drivers/infiniband/hw/ocrdma/ocrdma_verbs.c

   1 /*******************************************************************
   2  * This file is part of the Emulex RoCE Device Driver for          *
   3  * RoCE (RDMA over Converged Ethernet) adapters.                   *
   4  * Copyright (C) 2008-2012 Emulex. All rights reserved.            *
   5  * EMULEX and SLI are trademarks of Emulex.                        *
   6  * www.emulex.com                                                  *
   7  *                                                                 *
   8  * This program is free software; you can redistribute it and/or   *
   9  * modify it under the terms of version 2 of the GNU General       *
  10  * Public License as published by the Free Software Foundation.    *
  11  * This program is distributed in the hope that it will be useful. *
  12  * ALL EXPRESS OR IMPLIED CONDITIONS, REPRESENTATIONS AND          *
  13  * WARRANTIES, INCLUDING ANY IMPLIED WARRANTY OF MERCHANTABILITY,  *
  14  * FITNESS FOR A PARTICULAR PURPOSE, OR NON-INFRINGEMENT, ARE      *
  15  * DISCLAIMED, EXCEPT TO THE EXTENT THAT SUCH DISCLAIMERS ARE HELD *
  16  * TO BE LEGALLY INVALID.  See the GNU General Public License for  *
  17  * more details, a copy of which can be found in the file COPYING  *
  18  * included with this package.                                     *
  19  *
  20  * Contact Information:
  21  * linux-drivers@emulex.com
  22  *
  23  * Emulex
  24  * 3333 Susan Street
  25  * Costa Mesa, CA 92626
  26  *******************************************************************/
  27
  28 #include <linux/dma-mapping.h>
  29 #include <rdma/ib_verbs.h>
  30 #include <rdma/ib_user_verbs.h>
  31 #include <rdma/iw_cm.h>
  32 #include <rdma/ib_umem.h>
  33 #include <rdma/ib_addr.h>
  34
  35 #include "ocrdma.h"
  36 #include "ocrdma_hw.h"
  37 #include "ocrdma_verbs.h"
  38 #include "ocrdma_abi.h"
  39
  40 int ocrdma_query_pkey(struct ib_device *ibdev, u8 port, u16 index, u16 *pkey)
  41 {
  42         if (index > 1)
  43                 return -EINVAL;
  44
  45         *pkey = 0xffff;
  46         return 0;
  47 }
  48
  49 int ocrdma_query_gid(struct ib_device *ibdev, u8 port,
  50                      int index, union ib_gid *sgid)
  51 {
  52         struct ocrdma_dev *dev;
  53
  54         dev = get_ocrdma_dev(ibdev);
  55         memset(sgid, 0, sizeof(*sgid));
  56         if (index >= OCRDMA_MAX_SGID)
  57                 return -EINVAL;
  58
  59         memcpy(sgid, &dev->sgid_tbl[index], sizeof(*sgid));
  60
  61         return 0;
  62 }
  63
  64 int ocrdma_query_device(struct ib_device *ibdev, struct ib_device_attr *attr)
  65 {
  66         struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
  67
  68         memset(attr, 0, sizeof *attr);
  69         memcpy(&attr->fw_ver, &dev->attr.fw_ver[0],
  70                min(sizeof(dev->attr.fw_ver), sizeof(attr->fw_ver)));
  71         ocrdma_get_guid(dev, (u8 *)&attr->sys_image_guid);
  72         attr->max_mr_size = ~0ull;
  73         attr->page_size_cap = 0xffff000;
  74         attr->vendor_id = dev->nic_info.pdev->vendor;
  75         attr->vendor_part_id = dev->nic_info.pdev->device;
  76         attr->hw_ver = 0;
  77         attr->max_qp = dev->attr.max_qp;
  78         attr->max_ah = dev->attr.max_qp;
  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         attr->max_sge = min(dev->attr.max_send_sge, dev->attr.max_srq_sge);
  87         attr->max_sge_rd = 0;
  88         attr->max_cq = dev->attr.max_cq;
  89         attr->max_cqe = dev->attr.max_cqe;
  90         attr->max_mr = dev->attr.max_mr;
  91         attr->max_mw = 0;
  92         attr->max_pd = dev->attr.max_pd;
  93         attr->atomic_cap = 0;
  94         attr->max_fmr = 0;
  95         attr->max_map_per_fmr = 0;
  96         attr->max_qp_rd_atom =
  97             min(dev->attr.max_ord_per_qp, dev->attr.max_ird_per_qp);
  98         attr->max_qp_init_rd_atom = dev->attr.max_ord_per_qp;
  99         attr->max_srq = (dev->attr.max_qp - 1);
 100         attr->max_srq_sge = attr->max_srq_sge;
 101         attr->max_srq_wr = dev->attr.max_rqe;
 102         attr->local_ca_ack_delay = dev->attr.local_ca_ack_delay;
 103         attr->max_fast_reg_page_list_len = 0;
 104         attr->max_pkeys = 1;
 105         return 0;
 106 }
 107
 108 int ocrdma_query_port(struct ib_device *ibdev,
 109                       u8 port, struct ib_port_attr *props)
 110 {
 111         enum ib_port_state port_state;
 112         struct ocrdma_dev *dev;
 113         struct net_device *netdev;
 114
 115         dev = get_ocrdma_dev(ibdev);
 116         if (port > 1) {
 117                 ocrdma_err("%s(%d) invalid_port=0x%x\n", __func__,
 118                            dev->id, port);
 119                 return -EINVAL;
 120         }
 121         netdev = dev->nic_info.netdev;
 122         if (netif_running(netdev) && netif_oper_up(netdev)) {
 123                 port_state = IB_PORT_ACTIVE;
 124                 props->phys_state = 5;
 125         } else {
 126                 port_state = IB_PORT_DOWN;
 127                 props->phys_state = 3;
 128         }
 129         props->max_mtu = IB_MTU_4096;
 130         props->active_mtu = iboe_get_mtu(netdev->mtu);
 131         props->lid = 0;
 132         props->lmc = 0;
 133         props->sm_lid = 0;
 134         props->sm_sl = 0;
 135         props->state = port_state;
 136         props->port_cap_flags =
 137             IB_PORT_CM_SUP |
 138             IB_PORT_REINIT_SUP |
 139             IB_PORT_DEVICE_MGMT_SUP | IB_PORT_VENDOR_CLASS_SUP;
 140         props->gid_tbl_len = OCRDMA_MAX_SGID;
 141         props->pkey_tbl_len = 1;
 142         props->bad_pkey_cntr = 0;
 143         props->qkey_viol_cntr = 0;
 144         props->active_width = IB_WIDTH_1X;
 145         props->active_speed = 4;
 146         props->max_msg_sz = 0x80000000;
 147         props->max_vl_num = 4;
 148         return 0;
 149 }
 150
 151 int ocrdma_modify_port(struct ib_device *ibdev, u8 port, int mask,
 152                        struct ib_port_modify *props)
 153 {
 154         struct ocrdma_dev *dev;
 155
 156         dev = get_ocrdma_dev(ibdev);
 157         if (port > 1) {
 158                 ocrdma_err("%s(%d) invalid_port=0x%x\n", __func__,
 159                            dev->id, port);
 160                 return -EINVAL;
 161         }
 162         return 0;
 163 }
 164
 165 static int ocrdma_add_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
 166                            unsigned long len)
 167 {
 168         struct ocrdma_mm *mm;
 169
 170         mm = kzalloc(sizeof(*mm), GFP_KERNEL);
 171         if (mm == NULL)
 172                 return -ENOMEM;
 173         mm->key.phy_addr = phy_addr;
 174         mm->key.len = len;
 175         INIT_LIST_HEAD(&mm->entry);
 176
 177         mutex_lock(&uctx->mm_list_lock);
 178         list_add_tail(&mm->entry, &uctx->mm_head);
 179         mutex_unlock(&uctx->mm_list_lock);
 180         return 0;
 181 }
 182
 183 static void ocrdma_del_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
 184                             unsigned long len)
 185 {
 186         struct ocrdma_mm *mm, *tmp;
 187
 188         mutex_lock(&uctx->mm_list_lock);
 189         list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
 190                 if (len != mm->key.len || phy_addr != mm->key.phy_addr)
 191                         continue;
 192
 193                 list_del(&mm->entry);
 194                 kfree(mm);
 195                 break;
 196         }
 197         mutex_unlock(&uctx->mm_list_lock);
 198 }
 199
 200 static bool ocrdma_search_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
 201                               unsigned long len)
 202 {
 203         bool found = false;
 204         struct ocrdma_mm *mm;
 205
 206         mutex_lock(&uctx->mm_list_lock);
 207         list_for_each_entry(mm, &uctx->mm_head, entry) {
 208                 if (len != mm->key.len || phy_addr != mm->key.phy_addr)
 209                         continue;
 210
 211                 found = true;
 212                 break;
 213         }
 214         mutex_unlock(&uctx->mm_list_lock);
 215         return found;
 216 }
 217
 218 struct ib_ucontext *ocrdma_alloc_ucontext(struct ib_device *ibdev,
 219                                           struct ib_udata *udata)
 220 {
 221         int status;
 222         struct ocrdma_ucontext *ctx;
 223         struct ocrdma_alloc_ucontext_resp resp;
 224         struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
 225         struct pci_dev *pdev = dev->nic_info.pdev;
 226         u32 map_len = roundup(sizeof(u32) * 2048, PAGE_SIZE);
 227
 228         if (!udata)
 229                 return ERR_PTR(-EFAULT);
 230         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
 231         if (!ctx)
 232                 return ERR_PTR(-ENOMEM);
 233         ctx->dev = dev;
 234         INIT_LIST_HEAD(&ctx->mm_head);
 235         mutex_init(&ctx->mm_list_lock);
 236
 237         ctx->ah_tbl.va = dma_alloc_coherent(&pdev->dev, map_len,
 238                                             &ctx->ah_tbl.pa, GFP_KERNEL);
 239         if (!ctx->ah_tbl.va) {
 240                 kfree(ctx);
 241                 return ERR_PTR(-ENOMEM);
 242         }
 243         memset(ctx->ah_tbl.va, 0, map_len);
 244         ctx->ah_tbl.len = map_len;
 245
 246         resp.ah_tbl_len = ctx->ah_tbl.len;
 247         resp.ah_tbl_page = ctx->ah_tbl.pa;
 248
 249         status = ocrdma_add_mmap(ctx, resp.ah_tbl_page, resp.ah_tbl_len);
 250         if (status)
 251                 goto map_err;
 252         resp.dev_id = dev->id;
 253         resp.max_inline_data = dev->attr.max_inline_data;
 254         resp.wqe_size = dev->attr.wqe_size;
 255         resp.rqe_size = dev->attr.rqe_size;
 256         resp.dpp_wqe_size = dev->attr.wqe_size;
 257         resp.rsvd = 0;
 258
 259         memcpy(resp.fw_ver, dev->attr.fw_ver, sizeof(resp.fw_ver));
 260         status = ib_copy_to_udata(udata, &resp, sizeof(resp));
 261         if (status)
 262                 goto cpy_err;
 263         return &ctx->ibucontext;
 264
 265 cpy_err:
 266         ocrdma_del_mmap(ctx, ctx->ah_tbl.pa, ctx->ah_tbl.len);
 267 map_err:
 268         dma_free_coherent(&pdev->dev, ctx->ah_tbl.len, ctx->ah_tbl.va,
 269                           ctx->ah_tbl.pa);
 270         kfree(ctx);
 271         return ERR_PTR(status);
 272 }
 273
 274 int ocrdma_dealloc_ucontext(struct ib_ucontext *ibctx)
 275 {
 276         struct ocrdma_mm *mm, *tmp;
 277         struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ibctx);
 278         struct pci_dev *pdev = uctx->dev->nic_info.pdev;
 279
 280         ocrdma_del_mmap(uctx, uctx->ah_tbl.pa, uctx->ah_tbl.len);
 281         dma_free_coherent(&pdev->dev, uctx->ah_tbl.len, uctx->ah_tbl.va,
 282                           uctx->ah_tbl.pa);
 283
 284         list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
 285                 list_del(&mm->entry);
 286                 kfree(mm);
 287         }
 288         kfree(uctx);
 289         return 0;
 290 }
 291
 292 int ocrdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
 293 {
 294         struct ocrdma_ucontext *ucontext = get_ocrdma_ucontext(context);
 295         struct ocrdma_dev *dev = ucontext->dev;
 296         unsigned long vm_page = vma->vm_pgoff << PAGE_SHIFT;
 297         u64 unmapped_db = (u64) dev->nic_info.unmapped_db;
 298         unsigned long len = (vma->vm_end - vma->vm_start);
 299         int status = 0;
 300         bool found;
 301
 302         if (vma->vm_start & (PAGE_SIZE - 1))
 303                 return -EINVAL;
 304         found = ocrdma_search_mmap(ucontext, vma->vm_pgoff << PAGE_SHIFT, len);
 305         if (!found)
 306                 return -EINVAL;
 307
 308         if ((vm_page >= unmapped_db) && (vm_page <= (unmapped_db +
 309                 dev->nic_info.db_total_size)) &&
 310                 (len <= dev->nic_info.db_page_size)) {
 311                 /* doorbell mapping */
 312                 status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
 313                                             len, vma->vm_page_prot);
 314         } else if (dev->nic_info.dpp_unmapped_len &&
 315                 (vm_page >= (u64) dev->nic_info.dpp_unmapped_addr) &&
 316                 (vm_page <= (u64) (dev->nic_info.dpp_unmapped_addr +
 317                         dev->nic_info.dpp_unmapped_len)) &&
 318                 (len <= dev->nic_info.dpp_unmapped_len)) {
 319                 /* dpp area mapping */
 320                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 321                 status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
 322                                             len, vma->vm_page_prot);
 323         } else {
 324                 /* queue memory mapping */
 325                 status = remap_pfn_range(vma, vma->vm_start,
 326                                          vma->vm_pgoff, len, vma->vm_page_prot);
 327         }
 328         return status;
 329 }
 330
 331 static int ocrdma_copy_pd_uresp(struct ocrdma_pd *pd,
 332                                 struct ib_ucontext *ib_ctx,
 333                                 struct ib_udata *udata)
 334 {
 335         int status;
 336         u64 db_page_addr;
 337         u64 dpp_page_addr = 0;
 338         u32 db_page_size;
 339         struct ocrdma_alloc_pd_uresp rsp;
 340         struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
 341
 342         rsp.id = pd->id;
 343         rsp.dpp_enabled = pd->dpp_enabled;
 344         db_page_addr = pd->dev->nic_info.unmapped_db +
 345                         (pd->id * pd->dev->nic_info.db_page_size);
 346         db_page_size = pd->dev->nic_info.db_page_size;
 347
 348         status = ocrdma_add_mmap(uctx, db_page_addr, db_page_size);
 349         if (status)
 350                 return status;
 351
 352         if (pd->dpp_enabled) {
 353                 dpp_page_addr = pd->dev->nic_info.dpp_unmapped_addr +
 354                                 (pd->id * OCRDMA_DPP_PAGE_SIZE);
 355                 status = ocrdma_add_mmap(uctx, dpp_page_addr,
 356                                  OCRDMA_DPP_PAGE_SIZE);
 357                 if (status)
 358                         goto dpp_map_err;
 359                 rsp.dpp_page_addr_hi = upper_32_bits(dpp_page_addr);
 360                 rsp.dpp_page_addr_lo = dpp_page_addr;
 361         }
 362
 363         status = ib_copy_to_udata(udata, &rsp, sizeof(rsp));
 364         if (status)
 365                 goto ucopy_err;
 366
 367         pd->uctx = uctx;
 368         return 0;
 369
 370 ucopy_err:
 371         if (pd->dpp_enabled)
 372                 ocrdma_del_mmap(pd->uctx, dpp_page_addr, OCRDMA_DPP_PAGE_SIZE);
 373 dpp_map_err:
 374         ocrdma_del_mmap(pd->uctx, db_page_addr, db_page_size);
 375         return status;
 376 }
 377
 378 struct ib_pd *ocrdma_alloc_pd(struct ib_device *ibdev,
 379                               struct ib_ucontext *context,
 380                               struct ib_udata *udata)
 381 {
 382         struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
 383         struct ocrdma_pd *pd;
 384         int status;
 385
 386         pd = kzalloc(sizeof(*pd), GFP_KERNEL);
 387         if (!pd)
 388                 return ERR_PTR(-ENOMEM);
 389         pd->dev = dev;
 390         if (udata && context) {
 391                 pd->dpp_enabled = (dev->nic_info.dev_family ==
 392                                         OCRDMA_GEN2_FAMILY) ? true : false;
 393                 pd->num_dpp_qp =
 394                         pd->dpp_enabled ? OCRDMA_PD_MAX_DPP_ENABLED_QP : 0;
 395         }
 396         status = ocrdma_mbx_alloc_pd(dev, pd);
 397         if (status) {
 398                 kfree(pd);
 399                 return ERR_PTR(status);
 400         }
 401         atomic_set(&pd->use_cnt, 0);
 402
 403         if (udata && context) {
 404                 status = ocrdma_copy_pd_uresp(pd, context, udata);
 405                 if (status)
 406                         goto err;
 407         }
 408         return &pd->ibpd;
 409
 410 err:
 411         ocrdma_dealloc_pd(&pd->ibpd);
 412         return ERR_PTR(status);
 413 }
 414
 415 int ocrdma_dealloc_pd(struct ib_pd *ibpd)
 416 {
 417         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
 418         struct ocrdma_dev *dev = pd->dev;
 419         int status;
 420         u64 usr_db;
 421
 422         if (atomic_read(&pd->use_cnt)) {
 423                 ocrdma_err("%s(%d) pd=0x%x is in use.\n",
 424                            __func__, dev->id, pd->id);
 425                 status = -EFAULT;
 426                 goto dealloc_err;
 427         }
 428         status = ocrdma_mbx_dealloc_pd(dev, pd);
 429         if (pd->uctx) {
 430                 u64 dpp_db = dev->nic_info.dpp_unmapped_addr +
 431                     (pd->id * OCRDMA_DPP_PAGE_SIZE);
 432                 if (pd->dpp_enabled)
 433                         ocrdma_del_mmap(pd->uctx, dpp_db, OCRDMA_DPP_PAGE_SIZE);
 434                 usr_db = dev->nic_info.unmapped_db +
 435                     (pd->id * dev->nic_info.db_page_size);
 436                 ocrdma_del_mmap(pd->uctx, usr_db, dev->nic_info.db_page_size);
 437         }
 438         kfree(pd);
 439 dealloc_err:
 440         return status;
 441 }
 442
 443 static struct ocrdma_mr *ocrdma_alloc_lkey(struct ib_pd *ibpd,
 444                                            int acc, u32 num_pbls,
 445                                            u32 addr_check)
 446 {
 447         int status;
 448         struct ocrdma_mr *mr;
 449         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
 450         struct ocrdma_dev *dev = pd->dev;
 451
 452         if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE)) {
 453                 ocrdma_err("%s(%d) leaving err, invalid access rights\n",
 454                            __func__, dev->id);
 455                 return ERR_PTR(-EINVAL);
 456         }
 457
 458         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
 459         if (!mr)
 460                 return ERR_PTR(-ENOMEM);
 461         mr->hwmr.dev = dev;
 462         mr->hwmr.fr_mr = 0;
 463         mr->hwmr.local_rd = 1;
 464         mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
 465         mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
 466         mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
 467         mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
 468         mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
 469         mr->hwmr.num_pbls = num_pbls;
 470
 471         status = ocrdma_mbx_alloc_lkey(dev, &mr->hwmr, pd->id, addr_check);
 472         if (status) {
 473                 kfree(mr);
 474                 return ERR_PTR(-ENOMEM);
 475         }
 476         mr->pd = pd;
 477         atomic_inc(&pd->use_cnt);
 478         mr->ibmr.lkey = mr->hwmr.lkey;
 479         if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
 480                 mr->ibmr.rkey = mr->hwmr.lkey;
 481         return mr;
 482 }
 483
 484 struct ib_mr *ocrdma_get_dma_mr(struct ib_pd *ibpd, int acc)
 485 {
 486         struct ocrdma_mr *mr;
 487
 488         mr = ocrdma_alloc_lkey(ibpd, acc, 0, OCRDMA_ADDR_CHECK_DISABLE);
 489         if (IS_ERR(mr))
 490                 return ERR_CAST(mr);
 491
 492         return &mr->ibmr;
 493 }
 494
 495 static void ocrdma_free_mr_pbl_tbl(struct ocrdma_dev *dev,
 496                                    struct ocrdma_hw_mr *mr)
 497 {
 498         struct pci_dev *pdev = dev->nic_info.pdev;
 499         int i = 0;
 500
 501         if (mr->pbl_table) {
 502                 for (i = 0; i < mr->num_pbls; i++) {
 503                         if (!mr->pbl_table[i].va)
 504                                 continue;
 505                         dma_free_coherent(&pdev->dev, mr->pbl_size,
 506                                           mr->pbl_table[i].va,
 507                                           mr->pbl_table[i].pa);
 508                 }
 509                 kfree(mr->pbl_table);
 510                 mr->pbl_table = NULL;
 511         }
 512 }
 513
 514 static int ocrdma_get_pbl_info(struct ocrdma_mr *mr, u32 num_pbes)
 515 {
 516         u32 num_pbls = 0;
 517         u32 idx = 0;
 518         int status = 0;
 519         u32 pbl_size;
 520
 521         do {
 522                 pbl_size = OCRDMA_MIN_HPAGE_SIZE * (1 << idx);
 523                 if (pbl_size > MAX_OCRDMA_PBL_SIZE) {
 524                         status = -EFAULT;
 525                         break;
 526                 }
 527                 num_pbls = roundup(num_pbes, (pbl_size / sizeof(u64)));
 528                 num_pbls = num_pbls / (pbl_size / sizeof(u64));
 529                 idx++;
 530         } while (num_pbls >= mr->hwmr.dev->attr.max_num_mr_pbl);
 531
 532         mr->hwmr.num_pbes = num_pbes;
 533         mr->hwmr.num_pbls = num_pbls;
 534         mr->hwmr.pbl_size = pbl_size;
 535         return status;
 536 }
 537
 538 static int ocrdma_build_pbl_tbl(struct ocrdma_dev *dev, struct ocrdma_hw_mr *mr)
 539 {
 540         int status = 0;
 541         int i;
 542         u32 dma_len = mr->pbl_size;
 543         struct pci_dev *pdev = dev->nic_info.pdev;
 544         void *va;
 545         dma_addr_t pa;
 546
 547         mr->pbl_table = kzalloc(sizeof(struct ocrdma_pbl) *
 548                                 mr->num_pbls, GFP_KERNEL);
 549
 550         if (!mr->pbl_table)
 551                 return -ENOMEM;
 552
 553         for (i = 0; i < mr->num_pbls; i++) {
 554                 va = dma_alloc_coherent(&pdev->dev, dma_len, &pa, GFP_KERNEL);
 555                 if (!va) {
 556                         ocrdma_free_mr_pbl_tbl(dev, mr);
 557                         status = -ENOMEM;
 558                         break;
 559                 }
 560                 memset(va, 0, dma_len);
 561                 mr->pbl_table[i].va = va;
 562                 mr->pbl_table[i].pa = pa;
 563         }
 564         return status;
 565 }
 566
 567 static void build_user_pbes(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
 568                             u32 num_pbes)
 569 {
 570         struct ocrdma_pbe *pbe;
 571         struct ib_umem_chunk *chunk;
 572         struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
 573         struct ib_umem *umem = mr->umem;
 574         int i, shift, pg_cnt, pages, pbe_cnt, total_num_pbes = 0;
 575
 576         if (!mr->hwmr.num_pbes)
 577                 return;
 578
 579         pbe = (struct ocrdma_pbe *)pbl_tbl->va;
 580         pbe_cnt = 0;
 581
 582         shift = ilog2(umem->page_size);
 583
 584         list_for_each_entry(chunk, &umem->chunk_list, list) {
 585                 /* get all the dma regions from the chunk. */
 586                 for (i = 0; i < chunk->nmap; i++) {
 587                         pages = sg_dma_len(&chunk->page_list[i]) >> shift;
 588                         for (pg_cnt = 0; pg_cnt < pages; pg_cnt++) {
 589                                 /* store the page address in pbe */
 590                                 pbe->pa_lo =
 591                                     cpu_to_le32(sg_dma_address
 592                                                 (&chunk->page_list[i]) +
 593                                                 (umem->page_size * pg_cnt));
 594                                 pbe->pa_hi =
 595                                     cpu_to_le32(upper_32_bits
 596                                                 ((sg_dma_address
 597                                                   (&chunk->page_list[i]) +
 598                                                   umem->page_size * pg_cnt)));
 599                                 pbe_cnt += 1;
 600                                 total_num_pbes += 1;
 601                                 pbe++;
 602
 603                                 /* if done building pbes, issue the mbx cmd. */
 604                                 if (total_num_pbes == num_pbes)
 605                                         return;
 606
 607                                 /* if the given pbl is full storing the pbes,
 608                                  * move to next pbl.
 609                                  */
 610                                 if (pbe_cnt ==
 611                                         (mr->hwmr.pbl_size / sizeof(u64))) {
 612                                         pbl_tbl++;
 613                                         pbe = (struct ocrdma_pbe *)pbl_tbl->va;
 614                                         pbe_cnt = 0;
 615                                 }
 616                         }
 617                 }
 618         }
 619 }
 620
 621 struct ib_mr *ocrdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
 622                                  u64 usr_addr, int acc, struct ib_udata *udata)
 623 {
 624         int status = -ENOMEM;
 625         struct ocrdma_dev *dev;
 626         struct ocrdma_mr *mr;
 627         struct ocrdma_pd *pd;
 628         u32 num_pbes;
 629
 630         pd = get_ocrdma_pd(ibpd);
 631         dev = pd->dev;
 632
 633         if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
 634                 return ERR_PTR(-EINVAL);
 635
 636         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
 637         if (!mr)
 638                 return ERR_PTR(status);
 639         mr->hwmr.dev = dev;
 640         mr->umem = ib_umem_get(ibpd->uobject->context, start, len, acc, 0);
 641         if (IS_ERR(mr->umem)) {
 642                 status = -EFAULT;
 643                 goto umem_err;
 644         }
 645         num_pbes = ib_umem_page_count(mr->umem);
 646         status = ocrdma_get_pbl_info(mr, num_pbes);
 647         if (status)
 648                 goto umem_err;
 649
 650         mr->hwmr.pbe_size = mr->umem->page_size;
 651         mr->hwmr.fbo = mr->umem->offset;
 652         mr->hwmr.va = usr_addr;
 653         mr->hwmr.len = len;
 654         mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
 655         mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
 656         mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
 657         mr->hwmr.local_rd = 1;
 658         mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
 659         status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
 660         if (status)
 661                 goto umem_err;
 662         build_user_pbes(dev, mr, num_pbes);
 663         status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
 664         if (status)
 665                 goto mbx_err;
 666         mr->pd = pd;
 667         atomic_inc(&pd->use_cnt);
 668         mr->ibmr.lkey = mr->hwmr.lkey;
 669         if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
 670                 mr->ibmr.rkey = mr->hwmr.lkey;
 671
 672         return &mr->ibmr;
 673
 674 mbx_err:
 675         ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
 676 umem_err:
 677         kfree(mr);
 678         return ERR_PTR(status);
 679 }
 680
 681 int ocrdma_dereg_mr(struct ib_mr *ib_mr)
 682 {
 683         struct ocrdma_mr *mr = get_ocrdma_mr(ib_mr);
 684         struct ocrdma_dev *dev = mr->hwmr.dev;
 685         int status;
 686
 687         status = ocrdma_mbx_dealloc_lkey(dev, mr->hwmr.fr_mr, mr->hwmr.lkey);
 688
 689         if (mr->hwmr.fr_mr == 0)
 690                 ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
 691
 692         atomic_dec(&mr->pd->use_cnt);
 693         /* it could be user registered memory. */
 694         if (mr->umem)
 695                 ib_umem_release(mr->umem);
 696         kfree(mr);
 697         return status;
 698 }
 699
 700 static int ocrdma_copy_cq_uresp(struct ocrdma_cq *cq, struct ib_udata *udata,
 701                                 struct ib_ucontext *ib_ctx)
 702 {
 703         int status;
 704         struct ocrdma_ucontext *uctx;
 705         struct ocrdma_create_cq_uresp uresp;
 706
 707         uresp.cq_id = cq->id;
 708         uresp.page_size = cq->len;
 709         uresp.num_pages = 1;
 710         uresp.max_hw_cqe = cq->max_hw_cqe;
 711         uresp.page_addr[0] = cq->pa;
 712         uresp.db_page_addr = cq->dev->nic_info.unmapped_db;
 713         uresp.db_page_size = cq->dev->nic_info.db_page_size;
 714         uresp.phase_change = cq->phase_change ? 1 : 0;
 715         status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
 716         if (status) {
 717                 ocrdma_err("%s(%d) copy error cqid=0x%x.\n",
 718                            __func__, cq->dev->id, cq->id);
 719                 goto err;
 720         }
 721         uctx = get_ocrdma_ucontext(ib_ctx);
 722         status = ocrdma_add_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
 723         if (status)
 724                 goto err;
 725         status = ocrdma_add_mmap(uctx, uresp.page_addr[0], uresp.page_size);
 726         if (status) {
 727                 ocrdma_del_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
 728                 goto err;
 729         }
 730         cq->ucontext = uctx;
 731 err:
 732         return status;
 733 }
 734
 735 struct ib_cq *ocrdma_create_cq(struct ib_device *ibdev, int entries, int vector,
 736                                struct ib_ucontext *ib_ctx,
 737                                struct ib_udata *udata)
 738 {
 739         struct ocrdma_cq *cq;
 740         struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
 741         int status;
 742         struct ocrdma_create_cq_ureq ureq;
 743
 744         if (udata) {
 745                 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
 746                         return ERR_PTR(-EFAULT);
 747         } else
 748                 ureq.dpp_cq = 0;
 749         cq = kzalloc(sizeof(*cq), GFP_KERNEL);
 750         if (!cq)
 751                 return ERR_PTR(-ENOMEM);
 752
 753         spin_lock_init(&cq->cq_lock);
 754         spin_lock_init(&cq->comp_handler_lock);
 755         atomic_set(&cq->use_cnt, 0);
 756         INIT_LIST_HEAD(&cq->sq_head);
 757         INIT_LIST_HEAD(&cq->rq_head);
 758         cq->dev = dev;
 759
 760         status = ocrdma_mbx_create_cq(dev, cq, entries, ureq.dpp_cq);
 761         if (status) {
 762                 kfree(cq);
 763                 return ERR_PTR(status);
 764         }
 765         if (ib_ctx) {
 766                 status = ocrdma_copy_cq_uresp(cq, udata, ib_ctx);
 767                 if (status)
 768                         goto ctx_err;
 769         }
 770         cq->phase = OCRDMA_CQE_VALID;
 771         cq->arm_needed = true;
 772         dev->cq_tbl[cq->id] = cq;
 773
 774         return &cq->ibcq;
 775
 776 ctx_err:
 777         ocrdma_mbx_destroy_cq(dev, cq);
 778         kfree(cq);
 779         return ERR_PTR(status);
 780 }
 781
 782 int ocrdma_resize_cq(struct ib_cq *ibcq, int new_cnt,
 783                      struct ib_udata *udata)
 784 {
 785         int status = 0;
 786         struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
 787
 788         if (new_cnt < 1 || new_cnt > cq->max_hw_cqe) {
 789                 status = -EINVAL;
 790                 return status;
 791         }
 792         ibcq->cqe = new_cnt;
 793         return status;
 794 }
 795
 796 int ocrdma_destroy_cq(struct ib_cq *ibcq)
 797 {
 798         int status;
 799         struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
 800         struct ocrdma_dev *dev = cq->dev;
 801
 802         if (atomic_read(&cq->use_cnt))
 803                 return -EINVAL;
 804
 805         status = ocrdma_mbx_destroy_cq(dev, cq);
 806
 807         if (cq->ucontext) {
 808                 ocrdma_del_mmap(cq->ucontext, (u64) cq->pa, cq->len);
 809                 ocrdma_del_mmap(cq->ucontext, dev->nic_info.unmapped_db,
 810                                 dev->nic_info.db_page_size);
 811         }
 812         dev->cq_tbl[cq->id] = NULL;
 813
 814         kfree(cq);
 815         return status;
 816 }
 817
 818 static int ocrdma_add_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
 819 {
 820         int status = -EINVAL;
 821
 822         if (qp->id < OCRDMA_MAX_QP && dev->qp_tbl[qp->id] == NULL) {
 823                 dev->qp_tbl[qp->id] = qp;
 824                 status = 0;
 825         }
 826         return status;
 827 }
 828
 829 static void ocrdma_del_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
 830 {
 831         dev->qp_tbl[qp->id] = NULL;
 832 }
 833
 834 static int ocrdma_check_qp_params(struct ib_pd *ibpd, struct ocrdma_dev *dev,
 835                                   struct ib_qp_init_attr *attrs)
 836 {
 837         if (attrs->qp_type != IB_QPT_GSI &&
 838             attrs->qp_type != IB_QPT_RC &&
 839             attrs->qp_type != IB_QPT_UD) {
 840                 ocrdma_err("%s(%d) unsupported qp type=0x%x requested\n",
 841                            __func__, dev->id, attrs->qp_type);
 842                 return -EINVAL;
 843         }
 844         if (attrs->cap.max_send_wr > dev->attr.max_wqe) {
 845                 ocrdma_err("%s(%d) unsupported send_wr=0x%x requested\n",
 846                            __func__, dev->id, attrs->cap.max_send_wr);
 847                 ocrdma_err("%s(%d) supported send_wr=0x%x\n",
 848                            __func__, dev->id, dev->attr.max_wqe);
 849                 return -EINVAL;
 850         }
 851         if (!attrs->srq && (attrs->cap.max_recv_wr > dev->attr.max_rqe)) {
 852                 ocrdma_err("%s(%d) unsupported recv_wr=0x%x requested\n",
 853                            __func__, dev->id, attrs->cap.max_recv_wr);
 854                 ocrdma_err("%s(%d) supported recv_wr=0x%x\n",
 855                            __func__, dev->id, dev->attr.max_rqe);
 856                 return -EINVAL;
 857         }
 858         if (attrs->cap.max_inline_data > dev->attr.max_inline_data) {
 859                 ocrdma_err("%s(%d) unsupported inline data size=0x%x"
 860                            " requested\n", __func__, dev->id,
 861                            attrs->cap.max_inline_data);
 862                 ocrdma_err("%s(%d) supported inline data size=0x%x\n",
 863                            __func__, dev->id, dev->attr.max_inline_data);
 864                 return -EINVAL;
 865         }
 866         if (attrs->cap.max_send_sge > dev->attr.max_send_sge) {
 867                 ocrdma_err("%s(%d) unsupported send_sge=0x%x requested\n",
 868                            __func__, dev->id, attrs->cap.max_send_sge);
 869                 ocrdma_err("%s(%d) supported send_sge=0x%x\n",
 870                            __func__, dev->id, dev->attr.max_send_sge);
 871                 return -EINVAL;
 872         }
 873         if (attrs->cap.max_recv_sge > dev->attr.max_recv_sge) {
 874                 ocrdma_err("%s(%d) unsupported recv_sge=0x%x requested\n",
 875                            __func__, dev->id, attrs->cap.max_recv_sge);
 876                 ocrdma_err("%s(%d) supported recv_sge=0x%x\n",
 877                            __func__, dev->id, dev->attr.max_recv_sge);
 878                 return -EINVAL;
 879         }
 880         /* unprivileged user space cannot create special QP */
 881         if (ibpd->uobject && attrs->qp_type == IB_QPT_GSI) {
 882                 ocrdma_err
 883                     ("%s(%d) Userspace can't create special QPs of type=0x%x\n",
 884                      __func__, dev->id, attrs->qp_type);
 885                 return -EINVAL;
 886         }
 887         /* allow creating only one GSI type of QP */
 888         if (attrs->qp_type == IB_QPT_GSI && dev->gsi_qp_created) {
 889                 ocrdma_err("%s(%d) GSI special QPs already created.\n",
 890                            __func__, dev->id);
 891                 return -EINVAL;
 892         }
 893         /* verify consumer QPs are not trying to use GSI QP's CQ */
 894         if ((attrs->qp_type != IB_QPT_GSI) && (dev->gsi_qp_created)) {
 895                 if ((dev->gsi_sqcq == get_ocrdma_cq(attrs->send_cq)) ||
 896                     (dev->gsi_sqcq == get_ocrdma_cq(attrs->send_cq))) {
 897                         ocrdma_err("%s(%d) Consumer QP cannot use GSI CQs.\n",
 898                                    __func__, dev->id);
 899                         return -EINVAL;
 900                 }
 901         }
 902         return 0;
 903 }
 904
 905 static int ocrdma_copy_qp_uresp(struct ocrdma_qp *qp,
 906                                 struct ib_udata *udata, int dpp_offset,
 907                                 int dpp_credit_lmt, int srq)
 908 {
 909         int status = 0;
 910         u64 usr_db;
 911         struct ocrdma_create_qp_uresp uresp;
 912         struct ocrdma_dev *dev = qp->dev;
 913         struct ocrdma_pd *pd = qp->pd;
 914
 915         memset(&uresp, 0, sizeof(uresp));
 916         usr_db = dev->nic_info.unmapped_db +
 917                         (pd->id * dev->nic_info.db_page_size);
 918         uresp.qp_id = qp->id;
 919         uresp.sq_dbid = qp->sq.dbid;
 920         uresp.num_sq_pages = 1;
 921         uresp.sq_page_size = qp->sq.len;
 922         uresp.sq_page_addr[0] = qp->sq.pa;
 923         uresp.num_wqe_allocated = qp->sq.max_cnt;
 924         if (!srq) {
 925                 uresp.rq_dbid = qp->rq.dbid;
 926                 uresp.num_rq_pages = 1;
 927                 uresp.rq_page_size = qp->rq.len;
 928                 uresp.rq_page_addr[0] = qp->rq.pa;
 929                 uresp.num_rqe_allocated = qp->rq.max_cnt;
 930         }
 931         uresp.db_page_addr = usr_db;
 932         uresp.db_page_size = dev->nic_info.db_page_size;
 933         if (dev->nic_info.dev_family == OCRDMA_GEN2_FAMILY) {
 934                 uresp.db_sq_offset = OCRDMA_DB_GEN2_SQ_OFFSET;
 935                 uresp.db_rq_offset = ((qp->id & 0xFFFF) < 128) ?
 936                         OCRDMA_DB_GEN2_RQ1_OFFSET : OCRDMA_DB_GEN2_RQ2_OFFSET;
 937                 uresp.db_shift = (qp->id < 128) ? 24 : 16;
 938         } else {
 939                 uresp.db_sq_offset = OCRDMA_DB_SQ_OFFSET;
 940                 uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
 941                 uresp.db_shift = 16;
 942         }
 943
 944         if (qp->dpp_enabled) {
 945                 uresp.dpp_credit = dpp_credit_lmt;
 946                 uresp.dpp_offset = dpp_offset;
 947         }
 948         status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
 949         if (status) {
 950                 ocrdma_err("%s(%d) user copy error.\n", __func__, dev->id);
 951                 goto err;
 952         }
 953         status = ocrdma_add_mmap(pd->uctx, uresp.sq_page_addr[0],
 954                                  uresp.sq_page_size);
 955         if (status)
 956                 goto err;
 957
 958         if (!srq) {
 959                 status = ocrdma_add_mmap(pd->uctx, uresp.rq_page_addr[0],
 960                                          uresp.rq_page_size);
 961                 if (status)
 962                         goto rq_map_err;
 963         }
 964         return status;
 965 rq_map_err:
 966         ocrdma_del_mmap(pd->uctx, uresp.sq_page_addr[0], uresp.sq_page_size);
 967 err:
 968         return status;
 969 }
 970
 971 static void ocrdma_set_qp_db(struct ocrdma_dev *dev, struct ocrdma_qp *qp,
 972                              struct ocrdma_pd *pd)
 973 {
 974         if (dev->nic_info.dev_family == OCRDMA_GEN2_FAMILY) {
 975                 qp->sq_db = dev->nic_info.db +
 976                         (pd->id * dev->nic_info.db_page_size) +
 977                         OCRDMA_DB_GEN2_SQ_OFFSET;
 978                 qp->rq_db = dev->nic_info.db +
 979                         (pd->id * dev->nic_info.db_page_size) +
 980                         ((qp->id < 128) ?
 981                         OCRDMA_DB_GEN2_RQ1_OFFSET : OCRDMA_DB_GEN2_RQ2_OFFSET);
 982         } else {
 983                 qp->sq_db = dev->nic_info.db +
 984                         (pd->id * dev->nic_info.db_page_size) +
 985                         OCRDMA_DB_SQ_OFFSET;
 986                 qp->rq_db = dev->nic_info.db +
 987                         (pd->id * dev->nic_info.db_page_size) +
 988                         OCRDMA_DB_RQ_OFFSET;
 989         }
 990 }
 991
 992 static int ocrdma_alloc_wr_id_tbl(struct ocrdma_qp *qp)
 993 {
 994         qp->wqe_wr_id_tbl =
 995             kzalloc(sizeof(*(qp->wqe_wr_id_tbl)) * qp->sq.max_cnt,
 996                     GFP_KERNEL);
 997         if (qp->wqe_wr_id_tbl == NULL)
 998                 return -ENOMEM;
 999         qp->rqe_wr_id_tbl =
1000             kzalloc(sizeof(u64) * qp->rq.max_cnt, GFP_KERNEL);
1001         if (qp->rqe_wr_id_tbl == NULL)
1002                 return -ENOMEM;
1003
1004         return 0;
1005 }
1006
1007 static void ocrdma_set_qp_init_params(struct ocrdma_qp *qp,
1008                                       struct ocrdma_pd *pd,
1009                                       struct ib_qp_init_attr *attrs)
1010 {
1011         qp->pd = pd;
1012         spin_lock_init(&qp->q_lock);
1013         INIT_LIST_HEAD(&qp->sq_entry);
1014         INIT_LIST_HEAD(&qp->rq_entry);
1015
1016         qp->qp_type = attrs->qp_type;
1017         qp->cap_flags = OCRDMA_QP_INB_RD | OCRDMA_QP_INB_WR;
1018         qp->max_inline_data = attrs->cap.max_inline_data;
1019         qp->sq.max_sges = attrs->cap.max_send_sge;
1020         qp->rq.max_sges = attrs->cap.max_recv_sge;
1021         qp->state = OCRDMA_QPS_RST;
1022 }
1023
1024 static void ocrdma_set_qp_use_cnt(struct ocrdma_qp *qp, struct ocrdma_pd *pd)
1025 {
1026         atomic_inc(&pd->use_cnt);
1027         atomic_inc(&qp->sq_cq->use_cnt);
1028         atomic_inc(&qp->rq_cq->use_cnt);
1029         if (qp->srq)
1030                 atomic_inc(&qp->srq->use_cnt);
1031         qp->ibqp.qp_num = qp->id;
1032 }
1033
1034 static void ocrdma_store_gsi_qp_cq(struct ocrdma_dev *dev,
1035                                    struct ib_qp_init_attr *attrs)
1036 {
1037         if (attrs->qp_type == IB_QPT_GSI) {
1038                 dev->gsi_qp_created = 1;
1039                 dev->gsi_sqcq = get_ocrdma_cq(attrs->send_cq);
1040                 dev->gsi_rqcq = get_ocrdma_cq(attrs->recv_cq);
1041         }
1042 }
1043
1044 struct ib_qp *ocrdma_create_qp(struct ib_pd *ibpd,
1045                                struct ib_qp_init_attr *attrs,
1046                                struct ib_udata *udata)
1047 {
1048         int status;
1049         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1050         struct ocrdma_qp *qp;
1051         struct ocrdma_dev *dev = pd->dev;
1052         struct ocrdma_create_qp_ureq ureq;
1053         u16 dpp_credit_lmt, dpp_offset;
1054
1055         status = ocrdma_check_qp_params(ibpd, dev, attrs);
1056         if (status)
1057                 goto gen_err;
1058
1059         memset(&ureq, 0, sizeof(ureq));
1060         if (udata) {
1061                 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1062                         return ERR_PTR(-EFAULT);
1063         }
1064         qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1065         if (!qp) {
1066                 status = -ENOMEM;
1067                 goto gen_err;
1068         }
1069         qp->dev = dev;
1070         ocrdma_set_qp_init_params(qp, pd, attrs);
1071
1072         mutex_lock(&dev->dev_lock);
1073         status = ocrdma_mbx_create_qp(qp, attrs, ureq.enable_dpp_cq,
1074                                         ureq.dpp_cq_id,
1075                                         &dpp_offset, &dpp_credit_lmt);
1076         if (status)
1077                 goto mbx_err;
1078
1079         /* user space QP's wr_id table are managed in library */
1080         if (udata == NULL) {
1081                 qp->cap_flags |= (OCRDMA_QP_MW_BIND | OCRDMA_QP_LKEY0 |
1082                                   OCRDMA_QP_FAST_REG);
1083                 status = ocrdma_alloc_wr_id_tbl(qp);
1084                 if (status)
1085                         goto map_err;
1086         }
1087
1088         status = ocrdma_add_qpn_map(dev, qp);
1089         if (status)
1090                 goto map_err;
1091         ocrdma_set_qp_db(dev, qp, pd);
1092         if (udata) {
1093                 status = ocrdma_copy_qp_uresp(qp, udata, dpp_offset,
1094                                               dpp_credit_lmt,
1095                                               (attrs->srq != NULL));
1096                 if (status)
1097                         goto cpy_err;
1098         }
1099         ocrdma_store_gsi_qp_cq(dev, attrs);
1100         ocrdma_set_qp_use_cnt(qp, pd);
1101         mutex_unlock(&dev->dev_lock);
1102         return &qp->ibqp;
1103
1104 cpy_err:
1105         ocrdma_del_qpn_map(dev, qp);
1106 map_err:
1107         ocrdma_mbx_destroy_qp(dev, qp);
1108 mbx_err:
1109         mutex_unlock(&dev->dev_lock);
1110         kfree(qp->wqe_wr_id_tbl);
1111         kfree(qp->rqe_wr_id_tbl);
1112         kfree(qp);
1113         ocrdma_err("%s(%d) error=%d\n", __func__, dev->id, status);
1114 gen_err:
1115         return ERR_PTR(status);
1116 }
1117
1118 int _ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1119                       int attr_mask)
1120 {
1121         int status = 0;
1122         struct ocrdma_qp *qp;
1123         struct ocrdma_dev *dev;
1124         enum ib_qp_state old_qps;
1125
1126         qp = get_ocrdma_qp(ibqp);
1127         dev = qp->dev;
1128         if (attr_mask & IB_QP_STATE)
1129                 status = ocrdma_qp_state_machine(qp, attr->qp_state, &old_qps);
1130         /* if new and previous states are same hw doesn't need to
1131          * know about it.
1132          */
1133         if (status < 0)
1134                 return status;
1135         status = ocrdma_mbx_modify_qp(dev, qp, attr, attr_mask, old_qps);
1136         return status;
1137 }
1138
1139 int ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1140                      int attr_mask, struct ib_udata *udata)
1141 {
1142         unsigned long flags;
1143         int status = -EINVAL;
1144         struct ocrdma_qp *qp;
1145         struct ocrdma_dev *dev;
1146         enum ib_qp_state old_qps, new_qps;
1147
1148         qp = get_ocrdma_qp(ibqp);
1149         dev = qp->dev;
1150
1151         /* syncronize with multiple context trying to change, retrive qps */
1152         mutex_lock(&dev->dev_lock);
1153         /* syncronize with wqe, rqe posting and cqe processing contexts */
1154         spin_lock_irqsave(&qp->q_lock, flags);
1155         old_qps = get_ibqp_state(qp->state);
1156         if (attr_mask & IB_QP_STATE)
1157                 new_qps = attr->qp_state;
1158         else
1159                 new_qps = old_qps;
1160         spin_unlock_irqrestore(&qp->q_lock, flags);
1161
1162         if (!ib_modify_qp_is_ok(old_qps, new_qps, ibqp->qp_type, attr_mask)) {
1163                 ocrdma_err("%s(%d) invalid attribute mask=0x%x specified for "
1164                            "qpn=0x%x of type=0x%x old_qps=0x%x, new_qps=0x%x\n",
1165                            __func__, dev->id, attr_mask, qp->id, ibqp->qp_type,
1166                            old_qps, new_qps);
1167                 goto param_err;
1168         }
1169
1170         status = _ocrdma_modify_qp(ibqp, attr, attr_mask);
1171         if (status > 0)
1172                 status = 0;
1173 param_err:
1174         mutex_unlock(&dev->dev_lock);
1175         return status;
1176 }
1177
1178 static enum ib_mtu ocrdma_mtu_int_to_enum(u16 mtu)
1179 {
1180         switch (mtu) {
1181         case 256:
1182                 return IB_MTU_256;
1183         case 512:
1184                 return IB_MTU_512;
1185         case 1024:
1186                 return IB_MTU_1024;
1187         case 2048:
1188                 return IB_MTU_2048;
1189         case 4096:
1190                 return IB_MTU_4096;
1191         default:
1192                 return IB_MTU_1024;
1193         }
1194 }
1195
1196 static int ocrdma_to_ib_qp_acc_flags(int qp_cap_flags)
1197 {
1198         int ib_qp_acc_flags = 0;
1199
1200         if (qp_cap_flags & OCRDMA_QP_INB_WR)
1201                 ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
1202         if (qp_cap_flags & OCRDMA_QP_INB_RD)
1203                 ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
1204         return ib_qp_acc_flags;
1205 }
1206
1207 int ocrdma_query_qp(struct ib_qp *ibqp,
1208                     struct ib_qp_attr *qp_attr,
1209                     int attr_mask, struct ib_qp_init_attr *qp_init_attr)
1210 {
1211         int status;
1212         u32 qp_state;
1213         struct ocrdma_qp_params params;
1214         struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1215         struct ocrdma_dev *dev = qp->dev;
1216
1217         memset(&params, 0, sizeof(params));
1218         mutex_lock(&dev->dev_lock);
1219         status = ocrdma_mbx_query_qp(dev, qp, &params);
1220         mutex_unlock(&dev->dev_lock);
1221         if (status)
1222                 goto mbx_err;
1223         qp_attr->qp_state = get_ibqp_state(IB_QPS_INIT);
1224         qp_attr->cur_qp_state = get_ibqp_state(IB_QPS_INIT);
1225         qp_attr->path_mtu =
1226                 ocrdma_mtu_int_to_enum(params.path_mtu_pkey_indx &
1227                                 OCRDMA_QP_PARAMS_PATH_MTU_MASK) >>
1228                                 OCRDMA_QP_PARAMS_PATH_MTU_SHIFT;
1229         qp_attr->path_mig_state = IB_MIG_MIGRATED;
1230         qp_attr->rq_psn = params.hop_lmt_rq_psn & OCRDMA_QP_PARAMS_RQ_PSN_MASK;
1231         qp_attr->sq_psn = params.tclass_sq_psn & OCRDMA_QP_PARAMS_SQ_PSN_MASK;
1232         qp_attr->dest_qp_num =
1233             params.ack_to_rnr_rtc_dest_qpn & OCRDMA_QP_PARAMS_DEST_QPN_MASK;
1234
1235         qp_attr->qp_access_flags = ocrdma_to_ib_qp_acc_flags(qp->cap_flags);
1236         qp_attr->cap.max_send_wr = qp->sq.max_cnt - 1;
1237         qp_attr->cap.max_recv_wr = qp->rq.max_cnt - 1;
1238         qp_attr->cap.max_send_sge = qp->sq.max_sges;
1239         qp_attr->cap.max_recv_sge = qp->rq.max_sges;
1240         qp_attr->cap.max_inline_data = dev->attr.max_inline_data;
1241         qp_init_attr->cap = qp_attr->cap;
1242         memcpy(&qp_attr->ah_attr.grh.dgid, &params.dgid[0],
1243                sizeof(params.dgid));
1244         qp_attr->ah_attr.grh.flow_label = params.rnt_rc_sl_fl &
1245             OCRDMA_QP_PARAMS_FLOW_LABEL_MASK;
1246         qp_attr->ah_attr.grh.sgid_index = qp->sgid_idx;
1247         qp_attr->ah_attr.grh.hop_limit = (params.hop_lmt_rq_psn &
1248                                           OCRDMA_QP_PARAMS_HOP_LMT_MASK) >>
1249                                                 OCRDMA_QP_PARAMS_HOP_LMT_SHIFT;
1250         qp_attr->ah_attr.grh.traffic_class = (params.tclass_sq_psn &
1251                                               OCRDMA_QP_PARAMS_SQ_PSN_MASK) >>
1252                                                 OCRDMA_QP_PARAMS_TCLASS_SHIFT;
1253
1254         qp_attr->ah_attr.ah_flags = IB_AH_GRH;
1255         qp_attr->ah_attr.port_num = 1;
1256         qp_attr->ah_attr.sl = (params.rnt_rc_sl_fl &
1257                                OCRDMA_QP_PARAMS_SL_MASK) >>
1258                                 OCRDMA_QP_PARAMS_SL_SHIFT;
1259         qp_attr->timeout = (params.ack_to_rnr_rtc_dest_qpn &
1260                             OCRDMA_QP_PARAMS_ACK_TIMEOUT_MASK) >>
1261                                 OCRDMA_QP_PARAMS_ACK_TIMEOUT_SHIFT;
1262         qp_attr->rnr_retry = (params.ack_to_rnr_rtc_dest_qpn &
1263                               OCRDMA_QP_PARAMS_RNR_RETRY_CNT_MASK) >>
1264                                 OCRDMA_QP_PARAMS_RNR_RETRY_CNT_SHIFT;
1265         qp_attr->retry_cnt =
1266             (params.rnt_rc_sl_fl & OCRDMA_QP_PARAMS_RETRY_CNT_MASK) >>
1267                 OCRDMA_QP_PARAMS_RETRY_CNT_SHIFT;
1268         qp_attr->min_rnr_timer = 0;
1269         qp_attr->pkey_index = 0;
1270         qp_attr->port_num = 1;
1271         qp_attr->ah_attr.src_path_bits = 0;
1272         qp_attr->ah_attr.static_rate = 0;
1273         qp_attr->alt_pkey_index = 0;
1274         qp_attr->alt_port_num = 0;
1275         qp_attr->alt_timeout = 0;
1276         memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
1277         qp_state = (params.max_sge_recv_flags & OCRDMA_QP_PARAMS_STATE_MASK) >>
1278                     OCRDMA_QP_PARAMS_STATE_SHIFT;
1279         qp_attr->sq_draining = (qp_state == OCRDMA_QPS_SQ_DRAINING) ? 1 : 0;
1280         qp_attr->max_dest_rd_atomic =
1281             params.max_ord_ird >> OCRDMA_QP_PARAMS_MAX_ORD_SHIFT;
1282         qp_attr->max_rd_atomic =
1283             params.max_ord_ird & OCRDMA_QP_PARAMS_MAX_IRD_MASK;
1284         qp_attr->en_sqd_async_notify = (params.max_sge_recv_flags &
1285                                 OCRDMA_QP_PARAMS_FLAGS_SQD_ASYNC) ? 1 : 0;
1286 mbx_err:
1287         return status;
1288 }
1289
1290 static void ocrdma_srq_toggle_bit(struct ocrdma_srq *srq, int idx)
1291 {
1292         int i = idx / 32;
1293         unsigned int mask = (1 << (idx % 32));
1294
1295         if (srq->idx_bit_fields[i] & mask)
1296                 srq->idx_bit_fields[i] &= ~mask;
1297         else
1298                 srq->idx_bit_fields[i] |= mask;
1299 }
1300
1301 static int ocrdma_hwq_free_cnt(struct ocrdma_qp_hwq_info *q)
1302 {
1303         int free_cnt;
1304         if (q->head >= q->tail)
1305                 free_cnt = (q->max_cnt - q->head) + q->tail;
1306         else
1307                 free_cnt = q->tail - q->head;
1308         return free_cnt;
1309 }
1310
1311 static int is_hw_sq_empty(struct ocrdma_qp *qp)
1312 {
1313         return (qp->sq.tail == qp->sq.head &&
1314                 ocrdma_hwq_free_cnt(&qp->sq) ? 1 : 0);
1315 }
1316
1317 static int is_hw_rq_empty(struct ocrdma_qp *qp)
1318 {
1319         return (qp->rq.tail == qp->rq.head) ? 1 : 0;
1320 }
1321
1322 static void *ocrdma_hwq_head(struct ocrdma_qp_hwq_info *q)
1323 {
1324         return q->va + (q->head * q->entry_size);
1325 }
1326
1327 static void *ocrdma_hwq_head_from_idx(struct ocrdma_qp_hwq_info *q,
1328                                       u32 idx)
1329 {
1330         return q->va + (idx * q->entry_size);
1331 }
1332
1333 static void ocrdma_hwq_inc_head(struct ocrdma_qp_hwq_info *q)
1334 {
1335         q->head = (q->head + 1) & q->max_wqe_idx;
1336 }
1337
1338 static void ocrdma_hwq_inc_tail(struct ocrdma_qp_hwq_info *q)
1339 {
1340         q->tail = (q->tail + 1) & q->max_wqe_idx;
1341 }
1342
1343 /* discard the cqe for a given QP */
1344 static void ocrdma_discard_cqes(struct ocrdma_qp *qp, struct ocrdma_cq *cq)
1345 {
1346         unsigned long cq_flags;
1347         unsigned long flags;
1348         int discard_cnt = 0;
1349         u32 cur_getp, stop_getp;
1350         struct ocrdma_cqe *cqe;
1351         u32 qpn = 0;
1352
1353         spin_lock_irqsave(&cq->cq_lock, cq_flags);
1354
1355         /* traverse through the CQEs in the hw CQ,
1356          * find the matching CQE for a given qp,
1357          * mark the matching one discarded by clearing qpn.
1358          * ring the doorbell in the poll_cq() as
1359          * we don't complete out of order cqe.
1360          */
1361
1362         cur_getp = cq->getp;
1363         /* find upto when do we reap the cq. */
1364         stop_getp = cur_getp;
1365         do {
1366                 if (is_hw_sq_empty(qp) && (!qp->srq && is_hw_rq_empty(qp)))
1367                         break;
1368
1369                 cqe = cq->va + cur_getp;
1370                 /* if (a) done reaping whole hw cq, or
1371                  *    (b) qp_xq becomes empty.
1372                  * then exit
1373                  */
1374                 qpn = cqe->cmn.qpn & OCRDMA_CQE_QPN_MASK;
1375                 /* if previously discarded cqe found, skip that too. */
1376                 /* check for matching qp */
1377                 if (qpn == 0 || qpn != qp->id)
1378                         goto skip_cqe;
1379
1380                 /* mark cqe discarded so that it is not picked up later
1381                  * in the poll_cq().
1382                  */
1383                 discard_cnt += 1;
1384                 cqe->cmn.qpn = 0;
1385                 if (is_cqe_for_sq(cqe))
1386                         ocrdma_hwq_inc_tail(&qp->sq);
1387                 else {
1388                         if (qp->srq) {
1389                                 spin_lock_irqsave(&qp->srq->q_lock, flags);
1390                                 ocrdma_hwq_inc_tail(&qp->srq->rq);
1391                                 ocrdma_srq_toggle_bit(qp->srq, cur_getp);
1392                                 spin_unlock_irqrestore(&qp->srq->q_lock, flags);
1393
1394                         } else
1395                                 ocrdma_hwq_inc_tail(&qp->rq);
1396                 }
1397 skip_cqe:
1398                 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
1399         } while (cur_getp != stop_getp);
1400         spin_unlock_irqrestore(&cq->cq_lock, cq_flags);
1401 }
1402
1403 static void ocrdma_del_flush_qp(struct ocrdma_qp *qp)
1404 {
1405         int found = false;
1406         unsigned long flags;
1407         struct ocrdma_dev *dev = qp->dev;
1408         /* sync with any active CQ poll */
1409
1410         spin_lock_irqsave(&dev->flush_q_lock, flags);
1411         found = ocrdma_is_qp_in_sq_flushlist(qp->sq_cq, qp);
1412         if (found)
1413                 list_del(&qp->sq_entry);
1414         if (!qp->srq) {
1415                 found = ocrdma_is_qp_in_rq_flushlist(qp->rq_cq, qp);
1416                 if (found)
1417                         list_del(&qp->rq_entry);
1418         }
1419         spin_unlock_irqrestore(&dev->flush_q_lock, flags);
1420 }
1421
1422 int ocrdma_destroy_qp(struct ib_qp *ibqp)
1423 {
1424         int status;
1425         struct ocrdma_pd *pd;
1426         struct ocrdma_qp *qp;
1427         struct ocrdma_dev *dev;
1428         struct ib_qp_attr attrs;
1429         int attr_mask = IB_QP_STATE;
1430         unsigned long flags;
1431
1432         qp = get_ocrdma_qp(ibqp);
1433         dev = qp->dev;
1434
1435         attrs.qp_state = IB_QPS_ERR;
1436         pd = qp->pd;
1437
1438         /* change the QP state to ERROR */
1439         _ocrdma_modify_qp(ibqp, &attrs, attr_mask);
1440
1441         /* ensure that CQEs for newly created QP (whose id may be same with
1442          * one which just getting destroyed are same), dont get
1443          * discarded until the old CQEs are discarded.
1444          */
1445         mutex_lock(&dev->dev_lock);
1446         status = ocrdma_mbx_destroy_qp(dev, qp);
1447
1448         /*
1449          * acquire CQ lock while destroy is in progress, in order to
1450          * protect against proessing in-flight CQEs for this QP.
1451          */
1452         spin_lock_irqsave(&qp->sq_cq->cq_lock, flags);
1453         if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1454                 spin_lock(&qp->rq_cq->cq_lock);
1455
1456         ocrdma_del_qpn_map(dev, qp);
1457
1458         if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1459                 spin_unlock(&qp->rq_cq->cq_lock);
1460         spin_unlock_irqrestore(&qp->sq_cq->cq_lock, flags);
1461
1462         if (!pd->uctx) {
1463                 ocrdma_discard_cqes(qp, qp->sq_cq);
1464                 ocrdma_discard_cqes(qp, qp->rq_cq);
1465         }
1466         mutex_unlock(&dev->dev_lock);
1467
1468         if (pd->uctx) {
1469                 ocrdma_del_mmap(pd->uctx, (u64) qp->sq.pa, qp->sq.len);
1470                 if (!qp->srq)
1471                         ocrdma_del_mmap(pd->uctx, (u64) qp->rq.pa, qp->rq.len);
1472         }
1473
1474         ocrdma_del_flush_qp(qp);
1475
1476         atomic_dec(&qp->pd->use_cnt);
1477         atomic_dec(&qp->sq_cq->use_cnt);
1478         atomic_dec(&qp->rq_cq->use_cnt);
1479         if (qp->srq)
1480                 atomic_dec(&qp->srq->use_cnt);
1481         kfree(qp->wqe_wr_id_tbl);
1482         kfree(qp->rqe_wr_id_tbl);
1483         kfree(qp);
1484         return status;
1485 }
1486
1487 static int ocrdma_copy_srq_uresp(struct ocrdma_srq *srq, struct ib_udata *udata)
1488 {
1489         int status;
1490         struct ocrdma_create_srq_uresp uresp;
1491
1492         uresp.rq_dbid = srq->rq.dbid;
1493         uresp.num_rq_pages = 1;
1494         uresp.rq_page_addr[0] = srq->rq.pa;
1495         uresp.rq_page_size = srq->rq.len;
1496         uresp.db_page_addr = srq->dev->nic_info.unmapped_db +
1497             (srq->pd->id * srq->dev->nic_info.db_page_size);
1498         uresp.db_page_size = srq->dev->nic_info.db_page_size;
1499         uresp.num_rqe_allocated = srq->rq.max_cnt;
1500         if (srq->dev->nic_info.dev_family == OCRDMA_GEN2_FAMILY) {
1501                 uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ1_OFFSET;
1502                 uresp.db_shift = 24;
1503         } else {
1504                 uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
1505                 uresp.db_shift = 16;
1506         }
1507
1508         status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1509         if (status)
1510                 return status;
1511         status = ocrdma_add_mmap(srq->pd->uctx, uresp.rq_page_addr[0],
1512                                  uresp.rq_page_size);
1513         if (status)
1514                 return status;
1515         return status;
1516 }
1517
1518 struct ib_srq *ocrdma_create_srq(struct ib_pd *ibpd,
1519                                  struct ib_srq_init_attr *init_attr,
1520                                  struct ib_udata *udata)
1521 {
1522         int status = -ENOMEM;
1523         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1524         struct ocrdma_dev *dev = pd->dev;
1525         struct ocrdma_srq *srq;
1526
1527         if (init_attr->attr.max_sge > dev->attr.max_recv_sge)
1528                 return ERR_PTR(-EINVAL);
1529         if (init_attr->attr.max_wr > dev->attr.max_rqe)
1530                 return ERR_PTR(-EINVAL);
1531
1532         srq = kzalloc(sizeof(*srq), GFP_KERNEL);
1533         if (!srq)
1534                 return ERR_PTR(status);
1535
1536         spin_lock_init(&srq->q_lock);
1537         srq->dev = dev;
1538         srq->pd = pd;
1539         srq->db = dev->nic_info.db + (pd->id * dev->nic_info.db_page_size);
1540         status = ocrdma_mbx_create_srq(srq, init_attr, pd);
1541         if (status)
1542                 goto err;
1543
1544         if (udata == NULL) {
1545                 srq->rqe_wr_id_tbl = kzalloc(sizeof(u64) * srq->rq.max_cnt,
1546                             GFP_KERNEL);
1547                 if (srq->rqe_wr_id_tbl == NULL)
1548                         goto arm_err;
1549
1550                 srq->bit_fields_len = (srq->rq.max_cnt / 32) +
1551                     (srq->rq.max_cnt % 32 ? 1 : 0);
1552                 srq->idx_bit_fields =
1553                     kmalloc(srq->bit_fields_len * sizeof(u32), GFP_KERNEL);
1554                 if (srq->idx_bit_fields == NULL)
1555                         goto arm_err;
1556                 memset(srq->idx_bit_fields, 0xff,
1557                        srq->bit_fields_len * sizeof(u32));
1558         }
1559
1560         if (init_attr->attr.srq_limit) {
1561                 status = ocrdma_mbx_modify_srq(srq, &init_attr->attr);
1562                 if (status)
1563                         goto arm_err;
1564         }
1565
1566         atomic_set(&srq->use_cnt, 0);
1567         if (udata) {
1568                 status = ocrdma_copy_srq_uresp(srq, udata);
1569                 if (status)
1570                         goto arm_err;
1571         }
1572
1573         atomic_inc(&pd->use_cnt);
1574         return &srq->ibsrq;
1575
1576 arm_err:
1577         ocrdma_mbx_destroy_srq(dev, srq);
1578 err:
1579         kfree(srq->rqe_wr_id_tbl);
1580         kfree(srq->idx_bit_fields);
1581         kfree(srq);
1582         return ERR_PTR(status);
1583 }
1584
1585 int ocrdma_modify_srq(struct ib_srq *ibsrq,
1586                       struct ib_srq_attr *srq_attr,
1587                       enum ib_srq_attr_mask srq_attr_mask,
1588                       struct ib_udata *udata)
1589 {
1590         int status = 0;
1591         struct ocrdma_srq *srq;
1592
1593         srq = get_ocrdma_srq(ibsrq);
1594         if (srq_attr_mask & IB_SRQ_MAX_WR)
1595                 status = -EINVAL;
1596         else
1597                 status = ocrdma_mbx_modify_srq(srq, srq_attr);
1598         return status;
1599 }
1600
1601 int ocrdma_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
1602 {
1603         int status;
1604         struct ocrdma_srq *srq;
1605
1606         srq = get_ocrdma_srq(ibsrq);
1607         status = ocrdma_mbx_query_srq(srq, srq_attr);
1608         return status;
1609 }
1610
1611 int ocrdma_destroy_srq(struct ib_srq *ibsrq)
1612 {
1613         int status;
1614         struct ocrdma_srq *srq;
1615         struct ocrdma_dev *dev;
1616
1617         srq = get_ocrdma_srq(ibsrq);
1618         dev = srq->dev;
1619         if (atomic_read(&srq->use_cnt)) {
1620                 ocrdma_err("%s(%d) err, srq=0x%x in use\n",
1621                            __func__, dev->id, srq->id);
1622                 return -EAGAIN;
1623         }
1624
1625         status = ocrdma_mbx_destroy_srq(dev, srq);
1626
1627         if (srq->pd->uctx)
1628                 ocrdma_del_mmap(srq->pd->uctx, (u64) srq->rq.pa, srq->rq.len);
1629
1630         atomic_dec(&srq->pd->use_cnt);
1631         kfree(srq->idx_bit_fields);
1632         kfree(srq->rqe_wr_id_tbl);
1633         kfree(srq);
1634         return status;
1635 }
1636
1637 /* unprivileged verbs and their support functions. */
1638 static void ocrdma_build_ud_hdr(struct ocrdma_qp *qp,
1639                                 struct ocrdma_hdr_wqe *hdr,
1640                                 struct ib_send_wr *wr)
1641 {
1642         struct ocrdma_ewqe_ud_hdr *ud_hdr =
1643                 (struct ocrdma_ewqe_ud_hdr *)(hdr + 1);
1644         struct ocrdma_ah *ah = get_ocrdma_ah(wr->wr.ud.ah);
1645
1646         ud_hdr->rsvd_dest_qpn = wr->wr.ud.remote_qpn;
1647         if (qp->qp_type == IB_QPT_GSI)
1648                 ud_hdr->qkey = qp->qkey;
1649         else
1650                 ud_hdr->qkey = wr->wr.ud.remote_qkey;
1651         ud_hdr->rsvd_ahid = ah->id;
1652 }
1653
1654 static void ocrdma_build_sges(struct ocrdma_hdr_wqe *hdr,
1655                               struct ocrdma_sge *sge, int num_sge,
1656                               struct ib_sge *sg_list)
1657 {
1658         int i;
1659
1660         for (i = 0; i < num_sge; i++) {
1661                 sge[i].lrkey = sg_list[i].lkey;
1662                 sge[i].addr_lo = sg_list[i].addr;
1663                 sge[i].addr_hi = upper_32_bits(sg_list[i].addr);
1664                 sge[i].len = sg_list[i].length;
1665                 hdr->total_len += sg_list[i].length;
1666         }
1667         if (num_sge == 0)
1668                 memset(sge, 0, sizeof(*sge));
1669 }
1670
1671 static int ocrdma_build_inline_sges(struct ocrdma_qp *qp,
1672                                     struct ocrdma_hdr_wqe *hdr,
1673                                     struct ocrdma_sge *sge,
1674                                     struct ib_send_wr *wr, u32 wqe_size)
1675 {
1676         if (wr->send_flags & IB_SEND_INLINE) {
1677                 if (wr->sg_list[0].length > qp->max_inline_data) {
1678                         ocrdma_err("%s() supported_len=0x%x,"
1679                                 " unspported len req=0x%x\n", __func__,
1680                                 qp->max_inline_data, wr->sg_list[0].length);
1681                         return -EINVAL;
1682                 }
1683                 memcpy(sge,
1684                        (void *)(unsigned long)wr->sg_list[0].addr,
1685                        wr->sg_list[0].length);
1686                 hdr->total_len = wr->sg_list[0].length;
1687                 wqe_size += roundup(hdr->total_len, OCRDMA_WQE_ALIGN_BYTES);
1688                 hdr->cw |= (OCRDMA_TYPE_INLINE << OCRDMA_WQE_TYPE_SHIFT);
1689         } else {
1690                 ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
1691                 if (wr->num_sge)
1692                         wqe_size += (wr->num_sge * sizeof(struct ocrdma_sge));
1693                 else
1694                         wqe_size += sizeof(struct ocrdma_sge);
1695                 hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
1696         }
1697         hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
1698         return 0;
1699 }
1700
1701 static int ocrdma_build_send(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1702                              struct ib_send_wr *wr)
1703 {
1704         int status;
1705         struct ocrdma_sge *sge;
1706         u32 wqe_size = sizeof(*hdr);
1707
1708         if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
1709                 ocrdma_build_ud_hdr(qp, hdr, wr);
1710                 sge = (struct ocrdma_sge *)(hdr + 2);
1711                 wqe_size += sizeof(struct ocrdma_ewqe_ud_hdr);
1712         } else
1713                 sge = (struct ocrdma_sge *)(hdr + 1);
1714
1715         status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1716         return status;
1717 }
1718
1719 static int ocrdma_build_write(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1720                               struct ib_send_wr *wr)
1721 {
1722         int status;
1723         struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
1724         struct ocrdma_sge *sge = ext_rw + 1;
1725         u32 wqe_size = sizeof(*hdr) + sizeof(*ext_rw);
1726
1727         status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
1728         if (status)
1729                 return status;
1730         ext_rw->addr_lo = wr->wr.rdma.remote_addr;
1731         ext_rw->addr_hi = upper_32_bits(wr->wr.rdma.remote_addr);
1732         ext_rw->lrkey = wr->wr.rdma.rkey;
1733         ext_rw->len = hdr->total_len;
1734         return 0;
1735 }
1736
1737 static void ocrdma_build_read(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
1738                               struct ib_send_wr *wr)
1739 {
1740         struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
1741         struct ocrdma_sge *sge = ext_rw + 1;
1742         u32 wqe_size = ((wr->num_sge + 1) * sizeof(struct ocrdma_sge)) +
1743             sizeof(struct ocrdma_hdr_wqe);
1744
1745         ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
1746         hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
1747         hdr->cw |= (OCRDMA_READ << OCRDMA_WQE_OPCODE_SHIFT);
1748         hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
1749
1750         ext_rw->addr_lo = wr->wr.rdma.remote_addr;
1751         ext_rw->addr_hi = upper_32_bits(wr->wr.rdma.remote_addr);
1752         ext_rw->lrkey = wr->wr.rdma.rkey;
1753         ext_rw->len = hdr->total_len;
1754 }
1755
1756 static void ocrdma_ring_sq_db(struct ocrdma_qp *qp)
1757 {
1758         u32 val = qp->sq.dbid | (1 << 16);
1759
1760         iowrite32(val, qp->sq_db);
1761 }
1762
1763 int ocrdma_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1764                      struct ib_send_wr **bad_wr)
1765 {
1766         int status = 0;
1767         struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1768         struct ocrdma_hdr_wqe *hdr;
1769         unsigned long flags;
1770
1771         spin_lock_irqsave(&qp->q_lock, flags);
1772         if (qp->state != OCRDMA_QPS_RTS && qp->state != OCRDMA_QPS_SQD) {
1773                 spin_unlock_irqrestore(&qp->q_lock, flags);
1774                 return -EINVAL;
1775         }
1776
1777         while (wr) {
1778                 if (ocrdma_hwq_free_cnt(&qp->sq) == 0 ||
1779                     wr->num_sge > qp->sq.max_sges) {
1780                         status = -ENOMEM;
1781                         break;
1782                 }
1783                 hdr = ocrdma_hwq_head(&qp->sq);
1784                 hdr->cw = 0;
1785                 if (wr->send_flags & IB_SEND_SIGNALED)
1786                         hdr->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
1787                 if (wr->send_flags & IB_SEND_FENCE)
1788                         hdr->cw |=
1789                             (OCRDMA_FLAG_FENCE_L << OCRDMA_WQE_FLAGS_SHIFT);
1790                 if (wr->send_flags & IB_SEND_SOLICITED)
1791                         hdr->cw |=
1792                             (OCRDMA_FLAG_SOLICIT << OCRDMA_WQE_FLAGS_SHIFT);
1793                 hdr->total_len = 0;
1794                 switch (wr->opcode) {
1795                 case IB_WR_SEND_WITH_IMM:
1796                         hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
1797                         hdr->immdt = ntohl(wr->ex.imm_data);
1798                 case IB_WR_SEND:
1799                         hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
1800                         ocrdma_build_send(qp, hdr, wr);
1801                         break;
1802                 case IB_WR_SEND_WITH_INV:
1803                         hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
1804                         hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
1805                         hdr->lkey = wr->ex.invalidate_rkey;
1806                         status = ocrdma_build_send(qp, hdr, wr);
1807                         break;
1808                 case IB_WR_RDMA_WRITE_WITH_IMM:
1809                         hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
1810                         hdr->immdt = ntohl(wr->ex.imm_data);
1811                 case IB_WR_RDMA_WRITE:
1812                         hdr->cw |= (OCRDMA_WRITE << OCRDMA_WQE_OPCODE_SHIFT);
1813                         status = ocrdma_build_write(qp, hdr, wr);
1814                         break;
1815                 case IB_WR_RDMA_READ_WITH_INV:
1816                         hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
1817                 case IB_WR_RDMA_READ:
1818                         ocrdma_build_read(qp, hdr, wr);
1819                         break;
1820                 case IB_WR_LOCAL_INV:
1821                         hdr->cw |=
1822                             (OCRDMA_LKEY_INV << OCRDMA_WQE_OPCODE_SHIFT);
1823                         hdr->cw |= (sizeof(struct ocrdma_hdr_wqe) /
1824                                 OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT;
1825                         hdr->lkey = wr->ex.invalidate_rkey;
1826                         break;
1827                 default:
1828                         status = -EINVAL;
1829                         break;
1830                 }
1831                 if (status) {
1832                         *bad_wr = wr;
1833                         break;
1834                 }
1835                 if (wr->send_flags & IB_SEND_SIGNALED)
1836                         qp->wqe_wr_id_tbl[qp->sq.head].signaled = 1;
1837                 else
1838                         qp->wqe_wr_id_tbl[qp->sq.head].signaled = 0;
1839                 qp->wqe_wr_id_tbl[qp->sq.head].wrid = wr->wr_id;
1840                 ocrdma_cpu_to_le32(hdr, ((hdr->cw >> OCRDMA_WQE_SIZE_SHIFT) &
1841                                    OCRDMA_WQE_SIZE_MASK) * OCRDMA_WQE_STRIDE);
1842                 /* make sure wqe is written before adapter can access it */
1843                 wmb();
1844                 /* inform hw to start processing it */
1845                 ocrdma_ring_sq_db(qp);
1846
1847                 /* update pointer, counter for next wr */
1848                 ocrdma_hwq_inc_head(&qp->sq);
1849                 wr = wr->next;
1850         }
1851         spin_unlock_irqrestore(&qp->q_lock, flags);
1852         return status;
1853 }
1854
1855 static void ocrdma_ring_rq_db(struct ocrdma_qp *qp)
1856 {
1857         u32 val = qp->rq.dbid | (1 << OCRDMA_GET_NUM_POSTED_SHIFT_VAL(qp));
1858
1859         iowrite32(val, qp->rq_db);
1860 }
1861
1862 static void ocrdma_build_rqe(struct ocrdma_hdr_wqe *rqe, struct ib_recv_wr *wr,
1863                              u16 tag)
1864 {
1865         u32 wqe_size = 0;
1866         struct ocrdma_sge *sge;
1867         if (wr->num_sge)
1868                 wqe_size = (wr->num_sge * sizeof(*sge)) + sizeof(*rqe);
1869         else
1870                 wqe_size = sizeof(*sge) + sizeof(*rqe);
1871
1872         rqe->cw = ((wqe_size / OCRDMA_WQE_STRIDE) <<
1873                                 OCRDMA_WQE_SIZE_SHIFT);
1874         rqe->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
1875         rqe->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
1876         rqe->total_len = 0;
1877         rqe->rsvd_tag = tag;
1878         sge = (struct ocrdma_sge *)(rqe + 1);
1879         ocrdma_build_sges(rqe, sge, wr->num_sge, wr->sg_list);
1880         ocrdma_cpu_to_le32(rqe, wqe_size);
1881 }
1882
1883 int ocrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1884                      struct ib_recv_wr **bad_wr)
1885 {
1886         int status = 0;
1887         unsigned long flags;
1888         struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1889         struct ocrdma_hdr_wqe *rqe;
1890
1891         spin_lock_irqsave(&qp->q_lock, flags);
1892         if (qp->state == OCRDMA_QPS_RST || qp->state == OCRDMA_QPS_ERR) {
1893                 spin_unlock_irqrestore(&qp->q_lock, flags);
1894                 *bad_wr = wr;
1895                 return -EINVAL;
1896         }
1897         while (wr) {
1898                 if (ocrdma_hwq_free_cnt(&qp->rq) == 0 ||
1899                     wr->num_sge > qp->rq.max_sges) {
1900                         *bad_wr = wr;
1901                         status = -ENOMEM;
1902                         break;
1903                 }
1904                 rqe = ocrdma_hwq_head(&qp->rq);
1905                 ocrdma_build_rqe(rqe, wr, 0);
1906
1907                 qp->rqe_wr_id_tbl[qp->rq.head] = wr->wr_id;
1908                 /* make sure rqe is written before adapter can access it */
1909                 wmb();
1910
1911                 /* inform hw to start processing it */
1912                 ocrdma_ring_rq_db(qp);
1913
1914                 /* update pointer, counter for next wr */
1915                 ocrdma_hwq_inc_head(&qp->rq);
1916                 wr = wr->next;
1917         }
1918         spin_unlock_irqrestore(&qp->q_lock, flags);
1919         return status;
1920 }
1921
1922 /* cqe for srq's rqe can potentially arrive out of order.
1923  * index gives the entry in the shadow table where to store
1924  * the wr_id. tag/index is returned in cqe to reference back
1925  * for a given rqe.
1926  */
1927 static int ocrdma_srq_get_idx(struct ocrdma_srq *srq)
1928 {
1929         int row = 0;
1930         int indx = 0;
1931
1932         for (row = 0; row < srq->bit_fields_len; row++) {
1933                 if (srq->idx_bit_fields[row]) {
1934                         indx = ffs(srq->idx_bit_fields[row]);
1935                         indx = (row * 32) + (indx - 1);
1936                         if (indx >= srq->rq.max_cnt)
1937                                 BUG();
1938                         ocrdma_srq_toggle_bit(srq, indx);
1939                         break;
1940                 }
1941         }
1942
1943         if (row == srq->bit_fields_len)
1944                 BUG();
1945         return indx;
1946 }
1947
1948 static void ocrdma_ring_srq_db(struct ocrdma_srq *srq)
1949 {
1950         u32 val = srq->rq.dbid | (1 << 16);
1951
1952         iowrite32(val, srq->db + OCRDMA_DB_GEN2_SRQ_OFFSET);
1953 }
1954
1955 int ocrdma_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
1956                          struct ib_recv_wr **bad_wr)
1957 {
1958         int status = 0;
1959         unsigned long flags;
1960         struct ocrdma_srq *srq;
1961         struct ocrdma_hdr_wqe *rqe;
1962         u16 tag;
1963
1964         srq = get_ocrdma_srq(ibsrq);
1965
1966         spin_lock_irqsave(&srq->q_lock, flags);
1967         while (wr) {
1968                 if (ocrdma_hwq_free_cnt(&srq->rq) == 0 ||
1969                     wr->num_sge > srq->rq.max_sges) {
1970                         status = -ENOMEM;
1971                         *bad_wr = wr;
1972                         break;
1973                 }
1974                 tag = ocrdma_srq_get_idx(srq);
1975                 rqe = ocrdma_hwq_head(&srq->rq);
1976                 ocrdma_build_rqe(rqe, wr, tag);
1977
1978                 srq->rqe_wr_id_tbl[tag] = wr->wr_id;
1979                 /* make sure rqe is written before adapter can perform DMA */
1980                 wmb();
1981                 /* inform hw to start processing it */
1982                 ocrdma_ring_srq_db(srq);
1983                 /* update pointer, counter for next wr */
1984                 ocrdma_hwq_inc_head(&srq->rq);
1985                 wr = wr->next;
1986         }
1987         spin_unlock_irqrestore(&srq->q_lock, flags);
1988         return status;
1989 }
1990
1991 static enum ib_wc_status ocrdma_to_ibwc_err(u16 status)
1992 {
1993         enum ib_wc_status ibwc_status = IB_WC_GENERAL_ERR;
1994
1995         switch (status) {
1996         case OCRDMA_CQE_GENERAL_ERR:
1997                 ibwc_status = IB_WC_GENERAL_ERR;
1998                 break;
1999         case OCRDMA_CQE_LOC_LEN_ERR:
2000                 ibwc_status = IB_WC_LOC_LEN_ERR;
2001                 break;
2002         case OCRDMA_CQE_LOC_QP_OP_ERR:
2003                 ibwc_status = IB_WC_LOC_QP_OP_ERR;
2004                 break;
2005         case OCRDMA_CQE_LOC_EEC_OP_ERR:
2006                 ibwc_status = IB_WC_LOC_EEC_OP_ERR;
2007                 break;
2008         case OCRDMA_CQE_LOC_PROT_ERR:
2009                 ibwc_status = IB_WC_LOC_PROT_ERR;
2010                 break;
2011         case OCRDMA_CQE_WR_FLUSH_ERR:
2012                 ibwc_status = IB_WC_WR_FLUSH_ERR;
2013                 break;
2014         case OCRDMA_CQE_MW_BIND_ERR:
2015                 ibwc_status = IB_WC_MW_BIND_ERR;
2016                 break;
2017         case OCRDMA_CQE_BAD_RESP_ERR:
2018                 ibwc_status = IB_WC_BAD_RESP_ERR;
2019                 break;
2020         case OCRDMA_CQE_LOC_ACCESS_ERR:
2021                 ibwc_status = IB_WC_LOC_ACCESS_ERR;
2022                 break;
2023         case OCRDMA_CQE_REM_INV_REQ_ERR:
2024                 ibwc_status = IB_WC_REM_INV_REQ_ERR;
2025                 break;
2026         case OCRDMA_CQE_REM_ACCESS_ERR:
2027                 ibwc_status = IB_WC_REM_ACCESS_ERR;
2028                 break;
2029         case OCRDMA_CQE_REM_OP_ERR:
2030                 ibwc_status = IB_WC_REM_OP_ERR;
2031                 break;
2032         case OCRDMA_CQE_RETRY_EXC_ERR:
2033                 ibwc_status = IB_WC_RETRY_EXC_ERR;
2034                 break;
2035         case OCRDMA_CQE_RNR_RETRY_EXC_ERR:
2036                 ibwc_status = IB_WC_RNR_RETRY_EXC_ERR;
2037                 break;
2038         case OCRDMA_CQE_LOC_RDD_VIOL_ERR:
2039                 ibwc_status = IB_WC_LOC_RDD_VIOL_ERR;
2040                 break;
2041         case OCRDMA_CQE_REM_INV_RD_REQ_ERR:
2042                 ibwc_status = IB_WC_REM_INV_RD_REQ_ERR;
2043                 break;
2044         case OCRDMA_CQE_REM_ABORT_ERR:
2045                 ibwc_status = IB_WC_REM_ABORT_ERR;
2046                 break;
2047         case OCRDMA_CQE_INV_EECN_ERR:
2048                 ibwc_status = IB_WC_INV_EECN_ERR;
2049                 break;
2050         case OCRDMA_CQE_INV_EEC_STATE_ERR:
2051                 ibwc_status = IB_WC_INV_EEC_STATE_ERR;
2052                 break;
2053         case OCRDMA_CQE_FATAL_ERR:
2054                 ibwc_status = IB_WC_FATAL_ERR;
2055                 break;
2056         case OCRDMA_CQE_RESP_TIMEOUT_ERR:
2057                 ibwc_status = IB_WC_RESP_TIMEOUT_ERR;
2058                 break;
2059         default:
2060                 ibwc_status = IB_WC_GENERAL_ERR;
2061                 break;
2062         };
2063         return ibwc_status;
2064 }
2065
2066 static void ocrdma_update_wc(struct ocrdma_qp *qp, struct ib_wc *ibwc,
2067                       u32 wqe_idx)
2068 {
2069         struct ocrdma_hdr_wqe *hdr;
2070         struct ocrdma_sge *rw;
2071         int opcode;
2072
2073         hdr = ocrdma_hwq_head_from_idx(&qp->sq, wqe_idx);
2074
2075         ibwc->wr_id = qp->wqe_wr_id_tbl[wqe_idx].wrid;
2076         /* Undo the hdr->cw swap */
2077         opcode = le32_to_cpu(hdr->cw) & OCRDMA_WQE_OPCODE_MASK;
2078         switch (opcode) {
2079         case OCRDMA_WRITE:
2080                 ibwc->opcode = IB_WC_RDMA_WRITE;
2081                 break;
2082         case OCRDMA_READ:
2083                 rw = (struct ocrdma_sge *)(hdr + 1);
2084                 ibwc->opcode = IB_WC_RDMA_READ;
2085                 ibwc->byte_len = rw->len;
2086                 break;
2087         case OCRDMA_SEND:
2088                 ibwc->opcode = IB_WC_SEND;
2089                 break;
2090         case OCRDMA_LKEY_INV:
2091                 ibwc->opcode = IB_WC_LOCAL_INV;
2092                 break;
2093         default:
2094                 ibwc->status = IB_WC_GENERAL_ERR;
2095                 ocrdma_err("%s() invalid opcode received = 0x%x\n",
2096                            __func__, hdr->cw & OCRDMA_WQE_OPCODE_MASK);
2097                 break;
2098         };
2099 }
2100
2101 static void ocrdma_set_cqe_status_flushed(struct ocrdma_qp *qp,
2102                                                 struct ocrdma_cqe *cqe)
2103 {
2104         if (is_cqe_for_sq(cqe)) {
2105                 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2106                                 cqe->flags_status_srcqpn) &
2107                                         ~OCRDMA_CQE_STATUS_MASK);
2108                 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2109                                 cqe->flags_status_srcqpn) |
2110                                 (OCRDMA_CQE_WR_FLUSH_ERR <<
2111                                         OCRDMA_CQE_STATUS_SHIFT));
2112         } else {
2113                 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2114                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2115                                         cqe->flags_status_srcqpn) &
2116                                                 ~OCRDMA_CQE_UD_STATUS_MASK);
2117                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2118                                         cqe->flags_status_srcqpn) |
2119                                         (OCRDMA_CQE_WR_FLUSH_ERR <<
2120                                                 OCRDMA_CQE_UD_STATUS_SHIFT));
2121                 } else {
2122                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2123                                         cqe->flags_status_srcqpn) &
2124                                                 ~OCRDMA_CQE_STATUS_MASK);
2125                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2126                                         cqe->flags_status_srcqpn) |
2127                                         (OCRDMA_CQE_WR_FLUSH_ERR <<
2128                                                 OCRDMA_CQE_STATUS_SHIFT));
2129                 }
2130         }
2131 }
2132
2133 static bool ocrdma_update_err_cqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2134                                   struct ocrdma_qp *qp, int status)
2135 {
2136         bool expand = false;
2137
2138         ibwc->byte_len = 0;
2139         ibwc->qp = &qp->ibqp;
2140         ibwc->status = ocrdma_to_ibwc_err(status);
2141
2142         ocrdma_flush_qp(qp);
2143         ocrdma_qp_state_machine(qp, IB_QPS_ERR, NULL);
2144
2145         /* if wqe/rqe pending for which cqe needs to be returned,
2146          * trigger inflating it.
2147          */
2148         if (!is_hw_rq_empty(qp) || !is_hw_sq_empty(qp)) {
2149                 expand = true;
2150                 ocrdma_set_cqe_status_flushed(qp, cqe);
2151         }
2152         return expand;
2153 }
2154
2155 static int ocrdma_update_err_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2156                                   struct ocrdma_qp *qp, int status)
2157 {
2158         ibwc->opcode = IB_WC_RECV;
2159         ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2160         ocrdma_hwq_inc_tail(&qp->rq);
2161
2162         return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2163 }
2164
2165 static int ocrdma_update_err_scqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2166                                   struct ocrdma_qp *qp, int status)
2167 {
2168         ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2169         ocrdma_hwq_inc_tail(&qp->sq);
2170
2171         return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2172 }
2173
2174
2175 static bool ocrdma_poll_err_scqe(struct ocrdma_qp *qp,
2176                                  struct ocrdma_cqe *cqe, struct ib_wc *ibwc,
2177                                  bool *polled, bool *stop)
2178 {
2179         bool expand;
2180         int status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2181                 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2182
2183         /* when hw sq is empty, but rq is not empty, so we continue
2184          * to keep the cqe in order to get the cq event again.
2185          */
2186         if (is_hw_sq_empty(qp) && !is_hw_rq_empty(qp)) {
2187                 /* when cq for rq and sq is same, it is safe to return
2188                  * flush cqe for RQEs.
2189                  */
2190                 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2191                         *polled = true;
2192                         status = OCRDMA_CQE_WR_FLUSH_ERR;
2193                         expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2194                 } else {
2195                         /* stop processing further cqe as this cqe is used for
2196                          * triggering cq event on buddy cq of RQ.
2197                          * When QP is destroyed, this cqe will be removed
2198                          * from the cq's hardware q.
2199                          */
2200                         *polled = false;
2201                         *stop = true;
2202                         expand = false;
2203                 }
2204         } else {
2205                 *polled = true;
2206                 expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2207         }
2208         return expand;
2209 }
2210
2211 static bool ocrdma_poll_success_scqe(struct ocrdma_qp *qp,
2212                                      struct ocrdma_cqe *cqe,
2213                                      struct ib_wc *ibwc, bool *polled)
2214 {
2215         bool expand = false;
2216         int tail = qp->sq.tail;
2217         u32 wqe_idx;
2218
2219         if (!qp->wqe_wr_id_tbl[tail].signaled) {
2220                 expand = true;  /* CQE cannot be consumed yet */
2221                 *polled = false;    /* WC cannot be consumed yet */
2222         } else {
2223                 ibwc->status = IB_WC_SUCCESS;
2224                 ibwc->wc_flags = 0;
2225                 ibwc->qp = &qp->ibqp;
2226                 ocrdma_update_wc(qp, ibwc, tail);
2227                 *polled = true;
2228                 wqe_idx = le32_to_cpu(cqe->wq.wqeidx) & OCRDMA_CQE_WQEIDX_MASK;
2229                 if (tail != wqe_idx)
2230                         expand = true; /* Coalesced CQE can't be consumed yet */
2231         }
2232         ocrdma_hwq_inc_tail(&qp->sq);
2233         return expand;
2234 }
2235
2236 static bool ocrdma_poll_scqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2237                              struct ib_wc *ibwc, bool *polled, bool *stop)
2238 {
2239         int status;
2240         bool expand;
2241
2242         status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2243                 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2244
2245         if (status == OCRDMA_CQE_SUCCESS)
2246                 expand = ocrdma_poll_success_scqe(qp, cqe, ibwc, polled);
2247         else
2248                 expand = ocrdma_poll_err_scqe(qp, cqe, ibwc, polled, stop);
2249         return expand;
2250 }
2251
2252 static int ocrdma_update_ud_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe)
2253 {
2254         int status;
2255
2256         status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2257                 OCRDMA_CQE_UD_STATUS_MASK) >> OCRDMA_CQE_UD_STATUS_SHIFT;
2258         ibwc->src_qp = le32_to_cpu(cqe->flags_status_srcqpn) &
2259                                                 OCRDMA_CQE_SRCQP_MASK;
2260         ibwc->pkey_index = le32_to_cpu(cqe->ud.rxlen_pkey) &
2261                                                 OCRDMA_CQE_PKEY_MASK;
2262         ibwc->wc_flags = IB_WC_GRH;
2263         ibwc->byte_len = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2264                                         OCRDMA_CQE_UD_XFER_LEN_SHIFT);
2265         return status;
2266 }
2267
2268 static void ocrdma_update_free_srq_cqe(struct ib_wc *ibwc,
2269                                        struct ocrdma_cqe *cqe,
2270                                        struct ocrdma_qp *qp)
2271 {
2272         unsigned long flags;
2273         struct ocrdma_srq *srq;
2274         u32 wqe_idx;
2275
2276         srq = get_ocrdma_srq(qp->ibqp.srq);
2277         wqe_idx = le32_to_cpu(cqe->rq.buftag_qpn) >> OCRDMA_CQE_BUFTAG_SHIFT;
2278         ibwc->wr_id = srq->rqe_wr_id_tbl[wqe_idx];
2279         spin_lock_irqsave(&srq->q_lock, flags);
2280         ocrdma_srq_toggle_bit(srq, wqe_idx);
2281         spin_unlock_irqrestore(&srq->q_lock, flags);
2282         ocrdma_hwq_inc_tail(&srq->rq);
2283 }
2284
2285 static bool ocrdma_poll_err_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2286                                 struct ib_wc *ibwc, bool *polled, bool *stop,
2287                                 int status)
2288 {
2289         bool expand;
2290
2291         /* when hw_rq is empty, but wq is not empty, so continue
2292          * to keep the cqe to get the cq event again.
2293          */
2294         if (is_hw_rq_empty(qp) && !is_hw_sq_empty(qp)) {
2295                 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2296                         *polled = true;
2297                         status = OCRDMA_CQE_WR_FLUSH_ERR;
2298                         expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2299                 } else {
2300                         *polled = false;
2301                         *stop = true;
2302                         expand = false;
2303                 }
2304         } else {
2305                 *polled = true;
2306                 expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2307         }
2308         return expand;
2309 }
2310
2311 static void ocrdma_poll_success_rcqe(struct ocrdma_qp *qp,
2312                                      struct ocrdma_cqe *cqe, struct ib_wc *ibwc)
2313 {
2314         ibwc->opcode = IB_WC_RECV;
2315         ibwc->qp = &qp->ibqp;
2316         ibwc->status = IB_WC_SUCCESS;
2317
2318         if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI)
2319                 ocrdma_update_ud_rcqe(ibwc, cqe);
2320         else
2321                 ibwc->byte_len = le32_to_cpu(cqe->rq.rxlen);
2322
2323         if (is_cqe_imm(cqe)) {
2324                 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2325                 ibwc->wc_flags |= IB_WC_WITH_IMM;
2326         } else if (is_cqe_wr_imm(cqe)) {
2327                 ibwc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2328                 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2329                 ibwc->wc_flags |= IB_WC_WITH_IMM;
2330         } else if (is_cqe_invalidated(cqe)) {
2331                 ibwc->ex.invalidate_rkey = le32_to_cpu(cqe->rq.lkey_immdt);
2332                 ibwc->wc_flags |= IB_WC_WITH_INVALIDATE;
2333         }
2334         if (qp->ibqp.srq)
2335                 ocrdma_update_free_srq_cqe(ibwc, cqe, qp);
2336         else {
2337                 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2338                 ocrdma_hwq_inc_tail(&qp->rq);
2339         }
2340 }
2341
2342 static bool ocrdma_poll_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2343                              struct ib_wc *ibwc, bool *polled, bool *stop)
2344 {
2345         int status;
2346         bool expand = false;
2347
2348         ibwc->wc_flags = 0;
2349         if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI)
2350                 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2351                                         OCRDMA_CQE_UD_STATUS_MASK) >>
2352                                         OCRDMA_CQE_UD_STATUS_SHIFT;
2353         else
2354                 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2355                              OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2356
2357         if (status == OCRDMA_CQE_SUCCESS) {
2358                 *polled = true;
2359                 ocrdma_poll_success_rcqe(qp, cqe, ibwc);
2360         } else {
2361                 expand = ocrdma_poll_err_rcqe(qp, cqe, ibwc, polled, stop,
2362                                               status);
2363         }
2364         return expand;
2365 }
2366
2367 static void ocrdma_change_cq_phase(struct ocrdma_cq *cq, struct ocrdma_cqe *cqe,
2368                                    u16 cur_getp)
2369 {
2370         if (cq->phase_change) {
2371                 if (cur_getp == 0)
2372                         cq->phase = (~cq->phase & OCRDMA_CQE_VALID);
2373         } else
2374                 /* clear valid bit */
2375                 cqe->flags_status_srcqpn = 0;
2376 }
2377
2378 static int ocrdma_poll_hwcq(struct ocrdma_cq *cq, int num_entries,
2379                             struct ib_wc *ibwc)
2380 {
2381         u16 qpn = 0;
2382         int i = 0;
2383         bool expand = false;
2384         int polled_hw_cqes = 0;
2385         struct ocrdma_qp *qp = NULL;
2386         struct ocrdma_dev *dev = cq->dev;
2387         struct ocrdma_cqe *cqe;
2388         u16 cur_getp; bool polled = false; bool stop = false;
2389
2390         cur_getp = cq->getp;
2391         while (num_entries) {
2392                 cqe = cq->va + cur_getp;
2393                 /* check whether valid cqe or not */
2394                 if (!is_cqe_valid(cq, cqe))
2395                         break;
2396                 qpn = (le32_to_cpu(cqe->cmn.qpn) & OCRDMA_CQE_QPN_MASK);
2397                 /* ignore discarded cqe */
2398                 if (qpn == 0)
2399                         goto skip_cqe;
2400                 qp = dev->qp_tbl[qpn];
2401                 BUG_ON(qp == NULL);
2402
2403                 if (is_cqe_for_sq(cqe)) {
2404                         expand = ocrdma_poll_scqe(qp, cqe, ibwc, &polled,
2405                                                   &stop);
2406                 } else {
2407                         expand = ocrdma_poll_rcqe(qp, cqe, ibwc, &polled,
2408                                                   &stop);
2409                 }
2410                 if (expand)
2411                         goto expand_cqe;
2412                 if (stop)
2413                         goto stop_cqe;
2414                 /* clear qpn to avoid duplicate processing by discard_cqe() */
2415                 cqe->cmn.qpn = 0;
2416 skip_cqe:
2417                 polled_hw_cqes += 1;
2418                 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
2419                 ocrdma_change_cq_phase(cq, cqe, cur_getp);
2420 expand_cqe:
2421                 if (polled) {
2422                         num_entries -= 1;
2423                         i += 1;
2424                         ibwc = ibwc + 1;
2425                         polled = false;
2426                 }
2427         }
2428 stop_cqe:
2429         cq->getp = cur_getp;
2430         if (polled_hw_cqes || expand || stop) {
2431                 ocrdma_ring_cq_db(dev, cq->id, cq->armed, cq->solicited,
2432                                   polled_hw_cqes);
2433         }
2434         return i;
2435 }
2436
2437 /* insert error cqe if the QP's SQ or RQ's CQ matches the CQ under poll. */
2438 static int ocrdma_add_err_cqe(struct ocrdma_cq *cq, int num_entries,
2439                               struct ocrdma_qp *qp, struct ib_wc *ibwc)
2440 {
2441         int err_cqes = 0;
2442
2443         while (num_entries) {
2444                 if (is_hw_sq_empty(qp) && is_hw_rq_empty(qp))
2445                         break;
2446                 if (!is_hw_sq_empty(qp) && qp->sq_cq == cq) {
2447                         ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2448                         ocrdma_hwq_inc_tail(&qp->sq);
2449                 } else if (!is_hw_rq_empty(qp) && qp->rq_cq == cq) {
2450                         ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2451                         ocrdma_hwq_inc_tail(&qp->rq);
2452                 } else
2453                         return err_cqes;
2454                 ibwc->byte_len = 0;
2455                 ibwc->status = IB_WC_WR_FLUSH_ERR;
2456                 ibwc = ibwc + 1;
2457                 err_cqes += 1;
2458                 num_entries -= 1;
2459         }
2460         return err_cqes;
2461 }
2462
2463 int ocrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
2464 {
2465         int cqes_to_poll = num_entries;
2466         struct ocrdma_cq *cq = NULL;
2467         unsigned long flags;
2468         struct ocrdma_dev *dev;
2469         int num_os_cqe = 0, err_cqes = 0;
2470         struct ocrdma_qp *qp;
2471
2472         cq = get_ocrdma_cq(ibcq);
2473         dev = cq->dev;
2474
2475         /* poll cqes from adapter CQ */
2476         spin_lock_irqsave(&cq->cq_lock, flags);
2477         num_os_cqe = ocrdma_poll_hwcq(cq, cqes_to_poll, wc);
2478         spin_unlock_irqrestore(&cq->cq_lock, flags);
2479         cqes_to_poll -= num_os_cqe;
2480
2481         if (cqes_to_poll) {
2482                 wc = wc + num_os_cqe;
2483                 /* adapter returns single error cqe when qp moves to
2484                  * error state. So insert error cqes with wc_status as
2485                  * FLUSHED for pending WQEs and RQEs of QP's SQ and RQ
2486                  * respectively which uses this CQ.
2487                  */
2488                 spin_lock_irqsave(&dev->flush_q_lock, flags);
2489                 list_for_each_entry(qp, &cq->sq_head, sq_entry) {
2490                         if (cqes_to_poll == 0)
2491                                 break;
2492                         err_cqes = ocrdma_add_err_cqe(cq, cqes_to_poll, qp, wc);
2493                         cqes_to_poll -= err_cqes;
2494                         num_os_cqe += err_cqes;
2495                         wc = wc + err_cqes;
2496                 }
2497                 spin_unlock_irqrestore(&dev->flush_q_lock, flags);
2498         }
2499         return num_os_cqe;
2500 }
2501
2502 int ocrdma_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags cq_flags)
2503 {
2504         struct ocrdma_cq *cq;
2505         unsigned long flags;
2506         struct ocrdma_dev *dev;
2507         u16 cq_id;
2508         u16 cur_getp;
2509         struct ocrdma_cqe *cqe;
2510
2511         cq = get_ocrdma_cq(ibcq);
2512         cq_id = cq->id;
2513         dev = cq->dev;
2514
2515         spin_lock_irqsave(&cq->cq_lock, flags);
2516         if (cq_flags & IB_CQ_NEXT_COMP || cq_flags & IB_CQ_SOLICITED)
2517                 cq->armed = true;
2518         if (cq_flags & IB_CQ_SOLICITED)
2519                 cq->solicited = true;
2520
2521         cur_getp = cq->getp;
2522         cqe = cq->va + cur_getp;
2523
2524         /* check whether any valid cqe exist or not, if not then safe to
2525          * arm. If cqe is not yet consumed, then let it get consumed and then
2526          * we arm it to avoid false interrupts.
2527          */
2528         if (!is_cqe_valid(cq, cqe) || cq->arm_needed) {
2529                 cq->arm_needed = false;
2530                 ocrdma_ring_cq_db(dev, cq_id, cq->armed, cq->solicited, 0);
2531         }
2532         spin_unlock_irqrestore(&cq->cq_lock, flags);
2533         return 0;
2534 }