drivers/infiniband/hw/ocrdma/ocrdma_verbs.c

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