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