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