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