp->srq) {
1724                 found = ocrdma_is_qp_in_rq_flushlist(qp->rq_cq, qp);
1725                 if (found)
1726                         list_del(&qp->rq_entry);
1727         }
1728         spin_unlock_irqrestore(&dev->flush_q_lock, flags);
1729 }
1730
1731 int ocrdma_destroy_qp(struct ib_qp *ibqp)
1732 {
1733         struct ocrdma_pd *pd;
1734         struct ocrdma_qp *qp;
1735         struct ocrdma_dev *dev;
1736         struct ib_qp_attr attrs;
1737         int attr_mask;
1738         unsigned long flags;
1739
1740         qp = get_ocrdma_qp(ibqp);
1741         dev = get_ocrdma_dev(ibqp->device);
1742
1743         pd = qp->pd;
1744
1745         /* change the QP state to ERROR */
1746         if (qp->state != OCRDMA_QPS_RST) {
1747                 attrs.qp_state = IB_QPS_ERR;
1748                 attr_mask = IB_QP_STATE;
1749                 _ocrdma_modify_qp(ibqp, &attrs, attr_mask);
1750         }
1751         /* ensure that CQEs for newly created QP (whose id may be same with
1752          * one which just getting destroyed are same), dont get
1753          * discarded until the old CQEs are discarded.
1754          */
1755         mutex_lock(&dev->dev_lock);
1756         (void) ocrdma_mbx_destroy_qp(dev, qp);
1757
1758         /*
1759          * acquire CQ lock while destroy is in progress, in order to
1760          * protect against proessing in-flight CQEs for this QP.
1761          */
1762         spin_lock_irqsave(&qp->sq_cq->cq_lock, flags);
1763         if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1764                 spin_lock(&qp->rq_cq->cq_lock);
1765
1766         ocrdma_del_qpn_map(dev, qp);
1767
1768         if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1769                 spin_unlock(&qp->rq_cq->cq_lock);
1770         spin_unlock_irqrestore(&qp->sq_cq->cq_lock, flags);
1771
1772         if (!pd->uctx) {
1773                 ocrdma_discard_cqes(qp, qp->sq_cq);
1774                 ocrdma_discard_cqes(qp, qp->rq_cq);
1775         }
1776         mutex_unlock(&dev->dev_lock);
1777
1778         if (pd->uctx) {
1779                 ocrdma_del_mmap(pd->uctx, (u64) qp->sq.pa,
1780                                 PAGE_ALIGN(qp->sq.len));
1781                 if (!qp->srq)
1782                         ocrdma_del_mmap(pd->uctx, (u64) qp->rq.pa,
1783                                         PAGE_ALIGN(qp->rq.len));
1784         }
1785
1786         ocrdma_del_flush_qp(qp);
1787
1788         kfree(qp->wqe_wr_id_tbl);
1789         kfree(qp->rqe_wr_id_tbl);
1790         kfree(qp);
1791         return 0;
1792 }
1793
1794 static int ocrdma_copy_srq_uresp(struct ocrdma_dev *dev, struct ocrdma_srq *srq,
1795                                 struct ib_udata *udata)
1796 {
1797         int status;
1798         struct ocrdma_create_srq_uresp uresp;
1799
1800         memset(&uresp, 0, sizeof(uresp));
1801         uresp.rq_dbid = srq->rq.dbid;
1802         uresp.num_rq_pages = 1;
1803         uresp.rq_page_addr[0] = virt_to_phys(srq->rq.va);
1804         uresp.rq_page_size = srq->rq.len;
1805         uresp.db_page_addr = dev->nic_info.unmapped_db +
1806             (srq->pd->id * dev->nic_info.db_page_size);
1807         uresp.db_page_size = dev->nic_info.db_page_size;
1808         uresp.num_rqe_allocated = srq->rq.max_cnt;
1809         if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1810                 uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1811                 uresp.db_shift = 24;
1812         } else {
1813                 uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
1814                 uresp.db_shift = 16;
1815         }
1816
1817         status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1818         if (status)
1819                 return status;
1820         status = ocrdma_add_mmap(srq->pd->uctx, uresp.rq_page_addr[0],
1821                                  uresp.rq_page_size);
1822         if (status)
1823                 return status;
1824         return status;
1825 }
1826
1827 struct ib_srq *ocrdma_create_srq(struct ib_pd *ibpd,
1828                                  struct ib_srq_init_attr *init_attr,
1829                                  struct ib_udata *udata)
1830 {
1831         int status = -ENOMEM;
1832         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1833         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1834         struct ocrdma_srq *srq;
1835
1836         if (init_attr->attr.max_sge > dev->attr.max_recv_sge)
1837                 return ERR_PTR(-EINVAL);
1838         if (init_attr->attr.max_wr > dev->attr.max_rqe)
1839                 return ERR_PTR(-EINVAL);
1840
1841         srq = kzalloc(sizeof(*srq), GFP_KERNEL);
1842         if (!srq)
1843                 return ERR_PTR(status);
1844
1845         spin_lock_init(&srq->q_lock);
1846         srq->pd = pd;
1847         srq->db = dev->nic_info.db + (pd->id * dev->nic_info.db_page_size);
1848         status = ocrdma_mbx_create_srq(dev, srq, init_attr, pd);
1849         if (status)
1850                 goto err;
1851
1852         if (udata == NULL) {
1853                 srq->rqe_wr_id_tbl = kzalloc(sizeof(u64) * srq->rq.max_cnt,
1854                             GFP_KERNEL);
1855                 if (srq->rqe_wr_id_tbl == NULL)
1856                         goto arm_err;
1857
1858                 srq->bit_fields_len = (srq->rq.max_cnt / 32) +
1859                     (srq->rq.max_cnt % 32 ? 1 : 0);
1860                 srq->idx_bit_fields =
1861                     kmalloc(srq->bit_fields_len * sizeof(u32), GFP_KERNEL);
1862                 if (srq->idx_bit_fields == NULL)
1863                         goto arm_err;
1864                 memset(srq->idx_bit_fields, 0xff,
1865                        srq->bit_fields_len * sizeof(u32));
1866         }
1867
1868         if (init_attr->attr.srq_limit) {
1869                 status = ocrdma_mbx_modify_srq(srq, &init_attr->attr);
1870                 if (status)
1871                         goto arm_err;
1872         }
1873
1874         if (udata) {
1875                 status = ocrdma_copy_srq_uresp(dev, srq, udata);
1876                 if (status)
1877                         goto arm_err;
1878         }
1879
1880         return &srq->ibsrq;
1881
1882 arm_err:
1883         ocrdma_mbx_destroy_srq(dev, srq);
1884 err:
1885         kfree(srq->rqe_wr_id_tbl);
1886         kfree(srq->idx_bit_fields);
1887         kfree(srq);
1888         return ERR_PTR(status);
1889 }
1890
1891 int ocrdma_modify_srq(struct ib_srq *ibsrq,
1892                       struct ib_srq_attr *srq_attr,
1893                       enum ib_srq_attr_mask srq_attr_mask,
1894                       struct ib_udata *udata)
1895 {
1896         int status = 0;
1897         struct ocrdma_srq *srq;
1898
1899         srq = get_ocrdma_srq(ibsrq);
1900         if (srq_attr_mask & IB_SRQ_MAX_WR)
1901                 status = -EINVAL;
1902         else
1903                 status = ocrdma_mbx_modify_srq(srq, srq_attr);
1904         return status;
1905 }
1906
1907 int ocrdma_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
1908 {
1909         int status;
1910         struct ocrdma_srq *srq;
1911
1912         srq = get_ocrdma_srq(ibsrq);
1913         status = ocrdma_mbx_query_srq(srq, srq_attr);
1914         return status;
1915 }
1916
1917 int ocrdma_destroy_srq(struct ib_srq *ibsrq)
1918 {
1919         int status;
1920         struct ocrdma_srq *srq;
1921         struct ocrdma_dev *dev = get_ocrdma_dev(ibsrq->device);
1922
1923         srq = get_ocrdma_srq(ibsrq);
1924
1925         status = ocrdma_mbx_destroy_srq(dev, srq);
1926
1927         if (srq->pd->uctx)
1928                 ocrdma_del_mmap(srq->pd->uctx, (u64) srq->rq.pa,
1929                                 PAGE_ALIGN(srq->rq.len));
1930
1931         kfree(srq->idx_bit_fields);
1932         kfree(srq->rqe_wr_id_tbl);
1933         kfree(srq);
1934         return status;
1935 }
1936
1937 /* unprivileged verbs and their support functions. */
1938 static void ocrdma_build_ud_hdr(struct ocrdma_qp *qp,
1939                                 struct ocrdma_hdr_wqe *hdr,
1940                                 struct ib_send_wr *wr)
1941 {
1942         struct ocrdma_ewqe_ud_hdr *ud_hdr =
1943                 (struct ocrdma_ewqe_ud_hdr *)(hdr + 1);
1944         struct ocrdma_ah *ah = get_ocrdma_ah(wr->wr.ud.ah);
1945
1946         ud_hdr->rsvd_dest_qpn = wr->wr.ud.remote_qpn;
1947         if (qp->qp_type == IB_QPT_GSI)
1948                 ud_hdr->qkey = qp->qkey;
1949         else
1950                 ud_hdr->qkey = wr->wr.ud.remote_qkey;
1951         ud_hdr->rsvd_ahid = ah->id;
1952         if (ah->av->valid & OCRDMA_AV_VLAN_VALID)
1953                 hdr->cw |= (OCRDMA_FLAG_AH_VLAN_PR << OCRDMA_WQE_FLAGS_SHIFT);
1954 }
1955
1956 static void ocrdma_build_sges(struct ocrdma_hdr_wqe *hdr,
1957                               struct ocrdma_sge *sge, int num_sge,
1958                               struct ib_sge *sg_list)
1959 {
1960         int i;
1961
1962         for (i = 0; i < num_sge; i++) {
1963                 sge[i].lrkey = sg_list[i].lkey;
1964                 sge[i].addr_lo = sg_list[i].addr;
1965                 sge[i].addr_hi = upper_32_bits(sg_list[i].addr);
1966                 sge[i].len = sg_list[i].length;
1967                 hdr->total_len += sg_list[i].length;
1968         }
1969         if (num_sge == 0)
1970                 memset(sge, 0, sizeof(*sge));
1971 }
1972
1973 static inline uint32_t ocrdma_sglist_len(struct ib_sge *sg_list, int num_sge)
1974 {
1975         uint32_t total_len = 0, i;
1976
1977         for (i = 0; i < num_sge; i++)
1978                 total_len += sg_list[i].length;
1979         return total_len;
1980 }
1981
1982
1983 static int ocrdma_build_inline_sges(struct ocrdma_qp *qp,
1984                                     struct ocrdma_hdr_wqe *hdr,
1985                                     struct ocrdma_sge *sge,
1986                                     struct ib_send_wr *wr, u32 wqe_size)
1987 {
1988         int i;
1989         char *dpp_addr;
1990
1991         if (wr->send_flags & IB_SEND_INLINE && qp->qp_type != IB_QPT_UD) {
1992                 hdr->total_len = ocrdma_sglist_len(wr->sg_list, wr->num_sge);
1993                 if (unlikely(hdr->total_len > qp->max_inline_data)) {
1994                         pr_err("%s() supported_len=0x%x,\n"
1995                                " unsupported len req=0x%x\n", __func__,
1996                                 qp->max_inline_data, hdr->total_len);
1997                         return -EINVAL;
1998                 }
1999                 dpp_addr = (char *)sge;
2000                 for (i = 0; i < wr->num_sge; i++) {
2001                         memcpy(dpp_addr,
2002                                (void *)(unsigned long)wr->sg_list[i].addr,
2003                                wr->sg_list[i].length);
2004                         dpp_addr += wr->sg_list[i].length;
2005                 }
2006
2007                 wqe_size += roundup(hdr->total_len, OCRDMA_WQE_ALIGN_BYTES);
2008                 if (0 == hdr->total_len)
2009                         wqe_size += sizeof(struct ocrdma_sge);
2010                 hdr->cw |= (OCRDMA_TYPE_INLINE << OCRDMA_WQE_TYPE_SHIFT);
2011         } else {
2012                 ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
2013                 if (wr->num_sge)
2014                         wqe_size += (wr->num_sge * sizeof(struct ocrdma_sge));
2015                 else
2016                         wqe_size += sizeof(struct ocrdma_sge);
2017                 hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2018         }
2019         hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2020         return 0;
2021 }
2022
2023 static int ocrdma_build_send(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2024                              struct ib_send_wr *wr)
2025 {
2026         int status;
2027         struct ocrdma_sge *sge;
2028         u32 wqe_size = sizeof(*hdr);
2029
2030         if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2031                 ocrdma_build_ud_hdr(qp, hdr, wr);
2032                 sge = (struct ocrdma_sge *)(hdr + 2);
2033                 wqe_size += sizeof(struct ocrdma_ewqe_ud_hdr);
2034         } else {
2035                 sge = (struct ocrdma_sge *)(hdr + 1);
2036         }
2037
2038         status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
2039         return status;
2040 }
2041
2042 static int ocrdma_build_write(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2043                               struct ib_send_wr *wr)
2044 {
2045         int status;
2046         struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
2047         struct ocrdma_sge *sge = ext_rw + 1;
2048         u32 wqe_size = sizeof(*hdr) + sizeof(*ext_rw);
2049
2050         status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
2051         if (status)
2052                 return status;
2053         ext_rw->addr_lo = wr->wr.rdma.remote_addr;
2054         ext_rw->addr_hi = upper_32_bits(wr->wr.rdma.remote_addr);
2055         ext_rw->lrkey = wr->wr.rdma.rkey;
2056         ext_rw->len = hdr->total_len;
2057         return 0;
2058 }
2059
2060 static void ocrdma_build_read(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2061                               struct ib_send_wr *wr)
2062 {
2063         struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
2064         struct ocrdma_sge *sge = ext_rw + 1;
2065         u32 wqe_size = ((wr->num_sge + 1) * sizeof(struct ocrdma_sge)) +
2066             sizeof(struct ocrdma_hdr_wqe);
2067
2068         ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
2069         hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2070         hdr->cw |= (OCRDMA_READ << OCRDMA_WQE_OPCODE_SHIFT);
2071         hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2072
2073         ext_rw->addr_lo = wr->wr.rdma.remote_addr;
2074         ext_rw->addr_hi = upper_32_bits(wr->wr.rdma.remote_addr);
2075         ext_rw->lrkey = wr->wr.rdma.rkey;
2076         ext_rw->len = hdr->total_len;
2077 }
2078
2079 static void build_frmr_pbes(struct ib_send_wr *wr, struct ocrdma_pbl *pbl_tbl,
2080                             struct ocrdma_hw_mr *hwmr)
2081 {
2082         int i;
2083         u64 buf_addr = 0;
2084         int num_pbes;
2085         struct ocrdma_pbe *pbe;
2086
2087         pbe = (struct ocrdma_pbe *)pbl_tbl->va;
2088         num_pbes = 0;
2089
2090         /* go through the OS phy regions & fill hw pbe entries into pbls. */
2091         for (i = 0; i < wr->wr.fast_reg.page_list_len; i++) {
2092                 /* number of pbes can be more for one OS buf, when
2093                  * buffers are of different sizes.
2094                  * split the ib_buf to one or more pbes.
2095                  */
2096                 buf_addr = wr->wr.fast_reg.page_list->page_list[i];
2097                 pbe->pa_lo = cpu_to_le32((u32) (buf_addr & PAGE_MASK));
2098                 pbe->pa_hi = cpu_to_le32((u32) upper_32_bits(buf_addr));
2099                 num_pbes += 1;
2100                 pbe++;
2101
2102                 /* if the pbl is full storing the pbes,
2103                  * move to next pbl.
2104                 */
2105                 if (num_pbes == (hwmr->pbl_size/sizeof(u64))) {
2106                         pbl_tbl++;
2107                         pbe = (struct ocrdma_pbe *)pbl_tbl->va;
2108                 }
2109         }
2110         return;
2111 }
2112
2113 static int get_encoded_page_size(int pg_sz)
2114 {
2115         /* Max size is 256M 4096 << 16 */
2116         int i = 0;
2117         for (; i < 17; i++)
2118                 if (pg_sz == (4096 << i))
2119                         break;
2120         return i;
2121 }
2122
2123
2124 static int ocrdma_build_fr(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2125                            struct ib_send_wr *wr)
2126 {
2127         u64 fbo;
2128         struct ocrdma_ewqe_fr *fast_reg = (struct ocrdma_ewqe_fr *)(hdr + 1);
2129         struct ocrdma_mr *mr;
2130         struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2131         u32 wqe_size = sizeof(*fast_reg) + sizeof(*hdr);
2132
2133         wqe_size = roundup(wqe_size, OCRDMA_WQE_ALIGN_BYTES);
2134
2135         if (wr->wr.fast_reg.page_list_len > dev->attr.max_pages_per_frmr)
2136                 return -EINVAL;
2137
2138         hdr->cw |= (OCRDMA_FR_MR << OCRDMA_WQE_OPCODE_SHIFT);
2139         hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2140
2141         if (wr->wr.fast_reg.page_list_len == 0)
2142                 BUG();
2143         if (wr->wr.fast_reg.access_flags & IB_ACCESS_LOCAL_WRITE)
2144                 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_LOCAL_WR;
2145         if (wr->wr.fast_reg.access_flags & IB_ACCESS_REMOTE_WRITE)
2146                 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_WR;
2147         if (wr->wr.fast_reg.access_flags & IB_ACCESS_REMOTE_READ)
2148                 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_RD;
2149         hdr->lkey = wr->wr.fast_reg.rkey;
2150         hdr->total_len = wr->wr.fast_reg.length;
2151
2152         fbo = wr->wr.fast_reg.iova_start -
2153             (wr->wr.fast_reg.page_list->page_list[0] & PAGE_MASK);
2154
2155         fast_reg->va_hi = upper_32_bits(wr->wr.fast_reg.iova_start);
2156         fast_reg->va_lo = (u32) (wr->wr.fast_reg.iova_start & 0xffffffff);
2157         fast_reg->fbo_hi = upper_32_bits(fbo);
2158         fast_reg->fbo_lo = (u32) fbo & 0xffffffff;
2159         fast_reg->num_sges = wr->wr.fast_reg.page_list_len;
2160         fast_reg->size_sge =
2161                 get_encoded_page_size(1 << wr->wr.fast_reg.page_shift);
2162         mr = (struct ocrdma_mr *) (unsigned long)
2163                 dev->stag_arr[(hdr->lkey >> 8) & (OCRDMA_MAX_STAG - 1)];
2164         build_frmr_pbes(wr, mr->hwmr.pbl_table, &mr->hwmr);
2165         return 0;
2166 }
2167
2168 static void ocrdma_ring_sq_db(struct ocrdma_qp *qp)
2169 {
2170         u32 val = qp->sq.dbid | (1 << OCRDMA_DB_SQ_SHIFT);
2171
2172         iowrite32(val, qp->sq_db);
2173 }
2174
2175 int ocrdma_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2176                      struct ib_send_wr **bad_wr)
2177 {
2178         int status = 0;
2179         struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2180         struct ocrdma_hdr_wqe *hdr;
2181         unsigned long flags;
2182
2183         spin_lock_irqsave(&qp->q_lock, flags);
2184         if (qp->state != OCRDMA_QPS_RTS && qp->state != OCRDMA_QPS_SQD) {
2185                 spin_unlock_irqrestore(&qp->q_lock, flags);
2186                 *bad_wr = wr;
2187                 return -EINVAL;
2188         }
2189
2190         while (wr) {
2191                 if (qp->qp_type == IB_QPT_UD &&
2192                     (wr->opcode != IB_WR_SEND &&
2193                      wr->opcode != IB_WR_SEND_WITH_IMM)) {
2194                         *bad_wr = wr;
2195                         status = -EINVAL;
2196                         break;
2197                 }
2198                 if (ocrdma_hwq_free_cnt(&qp->sq) == 0 ||
2199                     wr->num_sge > qp->sq.max_sges) {
2200                         *bad_wr = wr;
2201                         status = -ENOMEM;
2202                         break;
2203                 }
2204                 hdr = ocrdma_hwq_head(&qp->sq);
2205                 hdr->cw = 0;
2206                 if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2207                         hdr->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2208                 if (wr->send_flags & IB_SEND_FENCE)
2209                         hdr->cw |=
2210                             (OCRDMA_FLAG_FENCE_L << OCRDMA_WQE_FLAGS_SHIFT);
2211                 if (wr->send_flags & IB_SEND_SOLICITED)
2212                         hdr->cw |=
2213                             (OCRDMA_FLAG_SOLICIT << OCRDMA_WQE_FLAGS_SHIFT);
2214                 hdr->total_len = 0;
2215                 switch (wr->opcode) {
2216                 case IB_WR_SEND_WITH_IMM:
2217                         hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2218                         hdr->immdt = ntohl(wr->ex.imm_data);
2219                 case IB_WR_SEND:
2220                         hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2221                         ocrdma_build_send(qp, hdr, wr);
2222                         break;
2223                 case IB_WR_SEND_WITH_INV:
2224                         hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
2225                         hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2226                         hdr->lkey = wr->ex.invalidate_rkey;
2227                         status = ocrdma_build_send(qp, hdr, wr);
2228                         break;
2229                 case IB_WR_RDMA_WRITE_WITH_IMM:
2230                         hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2231                         hdr->immdt = ntohl(wr->ex.imm_data);
2232                 case IB_WR_RDMA_WRITE:
2233                         hdr->cw |= (OCRDMA_WRITE << OCRDMA_WQE_OPCODE_SHIFT);
2234                         status = ocrdma_build_write(qp, hdr, wr);
2235                         break;
2236                 case IB_WR_RDMA_READ:
2237                         ocrdma_build_read(qp, hdr, wr);
2238                         break;
2239                 case IB_WR_LOCAL_INV:
2240                         hdr->cw |=
2241                             (OCRDMA_LKEY_INV << OCRDMA_WQE_OPCODE_SHIFT);
2242                         hdr->cw |= ((sizeof(struct ocrdma_hdr_wqe) +
2243                                         sizeof(struct ocrdma_sge)) /
2244                                 OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT;
2245                         hdr->lkey = wr->ex.invalidate_rkey;
2246                         break;
2247                 case IB_WR_FAST_REG_MR:
2248                         status = ocrdma_build_fr(qp, hdr, wr);
2249                         break;
2250                 default:
2251                         status = -EINVAL;
2252                         break;
2253                 }
2254                 if (status) {
2255                         *bad_wr = wr;
2256                         break;
2257                 }
2258                 if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2259                         qp->wqe_wr_id_tbl[qp->sq.head].signaled = 1;
2260                 else
2261                         qp->wqe_wr_id_tbl[qp->sq.head].signaled = 0;
2262                 qp->wqe_wr_id_tbl[qp->sq.head].wrid = wr->wr_id;
2263                 ocrdma_cpu_to_le32(hdr, ((hdr->cw >> OCRDMA_WQE_SIZE_SHIFT) &
2264                                    OCRDMA_WQE_SIZE_MASK) * OCRDMA_WQE_STRIDE);
2265                 /* make sure wqe is written before adapter can access it */
2266                 wmb();
2267                 /* inform hw to start processing it */
2268                 ocrdma_ring_sq_db(qp);
2269
2270                 /* update pointer, counter for next wr */
2271                 ocrdma_hwq_inc_head(&qp->sq);
2272                 wr = wr->next;
2273         }
2274         spin_unlock_irqrestore(&qp->q_lock, flags);
2275         return status;
2276 }
2277
2278 static void ocrdma_ring_rq_db(struct ocrdma_qp *qp)
2279 {
2280         u32 val = qp->rq.dbid | (1 << OCRDMA_DB_RQ_SHIFT);
2281
2282         iowrite32(val, qp->rq_db);
2283 }
2284
2285 static void ocrdma_build_rqe(struct ocrdma_hdr_wqe *rqe, struct ib_recv_wr *wr,
2286                              u16 tag)
2287 {
2288         u32 wqe_size = 0;
2289         struct ocrdma_sge *sge;
2290         if (wr->num_sge)
2291                 wqe_size = (wr->num_sge * sizeof(*sge)) + sizeof(*rqe);
2292         else
2293                 wqe_size = sizeof(*sge) + sizeof(*rqe);
2294
2295         rqe->cw = ((wqe_size / OCRDMA_WQE_STRIDE) <<
2296                                 OCRDMA_WQE_SIZE_SHIFT);
2297         rqe->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2298         rqe->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2299         rqe->total_len = 0;
2300         rqe->rsvd_tag = tag;
2301         sge = (struct ocrdma_sge *)(rqe + 1);
2302         ocrdma_build_sges(rqe, sge, wr->num_sge, wr->sg_list);
2303         ocrdma_cpu_to_le32(rqe, wqe_size);
2304 }
2305
2306 int ocrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2307                      struct ib_recv_wr **bad_wr)
2308 {
2309         int status = 0;
2310         unsigned long flags;
2311         struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2312         struct ocrdma_hdr_wqe *rqe;
2313
2314         spin_lock_irqsave(&qp->q_lock, flags);
2315         if (qp->state == OCRDMA_QPS_RST || qp->state == OCRDMA_QPS_ERR) {
2316                 spin_unlock_irqrestore(&qp->q_lock, flags);
2317                 *bad_wr = wr;
2318                 return -EINVAL;
2319         }
2320         while (wr) {
2321                 if (ocrdma_hwq_free_cnt(&qp->rq) == 0 ||
2322                     wr->num_sge > qp->rq.max_sges) {
2323                         *bad_wr = wr;
2324                         status = -ENOMEM;
2325                         break;
2326                 }
2327                 rqe = ocrdma_hwq_head(&qp->rq);
2328                 ocrdma_build_rqe(rqe, wr, 0);
2329
2330                 qp->rqe_wr_id_tbl[qp->rq.head] = wr->wr_id;
2331                 /* make sure rqe is written before adapter can access it */
2332                 wmb();
2333
2334                 /* inform hw to start processing it */
2335                 ocrdma_ring_rq_db(qp);
2336
2337                 /* update pointer, counter for next wr */
2338                 ocrdma_hwq_inc_head(&qp->rq);
2339                 wr = wr->next;
2340         }
2341         spin_unlock_irqrestore(&qp->q_lock, flags);
2342         return status;
2343 }
2344
2345 /* cqe for srq's rqe can potentially arrive out of order.
2346  * index gives the entry in the shadow table where to store
2347  * the wr_id. tag/index is returned in cqe to reference back
2348  * for a given rqe.
2349  */
2350 static int ocrdma_srq_get_idx(struct ocrdma_srq *srq)
2351 {
2352         int row = 0;
2353         int indx = 0;
2354
2355         for (row = 0; row < srq->bit_fields_len; row++) {
2356                 if (srq->idx_bit_fields[row]) {
2357                         indx = ffs(srq->idx_bit_fields[row]);
2358                         indx = (row * 32) + (indx - 1);
2359                         if (indx >= srq->rq.max_cnt)
2360                                 BUG();
2361                         ocrdma_srq_toggle_bit(srq, indx);
2362                         break;
2363                 }
2364         }
2365
2366         if (row == srq->bit_fields_len)
2367                 BUG();
2368         return indx + 1; /* Use from index 1 */
2369 }
2370
2371 static void ocrdma_ring_srq_db(struct ocrdma_srq *srq)
2372 {
2373         u32 val = srq->rq.dbid | (1 << 16);
2374
2375         iowrite32(val, srq->db + OCRDMA_DB_GEN2_SRQ_OFFSET);
2376 }
2377
2378 int ocrdma_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
2379                          struct ib_recv_wr **bad_wr)
2380 {
2381         int status = 0;
2382         unsigned long flags;
2383         struct ocrdma_srq *srq;
2384         struct ocrdma_hdr_wqe *rqe;
2385         u16 tag;
2386
2387         srq = get_ocrdma_srq(ibsrq);
2388
2389         spin_lock_irqsave(&srq->q_lock, flags);
2390         while (wr) {
2391                 if (ocrdma_hwq_free_cnt(&srq->rq) == 0 ||
2392                     wr->num_sge > srq->rq.max_sges) {
2393                         status = -ENOMEM;
2394                         *bad_wr = wr;
2395                         break;
2396                 }
2397                 tag = ocrdma_srq_get_idx(srq);
2398                 rqe = ocrdma_hwq_head(&srq->rq);
2399                 ocrdma_build_rqe(rqe, wr, tag);
2400
2401                 srq->rqe_wr_id_tbl[tag] = wr->wr_id;
2402                 /* make sure rqe is written before adapter can perform DMA */
2403                 wmb();
2404                 /* inform hw to start processing it */
2405                 ocrdma_ring_srq_db(srq);
2406                 /* update pointer, counter for next wr */
2407                 ocrdma_hwq_inc_head(&srq->rq);
2408                 wr = wr->next;
2409         }
2410         spin_unlock_irqrestore(&srq->q_lock, flags);
2411         return status;
2412 }
2413
2414 static enum ib_wc_status ocrdma_to_ibwc_err(u16 status)
2415 {
2416         enum ib_wc_status ibwc_status;
2417
2418         switch (status) {
2419         case OCRDMA_CQE_GENERAL_ERR:
2420                 ibwc_status = IB_WC_GENERAL_ERR;
2421                 break;
2422         case OCRDMA_CQE_LOC_LEN_ERR:
2423                 ibwc_status = IB_WC_LOC_LEN_ERR;
2424                 break;
2425         case OCRDMA_CQE_LOC_QP_OP_ERR:
2426                 ibwc_status = IB_WC_LOC_QP_OP_ERR;
2427                 break;
2428         case OCRDMA_CQE_LOC_EEC_OP_ERR:
2429                 ibwc_status = IB_WC_LOC_EEC_OP_ERR;
2430                 break;
2431         case OCRDMA_CQE_LOC_PROT_ERR:
2432                 ibwc_status = IB_WC_LOC_PROT_ERR;
2433                 break;
2434         case OCRDMA_CQE_WR_FLUSH_ERR:
2435                 ibwc_status = IB_WC_WR_FLUSH_ERR;
2436                 break;
2437         case OCRDMA_CQE_MW_BIND_ERR:
2438                 ibwc_status = IB_WC_MW_BIND_ERR;
2439                 break;
2440         case OCRDMA_CQE_BAD_RESP_ERR:
2441                 ibwc_status = IB_WC_BAD_RESP_ERR;
2442                 break;
2443         case OCRDMA_CQE_LOC_ACCESS_ERR:
2444                 ibwc_status = IB_WC_LOC_ACCESS_ERR;
2445                 break;
2446         case OCRDMA_CQE_REM_INV_REQ_ERR:
2447                 ibwc_status = IB_WC_REM_INV_REQ_ERR;
2448                 break;
2449         case OCRDMA_CQE_REM_ACCESS_ERR:
2450                 ibwc_status = IB_WC_REM_ACCESS_ERR;
2451                 break;
2452         case OCRDMA_CQE_REM_OP_ERR:
2453                 ibwc_status = IB_WC_REM_OP_ERR;
2454                 break;
2455         case OCRDMA_CQE_RETRY_EXC_ERR:
2456                 ibwc_status = IB_WC_RETRY_EXC_ERR;
2457                 break;
2458         case OCRDMA_CQE_RNR_RETRY_EXC_ERR:
2459                 ibwc_status = IB_WC_RNR_RETRY_EXC_ERR;
2460                 break;
2461         case OCRDMA_CQE_LOC_RDD_VIOL_ERR:
2462                 ibwc_status = IB_WC_LOC_RDD_VIOL_ERR;
2463                 break;
2464         case OCRDMA_CQE_REM_INV_RD_REQ_ERR:
2465                 ibwc_status = IB_WC_REM_INV_RD_REQ_ERR;
2466                 break;
2467         case OCRDMA_CQE_REM_ABORT_ERR:
2468                 ibwc_status = IB_WC_REM_ABORT_ERR;
2469                 break;
2470         case OCRDMA_CQE_INV_EECN_ERR:
2471                 ibwc_status = IB_WC_INV_EECN_ERR;
2472                 break;
2473         case OCRDMA_CQE_INV_EEC_STATE_ERR:
2474                 ibwc_status = IB_WC_INV_EEC_STATE_ERR;
2475                 break;
2476         case OCRDMA_CQE_FATAL_ERR:
2477                 ibwc_status = IB_WC_FATAL_ERR;
2478                 break;
2479         case OCRDMA_CQE_RESP_TIMEOUT_ERR:
2480                 ibwc_status = IB_WC_RESP_TIMEOUT_ERR;
2481                 break;
2482         default:
2483                 ibwc_status = IB_WC_GENERAL_ERR;
2484                 break;
2485         }
2486         return ibwc_status;
2487 }
2488
2489 static void ocrdma_update_wc(struct ocrdma_qp *qp, struct ib_wc *ibwc,
2490                       u32 wqe_idx)
2491 {
2492         struct ocrdma_hdr_wqe *hdr;
2493         struct ocrdma_sge *rw;
2494         int opcode;
2495
2496         hdr = ocrdma_hwq_head_from_idx(&qp->sq, wqe_idx);
2497
2498         ibwc->wr_id = qp->wqe_wr_id_tbl[wqe_idx].wrid;
2499         /* Undo the hdr->cw swap */
2500         opcode = le32_to_cpu(hdr->cw) & OCRDMA_WQE_OPCODE_MASK;
2501         switch (opcode) {
2502         case OCRDMA_WRITE:
2503                 ibwc->opcode = IB_WC_RDMA_WRITE;
2504                 break;
2505         case OCRDMA_READ:
2506                 rw = (struct ocrdma_sge *)(hdr + 1);
2507                 ibwc->opcode = IB_WC_RDMA_READ;
2508                 ibwc->byte_len = rw->len;
2509                 break;
2510         case OCRDMA_SEND:
2511                 ibwc->opcode = IB_WC_SEND;
2512                 break;
2513         case OCRDMA_FR_MR:
2514                 ibwc->opcode = IB_WC_FAST_REG_MR;
2515                 break;
2516         case OCRDMA_LKEY_INV:
2517                 ibwc->opcode = IB_WC_LOCAL_INV;
2518                 break;
2519         default:
2520                 ibwc->status = IB_WC_GENERAL_ERR;
2521                 pr_err("%s() invalid opcode received = 0x%x\n",
2522                        __func__, hdr->cw & OCRDMA_WQE_OPCODE_MASK);
2523                 break;
2524         }
2525 }
2526
2527 static void ocrdma_set_cqe_status_flushed(struct ocrdma_qp *qp,
2528                                                 struct ocrdma_cqe *cqe)
2529 {
2530         if (is_cqe_for_sq(cqe)) {
2531                 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2532                                 cqe->flags_status_srcqpn) &
2533                                         ~OCRDMA_CQE_STATUS_MASK);
2534                 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2535                                 cqe->flags_status_srcqpn) |
2536                                 (OCRDMA_CQE_WR_FLUSH_ERR <<
2537                                         OCRDMA_CQE_STATUS_SHIFT));
2538         } else {
2539                 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2540                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2541                                         cqe->flags_status_srcqpn) &
2542                                                 ~OCRDMA_CQE_UD_STATUS_MASK);
2543                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2544                                         cqe->flags_status_srcqpn) |
2545                                         (OCRDMA_CQE_WR_FLUSH_ERR <<
2546                                                 OCRDMA_CQE_UD_STATUS_SHIFT));
2547                 } else {
2548                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2549                                         cqe->flags_status_srcqpn) &
2550                                                 ~OCRDMA_CQE_STATUS_MASK);
2551                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2552                                         cqe->flags_status_srcqpn) |
2553                                         (OCRDMA_CQE_WR_FLUSH_ERR <<
2554                                                 OCRDMA_CQE_STATUS_SHIFT));
2555                 }
2556         }
2557 }
2558
2559 static bool ocrdma_update_err_cqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2560                                   struct ocrdma_qp *qp, int status)
2561 {
2562         bool expand = false;
2563
2564         ibwc->byte_len = 0;
2565         ibwc->qp = &qp->ibqp;
2566         ibwc->status = ocrdma_to_ibwc_err(status);
2567
2568         ocrdma_flush_qp(qp);
2569         ocrdma_qp_state_change(qp, IB_QPS_ERR, NULL);
2570
2571         /* if wqe/rqe pending for which cqe needs to be returned,
2572          * trigger inflating it.
2573          */
2574         if (!is_hw_rq_empty(qp) || !is_hw_sq_empty(qp)) {
2575                 expand = true;
2576                 ocrdma_set_cqe_status_flushed(qp, cqe);
2577         }
2578         return expand;
2579 }
2580
2581 static int ocrdma_update_err_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2582                                   struct ocrdma_qp *qp, int status)
2583 {
2584         ibwc->opcode = IB_WC_RECV;
2585         ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2586         ocrdma_hwq_inc_tail(&qp->rq);
2587
2588         return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2589 }
2590
2591 static int ocrdma_update_err_scqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2592                                   struct ocrdma_qp *qp, int status)
2593 {
2594         ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2595         ocrdma_hwq_inc_tail(&qp->sq);
2596
2597         return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2598 }
2599
2600
2601 static bool ocrdma_poll_err_scqe(struct ocrdma_qp *qp,
2602                                  struct ocrdma_cqe *cqe, struct ib_wc *ibwc,
2603                                  bool *polled, bool *stop)
2604 {
2605         bool expand;
2606         struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2607         int status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2608                 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2609         if (status < OCRDMA_MAX_CQE_ERR)
2610                 atomic_inc(&dev->cqe_err_stats[status]);
2611
2612         /* when hw sq is empty, but rq is not empty, so we continue
2613          * to keep the cqe in order to get the cq event again.
2614          */
2615         if (is_hw_sq_empty(qp) && !is_hw_rq_empty(qp)) {
2616                 /* when cq for rq and sq is same, it is safe to return
2617                  * flush cqe for RQEs.
2618                  */
2619                 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2620                         *polled = true;
2621                         status = OCRDMA_CQE_WR_FLUSH_ERR;
2622                         expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2623                 } else {
2624                         /* stop processing further cqe as this cqe is used for
2625                          * triggering cq event on buddy cq of RQ.
2626                          * When QP is destroyed, this cqe will be removed
2627                          * from the cq's hardware q.
2628                          */
2629                         *polled = false;
2630                         *stop = true;
2631                         expand = false;
2632                 }
2633         } else if (is_hw_sq_empty(qp)) {
2634                 /* Do nothing */
2635                 expand = false;
2636                 *polled = false;
2637                 *stop = false;
2638         } else {
2639                 *polled = true;
2640                 expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2641         }
2642         return expand;
2643 }
2644
2645 static bool ocrdma_poll_success_scqe(struct ocrdma_qp *qp,
2646                                      struct ocrdma_cqe *cqe,
2647                                      struct ib_wc *ibwc, bool *polled)
2648 {
2649         bool expand = false;
2650         int tail = qp->sq.tail;
2651         u32 wqe_idx;
2652
2653         if (!qp->wqe_wr_id_tbl[tail].signaled) {
2654                 *polled = false;    /* WC cannot be consumed yet */
2655         } else {
2656                 ibwc->status = IB_WC_SUCCESS;
2657                 ibwc->wc_flags = 0;
2658                 ibwc->qp = &qp->ibqp;
2659                 ocrdma_update_wc(qp, ibwc, tail);
2660                 *polled = true;
2661         }
2662         wqe_idx = (le32_to_cpu(cqe->wq.wqeidx) &
2663                         OCRDMA_CQE_WQEIDX_MASK) & qp->sq.max_wqe_idx;
2664         if (tail != wqe_idx)
2665                 expand = true; /* Coalesced CQE can't be consumed yet */
2666
2667         ocrdma_hwq_inc_tail(&qp->sq);
2668         return expand;
2669 }
2670
2671 static bool ocrdma_poll_scqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2672                              struct ib_wc *ibwc, bool *polled, bool *stop)
2673 {
2674         int status;
2675         bool expand;
2676
2677         status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2678                 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2679
2680         if (status == OCRDMA_CQE_SUCCESS)
2681                 expand = ocrdma_poll_success_scqe(qp, cqe, ibwc, polled);
2682         else
2683                 expand = ocrdma_poll_err_scqe(qp, cqe, ibwc, polled, stop);
2684         return expand;
2685 }
2686
2687 static int ocrdma_update_ud_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe)
2688 {
2689         int status;
2690
2691         status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2692                 OCRDMA_CQE_UD_STATUS_MASK) >> OCRDMA_CQE_UD_STATUS_SHIFT;
2693         ibwc->src_qp = le32_to_cpu(cqe->flags_status_srcqpn) &
2694                                                 OCRDMA_CQE_SRCQP_MASK;
2695         ibwc->pkey_index = le32_to_cpu(cqe->ud.rxlen_pkey) &
2696                                                 OCRDMA_CQE_PKEY_MASK;
2697         ibwc->wc_flags = IB_WC_GRH;
2698         ibwc->byte_len = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2699                                         OCRDMA_CQE_UD_XFER_LEN_SHIFT);
2700         return status;
2701 }
2702
2703 static void ocrdma_update_free_srq_cqe(struct ib_wc *ibwc,
2704                                        struct ocrdma_cqe *cqe,
2705                                        struct ocrdma_qp *qp)
2706 {
2707         unsigned long flags;
2708         struct ocrdma_srq *srq;
2709         u32 wqe_idx;
2710
2711         srq = get_ocrdma_srq(qp->ibqp.srq);
2712         wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
2713                 OCRDMA_CQE_BUFTAG_SHIFT) & srq->rq.max_wqe_idx;
2714         if (wqe_idx < 1)
2715                 BUG();
2716
2717         ibwc->wr_id = srq->rqe_wr_id_tbl[wqe_idx];
2718         spin_lock_irqsave(&srq->q_lock, flags);
2719         ocrdma_srq_toggle_bit(srq, wqe_idx - 1);
2720         spin_unlock_irqrestore(&srq->q_lock, flags);
2721         ocrdma_hwq_inc_tail(&srq->rq);
2722 }
2723
2724 static bool ocrdma_poll_err_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2725                                 struct ib_wc *ibwc, bool *polled, bool *stop,
2726                                 int status)
2727 {
2728         bool expand;
2729         struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2730
2731         if (status < OCRDMA_MAX_CQE_ERR)
2732                 atomic_inc(&dev->cqe_err_stats[status]);
2733
2734         /* when hw_rq is empty, but wq is not empty, so continue
2735          * to keep the cqe to get the cq event again.
2736          */
2737         if (is_hw_rq_empty(qp) && !is_hw_sq_empty(qp)) {
2738                 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2739                         *polled = true;
2740                         status = OCRDMA_CQE_WR_FLUSH_ERR;
2741                         expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2742                 } else {
2743                         *polled = false;
2744                         *stop = true;
2745                         expand = false;
2746                 }
2747         } else if (is_hw_rq_empty(qp)) {
2748                 /* Do nothing */
2749                 expand = false;
2750                 *polled = false;
2751                 *stop = false;
2752         } else {
2753                 *polled = true;
2754                 expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2755         }
2756         return expand;
2757 }
2758
2759 static void ocrdma_poll_success_rcqe(struct ocrdma_qp *qp,
2760                                      struct ocrdma_cqe *cqe, struct ib_wc *ibwc)
2761 {
2762         ibwc->opcode = IB_WC_RECV;
2763         ibwc->qp = &qp->ibqp;
2764         ibwc->status = IB_WC_SUCCESS;
2765
2766         if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI)
2767                 ocrdma_update_ud_rcqe(ibwc, cqe);
2768         else
2769                 ibwc->byte_len = le32_to_cpu(cqe->rq.rxlen);
2770
2771         if (is_cqe_imm(cqe)) {
2772                 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2773                 ibwc->wc_flags |= IB_WC_WITH_IMM;
2774         } else if (is_cqe_wr_imm(cqe)) {
2775                 ibwc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2776                 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2777                 ibwc->wc_flags |= IB_WC_WITH_IMM;
2778         } else if (is_cqe_invalidated(cqe)) {
2779                 ibwc->ex.invalidate_rkey = le32_to_cpu(cqe->rq.lkey_immdt);
2780                 ibwc->wc_flags |= IB_WC_WITH_INVALIDATE;
2781         }
2782         if (qp->ibqp.srq) {
2783                 ocrdma_update_free_srq_cqe(ibwc, cqe, qp);
2784         } else {
2785                 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2786                 ocrdma_hwq_inc_tail(&qp->rq);
2787         }
2788 }
2789
2790 static bool ocrdma_poll_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2791                              struct ib_wc *ibwc, bool *polled, bool *stop)
2792 {
2793         int status;
2794         bool expand = false;
2795
2796         ibwc->wc_flags = 0;
2797         if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2798                 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2799                                         OCRDMA_CQE_UD_STATUS_MASK) >>
2800                                         OCRDMA_CQE_UD_STATUS_SHIFT;
2801         } else {
2802                 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2803                              OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2804         }
2805
2806         if (status == OCRDMA_CQE_SUCCESS) {
2807                 *polled = true;
2808                 ocrdma_poll_success_rcqe(qp, cqe, ibwc);
2809         } else {
2810                 expand = ocrdma_poll_err_rcqe(qp, cqe, ibwc, polled, stop,
2811                                               status);
2812         }
2813         return expand;
2814 }
2815
2816 static void ocrdma_change_cq_phase(struct ocrdma_cq *cq, struct ocrdma_cqe *cqe,
2817                                    u16 cur_getp)
2818 {
2819         if (cq->phase_change) {
2820                 if (cur_getp == 0)
2821                         cq->phase = (~cq->phase & OCRDMA_CQE_VALID);
2822         } else {
2823                 /* clear valid bit */
2824                 cqe->flags_status_srcqpn = 0;
2825         }
2826 }
2827
2828 static int ocrdma_poll_hwcq(struct ocrdma_cq *cq, int num_entries,
2829                             struct ib_wc *ibwc)
2830 {
2831         u16 qpn = 0;
2832         int i = 0;
2833         bool expand = false;
2834         int polled_hw_cqes = 0;
2835         struct ocrdma_qp *qp = NULL;
2836         struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
2837         struct ocrdma_cqe *cqe;
2838         u16 cur_getp; bool polled = false; bool stop = false;
2839
2840         cur_getp = cq->getp;
2841         while (num_entries) {
2842                 cqe = cq->va + cur_getp;
2843                 /* check whether valid cqe or not */
2844                 if (!is_cqe_valid(cq, cqe))
2845                         break;
2846                 qpn = (le32_to_cpu(cqe->cmn.qpn) & OCRDMA_CQE_QPN_MASK);
2847                 /* ignore discarded cqe */
2848                 if (qpn == 0)
2849                         goto skip_cqe;
2850                 qp = dev->qp_tbl[qpn];
2851                 BUG_ON(qp == NULL);
2852
2853                 if (is_cqe_for_sq(cqe)) {
2854                         expand = ocrdma_poll_scqe(qp, cqe, ibwc, &polled,
2855                                                   &stop);
2856                 } else {
2857                         expand = ocrdma_poll_rcqe(qp, cqe, ibwc, &polled,
2858                                                   &stop);
2859                 }
2860                 if (expand)
2861                         goto expand_cqe;
2862                 if (stop)
2863                         goto stop_cqe;
2864                 /* clear qpn to avoid duplicate processing by discard_cqe() */
2865                 cqe->cmn.qpn = 0;
2866 skip_cqe:
2867                 polled_hw_cqes += 1;
2868                 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
2869                 ocrdma_change_cq_phase(cq, cqe, cur_getp);
2870 expand_cqe:
2871                 if (polled) {
2872                         num_entries -= 1;
2873                         i += 1;
2874                         ibwc = ibwc + 1;
2875                         polled = false;
2876                 }
2877         }
2878 stop_cqe:
2879         cq->getp = cur_getp;
2880         if (cq->deferred_arm) {
2881                 ocrdma_ring_cq_db(dev, cq->id, true, cq->deferred_sol,
2882                                   polled_hw_cqes);
2883                 cq->deferred_arm = false;
2884                 cq->deferred_sol = false;
2885         } else {
2886                 /* We need to pop the CQE. No need to arm */
2887                 ocrdma_ring_cq_db(dev, cq->id, false, cq->deferred_sol,
2888                                   polled_hw_cqes);
2889                 cq->deferred_sol = false;
2890         }
2891
2892         return i;
2893 }
2894
2895 /* insert error cqe if the QP's SQ or RQ's CQ matches the CQ under poll. */
2896 static int ocrdma_add_err_cqe(struct ocrdma_cq *cq, int num_entries,
2897                               struct ocrdma_qp *qp, struct ib_wc *ibwc)
2898 {
2899         int err_cqes = 0;
2900
2901         while (num_entries) {
2902                 if (is_hw_sq_empty(qp) && is_hw_rq_empty(qp))
2903                         break;
2904                 if (!is_hw_sq_empty(qp) && qp->sq_cq == cq) {
2905                         ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2906                         ocrdma_hwq_inc_tail(&qp->sq);
2907                 } else if (!is_hw_rq_empty(qp) && qp->rq_cq == cq) {
2908                         ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2909                         ocrdma_hwq_inc_tail(&qp->rq);
2910                 } else {
2911                         return err_cqes;
2912                 }
2913                 ibwc->byte_len = 0;
2914                 ibwc->status = IB_WC_WR_FLUSH_ERR;
2915                 ibwc = ibwc + 1;
2916                 err_cqes += 1;
2917                 num_entries -= 1;
2918         }
2919         return err_cqes;
2920 }
2921
2922 int ocrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
2923 {
2924         int cqes_to_poll = num_entries;
2925         struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2926         struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2927         int num_os_cqe = 0, err_cqes = 0;
2928         struct ocrdma_qp *qp;
2929         unsigned long flags;
2930
2931         /* poll cqes from adapter CQ */
2932         spin_lock_irqsave(&cq->cq_lock, flags);
2933         num_os_cqe = ocrdma_poll_hwcq(cq, cqes_to_poll, wc);
2934         spin_unlock_irqrestore(&cq->cq_lock, flags);
2935         cqes_to_poll -= num_os_cqe;
2936
2937         if (cqes_to_poll) {
2938                 wc = wc + num_os_cqe;
2939                 /* adapter returns single error cqe when qp moves to
2940                  * error state. So insert error cqes with wc_status as
2941                  * FLUSHED for pending WQEs and RQEs of QP's SQ and RQ
2942                  * respectively which uses this CQ.
2943                  */
2944                 spin_lock_irqsave(&dev->flush_q_lock, flags);
2945                 list_for_each_entry(qp, &cq->sq_head, sq_entry) {
2946                         if (cqes_to_poll == 0)
2947                                 break;
2948                         err_cqes = ocrdma_add_err_cqe(cq, cqes_to_poll, qp, wc);
2949                         cqes_to_poll -= err_cqes;
2950                         num_os_cqe += err_cqes;
2951                         wc = wc + err_cqes;
2952                 }
2953                 spin_unlock_irqrestore(&dev->flush_q_lock, flags);
2954         }
2955         return num_os_cqe;
2956 }
2957
2958 int ocrdma_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags cq_flags)
2959 {
2960         struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2961         struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2962         u16 cq_id;
2963         unsigned long flags;
2964         bool arm_needed = false, sol_needed = false;
2965
2966         cq_id = cq->id;
2967
2968         spin_lock_irqsave(&cq->cq_lock, flags);
2969         if (cq_flags & IB_CQ_NEXT_COMP || cq_flags & IB_CQ_SOLICITED)
2970                 arm_needed = true;
2971         if (cq_flags & IB_CQ_SOLICITED)
2972                 sol_needed = true;
2973
2974         if (cq->first_arm) {
2975                 ocrdma_ring_cq_db(dev, cq_id, arm_needed, sol_needed, 0);
2976                 cq->first_arm = false;
2977         }
2978
2979         cq->deferred_arm = true;
2980         cq->deferred_sol = sol_needed;
2981         spin_unlock_irqrestore(&cq->cq_lock, flags);
2982
2983         return 0;
2984 }
2985
2986 struct ib_mr *ocrdma_alloc_frmr(struct ib_pd *ibpd, int max_page_list_len)
2987 {
2988         int status;
2989         struct ocrdma_mr *mr;
2990         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
2991         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
2992
2993         if (max_page_list_len > dev->attr.max_pages_per_frmr)
2994                 return ERR_PTR(-EINVAL);
2995
2996         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2997         if (!mr)
2998                 return ERR_PTR(-ENOMEM);
2999
3000         status = ocrdma_get_pbl_info(dev, mr, max_page_list_len);
3001         if (status)
3002                 goto pbl_err;
3003         mr->hwmr.fr_mr = 1;
3004         mr->hwmr.remote_rd = 0;
3005         mr->hwmr.remote_wr = 0;
3006         mr->hwmr.local_rd = 0;
3007         mr->hwmr.local_wr = 0;
3008         mr->hwmr.mw_bind = 0;
3009         status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
3010         if (status)
3011                 goto pbl_err;
3012         status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, 0);
3013         if (status)
3014                 goto mbx_err;
3015         mr->ibmr.rkey = mr->hwmr.lkey;
3016         mr->ibmr.lkey = mr->hwmr.lkey;
3017         dev->stag_arr[(mr->hwmr.lkey >> 8) & (OCRDMA_MAX_STAG - 1)] =
3018                 (unsigned long) mr;
3019         return &mr->ibmr;
3020 mbx_err:
3021         ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
3022 pbl_err:
3023         kfree(mr);
3024         return ERR_PTR(-ENOMEM);
3025 }
3026
3027 struct ib_fast_reg_page_list *ocrdma_alloc_frmr_page_list(struct ib_device
3028                                                           *ibdev,
3029                                                           int page_list_len)
3030 {
3031         struct ib_fast_reg_page_list *frmr_list;
3032         int size;
3033
3034         size = sizeof(*frmr_list) + (page_list_len * sizeof(u64));
3035         frmr_list = kzalloc(size, GFP_KERNEL);
3036         if (!frmr_list)
3037                 return ERR_PTR(-ENOMEM);
3038         frmr_list->page_list = (u64 *)(frmr_list + 1);
3039         return frmr_list;
3040 }
3041
3042 void ocrdma_free_frmr_page_list(struct ib_fast_reg_page_list *page_list)
3043 {
3044         kfree(page_list);
3045 }
3046
3047 #define MAX_KERNEL_PBE_SIZE 65536
3048 static inline int count_kernel_pbes(struct ib_phys_buf *buf_list,
3049                                     int buf_cnt, u32 *pbe_size)
3050 {
3051         u64 total_size = 0;
3052         u64 buf_size = 0;
3053         int i;
3054         *pbe_size = roundup(buf_list[0].size, PAGE_SIZE);
3055         *pbe_size = roundup_pow_of_two(*pbe_size);
3056
3057         /* find the smallest PBE size that we can have */
3058         for (i = 0; i < buf_cnt; i++) {
3059                 /* first addr may not be page aligned, so ignore checking */
3060                 if ((i != 0) && ((buf_list[i].addr & ~PAGE_MASK) ||
3061                                  (buf_list[i].size & ~PAGE_MASK))) {
3062                         return 0;
3063                 }
3064
3065                 /* if configured PBE size is greater then the chosen one,
3066                  * reduce the PBE size.
3067                  */
3068                 buf_size = roundup(buf_list[i].size, PAGE_SIZE);
3069                 /* pbe_size has to be even multiple of 4K 1,2,4,8...*/
3070                 buf_size = roundup_pow_of_two(buf_size);
3071                 if (*pbe_size > buf_size)
3072                         *pbe_size = buf_size;
3073
3074                 total_size += buf_size;
3075         }
3076         *pbe_size = *pbe_size > MAX_KERNEL_PBE_SIZE ?
3077             (MAX_KERNEL_PBE_SIZE) : (*pbe_size);
3078
3079         /* num_pbes = total_size / (*pbe_size);  this is implemented below. */
3080
3081         return total_size >> ilog2(*pbe_size);
3082 }
3083
3084 static void build_kernel_pbes(struct ib_phys_buf *buf_list, int ib_buf_cnt,
3085                               u32 pbe_size, struct ocrdma_pbl *pbl_tbl,
3086                               struct ocrdma_hw_mr *hwmr)
3087 {
3088         int i;
3089         int idx;
3090         int pbes_per_buf = 0;
3091         u64 buf_addr = 0;
3092         int num_pbes;
3093         struct ocrdma_pbe *pbe;
3094         int total_num_pbes = 0;
3095
3096         if (!hwmr->num_pbes)
3097                 return;
3098
3099         pbe = (struct ocrdma_pbe *)pbl_tbl->va;
3100         num_pbes = 0;
3101
3102         /* go through the OS phy regions & fill hw pbe entries into pbls. */
3103         for (i = 0; i < ib_buf_cnt; i++) {
3104                 buf_addr = buf_list[i].addr;
3105                 pbes_per_buf =
3106                     roundup_pow_of_two(roundup(buf_list[i].size, PAGE_SIZE)) /
3107                     pbe_size;
3108                 hwmr->len += buf_list[i].size;
3109                 /* number of pbes can be more for one OS buf, when
3110                  * buffers are of different sizes.
3111                  * split the ib_buf to one or more pbes.
3112                  */
3113                 for (idx = 0; idx < pbes_per_buf; idx++) {
3114                         /* we program always page aligned addresses,
3115                          * first unaligned address is taken care by fbo.
3116                          */
3117                         if (i == 0) {
3118                                 /* for non zero fbo, assign the
3119                                  * start of the page.
3120                                  */
3121                                 pbe->pa_lo =
3122                                     cpu_to_le32((u32) (buf_addr & PAGE_MASK));
3123                                 pbe->pa_hi =
3124                                     cpu_to_le32((u32) upper_32_bits(buf_addr));
3125                         } else {
3126                                 pbe->pa_lo =
3127                                     cpu_to_le32((u32) (buf_addr & 0xffffffff));
3128                                 pbe->pa_hi =
3129                                     cpu_to_le32((u32) upper_32_bits(buf_addr));
3130                         }
3131                         buf_addr += pbe_size;
3132                         num_pbes += 1;
3133                         total_num_pbes += 1;
3134                         pbe++;
3135
3136                         if (total_num_pbes == hwmr->num_pbes)
3137                                 goto mr_tbl_done;
3138                         /* if the pbl is full storing the pbes,
3139                          * move to next pbl.
3140                          */
3141                         if (num_pbes == (hwmr->pbl_size/sizeof(u64))) {
3142                                 pbl_tbl++;
3143                                 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
3144                                 num_pbes = 0;
3145                         }
3146                 }
3147         }
3148 mr_tbl_done:
3149         return;
3150 }
3151
3152 struct ib_mr *ocrdma_reg_kernel_mr(struct ib_pd *ibpd,
3153                                    struct ib_phys_buf *buf_list,
3154                                    int buf_cnt, int acc, u64 *iova_start)
3155 {
3156         int status = -ENOMEM;
3157         struct ocrdma_mr *mr;
3158         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
3159         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
3160         u32 num_pbes;
3161         u32 pbe_size = 0;
3162
3163         if ((acc & IB_ACCESS_REMOTE_WRITE) && !(acc & IB_ACCESS_LOCAL_WRITE))
3164                 return ERR_PTR(-EINVAL);
3165
3166         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3167         if (!mr)
3168                 return ERR_PTR(status);
3169
3170         num_pbes = count_kernel_pbes(buf_list, buf_cnt, &pbe_size);
3171         if (num_pbes == 0) {
3172                 status = -EINVAL;
3173                 goto pbl_err;
3174         }
3175         status = ocrdma_get_pbl_info(dev, mr, num_pbes);
3176         if (status)
3177                 goto pbl_err;
3178
3179         mr->hwmr.pbe_size = pbe_size;
3180         mr->hwmr.fbo = *iova_start - (buf_list[0].addr & PAGE_MASK);
3181         mr->hwmr.va = *iova_start;
3182         mr->hwmr.local_rd = 1;
3183         mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
3184         mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
3185         mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
3186         mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
3187         mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
3188
3189         status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
3190         if (status)
3191                 goto pbl_err;
3192         build_kernel_pbes(buf_list, buf_cnt, pbe_size, mr->hwmr.pbl_table,
3193                           &mr->hwmr);
3194         status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
3195         if (status)
3196                 goto mbx_err;
3197
3198         mr->ibmr.lkey = mr->hwmr.lkey;
3199         if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
3200                 mr->ibmr.rkey = mr->hwmr.lkey;
3201         return &mr->ibmr;
3202
3203 mbx_err:
3204         ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
3205 pbl_err:
3206         kfree(mr);
3207         return ERR_PTR(status);
3208 }