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 = 0;
  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
 148 int ocrdma_query_port(struct ib_device *ibdev,
 149                       u8 port, struct ib_port_attr *props)
 150 {
 151         enum ib_port_state port_state;
 152         struct ocrdma_dev *dev;
 153         struct net_device *netdev;
 154
 155         dev = get_ocrdma_dev(ibdev);
 156         if (port > 1) {
 157                 pr_err("%s(%d) invalid_port=0x%x\n", __func__,
 158                        dev->id, port);
 159                 return -EINVAL;
 160         }
 161         netdev = dev->nic_info.netdev;
 162         if (netif_running(netdev) && netif_oper_up(netdev)) {
 163                 port_state = IB_PORT_ACTIVE;
 164                 props->phys_state = 5;
 165         } else {
 166                 port_state = IB_PORT_DOWN;
 167                 props->phys_state = 3;
 168         }
 169         props->max_mtu = IB_MTU_4096;
 170         props->active_mtu = iboe_get_mtu(netdev->mtu);
 171         props->lid = 0;
 172         props->lmc = 0;
 173         props->sm_lid = 0;
 174         props->sm_sl = 0;
 175         props->state = port_state;
 176         props->port_cap_flags =
 177             IB_PORT_CM_SUP |
 178             IB_PORT_REINIT_SUP |
 179             IB_PORT_DEVICE_MGMT_SUP | IB_PORT_VENDOR_CLASS_SUP;
 180         props->gid_tbl_len = OCRDMA_MAX_SGID;
 181         props->pkey_tbl_len = 1;
 182         props->bad_pkey_cntr = 0;
 183         props->qkey_viol_cntr = 0;
 184         get_link_speed_and_width(dev, &props->active_speed,
 185                                  &props->active_width);
 186         props->max_msg_sz = 0x80000000;
 187         props->max_vl_num = 4;
 188         return 0;
 189 }
 190
 191 int ocrdma_modify_port(struct ib_device *ibdev, u8 port, int mask,
 192                        struct ib_port_modify *props)
 193 {
 194         struct ocrdma_dev *dev;
 195
 196         dev = get_ocrdma_dev(ibdev);
 197         if (port > 1) {
 198                 pr_err("%s(%d) invalid_port=0x%x\n", __func__, dev->id, port);
 199                 return -EINVAL;
 200         }
 201         return 0;
 202 }
 203
 204 static int ocrdma_add_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
 205                            unsigned long len)
 206 {
 207         struct ocrdma_mm *mm;
 208
 209         mm = kzalloc(sizeof(*mm), GFP_KERNEL);
 210         if (mm == NULL)
 211                 return -ENOMEM;
 212         mm->key.phy_addr = phy_addr;
 213         mm->key.len = len;
 214         INIT_LIST_HEAD(&mm->entry);
 215
 216         mutex_lock(&uctx->mm_list_lock);
 217         list_add_tail(&mm->entry, &uctx->mm_head);
 218         mutex_unlock(&uctx->mm_list_lock);
 219         return 0;
 220 }
 221
 222 static void ocrdma_del_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
 223                             unsigned long len)
 224 {
 225         struct ocrdma_mm *mm, *tmp;
 226
 227         mutex_lock(&uctx->mm_list_lock);
 228         list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
 229                 if (len != mm->key.len && phy_addr != mm->key.phy_addr)
 230                         continue;
 231
 232                 list_del(&mm->entry);
 233                 kfree(mm);
 234                 break;
 235         }
 236         mutex_unlock(&uctx->mm_list_lock);
 237 }
 238
 239 static bool ocrdma_search_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
 240                               unsigned long len)
 241 {
 242         bool found = false;
 243         struct ocrdma_mm *mm;
 244
 245         mutex_lock(&uctx->mm_list_lock);
 246         list_for_each_entry(mm, &uctx->mm_head, entry) {
 247                 if (len != mm->key.len && phy_addr != mm->key.phy_addr)
 248                         continue;
 249
 250                 found = true;
 251                 break;
 252         }
 253         mutex_unlock(&uctx->mm_list_lock);
 254         return found;
 255 }
 256
 257 static struct ocrdma_pd *_ocrdma_alloc_pd(struct ocrdma_dev *dev,
 258                                           struct ocrdma_ucontext *uctx,
 259                                           struct ib_udata *udata)
 260 {
 261         struct ocrdma_pd *pd = NULL;
 262         int status = 0;
 263
 264         pd = kzalloc(sizeof(*pd), GFP_KERNEL);
 265         if (!pd)
 266                 return ERR_PTR(-ENOMEM);
 267
 268         if (udata && uctx) {
 269                 pd->dpp_enabled =
 270                         dev->nic_info.dev_family == OCRDMA_GEN2_FAMILY;
 271                 pd->num_dpp_qp =
 272                         pd->dpp_enabled ? OCRDMA_PD_MAX_DPP_ENABLED_QP : 0;
 273         }
 274
 275 retry:
 276         status = ocrdma_mbx_alloc_pd(dev, pd);
 277         if (status) {
 278                 if (pd->dpp_enabled) {
 279                         pd->dpp_enabled = false;
 280                         pd->num_dpp_qp = 0;
 281                         goto retry;
 282                 } else {
 283                         kfree(pd);
 284                         return ERR_PTR(status);
 285                 }
 286         }
 287
 288         return pd;
 289 }
 290
 291 static inline int is_ucontext_pd(struct ocrdma_ucontext *uctx,
 292                                  struct ocrdma_pd *pd)
 293 {
 294         return (uctx->cntxt_pd == pd ? true : false);
 295 }
 296
 297 static int _ocrdma_dealloc_pd(struct ocrdma_dev *dev,
 298                               struct ocrdma_pd *pd)
 299 {
 300         int status = 0;
 301
 302         status = ocrdma_mbx_dealloc_pd(dev, pd);
 303         kfree(pd);
 304         return status;
 305 }
 306
 307 static int ocrdma_alloc_ucontext_pd(struct ocrdma_dev *dev,
 308                                     struct ocrdma_ucontext *uctx,
 309                                     struct ib_udata *udata)
 310 {
 311         int status = 0;
 312
 313         uctx->cntxt_pd = _ocrdma_alloc_pd(dev, uctx, udata);
 314         if (IS_ERR(uctx->cntxt_pd)) {
 315                 status = PTR_ERR(uctx->cntxt_pd);
 316                 uctx->cntxt_pd = NULL;
 317                 goto err;
 318         }
 319
 320         uctx->cntxt_pd->uctx = uctx;
 321         uctx->cntxt_pd->ibpd.device = &dev->ibdev;
 322 err:
 323         return status;
 324 }
 325
 326 static int ocrdma_dealloc_ucontext_pd(struct ocrdma_ucontext *uctx)
 327 {
 328         int status = 0;
 329         struct ocrdma_pd *pd = uctx->cntxt_pd;
 330         struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
 331
 332         BUG_ON(uctx->pd_in_use);
 333         uctx->cntxt_pd = NULL;
 334         status = _ocrdma_dealloc_pd(dev, pd);
 335         return status;
 336 }
 337
 338 static struct ocrdma_pd *ocrdma_get_ucontext_pd(struct ocrdma_ucontext *uctx)
 339 {
 340         struct ocrdma_pd *pd = NULL;
 341
 342         mutex_lock(&uctx->mm_list_lock);
 343         if (!uctx->pd_in_use) {
 344                 uctx->pd_in_use = true;
 345                 pd = uctx->cntxt_pd;
 346         }
 347         mutex_unlock(&uctx->mm_list_lock);
 348
 349         return pd;
 350 }
 351
 352 static void ocrdma_release_ucontext_pd(struct ocrdma_ucontext *uctx)
 353 {
 354         mutex_lock(&uctx->mm_list_lock);
 355         uctx->pd_in_use = false;
 356         mutex_unlock(&uctx->mm_list_lock);
 357 }
 358
 359 struct ib_ucontext *ocrdma_alloc_ucontext(struct ib_device *ibdev,
 360                                           struct ib_udata *udata)
 361 {
 362         int status;
 363         struct ocrdma_ucontext *ctx;
 364         struct ocrdma_alloc_ucontext_resp resp;
 365         struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
 366         struct pci_dev *pdev = dev->nic_info.pdev;
 367         u32 map_len = roundup(sizeof(u32) * 2048, PAGE_SIZE);
 368
 369         if (!udata)
 370                 return ERR_PTR(-EFAULT);
 371         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
 372         if (!ctx)
 373                 return ERR_PTR(-ENOMEM);
 374         INIT_LIST_HEAD(&ctx->mm_head);
 375         mutex_init(&ctx->mm_list_lock);
 376
 377         ctx->ah_tbl.va = dma_alloc_coherent(&pdev->dev, map_len,
 378                                             &ctx->ah_tbl.pa, GFP_KERNEL);
 379         if (!ctx->ah_tbl.va) {
 380                 kfree(ctx);
 381                 return ERR_PTR(-ENOMEM);
 382         }
 383         memset(ctx->ah_tbl.va, 0, map_len);
 384         ctx->ah_tbl.len = map_len;
 385
 386         memset(&resp, 0, sizeof(resp));
 387         resp.ah_tbl_len = ctx->ah_tbl.len;
 388         resp.ah_tbl_page = ctx->ah_tbl.pa;
 389
 390         status = ocrdma_add_mmap(ctx, resp.ah_tbl_page, resp.ah_tbl_len);
 391         if (status)
 392                 goto map_err;
 393
 394         status = ocrdma_alloc_ucontext_pd(dev, ctx, udata);
 395         if (status)
 396                 goto pd_err;
 397
 398         resp.dev_id = dev->id;
 399         resp.max_inline_data = dev->attr.max_inline_data;
 400         resp.wqe_size = dev->attr.wqe_size;
 401         resp.rqe_size = dev->attr.rqe_size;
 402         resp.dpp_wqe_size = dev->attr.wqe_size;
 403
 404         memcpy(resp.fw_ver, dev->attr.fw_ver, sizeof(resp.fw_ver));
 405         status = ib_copy_to_udata(udata, &resp, sizeof(resp));
 406         if (status)
 407                 goto cpy_err;
 408         return &ctx->ibucontext;
 409
 410 cpy_err:
 411 pd_err:
 412         ocrdma_del_mmap(ctx, ctx->ah_tbl.pa, ctx->ah_tbl.len);
 413 map_err:
 414         dma_free_coherent(&pdev->dev, ctx->ah_tbl.len, ctx->ah_tbl.va,
 415                           ctx->ah_tbl.pa);
 416         kfree(ctx);
 417         return ERR_PTR(status);
 418 }
 419
 420 int ocrdma_dealloc_ucontext(struct ib_ucontext *ibctx)
 421 {
 422         int status = 0;
 423         struct ocrdma_mm *mm, *tmp;
 424         struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ibctx);
 425         struct ocrdma_dev *dev = get_ocrdma_dev(ibctx->device);
 426         struct pci_dev *pdev = dev->nic_info.pdev;
 427
 428         status = ocrdma_dealloc_ucontext_pd(uctx);
 429
 430         ocrdma_del_mmap(uctx, uctx->ah_tbl.pa, uctx->ah_tbl.len);
 431         dma_free_coherent(&pdev->dev, uctx->ah_tbl.len, uctx->ah_tbl.va,
 432                           uctx->ah_tbl.pa);
 433
 434         list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
 435                 list_del(&mm->entry);
 436                 kfree(mm);
 437         }
 438         kfree(uctx);
 439         return status;
 440 }
 441
 442 int ocrdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
 443 {
 444         struct ocrdma_ucontext *ucontext = get_ocrdma_ucontext(context);
 445         struct ocrdma_dev *dev = get_ocrdma_dev(context->device);
 446         unsigned long vm_page = vma->vm_pgoff << PAGE_SHIFT;
 447         u64 unmapped_db = (u64) dev->nic_info.unmapped_db;
 448         unsigned long len = (vma->vm_end - vma->vm_start);
 449         int status = 0;
 450         bool found;
 451
 452         if (vma->vm_start & (PAGE_SIZE - 1))
 453                 return -EINVAL;
 454         found = ocrdma_search_mmap(ucontext, vma->vm_pgoff << PAGE_SHIFT, len);
 455         if (!found)
 456                 return -EINVAL;
 457
 458         if ((vm_page >= unmapped_db) && (vm_page <= (unmapped_db +
 459                 dev->nic_info.db_total_size)) &&
 460                 (len <= dev->nic_info.db_page_size)) {
 461                 if (vma->vm_flags & VM_READ)
 462                         return -EPERM;
 463
 464                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 465                 status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
 466                                             len, vma->vm_page_prot);
 467         } else if (dev->nic_info.dpp_unmapped_len &&
 468                 (vm_page >= (u64) dev->nic_info.dpp_unmapped_addr) &&
 469                 (vm_page <= (u64) (dev->nic_info.dpp_unmapped_addr +
 470                         dev->nic_info.dpp_unmapped_len)) &&
 471                 (len <= dev->nic_info.dpp_unmapped_len)) {
 472                 if (vma->vm_flags & VM_READ)
 473                         return -EPERM;
 474
 475                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 476                 status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
 477                                             len, vma->vm_page_prot);
 478         } else {
 479                 status = remap_pfn_range(vma, vma->vm_start,
 480                                          vma->vm_pgoff, len, vma->vm_page_prot);
 481         }
 482         return status;
 483 }
 484
 485 static int ocrdma_copy_pd_uresp(struct ocrdma_dev *dev, struct ocrdma_pd *pd,
 486                                 struct ib_ucontext *ib_ctx,
 487                                 struct ib_udata *udata)
 488 {
 489         int status;
 490         u64 db_page_addr;
 491         u64 dpp_page_addr = 0;
 492         u32 db_page_size;
 493         struct ocrdma_alloc_pd_uresp rsp;
 494         struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
 495
 496         memset(&rsp, 0, sizeof(rsp));
 497         rsp.id = pd->id;
 498         rsp.dpp_enabled = pd->dpp_enabled;
 499         db_page_addr = ocrdma_get_db_addr(dev, pd->id);
 500         db_page_size = dev->nic_info.db_page_size;
 501
 502         status = ocrdma_add_mmap(uctx, db_page_addr, db_page_size);
 503         if (status)
 504                 return status;
 505
 506         if (pd->dpp_enabled) {
 507                 dpp_page_addr = dev->nic_info.dpp_unmapped_addr +
 508                                 (pd->id * PAGE_SIZE);
 509                 status = ocrdma_add_mmap(uctx, dpp_page_addr,
 510                                  PAGE_SIZE);
 511                 if (status)
 512                         goto dpp_map_err;
 513                 rsp.dpp_page_addr_hi = upper_32_bits(dpp_page_addr);
 514                 rsp.dpp_page_addr_lo = dpp_page_addr;
 515         }
 516
 517         status = ib_copy_to_udata(udata, &rsp, sizeof(rsp));
 518         if (status)
 519                 goto ucopy_err;
 520
 521         pd->uctx = uctx;
 522         return 0;
 523
 524 ucopy_err:
 525         if (pd->dpp_enabled)
 526                 ocrdma_del_mmap(pd->uctx, dpp_page_addr, PAGE_SIZE);
 527 dpp_map_err:
 528         ocrdma_del_mmap(pd->uctx, db_page_addr, db_page_size);
 529         return status;
 530 }
 531
 532 struct ib_pd *ocrdma_alloc_pd(struct ib_device *ibdev,
 533                               struct ib_ucontext *context,
 534                               struct ib_udata *udata)
 535 {
 536         struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
 537         struct ocrdma_pd *pd;
 538         struct ocrdma_ucontext *uctx = NULL;
 539         int status;
 540         u8 is_uctx_pd = false;
 541
 542         if (udata && context) {
 543                 uctx = get_ocrdma_ucontext(context);
 544                 pd = ocrdma_get_ucontext_pd(uctx);
 545                 if (pd) {
 546                         is_uctx_pd = true;
 547                         goto pd_mapping;
 548                 }
 549         }
 550
 551         pd = _ocrdma_alloc_pd(dev, uctx, udata);
 552         if (IS_ERR(pd)) {
 553                 status = PTR_ERR(pd);
 554                 goto exit;
 555         }
 556
 557 pd_mapping:
 558         if (udata && context) {
 559                 status = ocrdma_copy_pd_uresp(dev, pd, context, udata);
 560                 if (status)
 561                         goto err;
 562         }
 563         return &pd->ibpd;
 564
 565 err:
 566         if (is_uctx_pd) {
 567                 ocrdma_release_ucontext_pd(uctx);
 568         } else {
 569                 status = ocrdma_mbx_dealloc_pd(dev, pd);
 570                 kfree(pd);
 571         }
 572 exit:
 573         return ERR_PTR(status);
 574 }
 575
 576 int ocrdma_dealloc_pd(struct ib_pd *ibpd)
 577 {
 578         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
 579         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
 580         struct ocrdma_ucontext *uctx = NULL;
 581         int status = 0;
 582         u64 usr_db;
 583
 584         uctx = pd->uctx;
 585         if (uctx) {
 586                 u64 dpp_db = dev->nic_info.dpp_unmapped_addr +
 587                         (pd->id * PAGE_SIZE);
 588                 if (pd->dpp_enabled)
 589                         ocrdma_del_mmap(pd->uctx, dpp_db, PAGE_SIZE);
 590                 usr_db = ocrdma_get_db_addr(dev, pd->id);
 591                 ocrdma_del_mmap(pd->uctx, usr_db, dev->nic_info.db_page_size);
 592
 593                 if (is_ucontext_pd(uctx, pd)) {
 594                         ocrdma_release_ucontext_pd(uctx);
 595                         return status;
 596                 }
 597         }
 598         status = _ocrdma_dealloc_pd(dev, pd);
 599         return status;
 600 }
 601
 602 static int ocrdma_alloc_lkey(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
 603                             u32 pdid, int acc, u32 num_pbls, u32 addr_check)
 604 {
 605         int status;
 606
 607         mr->hwmr.fr_mr = 0;
 608         mr->hwmr.local_rd = 1;
 609         mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
 610         mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
 611         mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
 612         mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
 613         mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
 614         mr->hwmr.num_pbls = num_pbls;
 615
 616         status = ocrdma_mbx_alloc_lkey(dev, &mr->hwmr, pdid, addr_check);
 617         if (status)
 618                 return status;
 619
 620         mr->ibmr.lkey = mr->hwmr.lkey;
 621         if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
 622                 mr->ibmr.rkey = mr->hwmr.lkey;
 623         return 0;
 624 }
 625
 626 struct ib_mr *ocrdma_get_dma_mr(struct ib_pd *ibpd, int acc)
 627 {
 628         int status;
 629         struct ocrdma_mr *mr;
 630         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
 631         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
 632
 633         if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE)) {
 634                 pr_err("%s err, invalid access rights\n", __func__);
 635                 return ERR_PTR(-EINVAL);
 636         }
 637
 638         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
 639         if (!mr)
 640                 return ERR_PTR(-ENOMEM);
 641
 642         status = ocrdma_alloc_lkey(dev, mr, pd->id, acc, 0,
 643                                    OCRDMA_ADDR_CHECK_DISABLE);
 644         if (status) {
 645                 kfree(mr);
 646                 return ERR_PTR(status);
 647         }
 648
 649         return &mr->ibmr;
 650 }
 651
 652 static void ocrdma_free_mr_pbl_tbl(struct ocrdma_dev *dev,
 653                                    struct ocrdma_hw_mr *mr)
 654 {
 655         struct pci_dev *pdev = dev->nic_info.pdev;
 656         int i = 0;
 657
 658         if (mr->pbl_table) {
 659                 for (i = 0; i < mr->num_pbls; i++) {
 660                         if (!mr->pbl_table[i].va)
 661                                 continue;
 662                         dma_free_coherent(&pdev->dev, mr->pbl_size,
 663                                           mr->pbl_table[i].va,
 664                                           mr->pbl_table[i].pa);
 665                 }
 666                 kfree(mr->pbl_table);
 667                 mr->pbl_table = NULL;
 668         }
 669 }
 670
 671 static int ocrdma_get_pbl_info(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
 672                               u32 num_pbes)
 673 {
 674         u32 num_pbls = 0;
 675         u32 idx = 0;
 676         int status = 0;
 677         u32 pbl_size;
 678
 679         do {
 680                 pbl_size = OCRDMA_MIN_HPAGE_SIZE * (1 << idx);
 681                 if (pbl_size > MAX_OCRDMA_PBL_SIZE) {
 682                         status = -EFAULT;
 683                         break;
 684                 }
 685                 num_pbls = roundup(num_pbes, (pbl_size / sizeof(u64)));
 686                 num_pbls = num_pbls / (pbl_size / sizeof(u64));
 687                 idx++;
 688         } while (num_pbls >= dev->attr.max_num_mr_pbl);
 689
 690         mr->hwmr.num_pbes = num_pbes;
 691         mr->hwmr.num_pbls = num_pbls;
 692         mr->hwmr.pbl_size = pbl_size;
 693         return status;
 694 }
 695
 696 static int ocrdma_build_pbl_tbl(struct ocrdma_dev *dev, struct ocrdma_hw_mr *mr)
 697 {
 698         int status = 0;
 699         int i;
 700         u32 dma_len = mr->pbl_size;
 701         struct pci_dev *pdev = dev->nic_info.pdev;
 702         void *va;
 703         dma_addr_t pa;
 704
 705         mr->pbl_table = kzalloc(sizeof(struct ocrdma_pbl) *
 706                                 mr->num_pbls, GFP_KERNEL);
 707
 708         if (!mr->pbl_table)
 709                 return -ENOMEM;
 710
 711         for (i = 0; i < mr->num_pbls; i++) {
 712                 va = dma_alloc_coherent(&pdev->dev, dma_len, &pa, GFP_KERNEL);
 713                 if (!va) {
 714                         ocrdma_free_mr_pbl_tbl(dev, mr);
 715                         status = -ENOMEM;
 716                         break;
 717                 }
 718                 memset(va, 0, dma_len);
 719                 mr->pbl_table[i].va = va;
 720                 mr->pbl_table[i].pa = pa;
 721         }
 722         return status;
 723 }
 724
 725 static void build_user_pbes(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
 726                             u32 num_pbes)
 727 {
 728         struct ocrdma_pbe *pbe;
 729         struct ib_umem_chunk *chunk;
 730         struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
 731         struct ib_umem *umem = mr->umem;
 732         int i, shift, pg_cnt, pages, pbe_cnt, total_num_pbes = 0;
 733
 734         if (!mr->hwmr.num_pbes)
 735                 return;
 736
 737         pbe = (struct ocrdma_pbe *)pbl_tbl->va;
 738         pbe_cnt = 0;
 739
 740         shift = ilog2(umem->page_size);
 741
 742         list_for_each_entry(chunk, &umem->chunk_list, list) {
 743                 /* get all the dma regions from the chunk. */
 744                 for (i = 0; i < chunk->nmap; i++) {
 745                         pages = sg_dma_len(&chunk->page_list[i]) >> shift;
 746                         for (pg_cnt = 0; pg_cnt < pages; pg_cnt++) {
 747                                 /* store the page address in pbe */
 748                                 pbe->pa_lo =
 749                                     cpu_to_le32(sg_dma_address
 750                                                 (&chunk->page_list[i]) +
 751                                                 (umem->page_size * pg_cnt));
 752                                 pbe->pa_hi =
 753                                     cpu_to_le32(upper_32_bits
 754                                                 ((sg_dma_address
 755                                                   (&chunk->page_list[i]) +
 756                                                   umem->page_size * pg_cnt)));
 757                                 pbe_cnt += 1;
 758                                 total_num_pbes += 1;
 759                                 pbe++;
 760
 761                                 /* if done building pbes, issue the mbx cmd. */
 762                                 if (total_num_pbes == num_pbes)
 763                                         return;
 764
 765                                 /* if the given pbl is full storing the pbes,
 766                                  * move to next pbl.
 767                                  */
 768                                 if (pbe_cnt ==
 769                                         (mr->hwmr.pbl_size / sizeof(u64))) {
 770                                         pbl_tbl++;
 771                                         pbe = (struct ocrdma_pbe *)pbl_tbl->va;
 772                                         pbe_cnt = 0;
 773                                 }
 774                         }
 775                 }
 776         }
 777 }
 778
 779 struct ib_mr *ocrdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
 780                                  u64 usr_addr, int acc, struct ib_udata *udata)
 781 {
 782         int status = -ENOMEM;
 783         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
 784         struct ocrdma_mr *mr;
 785         struct ocrdma_pd *pd;
 786         u32 num_pbes;
 787
 788         pd = get_ocrdma_pd(ibpd);
 789
 790         if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
 791                 return ERR_PTR(-EINVAL);
 792
 793         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
 794         if (!mr)
 795                 return ERR_PTR(status);
 796         mr->umem = ib_umem_get(ibpd->uobject->context, start, len, acc, 0);
 797         if (IS_ERR(mr->umem)) {
 798                 status = -EFAULT;
 799                 goto umem_err;
 800         }
 801         num_pbes = ib_umem_page_count(mr->umem);
 802         status = ocrdma_get_pbl_info(dev, mr, num_pbes);
 803         if (status)
 804                 goto umem_err;
 805
 806         mr->hwmr.pbe_size = mr->umem->page_size;
 807         mr->hwmr.fbo = mr->umem->offset;
 808         mr->hwmr.va = usr_addr;
 809         mr->hwmr.len = len;
 810         mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
 811         mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
 812         mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
 813         mr->hwmr.local_rd = 1;
 814         mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
 815         status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
 816         if (status)
 817                 goto umem_err;
 818         build_user_pbes(dev, mr, num_pbes);
 819         status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
 820         if (status)
 821                 goto mbx_err;
 822         mr->ibmr.lkey = mr->hwmr.lkey;
 823         if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
 824                 mr->ibmr.rkey = mr->hwmr.lkey;
 825
 826         return &mr->ibmr;
 827
 828 mbx_err:
 829         ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
 830 umem_err:
 831         kfree(mr);
 832         return ERR_PTR(status);
 833 }
 834
 835 int ocrdma_dereg_mr(struct ib_mr *ib_mr)
 836 {
 837         struct ocrdma_mr *mr = get_ocrdma_mr(ib_mr);
 838         struct ocrdma_dev *dev = get_ocrdma_dev(ib_mr->device);
 839         int status;
 840
 841         status = ocrdma_mbx_dealloc_lkey(dev, mr->hwmr.fr_mr, mr->hwmr.lkey);
 842
 843         if (mr->hwmr.fr_mr == 0)
 844                 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
 845
 846         /* it could be user registered memory. */
 847         if (mr->umem)
 848                 ib_umem_release(mr->umem);
 849         kfree(mr);
 850         return status;
 851 }
 852
 853 static int ocrdma_copy_cq_uresp(struct ocrdma_dev *dev, struct ocrdma_cq *cq,
 854                                 struct ib_udata *udata,
 855                                 struct ib_ucontext *ib_ctx)
 856 {
 857         int status;
 858         struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
 859         struct ocrdma_create_cq_uresp uresp;
 860
 861         memset(&uresp, 0, sizeof(uresp));
 862         uresp.cq_id = cq->id;
 863         uresp.page_size = PAGE_ALIGN(cq->len);
 864         uresp.num_pages = 1;
 865         uresp.max_hw_cqe = cq->max_hw_cqe;
 866         uresp.page_addr[0] = cq->pa;
 867         uresp.db_page_addr =  ocrdma_get_db_addr(dev, uctx->cntxt_pd->id);
 868         uresp.db_page_size = dev->nic_info.db_page_size;
 869         uresp.phase_change = cq->phase_change ? 1 : 0;
 870         status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
 871         if (status) {
 872                 pr_err("%s(%d) copy error cqid=0x%x.\n",
 873                        __func__, dev->id, cq->id);
 874                 goto err;
 875         }
 876         status = ocrdma_add_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
 877         if (status)
 878                 goto err;
 879         status = ocrdma_add_mmap(uctx, uresp.page_addr[0], uresp.page_size);
 880         if (status) {
 881                 ocrdma_del_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
 882                 goto err;
 883         }
 884         cq->ucontext = uctx;
 885 err:
 886         return status;
 887 }
 888
 889 struct ib_cq *ocrdma_create_cq(struct ib_device *ibdev, int entries, int vector,
 890                                struct ib_ucontext *ib_ctx,
 891                                struct ib_udata *udata)
 892 {
 893         struct ocrdma_cq *cq;
 894         struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
 895         struct ocrdma_ucontext *uctx = NULL;
 896         u16 pd_id = 0;
 897         int status;
 898         struct ocrdma_create_cq_ureq ureq;
 899
 900         if (udata) {
 901                 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
 902                         return ERR_PTR(-EFAULT);
 903         } else
 904                 ureq.dpp_cq = 0;
 905         cq = kzalloc(sizeof(*cq), GFP_KERNEL);
 906         if (!cq)
 907                 return ERR_PTR(-ENOMEM);
 908
 909         spin_lock_init(&cq->cq_lock);
 910         spin_lock_init(&cq->comp_handler_lock);
 911         INIT_LIST_HEAD(&cq->sq_head);
 912         INIT_LIST_HEAD(&cq->rq_head);
 913
 914         if (ib_ctx) {
 915                 uctx = get_ocrdma_ucontext(ib_ctx);
 916                 pd_id = uctx->cntxt_pd->id;
 917         }
 918
 919         status = ocrdma_mbx_create_cq(dev, cq, entries, ureq.dpp_cq, pd_id);
 920         if (status) {
 921                 kfree(cq);
 922                 return ERR_PTR(status);
 923         }
 924         if (ib_ctx) {
 925                 status = ocrdma_copy_cq_uresp(dev, cq, udata, ib_ctx);
 926                 if (status)
 927                         goto ctx_err;
 928         }
 929         cq->phase = OCRDMA_CQE_VALID;
 930         cq->arm_needed = true;
 931         dev->cq_tbl[cq->id] = cq;
 932
 933         return &cq->ibcq;
 934
 935 ctx_err:
 936         ocrdma_mbx_destroy_cq(dev, cq);
 937         kfree(cq);
 938         return ERR_PTR(status);
 939 }
 940
 941 int ocrdma_resize_cq(struct ib_cq *ibcq, int new_cnt,
 942                      struct ib_udata *udata)
 943 {
 944         int status = 0;
 945         struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
 946
 947         if (new_cnt < 1 || new_cnt > cq->max_hw_cqe) {
 948                 status = -EINVAL;
 949                 return status;
 950         }
 951         ibcq->cqe = new_cnt;
 952         return status;
 953 }
 954
 955 int ocrdma_destroy_cq(struct ib_cq *ibcq)
 956 {
 957         int status;
 958         struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
 959         struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
 960         int pdid = 0;
 961
 962         status = ocrdma_mbx_destroy_cq(dev, cq);
 963
 964         if (cq->ucontext) {
 965                 pdid = cq->ucontext->cntxt_pd->id;
 966                 ocrdma_del_mmap(cq->ucontext, (u64) cq->pa,
 967                                 PAGE_ALIGN(cq->len));
 968                 ocrdma_del_mmap(cq->ucontext,
 969                                 ocrdma_get_db_addr(dev, pdid),
 970                                 dev->nic_info.db_page_size);
 971         }
 972         dev->cq_tbl[cq->id] = NULL;
 973
 974         kfree(cq);
 975         return status;
 976 }
 977
 978 static int ocrdma_add_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
 979 {
 980         int status = -EINVAL;
 981
 982         if (qp->id < OCRDMA_MAX_QP && dev->qp_tbl[qp->id] == NULL) {
 983                 dev->qp_tbl[qp->id] = qp;
 984                 status = 0;
 985         }
 986         return status;
 987 }
 988
 989 static void ocrdma_del_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
 990 {
 991         dev->qp_tbl[qp->id] = NULL;
 992 }
 993
 994 static int ocrdma_check_qp_params(struct ib_pd *ibpd, struct ocrdma_dev *dev,
 995                                   struct ib_qp_init_attr *attrs)
 996 {
 997         if ((attrs->qp_type != IB_QPT_GSI) &&
 998             (attrs->qp_type != IB_QPT_RC) &&
 999             (attrs->qp_type != IB_QPT_UC) &&
1000             (attrs->qp_type != IB_QPT_UD)) {
1001                 pr_err("%s(%d) unsupported qp type=0x%x requested\n",
1002                        __func__, dev->id, attrs->qp_type);
1003                 return -EINVAL;
1004         }
1005         /* Skip the check for QP1 to support CM size of 128 */
1006         if ((attrs->qp_type != IB_QPT_GSI) &&
1007             (attrs->cap.max_send_wr > dev->attr.max_wqe)) {
1008                 pr_err("%s(%d) unsupported send_wr=0x%x requested\n",
1009                        __func__, dev->id, attrs->cap.max_send_wr);
1010                 pr_err("%s(%d) supported send_wr=0x%x\n",
1011                        __func__, dev->id, dev->attr.max_wqe);
1012                 return -EINVAL;
1013         }
1014         if (!attrs->srq && (attrs->cap.max_recv_wr > dev->attr.max_rqe)) {
1015                 pr_err("%s(%d) unsupported recv_wr=0x%x requested\n",
1016                        __func__, dev->id, attrs->cap.max_recv_wr);
1017                 pr_err("%s(%d) supported recv_wr=0x%x\n",
1018                        __func__, dev->id, dev->attr.max_rqe);
1019                 return -EINVAL;
1020         }
1021         if (attrs->cap.max_inline_data > dev->attr.max_inline_data) {
1022                 pr_err("%s(%d) unsupported inline data size=0x%x requested\n",
1023                        __func__, dev->id, attrs->cap.max_inline_data);
1024                 pr_err("%s(%d) supported inline data size=0x%x\n",
1025                        __func__, dev->id, dev->attr.max_inline_data);
1026                 return -EINVAL;
1027         }
1028         if (attrs->cap.max_send_sge > dev->attr.max_send_sge) {
1029                 pr_err("%s(%d) unsupported send_sge=0x%x requested\n",
1030                        __func__, dev->id, attrs->cap.max_send_sge);
1031                 pr_err("%s(%d) supported send_sge=0x%x\n",
1032                        __func__, dev->id, dev->attr.max_send_sge);
1033                 return -EINVAL;
1034         }
1035         if (attrs->cap.max_recv_sge > dev->attr.max_recv_sge) {
1036                 pr_err("%s(%d) unsupported recv_sge=0x%x requested\n",
1037                        __func__, dev->id, attrs->cap.max_recv_sge);
1038                 pr_err("%s(%d) supported recv_sge=0x%x\n",
1039                        __func__, dev->id, dev->attr.max_recv_sge);
1040                 return -EINVAL;
1041         }
1042         /* unprivileged user space cannot create special QP */
1043         if (ibpd->uobject && attrs->qp_type == IB_QPT_GSI) {
1044                 pr_err
1045                     ("%s(%d) Userspace can't create special QPs of type=0x%x\n",
1046                      __func__, dev->id, attrs->qp_type);
1047                 return -EINVAL;
1048         }
1049         /* allow creating only one GSI type of QP */
1050         if (attrs->qp_type == IB_QPT_GSI && dev->gsi_qp_created) {
1051                 pr_err("%s(%d) GSI special QPs already created.\n",
1052                        __func__, dev->id);
1053                 return -EINVAL;
1054         }
1055         /* verify consumer QPs are not trying to use GSI QP's CQ */
1056         if ((attrs->qp_type != IB_QPT_GSI) && (dev->gsi_qp_created)) {
1057                 if ((dev->gsi_sqcq == get_ocrdma_cq(attrs->send_cq)) ||
1058                         (dev->gsi_rqcq == get_ocrdma_cq(attrs->recv_cq))) {
1059                         pr_err("%s(%d) Consumer QP cannot use GSI CQs.\n",
1060                                 __func__, dev->id);
1061                         return -EINVAL;
1062                 }
1063         }
1064         return 0;
1065 }
1066
1067 static int ocrdma_copy_qp_uresp(struct ocrdma_qp *qp,
1068                                 struct ib_udata *udata, int dpp_offset,
1069                                 int dpp_credit_lmt, int srq)
1070 {
1071         int status = 0;
1072         u64 usr_db;
1073         struct ocrdma_create_qp_uresp uresp;
1074         struct ocrdma_dev *dev = qp->dev;
1075         struct ocrdma_pd *pd = qp->pd;
1076
1077         memset(&uresp, 0, sizeof(uresp));
1078         usr_db = dev->nic_info.unmapped_db +
1079                         (pd->id * dev->nic_info.db_page_size);
1080         uresp.qp_id = qp->id;
1081         uresp.sq_dbid = qp->sq.dbid;
1082         uresp.num_sq_pages = 1;
1083         uresp.sq_page_size = PAGE_ALIGN(qp->sq.len);
1084         uresp.sq_page_addr[0] = qp->sq.pa;
1085         uresp.num_wqe_allocated = qp->sq.max_cnt;
1086         if (!srq) {
1087                 uresp.rq_dbid = qp->rq.dbid;
1088                 uresp.num_rq_pages = 1;
1089                 uresp.rq_page_size = PAGE_ALIGN(qp->rq.len);
1090                 uresp.rq_page_addr[0] = qp->rq.pa;
1091                 uresp.num_rqe_allocated = qp->rq.max_cnt;
1092         }
1093         uresp.db_page_addr = usr_db;
1094         uresp.db_page_size = dev->nic_info.db_page_size;
1095         if (dev->nic_info.dev_family == OCRDMA_GEN2_FAMILY) {
1096                 uresp.db_sq_offset = OCRDMA_DB_GEN2_SQ_OFFSET;
1097                 uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1098                 uresp.db_shift = 24;
1099         } else {
1100                 uresp.db_sq_offset = OCRDMA_DB_SQ_OFFSET;
1101                 uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
1102                 uresp.db_shift = 16;
1103         }
1104
1105         if (qp->dpp_enabled) {
1106                 uresp.dpp_credit = dpp_credit_lmt;
1107                 uresp.dpp_offset = dpp_offset;
1108         }
1109         status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1110         if (status) {
1111                 pr_err("%s(%d) user copy error.\n", __func__, dev->id);
1112                 goto err;
1113         }
1114         status = ocrdma_add_mmap(pd->uctx, uresp.sq_page_addr[0],
1115                                  uresp.sq_page_size);
1116         if (status)
1117                 goto err;
1118
1119         if (!srq) {
1120                 status = ocrdma_add_mmap(pd->uctx, uresp.rq_page_addr[0],
1121                                          uresp.rq_page_size);
1122                 if (status)
1123                         goto rq_map_err;
1124         }
1125         return status;
1126 rq_map_err:
1127         ocrdma_del_mmap(pd->uctx, uresp.sq_page_addr[0], uresp.sq_page_size);
1128 err:
1129         return status;
1130 }
1131
1132 static void ocrdma_set_qp_db(struct ocrdma_dev *dev, struct ocrdma_qp *qp,
1133                              struct ocrdma_pd *pd)
1134 {
1135         if (dev->nic_info.dev_family == OCRDMA_GEN2_FAMILY) {
1136                 qp->sq_db = dev->nic_info.db +
1137                         (pd->id * dev->nic_info.db_page_size) +
1138                         OCRDMA_DB_GEN2_SQ_OFFSET;
1139                 qp->rq_db = dev->nic_info.db +
1140                         (pd->id * dev->nic_info.db_page_size) +
1141                         OCRDMA_DB_GEN2_RQ_OFFSET;
1142         } else {
1143                 qp->sq_db = dev->nic_info.db +
1144                         (pd->id * dev->nic_info.db_page_size) +
1145                         OCRDMA_DB_SQ_OFFSET;
1146                 qp->rq_db = dev->nic_info.db +
1147                         (pd->id * dev->nic_info.db_page_size) +
1148                         OCRDMA_DB_RQ_OFFSET;
1149         }
1150 }
1151
1152 static int ocrdma_alloc_wr_id_tbl(struct ocrdma_qp *qp)
1153 {
1154         qp->wqe_wr_id_tbl =
1155             kzalloc(sizeof(*(qp->wqe_wr_id_tbl)) * qp->sq.max_cnt,
1156                     GFP_KERNEL);
1157         if (qp->wqe_wr_id_tbl == NULL)
1158                 return -ENOMEM;
1159         qp->rqe_wr_id_tbl =
1160             kzalloc(sizeof(u64) * qp->rq.max_cnt, GFP_KERNEL);
1161         if (qp->rqe_wr_id_tbl == NULL)
1162                 return -ENOMEM;
1163
1164         return 0;
1165 }
1166
1167 static void ocrdma_set_qp_init_params(struct ocrdma_qp *qp,
1168                                       struct ocrdma_pd *pd,
1169                                       struct ib_qp_init_attr *attrs)
1170 {
1171         qp->pd = pd;
1172         spin_lock_init(&qp->q_lock);
1173         INIT_LIST_HEAD(&qp->sq_entry);
1174         INIT_LIST_HEAD(&qp->rq_entry);
1175
1176         qp->qp_type = attrs->qp_type;
1177         qp->cap_flags = OCRDMA_QP_INB_RD | OCRDMA_QP_INB_WR;
1178         qp->max_inline_data = attrs->cap.max_inline_data;
1179         qp->sq.max_sges = attrs->cap.max_send_sge;
1180         qp->rq.max_sges = attrs->cap.max_recv_sge;
1181         qp->state = OCRDMA_QPS_RST;
1182         qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
1183 }
1184
1185
1186 static void ocrdma_store_gsi_qp_cq(struct ocrdma_dev *dev,
1187                                    struct ib_qp_init_attr *attrs)
1188 {
1189         if (attrs->qp_type == IB_QPT_GSI) {
1190                 dev->gsi_qp_created = 1;
1191                 dev->gsi_sqcq = get_ocrdma_cq(attrs->send_cq);
1192                 dev->gsi_rqcq = get_ocrdma_cq(attrs->recv_cq);
1193         }
1194 }
1195
1196 struct ib_qp *ocrdma_create_qp(struct ib_pd *ibpd,
1197                                struct ib_qp_init_attr *attrs,
1198                                struct ib_udata *udata)
1199 {
1200         int status;
1201         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1202         struct ocrdma_qp *qp;
1203         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1204         struct ocrdma_create_qp_ureq ureq;
1205         u16 dpp_credit_lmt, dpp_offset;
1206
1207         status = ocrdma_check_qp_params(ibpd, dev, attrs);
1208         if (status)
1209                 goto gen_err;
1210
1211         memset(&ureq, 0, sizeof(ureq));
1212         if (udata) {
1213                 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1214                         return ERR_PTR(-EFAULT);
1215         }
1216         qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1217         if (!qp) {
1218                 status = -ENOMEM;
1219                 goto gen_err;
1220         }
1221         qp->dev = dev;
1222         ocrdma_set_qp_init_params(qp, pd, attrs);
1223         if (udata == NULL)
1224                 qp->cap_flags |= (OCRDMA_QP_MW_BIND | OCRDMA_QP_LKEY0 |
1225                                         OCRDMA_QP_FAST_REG);
1226
1227         mutex_lock(&dev->dev_lock);
1228         status = ocrdma_mbx_create_qp(qp, attrs, ureq.enable_dpp_cq,
1229                                         ureq.dpp_cq_id,
1230                                         &dpp_offset, &dpp_credit_lmt);
1231         if (status)
1232                 goto mbx_err;
1233
1234         /* user space QP's wr_id table are managed in library */
1235         if (udata == NULL) {
1236                 status = ocrdma_alloc_wr_id_tbl(qp);
1237                 if (status)
1238                         goto map_err;
1239         }
1240
1241         status = ocrdma_add_qpn_map(dev, qp);
1242         if (status)
1243                 goto map_err;
1244         ocrdma_set_qp_db(dev, qp, pd);
1245         if (udata) {
1246                 status = ocrdma_copy_qp_uresp(qp, udata, dpp_offset,
1247                                               dpp_credit_lmt,
1248                                               (attrs->srq != NULL));
1249                 if (status)
1250                         goto cpy_err;
1251         }
1252         ocrdma_store_gsi_qp_cq(dev, attrs);
1253         qp->ibqp.qp_num = qp->id;
1254         mutex_unlock(&dev->dev_lock);
1255         return &qp->ibqp;
1256
1257 cpy_err:
1258         ocrdma_del_qpn_map(dev, qp);
1259 map_err:
1260         ocrdma_mbx_destroy_qp(dev, qp);
1261 mbx_err:
1262         mutex_unlock(&dev->dev_lock);
1263         kfree(qp->wqe_wr_id_tbl);
1264         kfree(qp->rqe_wr_id_tbl);
1265         kfree(qp);
1266         pr_err("%s(%d) error=%d\n", __func__, dev->id, status);
1267 gen_err:
1268         return ERR_PTR(status);
1269 }
1270
1271
1272 static void ocrdma_flush_rq_db(struct ocrdma_qp *qp)
1273 {
1274         if (qp->db_cache) {
1275                 u32 val = qp->rq.dbid | (qp->db_cache <<
1276                                 ocrdma_get_num_posted_shift(qp));
1277                 iowrite32(val, qp->rq_db);
1278                 qp->db_cache = 0;
1279         }
1280 }
1281
1282 int _ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1283                       int attr_mask)
1284 {
1285         int status = 0;
1286         struct ocrdma_qp *qp;
1287         struct ocrdma_dev *dev;
1288         enum ib_qp_state old_qps;
1289
1290         qp = get_ocrdma_qp(ibqp);
1291         dev = qp->dev;
1292         if (attr_mask & IB_QP_STATE)
1293                 status = ocrdma_qp_state_change(qp, attr->qp_state, &old_qps);
1294         /* if new and previous states are same hw doesn't need to
1295          * know about it.
1296          */
1297         if (status < 0)
1298                 return status;
1299         status = ocrdma_mbx_modify_qp(dev, qp, attr, attr_mask, old_qps);
1300         if (!status && attr_mask & IB_QP_STATE && attr->qp_state == IB_QPS_RTR)
1301                 ocrdma_flush_rq_db(qp);
1302
1303         return status;
1304 }
1305
1306 int ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1307                      int attr_mask, struct ib_udata *udata)
1308 {
1309         unsigned long flags;
1310         int status = -EINVAL;
1311         struct ocrdma_qp *qp;
1312         struct ocrdma_dev *dev;
1313         enum ib_qp_state old_qps, new_qps;
1314
1315         qp = get_ocrdma_qp(ibqp);
1316         dev = qp->dev;
1317
1318         /* syncronize with multiple context trying to change, retrive qps */
1319         mutex_lock(&dev->dev_lock);
1320         /* syncronize with wqe, rqe posting and cqe processing contexts */
1321         spin_lock_irqsave(&qp->q_lock, flags);
1322         old_qps = get_ibqp_state(qp->state);
1323         if (attr_mask & IB_QP_STATE)
1324                 new_qps = attr->qp_state;
1325         else
1326                 new_qps = old_qps;
1327         spin_unlock_irqrestore(&qp->q_lock, flags);
1328
1329         if (!ib_modify_qp_is_ok(old_qps, new_qps, ibqp->qp_type, attr_mask,
1330                                 IB_LINK_LAYER_ETHERNET)) {
1331                 pr_err("%s(%d) invalid attribute mask=0x%x specified for\n"
1332                        "qpn=0x%x of type=0x%x old_qps=0x%x, new_qps=0x%x\n",
1333                        __func__, dev->id, attr_mask, qp->id, ibqp->qp_type,
1334                        old_qps, new_qps);
1335                 goto param_err;
1336         }
1337
1338         status = _ocrdma_modify_qp(ibqp, attr, attr_mask);
1339         if (status > 0)
1340                 status = 0;
1341 param_err:
1342         mutex_unlock(&dev->dev_lock);
1343         return status;
1344 }
1345
1346 static enum ib_mtu ocrdma_mtu_int_to_enum(u16 mtu)
1347 {
1348         switch (mtu) {
1349         case 256:
1350                 return IB_MTU_256;
1351         case 512:
1352                 return IB_MTU_512;
1353         case 1024:
1354                 return IB_MTU_1024;
1355         case 2048:
1356                 return IB_MTU_2048;
1357         case 4096:
1358                 return IB_MTU_4096;
1359         default:
1360                 return IB_MTU_1024;
1361         }
1362 }
1363
1364 static int ocrdma_to_ib_qp_acc_flags(int qp_cap_flags)
1365 {
1366         int ib_qp_acc_flags = 0;
1367
1368         if (qp_cap_flags & OCRDMA_QP_INB_WR)
1369                 ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
1370         if (qp_cap_flags & OCRDMA_QP_INB_RD)
1371                 ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
1372         return ib_qp_acc_flags;
1373 }
1374
1375 int ocrdma_query_qp(struct ib_qp *ibqp,
1376                     struct ib_qp_attr *qp_attr,
1377                     int attr_mask, struct ib_qp_init_attr *qp_init_attr)
1378 {
1379         int status;
1380         u32 qp_state;
1381         struct ocrdma_qp_params params;
1382         struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1383         struct ocrdma_dev *dev = qp->dev;
1384
1385         memset(&params, 0, sizeof(params));
1386         mutex_lock(&dev->dev_lock);
1387         status = ocrdma_mbx_query_qp(dev, qp, &params);
1388         mutex_unlock(&dev->dev_lock);
1389         if (status)
1390                 goto mbx_err;
1391         qp_attr->qp_state = get_ibqp_state(IB_QPS_INIT);
1392         qp_attr->cur_qp_state = get_ibqp_state(IB_QPS_INIT);
1393         qp_attr->path_mtu =
1394                 ocrdma_mtu_int_to_enum(params.path_mtu_pkey_indx &
1395                                 OCRDMA_QP_PARAMS_PATH_MTU_MASK) >>
1396                                 OCRDMA_QP_PARAMS_PATH_MTU_SHIFT;
1397         qp_attr->path_mig_state = IB_MIG_MIGRATED;
1398         qp_attr->rq_psn = params.hop_lmt_rq_psn & OCRDMA_QP_PARAMS_RQ_PSN_MASK;
1399         qp_attr->sq_psn = params.tclass_sq_psn & OCRDMA_QP_PARAMS_SQ_PSN_MASK;
1400         qp_attr->dest_qp_num =
1401             params.ack_to_rnr_rtc_dest_qpn & OCRDMA_QP_PARAMS_DEST_QPN_MASK;
1402
1403         qp_attr->qp_access_flags = ocrdma_to_ib_qp_acc_flags(qp->cap_flags);
1404         qp_attr->cap.max_send_wr = qp->sq.max_cnt - 1;
1405         qp_attr->cap.max_recv_wr = qp->rq.max_cnt - 1;
1406         qp_attr->cap.max_send_sge = qp->sq.max_sges;
1407         qp_attr->cap.max_recv_sge = qp->rq.max_sges;
1408         qp_attr->cap.max_inline_data = qp->max_inline_data;
1409         qp_init_attr->cap = qp_attr->cap;
1410         memcpy(&qp_attr->ah_attr.grh.dgid, &params.dgid[0],
1411                sizeof(params.dgid));
1412         qp_attr->ah_attr.grh.flow_label = params.rnt_rc_sl_fl &
1413             OCRDMA_QP_PARAMS_FLOW_LABEL_MASK;
1414         qp_attr->ah_attr.grh.sgid_index = qp->sgid_idx;
1415         qp_attr->ah_attr.grh.hop_limit = (params.hop_lmt_rq_psn &
1416                                           OCRDMA_QP_PARAMS_HOP_LMT_MASK) >>
1417                                                 OCRDMA_QP_PARAMS_HOP_LMT_SHIFT;
1418         qp_attr->ah_attr.grh.traffic_class = (params.tclass_sq_psn &
1419                                               OCRDMA_QP_PARAMS_SQ_PSN_MASK) >>
1420                                                 OCRDMA_QP_PARAMS_TCLASS_SHIFT;
1421
1422         qp_attr->ah_attr.ah_flags = IB_AH_GRH;
1423         qp_attr->ah_attr.port_num = 1;
1424         qp_attr->ah_attr.sl = (params.rnt_rc_sl_fl &
1425                                OCRDMA_QP_PARAMS_SL_MASK) >>
1426                                 OCRDMA_QP_PARAMS_SL_SHIFT;
1427         qp_attr->timeout = (params.ack_to_rnr_rtc_dest_qpn &
1428                             OCRDMA_QP_PARAMS_ACK_TIMEOUT_MASK) >>
1429                                 OCRDMA_QP_PARAMS_ACK_TIMEOUT_SHIFT;
1430         qp_attr->rnr_retry = (params.ack_to_rnr_rtc_dest_qpn &
1431                               OCRDMA_QP_PARAMS_RNR_RETRY_CNT_MASK) >>
1432                                 OCRDMA_QP_PARAMS_RNR_RETRY_CNT_SHIFT;
1433         qp_attr->retry_cnt =
1434             (params.rnt_rc_sl_fl & OCRDMA_QP_PARAMS_RETRY_CNT_MASK) >>
1435                 OCRDMA_QP_PARAMS_RETRY_CNT_SHIFT;
1436         qp_attr->min_rnr_timer = 0;
1437         qp_attr->pkey_index = 0;
1438         qp_attr->port_num = 1;
1439         qp_attr->ah_attr.src_path_bits = 0;
1440         qp_attr->ah_attr.static_rate = 0;
1441         qp_attr->alt_pkey_index = 0;
1442         qp_attr->alt_port_num = 0;
1443         qp_attr->alt_timeout = 0;
1444         memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
1445         qp_state = (params.max_sge_recv_flags & OCRDMA_QP_PARAMS_STATE_MASK) >>
1446                     OCRDMA_QP_PARAMS_STATE_SHIFT;
1447         qp_attr->sq_draining = (qp_state == OCRDMA_QPS_SQ_DRAINING) ? 1 : 0;
1448         qp_attr->max_dest_rd_atomic =
1449             params.max_ord_ird >> OCRDMA_QP_PARAMS_MAX_ORD_SHIFT;
1450         qp_attr->max_rd_atomic =
1451             params.max_ord_ird & OCRDMA_QP_PARAMS_MAX_IRD_MASK;
1452         qp_attr->en_sqd_async_notify = (params.max_sge_recv_flags &
1453                                 OCRDMA_QP_PARAMS_FLAGS_SQD_ASYNC) ? 1 : 0;
1454 mbx_err:
1455         return status;
1456 }
1457
1458 static void ocrdma_srq_toggle_bit(struct ocrdma_srq *srq, int idx)
1459 {
1460         int i = idx / 32;
1461         unsigned int mask = (1 << (idx % 32));
1462
1463         if (srq->idx_bit_fields[i] & mask)
1464                 srq->idx_bit_fields[i] &= ~mask;
1465         else
1466                 srq->idx_bit_fields[i] |= mask;
1467 }
1468
1469 static int ocrdma_hwq_free_cnt(struct ocrdma_qp_hwq_info *q)
1470 {
1471         return ((q->max_wqe_idx - q->head) + q->tail) % q->max_cnt;
1472 }
1473
1474 static int is_hw_sq_empty(struct ocrdma_qp *qp)
1475 {
1476         return (qp->sq.tail == qp->sq.head);
1477 }
1478
1479 static int is_hw_rq_empty(struct ocrdma_qp *qp)
1480 {
1481         return (qp->rq.tail == qp->rq.head);
1482 }
1483
1484 static void *ocrdma_hwq_head(struct ocrdma_qp_hwq_info *q)
1485 {
1486         return q->va + (q->head * q->entry_size);
1487 }
1488
1489 static void *ocrdma_hwq_head_from_idx(struct ocrdma_qp_hwq_info *q,
1490                                       u32 idx)
1491 {
1492         return q->va + (idx * q->entry_size);
1493 }
1494
1495 static void ocrdma_hwq_inc_head(struct ocrdma_qp_hwq_info *q)
1496 {
1497         q->head = (q->head + 1) & q->max_wqe_idx;
1498 }
1499
1500 static void ocrdma_hwq_inc_tail(struct ocrdma_qp_hwq_info *q)
1501 {
1502         q->tail = (q->tail + 1) & q->max_wqe_idx;
1503 }
1504
1505 /* discard the cqe for a given QP */
1506 static void ocrdma_discard_cqes(struct ocrdma_qp *qp, struct ocrdma_cq *cq)
1507 {
1508         unsigned long cq_flags;
1509         unsigned long flags;
1510         int discard_cnt = 0;
1511         u32 cur_getp, stop_getp;
1512         struct ocrdma_cqe *cqe;
1513         u32 qpn = 0;
1514
1515         spin_lock_irqsave(&cq->cq_lock, cq_flags);
1516
1517         /* traverse through the CQEs in the hw CQ,
1518          * find the matching CQE for a given qp,
1519          * mark the matching one discarded by clearing qpn.
1520          * ring the doorbell in the poll_cq() as
1521          * we don't complete out of order cqe.
1522          */
1523
1524         cur_getp = cq->getp;
1525         /* find upto when do we reap the cq. */
1526         stop_getp = cur_getp;
1527         do {
1528                 if (is_hw_sq_empty(qp) && (!qp->srq && is_hw_rq_empty(qp)))
1529                         break;
1530
1531                 cqe = cq->va + cur_getp;
1532                 /* if (a) done reaping whole hw cq, or
1533                  *    (b) qp_xq becomes empty.
1534                  * then exit
1535                  */
1536                 qpn = cqe->cmn.qpn & OCRDMA_CQE_QPN_MASK;
1537                 /* if previously discarded cqe found, skip that too. */
1538                 /* check for matching qp */
1539                 if (qpn == 0 || qpn != qp->id)
1540                         goto skip_cqe;
1541
1542                 /* mark cqe discarded so that it is not picked up later
1543                  * in the poll_cq().
1544                  */
1545                 discard_cnt += 1;
1546                 cqe->cmn.qpn = 0;
1547                 if (is_cqe_for_sq(cqe)) {
1548                         ocrdma_hwq_inc_tail(&qp->sq);
1549                 } else {
1550                         if (qp->srq) {
1551                                 spin_lock_irqsave(&qp->srq->q_lock, flags);
1552                                 ocrdma_hwq_inc_tail(&qp->srq->rq);
1553                                 ocrdma_srq_toggle_bit(qp->srq, cur_getp);
1554                                 spin_unlock_irqrestore(&qp->srq->q_lock, flags);
1555
1556                         } else {
1557                                 ocrdma_hwq_inc_tail(&qp->rq);
1558                         }
1559                 }
1560 skip_cqe:
1561                 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
1562         } while (cur_getp != stop_getp);
1563         spin_unlock_irqrestore(&cq->cq_lock, cq_flags);
1564 }
1565
1566 void ocrdma_del_flush_qp(struct ocrdma_qp *qp)
1567 {
1568         int found = false;
1569         unsigned long flags;
1570         struct ocrdma_dev *dev = qp->dev;
1571         /* sync with any active CQ poll */
1572
1573         spin_lock_irqsave(&dev->flush_q_lock, flags);
1574         found = ocrdma_is_qp_in_sq_flushlist(qp->sq_cq, qp);
1575         if (found)
1576                 list_del(&qp->sq_entry);
1577         if (!qp->srq) {
1578                 found = ocrdma_is_qp_in_rq_flushlist(qp->rq_cq, qp);
1579                 if (found)
1580                         list_del(&qp->rq_entry);
1581         }
1582         spin_unlock_irqrestore(&dev->flush_q_lock, flags);
1583 }
1584
1585 int ocrdma_destroy_qp(struct ib_qp *ibqp)
1586 {
1587         int status;
1588         struct ocrdma_pd *pd;
1589         struct ocrdma_qp *qp;
1590         struct ocrdma_dev *dev;
1591         struct ib_qp_attr attrs;
1592         int attr_mask = IB_QP_STATE;
1593         unsigned long flags;
1594
1595         qp = get_ocrdma_qp(ibqp);
1596         dev = qp->dev;
1597
1598         attrs.qp_state = IB_QPS_ERR;
1599         pd = qp->pd;
1600
1601         /* change the QP state to ERROR */
1602         _ocrdma_modify_qp(ibqp, &attrs, attr_mask);
1603
1604         /* ensure that CQEs for newly created QP (whose id may be same with
1605          * one which just getting destroyed are same), dont get
1606          * discarded until the old CQEs are discarded.
1607          */
1608         mutex_lock(&dev->dev_lock);
1609         status = ocrdma_mbx_destroy_qp(dev, qp);
1610
1611         /*
1612          * acquire CQ lock while destroy is in progress, in order to
1613          * protect against proessing in-flight CQEs for this QP.
1614          */
1615         spin_lock_irqsave(&qp->sq_cq->cq_lock, flags);
1616         if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1617                 spin_lock(&qp->rq_cq->cq_lock);
1618
1619         ocrdma_del_qpn_map(dev, qp);
1620
1621         if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1622                 spin_unlock(&qp->rq_cq->cq_lock);
1623         spin_unlock_irqrestore(&qp->sq_cq->cq_lock, flags);
1624
1625         if (!pd->uctx) {
1626                 ocrdma_discard_cqes(qp, qp->sq_cq);
1627                 ocrdma_discard_cqes(qp, qp->rq_cq);
1628         }
1629         mutex_unlock(&dev->dev_lock);
1630
1631         if (pd->uctx) {
1632                 ocrdma_del_mmap(pd->uctx, (u64) qp->sq.pa,
1633                                 PAGE_ALIGN(qp->sq.len));
1634                 if (!qp->srq)
1635                         ocrdma_del_mmap(pd->uctx, (u64) qp->rq.pa,
1636                                         PAGE_ALIGN(qp->rq.len));
1637         }
1638
1639         ocrdma_del_flush_qp(qp);
1640
1641         kfree(qp->wqe_wr_id_tbl);
1642         kfree(qp->rqe_wr_id_tbl);
1643         kfree(qp);
1644         return status;
1645 }
1646
1647 static int ocrdma_copy_srq_uresp(struct ocrdma_dev *dev, struct ocrdma_srq *srq,
1648                                 struct ib_udata *udata)
1649 {
1650         int status;
1651         struct ocrdma_create_srq_uresp uresp;
1652
1653         memset(&uresp, 0, sizeof(uresp));
1654         uresp.rq_dbid = srq->rq.dbid;
1655         uresp.num_rq_pages = 1;
1656         uresp.rq_page_addr[0] = srq->rq.pa;
1657         uresp.rq_page_size = srq->rq.len;
1658         uresp.db_page_addr = dev->nic_info.unmapped_db +
1659             (srq->pd->id * dev->nic_info.db_page_size);
1660         uresp.db_page_size = dev->nic_info.db_page_size;
1661         uresp.num_rqe_allocated = srq->rq.max_cnt;
1662         if (dev->nic_info.dev_family == OCRDMA_GEN2_FAMILY) {
1663                 uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1664                 uresp.db_shift = 24;
1665         } else {
1666                 uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
1667                 uresp.db_shift = 16;
1668         }
1669
1670         status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1671         if (status)
1672                 return status;
1673         status = ocrdma_add_mmap(srq->pd->uctx, uresp.rq_page_addr[0],
1674                                  uresp.rq_page_size);
1675         if (status)
1676                 return status;
1677         return status;
1678 }
1679
1680 struct ib_srq *ocrdma_create_srq(struct ib_pd *ibpd,
1681                                  struct ib_srq_init_attr *init_attr,
1682                                  struct ib_udata *udata)
1683 {
1684         int status = -ENOMEM;
1685         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1686         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1687         struct ocrdma_srq *srq;
1688
1689         if (init_attr->attr.max_sge > dev->attr.max_recv_sge)
1690                 return ERR_PTR(-EINVAL);
1691         if (init_attr->attr.max_wr > dev->attr.max_rqe)
1692                 return ERR_PTR(-EINVAL);
1693
1694         srq = kzalloc(sizeof(*srq), GFP_KERNEL);
1695         if (!srq)
1696                 return ERR_PTR(status);
1697
1698         spin_lock_init(&srq->q_lock);
1699         srq->pd = pd;
1700         srq->db = dev->nic_info.db + (pd->id * dev->nic_info.db_page_size);
1701         status = ocrdma_mbx_create_srq(dev, srq, init_attr, pd);
1702         if (status)
1703                 goto err;
1704
1705         if (udata == NULL) {
1706                 srq->rqe_wr_id_tbl = kzalloc(sizeof(u64) * srq->rq.max_cnt,
1707                             GFP_KERNEL);
1708                 if (srq->rqe_wr_id_tbl == NULL)
1709                         goto arm_err;
1710
1711                 srq->bit_fields_len = (srq->rq.max_cnt / 32) +
1712                     (srq->rq.max_cnt % 32 ? 1 : 0);
1713                 srq->idx_bit_fields =
1714                     kmalloc(srq->bit_fields_len * sizeof(u32), GFP_KERNEL);
1715                 if (srq->idx_bit_fields == NULL)
1716                         goto arm_err;
1717                 memset(srq->idx_bit_fields, 0xff,
1718                        srq->bit_fields_len * sizeof(u32));
1719         }
1720
1721         if (init_attr->attr.srq_limit) {
1722                 status = ocrdma_mbx_modify_srq(srq, &init_attr->attr);
1723                 if (status)
1724                         goto arm_err;
1725         }
1726
1727         if (udata) {
1728                 status = ocrdma_copy_srq_uresp(dev, srq, udata);
1729                 if (status)
1730                         goto arm_err;
1731         }
1732
1733         return &srq->ibsrq;
1734
1735 arm_err:
1736         ocrdma_mbx_destroy_srq(dev, srq);
1737 err:
1738         kfree(srq->rqe_wr_id_tbl);
1739         kfree(srq->idx_bit_fields);
1740         kfree(srq);
1741         return ERR_PTR(status);
1742 }
1743
1744 int ocrdma_modify_srq(struct ib_srq *ibsrq,
1745                       struct ib_srq_attr *srq_attr,
1746                       enum ib_srq_attr_mask srq_attr_mask,
1747                       struct ib_udata *udata)
1748 {
1749         int status = 0;
1750         struct ocrdma_srq *srq;
1751
1752         srq = get_ocrdma_srq(ibsrq);
1753         if (srq_attr_mask & IB_SRQ_MAX_WR)
1754                 status = -EINVAL;
1755         else
1756                 status = ocrdma_mbx_modify_srq(srq, srq_attr);
1757         return status;
1758 }
1759
1760 int ocrdma_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
1761 {
1762         int status;
1763         struct ocrdma_srq *srq;
1764
1765         srq = get_ocrdma_srq(ibsrq);
1766         status = ocrdma_mbx_query_srq(srq, srq_attr);
1767         return status;
1768 }
1769
1770 int ocrdma_destroy_srq(struct ib_srq *ibsrq)
1771 {
1772         int status;
1773         struct ocrdma_srq *srq;
1774         struct ocrdma_dev *dev = get_ocrdma_dev(ibsrq->device);
1775
1776         srq = get_ocrdma_srq(ibsrq);
1777
1778         status = ocrdma_mbx_destroy_srq(dev, srq);
1779
1780         if (srq->pd->uctx)
1781                 ocrdma_del_mmap(srq->pd->uctx, (u64) srq->rq.pa,
1782                                 PAGE_ALIGN(srq->rq.len));
1783
1784         kfree(srq->idx_bit_fields);
1785         kfree(srq->rqe_wr_id_tbl);
1786         kfree(srq);
1787         return status;
1788 }
1789
1790 /* unprivileged verbs and their support functions. */
1791 static void ocrdma_build_ud_hdr(struct ocrdma_qp *qp,
1792                                 struct ocrdma_hdr_wqe *hdr,
1793                                 struct ib_send_wr *wr)
1794 {
1795         struct ocrdma_ewqe_ud_hdr *ud_hdr =
1796                 (struct ocrdma_ewqe_ud_hdr *)(hdr + 1);
1797         struct ocrdma_ah *ah = get_ocrdma_ah(wr->wr.ud.ah);
1798
1799         ud_hdr->rsvd_dest_qpn = wr->wr.ud.remote_qpn;
1800         if (qp->qp_type == IB_QPT_GSI)
1801                 ud_hdr->qkey = qp->qkey;
1802         else
1803                 ud_hdr->qkey = wr->wr.ud.remote_qkey;
1804         ud_hdr->rsvd_ahid = ah->id;
1805 }
1806
1807 static void ocrdma_build_sges(struct ocrdma_hdr_wqe *hdr,
1808                               struct ocrdma_sge *sge, int num_sge,
1809                               struct ib_sge *sg_list)
1810 {
1811         int i;
1812
1813         for (i = 0; i < num_sge; i++) {
1814                 sge[i].lrkey = sg_list[i].lkey;
1815                 sge[i].addr_lo = sg_list[i].addr;
1816                 sge[i].addr_hi = upper_32_bits(sg_list[i].addr);
1817                 sge[i].len = sg_list[i].length;
1818                 hdr->total_len += sg_list[i].length;
1819         }
1820         if (num_sge == 0)
1821                 memset(sge, 0, sizeof(*sge));
1822 }
1823
1824 static inline uint32_t ocrdma_sglist_len(struct ib_sge *sg_list, int num_sge)
1825 {
1826         uint32_t total_len = 0, i;
1827
1828         for (i = 0; i < num_sge; i++)
1829                 total_len += sg_list[i].length;
1830         return total_len;
1831 }
1832
1833
1834 static int ocrdma_build_inline_sges(struct ocrdma_qp *qp,
1835                                     struct ocrdma_hdr_wqe *hdr,
1836                                     struct ocrdma_sge *sge,
1837                                     struct ib_send_wr *wr, u32 wqe_size)
1838 {
1839         int i;
1840         char *dpp_addr;
1841
1842         if (wr->send_flags & IB_SEND_INLINE && qp->qp_type != IB_QPT_UD) {
1843                 hdr->total_len = ocrdma_sglist_len(wr->sg_list, wr->num_sge);
1844                 if (unlikely(hdr->total_len > qp->max_inline_data)) {
1845                         pr_err("%s() supported_len=0x%x,\n"
1846                                " unspported len req=0x%x\n", __func__,
1847                                 qp->max_inline_data, hdr->total_len);
1848                         return -EINVAL;
1849                 }
1850                 dpp_addr = (char *)sge;
1851                 for (i = 0; i < wr->num_sge; i++) {
1852                         memcpy(dpp_addr,
1853                                (void *)(unsigned long)wr->sg_list[i].addr,
1854                                wr->sg_list[i].length);
1855                         dpp_addr += wr->sg_list[i].length;
1856                 }
1857
1858                 wqe_size += roundup(hdr->total_len, OCRDMA_WQE_ALIGN_BYTES);
1859                 if (0 == hdr->total_len)
1860                         wqe_size += sizeof(struct ocrdma_sge);
1861                 hdr->cw |= (OCRDMA_TYPE_INLINE << OCRDMA_WQE_TYPE_SHIFT);
1862         } else {
1863                 ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
1864                 if (wr->num_sge)
1865                         wqe_size += (wr->num_sge * sizeof(struct ocrdma_sge));
1866                 else
1867                         wqe_size += sizeof(struct ocrdma_sge);
1868                 hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
1869         }
1870         hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
1871         return 0;
1872 }
1873
1874 static int ocrdma_build_send(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1875                              struct ib_send_wr *wr)
1876 {
1877         int status;
1878         struct ocrdma_sge *sge;
1879         u32 wqe_size = sizeof(*hdr);
1880
1881         if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
1882                 ocrdma_build_ud_hdr(qp, hdr, wr);
1883                 sge = (struct ocrdma_sge *)(hdr + 2);
1884                 wqe_size += sizeof(struct ocrdma_ewqe_ud_hdr);
1885         } else {
1886                 sge = (struct ocrdma_sge *)(hdr + 1);
1887         }
1888
1889         status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1890         return status;
1891 }
1892
1893 static int ocrdma_build_write(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1894                               struct ib_send_wr *wr)
1895 {
1896         int status;
1897         struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
1898         struct ocrdma_sge *sge = ext_rw + 1;
1899         u32 wqe_size = sizeof(*hdr) + sizeof(*ext_rw);
1900
1901         status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1902         if (status)
1903                 return status;
1904         ext_rw->addr_lo = wr->wr.rdma.remote_addr;
1905         ext_rw->addr_hi = upper_32_bits(wr->wr.rdma.remote_addr);
1906         ext_rw->lrkey = wr->wr.rdma.rkey;
1907         ext_rw->len = hdr->total_len;
1908         return 0;
1909 }
1910
1911 static void ocrdma_build_read(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1912                               struct ib_send_wr *wr)
1913 {
1914         struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
1915         struct ocrdma_sge *sge = ext_rw + 1;
1916         u32 wqe_size = ((wr->num_sge + 1) * sizeof(struct ocrdma_sge)) +
1917             sizeof(struct ocrdma_hdr_wqe);
1918
1919         ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
1920         hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
1921         hdr->cw |= (OCRDMA_READ << OCRDMA_WQE_OPCODE_SHIFT);
1922         hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
1923
1924         ext_rw->addr_lo = wr->wr.rdma.remote_addr;
1925         ext_rw->addr_hi = upper_32_bits(wr->wr.rdma.remote_addr);
1926         ext_rw->lrkey = wr->wr.rdma.rkey;
1927         ext_rw->len = hdr->total_len;
1928 }
1929
1930 static void build_frmr_pbes(struct ib_send_wr *wr, struct ocrdma_pbl *pbl_tbl,
1931                             struct ocrdma_hw_mr *hwmr)
1932 {
1933         int i;
1934         u64 buf_addr = 0;
1935         int num_pbes;
1936         struct ocrdma_pbe *pbe;
1937
1938         pbe = (struct ocrdma_pbe *)pbl_tbl->va;
1939         num_pbes = 0;
1940
1941         /* go through the OS phy regions & fill hw pbe entries into pbls. */
1942         for (i = 0; i < wr->wr.fast_reg.page_list_len; i++) {
1943                 /* number of pbes can be more for one OS buf, when
1944                  * buffers are of different sizes.
1945                  * split the ib_buf to one or more pbes.
1946                  */
1947                 buf_addr = wr->wr.fast_reg.page_list->page_list[i];
1948                 pbe->pa_lo = cpu_to_le32((u32) (buf_addr & PAGE_MASK));
1949                 pbe->pa_hi = cpu_to_le32((u32) upper_32_bits(buf_addr));
1950                 num_pbes += 1;
1951                 pbe++;
1952
1953                 /* if the pbl is full storing the pbes,
1954                  * move to next pbl.
1955                 */
1956                 if (num_pbes == (hwmr->pbl_size/sizeof(u64))) {
1957                         pbl_tbl++;
1958                         pbe = (struct ocrdma_pbe *)pbl_tbl->va;
1959                 }
1960         }
1961         return;
1962 }
1963
1964 static int get_encoded_page_size(int pg_sz)
1965 {
1966         /* Max size is 256M 4096 << 16 */
1967         int i = 0;
1968         for (; i < 17; i++)
1969                 if (pg_sz == (4096 << i))
1970                         break;
1971         return i;
1972 }
1973
1974
1975 static int ocrdma_build_fr(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1976                            struct ib_send_wr *wr)
1977 {
1978         u64 fbo;
1979         struct ocrdma_ewqe_fr *fast_reg = (struct ocrdma_ewqe_fr *)(hdr + 1);
1980         struct ocrdma_mr *mr;
1981         u32 wqe_size = sizeof(*fast_reg) + sizeof(*hdr);
1982
1983         wqe_size = roundup(wqe_size, OCRDMA_WQE_ALIGN_BYTES);
1984
1985         if (wr->wr.fast_reg.page_list_len > qp->dev->attr.max_pages_per_frmr)
1986                 return -EINVAL;
1987
1988         hdr->cw |= (OCRDMA_FR_MR << OCRDMA_WQE_OPCODE_SHIFT);
1989         hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
1990
1991         if (wr->wr.fast_reg.page_list_len == 0)
1992                 BUG();
1993         if (wr->wr.fast_reg.access_flags & IB_ACCESS_LOCAL_WRITE)
1994                 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_LOCAL_WR;
1995         if (wr->wr.fast_reg.access_flags & IB_ACCESS_REMOTE_WRITE)
1996                 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_WR;
1997         if (wr->wr.fast_reg.access_flags & IB_ACCESS_REMOTE_READ)
1998                 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_RD;
1999         hdr->lkey = wr->wr.fast_reg.rkey;
2000         hdr->total_len = wr->wr.fast_reg.length;
2001
2002         fbo = wr->wr.fast_reg.iova_start -
2003             (wr->wr.fast_reg.page_list->page_list[0] & PAGE_MASK);
2004
2005         fast_reg->va_hi = upper_32_bits(wr->wr.fast_reg.iova_start);
2006         fast_reg->va_lo = (u32) (wr->wr.fast_reg.iova_start & 0xffffffff);
2007         fast_reg->fbo_hi = upper_32_bits(fbo);
2008         fast_reg->fbo_lo = (u32) fbo & 0xffffffff;
2009         fast_reg->num_sges = wr->wr.fast_reg.page_list_len;
2010         fast_reg->size_sge =
2011                 get_encoded_page_size(1 << wr->wr.fast_reg.page_shift);
2012         mr = (struct ocrdma_mr *) (unsigned long) qp->dev->stag_arr[(hdr->lkey >> 8) &
2013                 (OCRDMA_MAX_STAG - 1)];
2014         build_frmr_pbes(wr, mr->hwmr.pbl_table, &mr->hwmr);
2015         return 0;
2016 }
2017
2018 static void ocrdma_ring_sq_db(struct ocrdma_qp *qp)
2019 {
2020         u32 val = qp->sq.dbid | (1 << 16);
2021
2022         iowrite32(val, qp->sq_db);
2023 }
2024
2025 int ocrdma_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2026                      struct ib_send_wr **bad_wr)
2027 {
2028         int status = 0;
2029         struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2030         struct ocrdma_hdr_wqe *hdr;
2031         unsigned long flags;
2032
2033         spin_lock_irqsave(&qp->q_lock, flags);
2034         if (qp->state != OCRDMA_QPS_RTS && qp->state != OCRDMA_QPS_SQD) {
2035                 spin_unlock_irqrestore(&qp->q_lock, flags);
2036                 *bad_wr = wr;
2037                 return -EINVAL;
2038         }
2039
2040         while (wr) {
2041                 if (ocrdma_hwq_free_cnt(&qp->sq) == 0 ||
2042                     wr->num_sge > qp->sq.max_sges) {
2043                         *bad_wr = wr;
2044                         status = -ENOMEM;
2045                         break;
2046                 }
2047                 hdr = ocrdma_hwq_head(&qp->sq);
2048                 hdr->cw = 0;
2049                 if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2050                         hdr->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2051                 if (wr->send_flags & IB_SEND_FENCE)
2052                         hdr->cw |=
2053                             (OCRDMA_FLAG_FENCE_L << OCRDMA_WQE_FLAGS_SHIFT);
2054                 if (wr->send_flags & IB_SEND_SOLICITED)
2055                         hdr->cw |=
2056                             (OCRDMA_FLAG_SOLICIT << OCRDMA_WQE_FLAGS_SHIFT);
2057                 hdr->total_len = 0;
2058                 switch (wr->opcode) {
2059                 case IB_WR_SEND_WITH_IMM:
2060                         hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2061                         hdr->immdt = ntohl(wr->ex.imm_data);
2062                 case IB_WR_SEND:
2063                         hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2064                         ocrdma_build_send(qp, hdr, wr);
2065                         break;
2066                 case IB_WR_SEND_WITH_INV:
2067                         hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
2068                         hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2069                         hdr->lkey = wr->ex.invalidate_rkey;
2070                         status = ocrdma_build_send(qp, hdr, wr);
2071                         break;
2072                 case IB_WR_RDMA_WRITE_WITH_IMM:
2073                         hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2074                         hdr->immdt = ntohl(wr->ex.imm_data);
2075                 case IB_WR_RDMA_WRITE:
2076                         hdr->cw |= (OCRDMA_WRITE << OCRDMA_WQE_OPCODE_SHIFT);
2077                         status = ocrdma_build_write(qp, hdr, wr);
2078                         break;
2079                 case IB_WR_RDMA_READ_WITH_INV:
2080                         hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
2081                 case IB_WR_RDMA_READ:
2082                         ocrdma_build_read(qp, hdr, wr);
2083                         break;
2084                 case IB_WR_LOCAL_INV:
2085                         hdr->cw |=
2086                             (OCRDMA_LKEY_INV << OCRDMA_WQE_OPCODE_SHIFT);
2087                         hdr->cw |= ((sizeof(struct ocrdma_hdr_wqe) +
2088                                         sizeof(struct ocrdma_sge)) /
2089                                 OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT;
2090                         hdr->lkey = wr->ex.invalidate_rkey;
2091                         break;
2092                 case IB_WR_FAST_REG_MR:
2093                         status = ocrdma_build_fr(qp, hdr, wr);
2094                         break;
2095                 default:
2096                         status = -EINVAL;
2097                         break;
2098                 }
2099                 if (status) {
2100                         *bad_wr = wr;
2101                         break;
2102                 }
2103                 if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2104                         qp->wqe_wr_id_tbl[qp->sq.head].signaled = 1;
2105                 else
2106                         qp->wqe_wr_id_tbl[qp->sq.head].signaled = 0;
2107                 qp->wqe_wr_id_tbl[qp->sq.head].wrid = wr->wr_id;
2108                 ocrdma_cpu_to_le32(hdr, ((hdr->cw >> OCRDMA_WQE_SIZE_SHIFT) &
2109                                    OCRDMA_WQE_SIZE_MASK) * OCRDMA_WQE_STRIDE);
2110                 /* make sure wqe is written before adapter can access it */
2111                 wmb();
2112                 /* inform hw to start processing it */
2113                 ocrdma_ring_sq_db(qp);
2114
2115                 /* update pointer, counter for next wr */
2116                 ocrdma_hwq_inc_head(&qp->sq);
2117                 wr = wr->next;
2118         }
2119         spin_unlock_irqrestore(&qp->q_lock, flags);
2120         return status;
2121 }
2122
2123 static void ocrdma_ring_rq_db(struct ocrdma_qp *qp)
2124 {
2125         u32 val = qp->rq.dbid | (1 << ocrdma_get_num_posted_shift(qp));
2126
2127         if (qp->state != OCRDMA_QPS_INIT)
2128                 iowrite32(val, qp->rq_db);
2129         else
2130                 qp->db_cache++;
2131 }
2132
2133 static void ocrdma_build_rqe(struct ocrdma_hdr_wqe *rqe, struct ib_recv_wr *wr,
2134                              u16 tag)
2135 {
2136         u32 wqe_size = 0;
2137         struct ocrdma_sge *sge;
2138         if (wr->num_sge)
2139                 wqe_size = (wr->num_sge * sizeof(*sge)) + sizeof(*rqe);
2140         else
2141                 wqe_size = sizeof(*sge) + sizeof(*rqe);
2142
2143         rqe->cw = ((wqe_size / OCRDMA_WQE_STRIDE) <<
2144                                 OCRDMA_WQE_SIZE_SHIFT);
2145         rqe->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2146         rqe->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2147         rqe->total_len = 0;
2148         rqe->rsvd_tag = tag;
2149         sge = (struct ocrdma_sge *)(rqe + 1);
2150         ocrdma_build_sges(rqe, sge, wr->num_sge, wr->sg_list);
2151         ocrdma_cpu_to_le32(rqe, wqe_size);
2152 }
2153
2154 int ocrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2155                      struct ib_recv_wr **bad_wr)
2156 {
2157         int status = 0;
2158         unsigned long flags;
2159         struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2160         struct ocrdma_hdr_wqe *rqe;
2161
2162         spin_lock_irqsave(&qp->q_lock, flags);
2163         if (qp->state == OCRDMA_QPS_RST || qp->state == OCRDMA_QPS_ERR) {
2164                 spin_unlock_irqrestore(&qp->q_lock, flags);
2165                 *bad_wr = wr;
2166                 return -EINVAL;
2167         }
2168         while (wr) {
2169                 if (ocrdma_hwq_free_cnt(&qp->rq) == 0 ||
2170                     wr->num_sge > qp->rq.max_sges) {
2171                         *bad_wr = wr;
2172                         status = -ENOMEM;
2173                         break;
2174                 }
2175                 rqe = ocrdma_hwq_head(&qp->rq);
2176                 ocrdma_build_rqe(rqe, wr, 0);
2177
2178                 qp->rqe_wr_id_tbl[qp->rq.head] = wr->wr_id;
2179                 /* make sure rqe is written before adapter can access it */
2180                 wmb();
2181
2182                 /* inform hw to start processing it */
2183                 ocrdma_ring_rq_db(qp);
2184
2185                 /* update pointer, counter for next wr */
2186                 ocrdma_hwq_inc_head(&qp->rq);
2187                 wr = wr->next;
2188         }
2189         spin_unlock_irqrestore(&qp->q_lock, flags);
2190         return status;
2191 }
2192
2193 /* cqe for srq's rqe can potentially arrive out of order.
2194  * index gives the entry in the shadow table where to store
2195  * the wr_id. tag/index is returned in cqe to reference back
2196  * for a given rqe.
2197  */
2198 static int ocrdma_srq_get_idx(struct ocrdma_srq *srq)
2199 {
2200         int row = 0;
2201         int indx = 0;
2202
2203         for (row = 0; row < srq->bit_fields_len; row++) {
2204                 if (srq->idx_bit_fields[row]) {
2205                         indx = ffs(srq->idx_bit_fields[row]);
2206                         indx = (row * 32) + (indx - 1);
2207                         if (indx >= srq->rq.max_cnt)
2208                                 BUG();
2209                         ocrdma_srq_toggle_bit(srq, indx);
2210                         break;
2211                 }
2212         }
2213
2214         if (row == srq->bit_fields_len)
2215                 BUG();
2216         return indx;
2217 }
2218
2219 static void ocrdma_ring_srq_db(struct ocrdma_srq *srq)
2220 {
2221         u32 val = srq->rq.dbid | (1 << 16);
2222
2223         iowrite32(val, srq->db + OCRDMA_DB_GEN2_SRQ_OFFSET);
2224 }
2225
2226 int ocrdma_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
2227                          struct ib_recv_wr **bad_wr)
2228 {
2229         int status = 0;
2230         unsigned long flags;
2231         struct ocrdma_srq *srq;
2232         struct ocrdma_hdr_wqe *rqe;
2233         u16 tag;
2234
2235         srq = get_ocrdma_srq(ibsrq);
2236
2237         spin_lock_irqsave(&srq->q_lock, flags);
2238         while (wr) {
2239                 if (ocrdma_hwq_free_cnt(&srq->rq) == 0 ||
2240                     wr->num_sge > srq->rq.max_sges) {
2241                         status = -ENOMEM;
2242                         *bad_wr = wr;
2243                         break;
2244                 }
2245                 tag = ocrdma_srq_get_idx(srq);
2246                 rqe = ocrdma_hwq_head(&srq->rq);
2247                 ocrdma_build_rqe(rqe, wr, tag);
2248
2249                 srq->rqe_wr_id_tbl[tag] = wr->wr_id;
2250                 /* make sure rqe is written before adapter can perform DMA */
2251                 wmb();
2252                 /* inform hw to start processing it */
2253                 ocrdma_ring_srq_db(srq);
2254                 /* update pointer, counter for next wr */
2255                 ocrdma_hwq_inc_head(&srq->rq);
2256                 wr = wr->next;
2257         }
2258         spin_unlock_irqrestore(&srq->q_lock, flags);
2259         return status;
2260 }
2261
2262 static enum ib_wc_status ocrdma_to_ibwc_err(u16 status)
2263 {
2264         enum ib_wc_status ibwc_status;
2265
2266         switch (status) {
2267         case OCRDMA_CQE_GENERAL_ERR:
2268                 ibwc_status = IB_WC_GENERAL_ERR;
2269                 break;
2270         case OCRDMA_CQE_LOC_LEN_ERR:
2271                 ibwc_status = IB_WC_LOC_LEN_ERR;
2272                 break;
2273         case OCRDMA_CQE_LOC_QP_OP_ERR:
2274                 ibwc_status = IB_WC_LOC_QP_OP_ERR;
2275                 break;
2276         case OCRDMA_CQE_LOC_EEC_OP_ERR:
2277                 ibwc_status = IB_WC_LOC_EEC_OP_ERR;
2278                 break;
2279         case OCRDMA_CQE_LOC_PROT_ERR:
2280                 ibwc_status = IB_WC_LOC_PROT_ERR;
2281                 break;
2282         case OCRDMA_CQE_WR_FLUSH_ERR:
2283                 ibwc_status = IB_WC_WR_FLUSH_ERR;
2284                 break;
2285         case OCRDMA_CQE_MW_BIND_ERR:
2286                 ibwc_status = IB_WC_MW_BIND_ERR;
2287                 break;
2288         case OCRDMA_CQE_BAD_RESP_ERR:
2289                 ibwc_status = IB_WC_BAD_RESP_ERR;
2290                 break;
2291         case OCRDMA_CQE_LOC_ACCESS_ERR:
2292                 ibwc_status = IB_WC_LOC_ACCESS_ERR;
2293                 break;
2294         case OCRDMA_CQE_REM_INV_REQ_ERR:
2295                 ibwc_status = IB_WC_REM_INV_REQ_ERR;
2296                 break;
2297         case OCRDMA_CQE_REM_ACCESS_ERR:
2298                 ibwc_status = IB_WC_REM_ACCESS_ERR;
2299                 break;
2300         case OCRDMA_CQE_REM_OP_ERR:
2301                 ibwc_status = IB_WC_REM_OP_ERR;
2302                 break;
2303         case OCRDMA_CQE_RETRY_EXC_ERR:
2304                 ibwc_status = IB_WC_RETRY_EXC_ERR;
2305                 break;
2306         case OCRDMA_CQE_RNR_RETRY_EXC_ERR:
2307                 ibwc_status = IB_WC_RNR_RETRY_EXC_ERR;
2308                 break;
2309         case OCRDMA_CQE_LOC_RDD_VIOL_ERR:
2310                 ibwc_status = IB_WC_LOC_RDD_VIOL_ERR;
2311                 break;
2312         case OCRDMA_CQE_REM_INV_RD_REQ_ERR:
2313                 ibwc_status = IB_WC_REM_INV_RD_REQ_ERR;
2314                 break;
2315         case OCRDMA_CQE_REM_ABORT_ERR:
2316                 ibwc_status = IB_WC_REM_ABORT_ERR;
2317                 break;
2318         case OCRDMA_CQE_INV_EECN_ERR:
2319                 ibwc_status = IB_WC_INV_EECN_ERR;
2320                 break;
2321         case OCRDMA_CQE_INV_EEC_STATE_ERR:
2322                 ibwc_status = IB_WC_INV_EEC_STATE_ERR;
2323                 break;
2324         case OCRDMA_CQE_FATAL_ERR:
2325                 ibwc_status = IB_WC_FATAL_ERR;
2326                 break;
2327         case OCRDMA_CQE_RESP_TIMEOUT_ERR:
2328                 ibwc_status = IB_WC_RESP_TIMEOUT_ERR;
2329                 break;
2330         default:
2331                 ibwc_status = IB_WC_GENERAL_ERR;
2332                 break;
2333         }
2334         return ibwc_status;
2335 }
2336
2337 static void ocrdma_update_wc(struct ocrdma_qp *qp, struct ib_wc *ibwc,
2338                       u32 wqe_idx)
2339 {
2340         struct ocrdma_hdr_wqe *hdr;
2341         struct ocrdma_sge *rw;
2342         int opcode;
2343
2344         hdr = ocrdma_hwq_head_from_idx(&qp->sq, wqe_idx);
2345
2346         ibwc->wr_id = qp->wqe_wr_id_tbl[wqe_idx].wrid;
2347         /* Undo the hdr->cw swap */
2348         opcode = le32_to_cpu(hdr->cw) & OCRDMA_WQE_OPCODE_MASK;
2349         switch (opcode) {
2350         case OCRDMA_WRITE:
2351                 ibwc->opcode = IB_WC_RDMA_WRITE;
2352                 break;
2353         case OCRDMA_READ:
2354                 rw = (struct ocrdma_sge *)(hdr + 1);
2355                 ibwc->opcode = IB_WC_RDMA_READ;
2356                 ibwc->byte_len = rw->len;
2357                 break;
2358         case OCRDMA_SEND:
2359                 ibwc->opcode = IB_WC_SEND;
2360                 break;
2361         case OCRDMA_FR_MR:
2362                 ibwc->opcode = IB_WC_FAST_REG_MR;
2363                 break;
2364         case OCRDMA_LKEY_INV:
2365                 ibwc->opcode = IB_WC_LOCAL_INV;
2366                 break;
2367         default:
2368                 ibwc->status = IB_WC_GENERAL_ERR;
2369                 pr_err("%s() invalid opcode received = 0x%x\n",
2370                        __func__, hdr->cw & OCRDMA_WQE_OPCODE_MASK);
2371                 break;
2372         }
2373 }
2374
2375 static void ocrdma_set_cqe_status_flushed(struct ocrdma_qp *qp,
2376                                                 struct ocrdma_cqe *cqe)
2377 {
2378         if (is_cqe_for_sq(cqe)) {
2379                 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2380                                 cqe->flags_status_srcqpn) &
2381                                         ~OCRDMA_CQE_STATUS_MASK);
2382                 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2383                                 cqe->flags_status_srcqpn) |
2384                                 (OCRDMA_CQE_WR_FLUSH_ERR <<
2385                                         OCRDMA_CQE_STATUS_SHIFT));
2386         } else {
2387                 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2388                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2389                                         cqe->flags_status_srcqpn) &
2390                                                 ~OCRDMA_CQE_UD_STATUS_MASK);
2391                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2392                                         cqe->flags_status_srcqpn) |
2393                                         (OCRDMA_CQE_WR_FLUSH_ERR <<
2394                                                 OCRDMA_CQE_UD_STATUS_SHIFT));
2395                 } else {
2396                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2397                                         cqe->flags_status_srcqpn) &
2398                                                 ~OCRDMA_CQE_STATUS_MASK);
2399                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2400                                         cqe->flags_status_srcqpn) |
2401                                         (OCRDMA_CQE_WR_FLUSH_ERR <<
2402                                                 OCRDMA_CQE_STATUS_SHIFT));
2403                 }
2404         }
2405 }
2406
2407 static bool ocrdma_update_err_cqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2408                                   struct ocrdma_qp *qp, int status)
2409 {
2410         bool expand = false;
2411
2412         ibwc->byte_len = 0;
2413         ibwc->qp = &qp->ibqp;
2414         ibwc->status = ocrdma_to_ibwc_err(status);
2415
2416         ocrdma_flush_qp(qp);
2417         ocrdma_qp_state_change(qp, IB_QPS_ERR, NULL);
2418
2419         /* if wqe/rqe pending for which cqe needs to be returned,
2420          * trigger inflating it.
2421          */
2422         if (!is_hw_rq_empty(qp) || !is_hw_sq_empty(qp)) {
2423                 expand = true;
2424                 ocrdma_set_cqe_status_flushed(qp, cqe);
2425         }
2426         return expand;
2427 }
2428
2429 static int ocrdma_update_err_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2430                                   struct ocrdma_qp *qp, int status)
2431 {
2432         ibwc->opcode = IB_WC_RECV;
2433         ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2434         ocrdma_hwq_inc_tail(&qp->rq);
2435
2436         return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2437 }
2438
2439 static int ocrdma_update_err_scqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2440                                   struct ocrdma_qp *qp, int status)
2441 {
2442         ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2443         ocrdma_hwq_inc_tail(&qp->sq);
2444
2445         return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2446 }
2447
2448
2449 static bool ocrdma_poll_err_scqe(struct ocrdma_qp *qp,
2450                                  struct ocrdma_cqe *cqe, struct ib_wc *ibwc,
2451                                  bool *polled, bool *stop)
2452 {
2453         bool expand;
2454         int status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2455                 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2456
2457         /* when hw sq is empty, but rq is not empty, so we continue
2458          * to keep the cqe in order to get the cq event again.
2459          */
2460         if (is_hw_sq_empty(qp) && !is_hw_rq_empty(qp)) {
2461                 /* when cq for rq and sq is same, it is safe to return
2462                  * flush cqe for RQEs.
2463                  */
2464                 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2465                         *polled = true;
2466                         status = OCRDMA_CQE_WR_FLUSH_ERR;
2467                         expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2468                 } else {
2469                         /* stop processing further cqe as this cqe is used for
2470                          * triggering cq event on buddy cq of RQ.
2471                          * When QP is destroyed, this cqe will be removed
2472                          * from the cq's hardware q.
2473                          */
2474                         *polled = false;
2475                         *stop = true;
2476                         expand = false;
2477                 }
2478         } else {
2479                 *polled = true;
2480                 expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2481         }
2482         return expand;
2483 }
2484
2485 static bool ocrdma_poll_success_scqe(struct ocrdma_qp *qp,
2486                                      struct ocrdma_cqe *cqe,
2487                                      struct ib_wc *ibwc, bool *polled)
2488 {
2489         bool expand = false;
2490         int tail = qp->sq.tail;
2491         u32 wqe_idx;
2492
2493         if (!qp->wqe_wr_id_tbl[tail].signaled) {
2494                 *polled = false;    /* WC cannot be consumed yet */
2495         } else {
2496                 ibwc->status = IB_WC_SUCCESS;
2497                 ibwc->wc_flags = 0;
2498                 ibwc->qp = &qp->ibqp;
2499                 ocrdma_update_wc(qp, ibwc, tail);
2500                 *polled = true;
2501         }
2502         wqe_idx = (le32_to_cpu(cqe->wq.wqeidx) &
2503                         OCRDMA_CQE_WQEIDX_MASK) & qp->sq.max_wqe_idx;
2504         if (tail != wqe_idx)
2505                 expand = true; /* Coalesced CQE can't be consumed yet */
2506
2507         ocrdma_hwq_inc_tail(&qp->sq);
2508         return expand;
2509 }
2510
2511 static bool ocrdma_poll_scqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2512                              struct ib_wc *ibwc, bool *polled, bool *stop)
2513 {
2514         int status;
2515         bool expand;
2516
2517         status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2518                 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2519
2520         if (status == OCRDMA_CQE_SUCCESS)
2521                 expand = ocrdma_poll_success_scqe(qp, cqe, ibwc, polled);
2522         else
2523                 expand = ocrdma_poll_err_scqe(qp, cqe, ibwc, polled, stop);
2524         return expand;
2525 }
2526
2527 static int ocrdma_update_ud_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe)
2528 {
2529         int status;
2530
2531         status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2532                 OCRDMA_CQE_UD_STATUS_MASK) >> OCRDMA_CQE_UD_STATUS_SHIFT;
2533         ibwc->src_qp = le32_to_cpu(cqe->flags_status_srcqpn) &
2534                                                 OCRDMA_CQE_SRCQP_MASK;
2535         ibwc->pkey_index = le32_to_cpu(cqe->ud.rxlen_pkey) &
2536                                                 OCRDMA_CQE_PKEY_MASK;
2537         ibwc->wc_flags = IB_WC_GRH;
2538         ibwc->byte_len = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2539                                         OCRDMA_CQE_UD_XFER_LEN_SHIFT);
2540         return status;
2541 }
2542
2543 static void ocrdma_update_free_srq_cqe(struct ib_wc *ibwc,
2544                                        struct ocrdma_cqe *cqe,
2545                                        struct ocrdma_qp *qp)
2546 {
2547         unsigned long flags;
2548         struct ocrdma_srq *srq;
2549         u32 wqe_idx;
2550
2551         srq = get_ocrdma_srq(qp->ibqp.srq);
2552         wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
2553                         OCRDMA_CQE_BUFTAG_SHIFT) & srq->rq.max_wqe_idx;
2554         ibwc->wr_id = srq->rqe_wr_id_tbl[wqe_idx];
2555         spin_lock_irqsave(&srq->q_lock, flags);
2556         ocrdma_srq_toggle_bit(srq, wqe_idx);
2557         spin_unlock_irqrestore(&srq->q_lock, flags);
2558         ocrdma_hwq_inc_tail(&srq->rq);
2559 }
2560
2561 static bool ocrdma_poll_err_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2562                                 struct ib_wc *ibwc, bool *polled, bool *stop,
2563                                 int status)
2564 {
2565         bool expand;
2566
2567         /* when hw_rq is empty, but wq is not empty, so continue
2568          * to keep the cqe to get the cq event again.
2569          */
2570         if (is_hw_rq_empty(qp) && !is_hw_sq_empty(qp)) {
2571                 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2572                         *polled = true;
2573                         status = OCRDMA_CQE_WR_FLUSH_ERR;
2574                         expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2575                 } else {
2576                         *polled = false;
2577                         *stop = true;
2578                         expand = false;
2579                 }
2580         } else {
2581                 *polled = true;
2582                 expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2583         }
2584         return expand;
2585 }
2586
2587 static void ocrdma_poll_success_rcqe(struct ocrdma_qp *qp,
2588                                      struct ocrdma_cqe *cqe, struct ib_wc *ibwc)
2589 {
2590         ibwc->opcode = IB_WC_RECV;
2591         ibwc->qp = &qp->ibqp;
2592         ibwc->status = IB_WC_SUCCESS;
2593
2594         if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI)
2595                 ocrdma_update_ud_rcqe(ibwc, cqe);
2596         else
2597                 ibwc->byte_len = le32_to_cpu(cqe->rq.rxlen);
2598
2599         if (is_cqe_imm(cqe)) {
2600                 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2601                 ibwc->wc_flags |= IB_WC_WITH_IMM;
2602         } else if (is_cqe_wr_imm(cqe)) {
2603                 ibwc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2604                 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2605                 ibwc->wc_flags |= IB_WC_WITH_IMM;
2606         } else if (is_cqe_invalidated(cqe)) {
2607                 ibwc->ex.invalidate_rkey = le32_to_cpu(cqe->rq.lkey_immdt);
2608                 ibwc->wc_flags |= IB_WC_WITH_INVALIDATE;
2609         }
2610         if (qp->ibqp.srq) {
2611                 ocrdma_update_free_srq_cqe(ibwc, cqe, qp);
2612         } else {
2613                 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2614                 ocrdma_hwq_inc_tail(&qp->rq);
2615         }
2616 }
2617
2618 static bool ocrdma_poll_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2619                              struct ib_wc *ibwc, bool *polled, bool *stop)
2620 {
2621         int status;
2622         bool expand = false;
2623
2624         ibwc->wc_flags = 0;
2625         if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2626                 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2627                                         OCRDMA_CQE_UD_STATUS_MASK) >>
2628                                         OCRDMA_CQE_UD_STATUS_SHIFT;
2629         } else {
2630                 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2631                              OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2632         }
2633
2634         if (status == OCRDMA_CQE_SUCCESS) {
2635                 *polled = true;
2636                 ocrdma_poll_success_rcqe(qp, cqe, ibwc);
2637         } else {
2638                 expand = ocrdma_poll_err_rcqe(qp, cqe, ibwc, polled, stop,
2639                                               status);
2640         }
2641         return expand;
2642 }
2643
2644 static void ocrdma_change_cq_phase(struct ocrdma_cq *cq, struct ocrdma_cqe *cqe,
2645                                    u16 cur_getp)
2646 {
2647         if (cq->phase_change) {
2648                 if (cur_getp == 0)
2649                         cq->phase = (~cq->phase & OCRDMA_CQE_VALID);
2650         } else {
2651                 /* clear valid bit */
2652                 cqe->flags_status_srcqpn = 0;
2653         }
2654 }
2655
2656 static int ocrdma_poll_hwcq(struct ocrdma_cq *cq, int num_entries,
2657                             struct ib_wc *ibwc)
2658 {
2659         u16 qpn = 0;
2660         int i = 0;
2661         bool expand = false;
2662         int polled_hw_cqes = 0;
2663         struct ocrdma_qp *qp = NULL;
2664         struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
2665         struct ocrdma_cqe *cqe;
2666         u16 cur_getp; bool polled = false; bool stop = false;
2667
2668         cur_getp = cq->getp;
2669         while (num_entries) {
2670                 cqe = cq->va + cur_getp;
2671                 /* check whether valid cqe or not */
2672                 if (!is_cqe_valid(cq, cqe))
2673                         break;
2674                 qpn = (le32_to_cpu(cqe->cmn.qpn) & OCRDMA_CQE_QPN_MASK);
2675                 /* ignore discarded cqe */
2676                 if (qpn == 0)
2677                         goto skip_cqe;
2678                 qp = dev->qp_tbl[qpn];
2679                 BUG_ON(qp == NULL);
2680
2681                 if (is_cqe_for_sq(cqe)) {
2682                         expand = ocrdma_poll_scqe(qp, cqe, ibwc, &polled,
2683                                                   &stop);
2684                 } else {
2685                         expand = ocrdma_poll_rcqe(qp, cqe, ibwc, &polled,
2686                                                   &stop);
2687                 }
2688                 if (expand)
2689                         goto expand_cqe;
2690                 if (stop)
2691                         goto stop_cqe;
2692                 /* clear qpn to avoid duplicate processing by discard_cqe() */
2693                 cqe->cmn.qpn = 0;
2694 skip_cqe:
2695                 polled_hw_cqes += 1;
2696                 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
2697                 ocrdma_change_cq_phase(cq, cqe, cur_getp);
2698 expand_cqe:
2699                 if (polled) {
2700                         num_entries -= 1;
2701                         i += 1;
2702                         ibwc = ibwc + 1;
2703                         polled = false;
2704                 }
2705         }
2706 stop_cqe:
2707         cq->getp = cur_getp;
2708         if (polled_hw_cqes || expand || stop) {
2709                 ocrdma_ring_cq_db(dev, cq->id, cq->armed, cq->solicited,
2710                                   polled_hw_cqes);
2711         }
2712         return i;
2713 }
2714
2715 /* insert error cqe if the QP's SQ or RQ's CQ matches the CQ under poll. */
2716 static int ocrdma_add_err_cqe(struct ocrdma_cq *cq, int num_entries,
2717                               struct ocrdma_qp *qp, struct ib_wc *ibwc)
2718 {
2719         int err_cqes = 0;
2720
2721         while (num_entries) {
2722                 if (is_hw_sq_empty(qp) && is_hw_rq_empty(qp))
2723                         break;
2724                 if (!is_hw_sq_empty(qp) && qp->sq_cq == cq) {
2725                         ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2726                         ocrdma_hwq_inc_tail(&qp->sq);
2727                 } else if (!is_hw_rq_empty(qp) && qp->rq_cq == cq) {
2728                         ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2729                         ocrdma_hwq_inc_tail(&qp->rq);
2730                 } else {
2731                         return err_cqes;
2732                 }
2733                 ibwc->byte_len = 0;
2734                 ibwc->status = IB_WC_WR_FLUSH_ERR;
2735                 ibwc = ibwc + 1;
2736                 err_cqes += 1;
2737                 num_entries -= 1;
2738         }
2739         return err_cqes;
2740 }
2741
2742 int ocrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
2743 {
2744         int cqes_to_poll = num_entries;
2745         struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2746         struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2747         int num_os_cqe = 0, err_cqes = 0;
2748         struct ocrdma_qp *qp;
2749         unsigned long flags;
2750
2751         /* poll cqes from adapter CQ */
2752         spin_lock_irqsave(&cq->cq_lock, flags);
2753         num_os_cqe = ocrdma_poll_hwcq(cq, cqes_to_poll, wc);
2754         spin_unlock_irqrestore(&cq->cq_lock, flags);
2755         cqes_to_poll -= num_os_cqe;
2756
2757         if (cqes_to_poll) {
2758                 wc = wc + num_os_cqe;
2759                 /* adapter returns single error cqe when qp moves to
2760                  * error state. So insert error cqes with wc_status as
2761                  * FLUSHED for pending WQEs and RQEs of QP's SQ and RQ
2762                  * respectively which uses this CQ.
2763                  */
2764                 spin_lock_irqsave(&dev->flush_q_lock, flags);
2765                 list_for_each_entry(qp, &cq->sq_head, sq_entry) {
2766                         if (cqes_to_poll == 0)
2767                                 break;
2768                         err_cqes = ocrdma_add_err_cqe(cq, cqes_to_poll, qp, wc);
2769                         cqes_to_poll -= err_cqes;
2770                         num_os_cqe += err_cqes;
2771                         wc = wc + err_cqes;
2772                 }
2773                 spin_unlock_irqrestore(&dev->flush_q_lock, flags);
2774         }
2775         return num_os_cqe;
2776 }
2777
2778 int ocrdma_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags cq_flags)
2779 {
2780         struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2781         struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2782         u16 cq_id;
2783         u16 cur_getp;
2784         struct ocrdma_cqe *cqe;
2785         unsigned long flags;
2786
2787         cq_id = cq->id;
2788
2789         spin_lock_irqsave(&cq->cq_lock, flags);
2790         if (cq_flags & IB_CQ_NEXT_COMP || cq_flags & IB_CQ_SOLICITED)
2791                 cq->armed = true;
2792         if (cq_flags & IB_CQ_SOLICITED)
2793                 cq->solicited = true;
2794
2795         cur_getp = cq->getp;
2796         cqe = cq->va + cur_getp;
2797
2798         /* check whether any valid cqe exist or not, if not then safe to
2799          * arm. If cqe is not yet consumed, then let it get consumed and then
2800          * we arm it to avoid false interrupts.
2801          */
2802         if (!is_cqe_valid(cq, cqe) || cq->arm_needed) {
2803                 cq->arm_needed = false;
2804                 ocrdma_ring_cq_db(dev, cq_id, cq->armed, cq->solicited, 0);
2805         }
2806         spin_unlock_irqrestore(&cq->cq_lock, flags);
2807         return 0;
2808 }
2809
2810 struct ib_mr *ocrdma_alloc_frmr(struct ib_pd *ibpd, int max_page_list_len)
2811 {
2812         int status;
2813         struct ocrdma_mr *mr;
2814         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
2815         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
2816
2817         if (max_page_list_len > dev->attr.max_pages_per_frmr)
2818                 return ERR_PTR(-EINVAL);
2819
2820         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2821         if (!mr)
2822                 return ERR_PTR(-ENOMEM);
2823
2824         status = ocrdma_get_pbl_info(dev, mr, max_page_list_len);
2825         if (status)
2826                 goto pbl_err;
2827         mr->hwmr.fr_mr = 1;
2828         mr->hwmr.remote_rd = 0;
2829         mr->hwmr.remote_wr = 0;
2830         mr->hwmr.local_rd = 0;
2831         mr->hwmr.local_wr = 0;
2832         mr->hwmr.mw_bind = 0;
2833         status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
2834         if (status)
2835                 goto pbl_err;
2836         status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, 0);
2837         if (status)
2838                 goto mbx_err;
2839         mr->ibmr.rkey = mr->hwmr.lkey;
2840         mr->ibmr.lkey = mr->hwmr.lkey;
2841         dev->stag_arr[(mr->hwmr.lkey >> 8) & (OCRDMA_MAX_STAG - 1)] = mr;
2842         return &mr->ibmr;
2843 mbx_err:
2844         ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
2845 pbl_err:
2846         kfree(mr);
2847         return ERR_PTR(-ENOMEM);
2848 }
2849
2850 struct ib_fast_reg_page_list *ocrdma_alloc_frmr_page_list(struct ib_device
2851                                                           *ibdev,
2852                                                           int page_list_len)
2853 {
2854         struct ib_fast_reg_page_list *frmr_list;
2855         int size;
2856
2857         size = sizeof(*frmr_list) + (page_list_len * sizeof(u64));
2858         frmr_list = kzalloc(size, GFP_KERNEL);
2859         if (!frmr_list)
2860                 return ERR_PTR(-ENOMEM);
2861         frmr_list->page_list = (u64 *)(frmr_list + 1);
2862         return frmr_list;
2863 }
2864
2865 void ocrdma_free_frmr_page_list(struct ib_fast_reg_page_list *page_list)
2866 {
2867         kfree(page_list);
2868 }
2869
2870 #define MAX_KERNEL_PBE_SIZE 65536
2871 static inline int count_kernel_pbes(struct ib_phys_buf *buf_list,
2872                                     int buf_cnt, u32 *pbe_size)
2873 {
2874         u64 total_size = 0;
2875         u64 buf_size = 0;
2876         int i;
2877         *pbe_size = roundup(buf_list[0].size, PAGE_SIZE);
2878         *pbe_size = roundup_pow_of_two(*pbe_size);
2879
2880         /* find the smallest PBE size that we can have */
2881         for (i = 0; i < buf_cnt; i++) {
2882                 /* first addr may not be page aligned, so ignore checking */
2883                 if ((i != 0) && ((buf_list[i].addr & ~PAGE_MASK) ||
2884                                  (buf_list[i].size & ~PAGE_MASK))) {
2885                         return 0;
2886                 }
2887
2888                 /* if configured PBE size is greater then the chosen one,
2889                  * reduce the PBE size.
2890                  */
2891                 buf_size = roundup(buf_list[i].size, PAGE_SIZE);
2892                 /* pbe_size has to be even multiple of 4K 1,2,4,8...*/
2893                 buf_size = roundup_pow_of_two(buf_size);
2894                 if (*pbe_size > buf_size)
2895                         *pbe_size = buf_size;
2896
2897                 total_size += buf_size;
2898         }
2899         *pbe_size = *pbe_size > MAX_KERNEL_PBE_SIZE ?
2900             (MAX_KERNEL_PBE_SIZE) : (*pbe_size);
2901
2902         /* num_pbes = total_size / (*pbe_size);  this is implemented below. */
2903
2904         return total_size >> ilog2(*pbe_size);
2905 }
2906
2907 static void build_kernel_pbes(struct ib_phys_buf *buf_list, int ib_buf_cnt,
2908                               u32 pbe_size, struct ocrdma_pbl *pbl_tbl,
2909                               struct ocrdma_hw_mr *hwmr)
2910 {
2911         int i;
2912         int idx;
2913         int pbes_per_buf = 0;
2914         u64 buf_addr = 0;
2915         int num_pbes;
2916         struct ocrdma_pbe *pbe;
2917         int total_num_pbes = 0;
2918
2919         if (!hwmr->num_pbes)
2920                 return;
2921
2922         pbe = (struct ocrdma_pbe *)pbl_tbl->va;
2923         num_pbes = 0;
2924
2925         /* go through the OS phy regions & fill hw pbe entries into pbls. */
2926         for (i = 0; i < ib_buf_cnt; i++) {
2927                 buf_addr = buf_list[i].addr;
2928                 pbes_per_buf =
2929                     roundup_pow_of_two(roundup(buf_list[i].size, PAGE_SIZE)) /
2930                     pbe_size;
2931                 hwmr->len += buf_list[i].size;
2932                 /* number of pbes can be more for one OS buf, when
2933                  * buffers are of different sizes.
2934                  * split the ib_buf to one or more pbes.
2935                  */
2936                 for (idx = 0; idx < pbes_per_buf; idx++) {
2937                         /* we program always page aligned addresses,
2938                          * first unaligned address is taken care by fbo.
2939                          */
2940                         if (i == 0) {
2941                                 /* for non zero fbo, assign the
2942                                  * start of the page.
2943                                  */
2944                                 pbe->pa_lo =
2945                                     cpu_to_le32((u32) (buf_addr & PAGE_MASK));
2946                                 pbe->pa_hi =
2947                                     cpu_to_le32((u32) upper_32_bits(buf_addr));
2948                         } else {
2949                                 pbe->pa_lo =
2950                                     cpu_to_le32((u32) (buf_addr & 0xffffffff));
2951                                 pbe->pa_hi =
2952                                     cpu_to_le32((u32) upper_32_bits(buf_addr));
2953                         }
2954                         buf_addr += pbe_size;
2955                         num_pbes += 1;
2956                         total_num_pbes += 1;
2957                         pbe++;
2958
2959                         if (total_num_pbes == hwmr->num_pbes)
2960                                 goto mr_tbl_done;
2961                         /* if the pbl is full storing the pbes,
2962                          * move to next pbl.
2963                          */
2964                         if (num_pbes == (hwmr->pbl_size/sizeof(u64))) {
2965                                 pbl_tbl++;
2966                                 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
2967                                 num_pbes = 0;
2968                         }
2969                 }
2970         }
2971 mr_tbl_done:
2972         return;
2973 }
2974
2975 struct ib_mr *ocrdma_reg_kernel_mr(struct ib_pd *ibpd,
2976                                    struct ib_phys_buf *buf_list,
2977                                    int buf_cnt, int acc, u64 *iova_start)
2978 {
2979         int status = -ENOMEM;
2980         struct ocrdma_mr *mr;
2981         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
2982         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
2983         u32 num_pbes;
2984         u32 pbe_size = 0;
2985
2986         if ((acc & IB_ACCESS_REMOTE_WRITE) && !(acc & IB_ACCESS_LOCAL_WRITE))
2987                 return ERR_PTR(-EINVAL);
2988
2989         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2990         if (!mr)
2991                 return ERR_PTR(status);
2992
2993         num_pbes = count_kernel_pbes(buf_list, buf_cnt, &pbe_size);
2994         if (num_pbes == 0) {
2995                 status = -EINVAL;
2996                 goto pbl_err;
2997         }
2998         status = ocrdma_get_pbl_info(dev, mr, num_pbes);
2999         if (status)
3000                 goto pbl_err;
3001
3002         mr->hwmr.pbe_size = pbe_size;
3003         mr->hwmr.fbo = *iova_start - (buf_list[0].addr & PAGE_MASK);
3004         mr->hwmr.va = *iova_start;
3005         mr->hwmr.local_rd = 1;
3006         mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
3007         mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
3008         mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
3009         mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
3010         mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
3011
3012         status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
3013         if (status)
3014                 goto pbl_err;
3015         build_kernel_pbes(buf_list, buf_cnt, pbe_size, mr->hwmr.pbl_table,
3016                           &mr->hwmr);
3017         status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
3018         if (status)
3019                 goto mbx_err;
3020
3021         mr->ibmr.lkey = mr->hwmr.lkey;
3022         if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
3023                 mr->ibmr.rkey = mr->hwmr.lkey;
3024         return &mr->ibmr;
3025
3026 mbx_err:
3027         ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
3028 pbl_err:
3029         kfree(mr);
3030         return ERR_PTR(status);
3031 }