drivers/infiniband/hw/ocrdma/ocrdma_verbs.c

   1 /* This file is part of the Emulex RoCE Device Driver for
   2  * RoCE (RDMA over Converged Ethernet) adapters.
   3  * Copyright (C) 2012-2015 Emulex. All rights reserved.
   4  * EMULEX and SLI are trademarks of Emulex.
   5  * www.emulex.com
   6  *
   7  * This software is available to you under a choice of one of two licenses.
   8  * You may choose to be licensed under the terms of the GNU General Public
   9  * License (GPL) Version 2, available from the file COPYING in the main
  10  * directory of this source tree, or the BSD license below:
  11  *
  12  * Redistribution and use in source and binary forms, with or without
  13  * modification, are permitted provided that the following conditions
  14  * are met:
  15  *
  16  * - Redistributions of source code must retain the above copyright notice,
  17  *   this list of conditions and the following disclaimer.
  18  *
  19  * - Redistributions in binary form must reproduce the above copyright
  20  *   notice, this list of conditions and the following disclaimer in
  21  *   the documentation and/or other materials provided with the distribution.
  22  *
  23  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
  24  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,THE
  25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  26  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
  27  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
  28  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
  29  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
  30  * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
  31  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
  32  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
  33  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  34  *
  35  * Contact Information:
  36  * linux-drivers@emulex.com
  37  *
  38  * Emulex
  39  * 3333 Susan Street
  40  * Costa Mesa, CA 92626
  41  */
  42
  43 #include <linux/dma-mapping.h>
  44 #include <rdma/ib_verbs.h>
  45 #include <rdma/ib_user_verbs.h>
  46 #include <rdma/iw_cm.h>
  47 #include <rdma/ib_umem.h>
  48 #include <rdma/ib_addr.h>
  49 #include <rdma/ib_cache.h>
  50
  51 #include "ocrdma.h"
  52 #include "ocrdma_hw.h"
  53 #include "ocrdma_verbs.h"
  54 #include <rdma/ocrdma-abi.h>
  55
  56 int ocrdma_query_pkey(struct ib_device *ibdev, u8 port, u16 index, u16 *pkey)
  57 {
  58         if (index > 1)
  59                 return -EINVAL;
  60
  61         *pkey = 0xffff;
  62         return 0;
  63 }
  64
  65 int ocrdma_query_gid(struct ib_device *ibdev, u8 port,
  66                      int index, union ib_gid *sgid)
  67 {
  68         int ret;
  69         struct ocrdma_dev *dev;
  70
  71         dev = get_ocrdma_dev(ibdev);
  72         memset(sgid, 0, sizeof(*sgid));
  73         if (index >= OCRDMA_MAX_SGID)
  74                 return -EINVAL;
  75
  76         ret = ib_get_cached_gid(ibdev, port, index, sgid, NULL);
  77         if (ret == -EAGAIN) {
  78                 memcpy(sgid, &zgid, sizeof(*sgid));
  79                 return 0;
  80         }
  81
  82         return ret;
  83 }
  84
  85 int ocrdma_add_gid(struct ib_device *device,
  86                    u8 port_num,
  87                    unsigned int index,
  88                    const union ib_gid *gid,
  89                    const struct ib_gid_attr *attr,
  90                    void **context) {
  91         return  0;
  92 }
  93
  94 int  ocrdma_del_gid(struct ib_device *device,
  95                     u8 port_num,
  96                     unsigned int index,
  97                     void **context) {
  98         return 0;
  99 }
 100
 101 int ocrdma_query_device(struct ib_device *ibdev, struct ib_device_attr *attr,
 102                         struct ib_udata *uhw)
 103 {
 104         struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
 105
 106         if (uhw->inlen || uhw->outlen)
 107                 return -EINVAL;
 108
 109         memset(attr, 0, sizeof *attr);
 110         memcpy(&attr->fw_ver, &dev->attr.fw_ver[0],
 111                min(sizeof(dev->attr.fw_ver), sizeof(attr->fw_ver)));
 112         ocrdma_get_guid(dev, (u8 *)&attr->sys_image_guid);
 113         attr->max_mr_size = dev->attr.max_mr_size;
 114         attr->page_size_cap = 0xffff000;
 115         attr->vendor_id = dev->nic_info.pdev->vendor;
 116         attr->vendor_part_id = dev->nic_info.pdev->device;
 117         attr->hw_ver = dev->asic_id;
 118         attr->max_qp = dev->attr.max_qp;
 119         attr->max_ah = OCRDMA_MAX_AH;
 120         attr->max_qp_wr = dev->attr.max_wqe;
 121
 122         attr->device_cap_flags = IB_DEVICE_CURR_QP_STATE_MOD |
 123                                         IB_DEVICE_RC_RNR_NAK_GEN |
 124                                         IB_DEVICE_SHUTDOWN_PORT |
 125                                         IB_DEVICE_SYS_IMAGE_GUID |
 126                                         IB_DEVICE_LOCAL_DMA_LKEY |
 127                                         IB_DEVICE_MEM_MGT_EXTENSIONS;
 128         attr->max_sge = min(dev->attr.max_send_sge, dev->attr.max_recv_sge);
 129         attr->max_sge_rd = dev->attr.max_rdma_sge;
 130         attr->max_cq = dev->attr.max_cq;
 131         attr->max_cqe = dev->attr.max_cqe;
 132         attr->max_mr = dev->attr.max_mr;
 133         attr->max_mw = dev->attr.max_mw;
 134         attr->max_pd = dev->attr.max_pd;
 135         attr->atomic_cap = 0;
 136         attr->max_fmr = 0;
 137         attr->max_map_per_fmr = 0;
 138         attr->max_qp_rd_atom =
 139             min(dev->attr.max_ord_per_qp, dev->attr.max_ird_per_qp);
 140         attr->max_qp_init_rd_atom = dev->attr.max_ord_per_qp;
 141         attr->max_srq = dev->attr.max_srq;
 142         attr->max_srq_sge = dev->attr.max_srq_sge;
 143         attr->max_srq_wr = dev->attr.max_rqe;
 144         attr->local_ca_ack_delay = dev->attr.local_ca_ack_delay;
 145         attr->max_fast_reg_page_list_len = dev->attr.max_pages_per_frmr;
 146         attr->max_pkeys = 1;
 147         return 0;
 148 }
 149
 150 struct net_device *ocrdma_get_netdev(struct ib_device *ibdev, u8 port_num)
 151 {
 152         struct ocrdma_dev *dev;
 153         struct net_device *ndev = NULL;
 154
 155         rcu_read_lock();
 156
 157         dev = get_ocrdma_dev(ibdev);
 158         if (dev)
 159                 ndev = dev->nic_info.netdev;
 160         if (ndev)
 161                 dev_hold(ndev);
 162
 163         rcu_read_unlock();
 164
 165         return ndev;
 166 }
 167
 168 static inline void get_link_speed_and_width(struct ocrdma_dev *dev,
 169                                             u8 *ib_speed, u8 *ib_width)
 170 {
 171         int status;
 172         u8 speed;
 173
 174         status = ocrdma_mbx_get_link_speed(dev, &speed, NULL);
 175         if (status)
 176                 speed = OCRDMA_PHYS_LINK_SPEED_ZERO;
 177
 178         switch (speed) {
 179         case OCRDMA_PHYS_LINK_SPEED_1GBPS:
 180                 *ib_speed = IB_SPEED_SDR;
 181                 *ib_width = IB_WIDTH_1X;
 182                 break;
 183
 184         case OCRDMA_PHYS_LINK_SPEED_10GBPS:
 185                 *ib_speed = IB_SPEED_QDR;
 186                 *ib_width = IB_WIDTH_1X;
 187                 break;
 188
 189         case OCRDMA_PHYS_LINK_SPEED_20GBPS:
 190                 *ib_speed = IB_SPEED_DDR;
 191                 *ib_width = IB_WIDTH_4X;
 192                 break;
 193
 194         case OCRDMA_PHYS_LINK_SPEED_40GBPS:
 195                 *ib_speed = IB_SPEED_QDR;
 196                 *ib_width = IB_WIDTH_4X;
 197                 break;
 198
 199         default:
 200                 /* Unsupported */
 201                 *ib_speed = IB_SPEED_SDR;
 202                 *ib_width = IB_WIDTH_1X;
 203         }
 204 }
 205
 206 int ocrdma_query_port(struct ib_device *ibdev,
 207                       u8 port, struct ib_port_attr *props)
 208 {
 209         enum ib_port_state port_state;
 210         struct ocrdma_dev *dev;
 211         struct net_device *netdev;
 212
 213         /* props being zeroed by the caller, avoid zeroing it here */
 214         dev = get_ocrdma_dev(ibdev);
 215         if (port > 1) {
 216                 pr_err("%s(%d) invalid_port=0x%x\n", __func__,
 217                        dev->id, port);
 218                 return -EINVAL;
 219         }
 220         netdev = dev->nic_info.netdev;
 221         if (netif_running(netdev) && netif_oper_up(netdev)) {
 222                 port_state = IB_PORT_ACTIVE;
 223                 props->phys_state = 5;
 224         } else {
 225                 port_state = IB_PORT_DOWN;
 226                 props->phys_state = 3;
 227         }
 228         props->max_mtu = IB_MTU_4096;
 229         props->active_mtu = iboe_get_mtu(netdev->mtu);
 230         props->lid = 0;
 231         props->lmc = 0;
 232         props->sm_lid = 0;
 233         props->sm_sl = 0;
 234         props->state = port_state;
 235         props->port_cap_flags =
 236             IB_PORT_CM_SUP |
 237             IB_PORT_REINIT_SUP |
 238             IB_PORT_DEVICE_MGMT_SUP | IB_PORT_VENDOR_CLASS_SUP |
 239             IB_PORT_IP_BASED_GIDS;
 240         props->gid_tbl_len = OCRDMA_MAX_SGID;
 241         props->pkey_tbl_len = 1;
 242         props->bad_pkey_cntr = 0;
 243         props->qkey_viol_cntr = 0;
 244         get_link_speed_and_width(dev, &props->active_speed,
 245                                  &props->active_width);
 246         props->max_msg_sz = 0x80000000;
 247         props->max_vl_num = 4;
 248         return 0;
 249 }
 250
 251 int ocrdma_modify_port(struct ib_device *ibdev, u8 port, int mask,
 252                        struct ib_port_modify *props)
 253 {
 254         struct ocrdma_dev *dev;
 255
 256         dev = get_ocrdma_dev(ibdev);
 257         if (port > 1) {
 258                 pr_err("%s(%d) invalid_port=0x%x\n", __func__, dev->id, port);
 259                 return -EINVAL;
 260         }
 261         return 0;
 262 }
 263
 264 static int ocrdma_add_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
 265                            unsigned long len)
 266 {
 267         struct ocrdma_mm *mm;
 268
 269         mm = kzalloc(sizeof(*mm), GFP_KERNEL);
 270         if (mm == NULL)
 271                 return -ENOMEM;
 272         mm->key.phy_addr = phy_addr;
 273         mm->key.len = len;
 274         INIT_LIST_HEAD(&mm->entry);
 275
 276         mutex_lock(&uctx->mm_list_lock);
 277         list_add_tail(&mm->entry, &uctx->mm_head);
 278         mutex_unlock(&uctx->mm_list_lock);
 279         return 0;
 280 }
 281
 282 static void ocrdma_del_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
 283                             unsigned long len)
 284 {
 285         struct ocrdma_mm *mm, *tmp;
 286
 287         mutex_lock(&uctx->mm_list_lock);
 288         list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
 289                 if (len != mm->key.len && phy_addr != mm->key.phy_addr)
 290                         continue;
 291
 292                 list_del(&mm->entry);
 293                 kfree(mm);
 294                 break;
 295         }
 296         mutex_unlock(&uctx->mm_list_lock);
 297 }
 298
 299 static bool ocrdma_search_mmap(struct ocrdma_ucontext *uctx, u64 phy_addr,
 300                               unsigned long len)
 301 {
 302         bool found = false;
 303         struct ocrdma_mm *mm;
 304
 305         mutex_lock(&uctx->mm_list_lock);
 306         list_for_each_entry(mm, &uctx->mm_head, entry) {
 307                 if (len != mm->key.len && phy_addr != mm->key.phy_addr)
 308                         continue;
 309
 310                 found = true;
 311                 break;
 312         }
 313         mutex_unlock(&uctx->mm_list_lock);
 314         return found;
 315 }
 316
 317
 318 static u16 _ocrdma_pd_mgr_get_bitmap(struct ocrdma_dev *dev, bool dpp_pool)
 319 {
 320         u16 pd_bitmap_idx = 0;
 321         const unsigned long *pd_bitmap;
 322
 323         if (dpp_pool) {
 324                 pd_bitmap = dev->pd_mgr->pd_dpp_bitmap;
 325                 pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
 326                                                     dev->pd_mgr->max_dpp_pd);
 327                 __set_bit(pd_bitmap_idx, dev->pd_mgr->pd_dpp_bitmap);
 328                 dev->pd_mgr->pd_dpp_count++;
 329                 if (dev->pd_mgr->pd_dpp_count > dev->pd_mgr->pd_dpp_thrsh)
 330                         dev->pd_mgr->pd_dpp_thrsh = dev->pd_mgr->pd_dpp_count;
 331         } else {
 332                 pd_bitmap = dev->pd_mgr->pd_norm_bitmap;
 333                 pd_bitmap_idx = find_first_zero_bit(pd_bitmap,
 334                                                     dev->pd_mgr->max_normal_pd);
 335                 __set_bit(pd_bitmap_idx, dev->pd_mgr->pd_norm_bitmap);
 336                 dev->pd_mgr->pd_norm_count++;
 337                 if (dev->pd_mgr->pd_norm_count > dev->pd_mgr->pd_norm_thrsh)
 338                         dev->pd_mgr->pd_norm_thrsh = dev->pd_mgr->pd_norm_count;
 339         }
 340         return pd_bitmap_idx;
 341 }
 342
 343 static int _ocrdma_pd_mgr_put_bitmap(struct ocrdma_dev *dev, u16 pd_id,
 344                                         bool dpp_pool)
 345 {
 346         u16 pd_count;
 347         u16 pd_bit_index;
 348
 349         pd_count = dpp_pool ? dev->pd_mgr->pd_dpp_count :
 350                               dev->pd_mgr->pd_norm_count;
 351         if (pd_count == 0)
 352                 return -EINVAL;
 353
 354         if (dpp_pool) {
 355                 pd_bit_index = pd_id - dev->pd_mgr->pd_dpp_start;
 356                 if (pd_bit_index >= dev->pd_mgr->max_dpp_pd) {
 357                         return -EINVAL;
 358                 } else {
 359                         __clear_bit(pd_bit_index, dev->pd_mgr->pd_dpp_bitmap);
 360                         dev->pd_mgr->pd_dpp_count--;
 361                 }
 362         } else {
 363                 pd_bit_index = pd_id - dev->pd_mgr->pd_norm_start;
 364                 if (pd_bit_index >= dev->pd_mgr->max_normal_pd) {
 365                         return -EINVAL;
 366                 } else {
 367                         __clear_bit(pd_bit_index, dev->pd_mgr->pd_norm_bitmap);
 368                         dev->pd_mgr->pd_norm_count--;
 369                 }
 370         }
 371
 372         return 0;
 373 }
 374
 375 static int ocrdma_put_pd_num(struct ocrdma_dev *dev, u16 pd_id,
 376                                    bool dpp_pool)
 377 {
 378         int status;
 379
 380         mutex_lock(&dev->dev_lock);
 381         status = _ocrdma_pd_mgr_put_bitmap(dev, pd_id, dpp_pool);
 382         mutex_unlock(&dev->dev_lock);
 383         return status;
 384 }
 385
 386 static int ocrdma_get_pd_num(struct ocrdma_dev *dev, struct ocrdma_pd *pd)
 387 {
 388         u16 pd_idx = 0;
 389         int status = 0;
 390
 391         mutex_lock(&dev->dev_lock);
 392         if (pd->dpp_enabled) {
 393                 /* try allocating DPP PD, if not available then normal PD */
 394                 if (dev->pd_mgr->pd_dpp_count < dev->pd_mgr->max_dpp_pd) {
 395                         pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, true);
 396                         pd->id = dev->pd_mgr->pd_dpp_start + pd_idx;
 397                         pd->dpp_page = dev->pd_mgr->dpp_page_index + pd_idx;
 398                 } else if (dev->pd_mgr->pd_norm_count <
 399                            dev->pd_mgr->max_normal_pd) {
 400                         pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
 401                         pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
 402                         pd->dpp_enabled = false;
 403                 } else {
 404                         status = -EINVAL;
 405                 }
 406         } else {
 407                 if (dev->pd_mgr->pd_norm_count < dev->pd_mgr->max_normal_pd) {
 408                         pd_idx = _ocrdma_pd_mgr_get_bitmap(dev, false);
 409                         pd->id = dev->pd_mgr->pd_norm_start + pd_idx;
 410                 } else {
 411                         status = -EINVAL;
 412                 }
 413         }
 414         mutex_unlock(&dev->dev_lock);
 415         return status;
 416 }
 417
 418 static struct ocrdma_pd *_ocrdma_alloc_pd(struct ocrdma_dev *dev,
 419                                           struct ocrdma_ucontext *uctx,
 420                                           struct ib_udata *udata)
 421 {
 422         struct ocrdma_pd *pd = NULL;
 423         int status;
 424
 425         pd = kzalloc(sizeof(*pd), GFP_KERNEL);
 426         if (!pd)
 427                 return ERR_PTR(-ENOMEM);
 428
 429         if (udata && uctx && dev->attr.max_dpp_pds) {
 430                 pd->dpp_enabled =
 431                         ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R;
 432                 pd->num_dpp_qp =
 433                         pd->dpp_enabled ? (dev->nic_info.db_page_size /
 434                                            dev->attr.wqe_size) : 0;
 435         }
 436
 437         if (dev->pd_mgr->pd_prealloc_valid) {
 438                 status = ocrdma_get_pd_num(dev, pd);
 439                 if (status == 0) {
 440                         return pd;
 441                 } else {
 442                         kfree(pd);
 443                         return ERR_PTR(status);
 444                 }
 445         }
 446
 447 retry:
 448         status = ocrdma_mbx_alloc_pd(dev, pd);
 449         if (status) {
 450                 if (pd->dpp_enabled) {
 451                         pd->dpp_enabled = false;
 452                         pd->num_dpp_qp = 0;
 453                         goto retry;
 454                 } else {
 455                         kfree(pd);
 456                         return ERR_PTR(status);
 457                 }
 458         }
 459
 460         return pd;
 461 }
 462
 463 static inline int is_ucontext_pd(struct ocrdma_ucontext *uctx,
 464                                  struct ocrdma_pd *pd)
 465 {
 466         return (uctx->cntxt_pd == pd ? true : false);
 467 }
 468
 469 static int _ocrdma_dealloc_pd(struct ocrdma_dev *dev,
 470                               struct ocrdma_pd *pd)
 471 {
 472         int status;
 473
 474         if (dev->pd_mgr->pd_prealloc_valid)
 475                 status = ocrdma_put_pd_num(dev, pd->id, pd->dpp_enabled);
 476         else
 477                 status = ocrdma_mbx_dealloc_pd(dev, pd);
 478
 479         kfree(pd);
 480         return status;
 481 }
 482
 483 static int ocrdma_alloc_ucontext_pd(struct ocrdma_dev *dev,
 484                                     struct ocrdma_ucontext *uctx,
 485                                     struct ib_udata *udata)
 486 {
 487         int status = 0;
 488
 489         uctx->cntxt_pd = _ocrdma_alloc_pd(dev, uctx, udata);
 490         if (IS_ERR(uctx->cntxt_pd)) {
 491                 status = PTR_ERR(uctx->cntxt_pd);
 492                 uctx->cntxt_pd = NULL;
 493                 goto err;
 494         }
 495
 496         uctx->cntxt_pd->uctx = uctx;
 497         uctx->cntxt_pd->ibpd.device = &dev->ibdev;
 498 err:
 499         return status;
 500 }
 501
 502 static int ocrdma_dealloc_ucontext_pd(struct ocrdma_ucontext *uctx)
 503 {
 504         struct ocrdma_pd *pd = uctx->cntxt_pd;
 505         struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
 506
 507         if (uctx->pd_in_use) {
 508                 pr_err("%s(%d) Freeing in use pdid=0x%x.\n",
 509                        __func__, dev->id, pd->id);
 510         }
 511         uctx->cntxt_pd = NULL;
 512         (void)_ocrdma_dealloc_pd(dev, pd);
 513         return 0;
 514 }
 515
 516 static struct ocrdma_pd *ocrdma_get_ucontext_pd(struct ocrdma_ucontext *uctx)
 517 {
 518         struct ocrdma_pd *pd = NULL;
 519
 520         mutex_lock(&uctx->mm_list_lock);
 521         if (!uctx->pd_in_use) {
 522                 uctx->pd_in_use = true;
 523                 pd = uctx->cntxt_pd;
 524         }
 525         mutex_unlock(&uctx->mm_list_lock);
 526
 527         return pd;
 528 }
 529
 530 static void ocrdma_release_ucontext_pd(struct ocrdma_ucontext *uctx)
 531 {
 532         mutex_lock(&uctx->mm_list_lock);
 533         uctx->pd_in_use = false;
 534         mutex_unlock(&uctx->mm_list_lock);
 535 }
 536
 537 struct ib_ucontext *ocrdma_alloc_ucontext(struct ib_device *ibdev,
 538                                           struct ib_udata *udata)
 539 {
 540         int status;
 541         struct ocrdma_ucontext *ctx;
 542         struct ocrdma_alloc_ucontext_resp resp;
 543         struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
 544         struct pci_dev *pdev = dev->nic_info.pdev;
 545         u32 map_len = roundup(sizeof(u32) * 2048, PAGE_SIZE);
 546
 547         if (!udata)
 548                 return ERR_PTR(-EFAULT);
 549         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
 550         if (!ctx)
 551                 return ERR_PTR(-ENOMEM);
 552         INIT_LIST_HEAD(&ctx->mm_head);
 553         mutex_init(&ctx->mm_list_lock);
 554
 555         ctx->ah_tbl.va = dma_alloc_coherent(&pdev->dev, map_len,
 556                                             &ctx->ah_tbl.pa, GFP_KERNEL);
 557         if (!ctx->ah_tbl.va) {
 558                 kfree(ctx);
 559                 return ERR_PTR(-ENOMEM);
 560         }
 561         memset(ctx->ah_tbl.va, 0, map_len);
 562         ctx->ah_tbl.len = map_len;
 563
 564         memset(&resp, 0, sizeof(resp));
 565         resp.ah_tbl_len = ctx->ah_tbl.len;
 566         resp.ah_tbl_page = virt_to_phys(ctx->ah_tbl.va);
 567
 568         status = ocrdma_add_mmap(ctx, resp.ah_tbl_page, resp.ah_tbl_len);
 569         if (status)
 570                 goto map_err;
 571
 572         status = ocrdma_alloc_ucontext_pd(dev, ctx, udata);
 573         if (status)
 574                 goto pd_err;
 575
 576         resp.dev_id = dev->id;
 577         resp.max_inline_data = dev->attr.max_inline_data;
 578         resp.wqe_size = dev->attr.wqe_size;
 579         resp.rqe_size = dev->attr.rqe_size;
 580         resp.dpp_wqe_size = dev->attr.wqe_size;
 581
 582         memcpy(resp.fw_ver, dev->attr.fw_ver, sizeof(resp.fw_ver));
 583         status = ib_copy_to_udata(udata, &resp, sizeof(resp));
 584         if (status)
 585                 goto cpy_err;
 586         return &ctx->ibucontext;
 587
 588 cpy_err:
 589 pd_err:
 590         ocrdma_del_mmap(ctx, ctx->ah_tbl.pa, ctx->ah_tbl.len);
 591 map_err:
 592         dma_free_coherent(&pdev->dev, ctx->ah_tbl.len, ctx->ah_tbl.va,
 593                           ctx->ah_tbl.pa);
 594         kfree(ctx);
 595         return ERR_PTR(status);
 596 }
 597
 598 int ocrdma_dealloc_ucontext(struct ib_ucontext *ibctx)
 599 {
 600         int status;
 601         struct ocrdma_mm *mm, *tmp;
 602         struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ibctx);
 603         struct ocrdma_dev *dev = get_ocrdma_dev(ibctx->device);
 604         struct pci_dev *pdev = dev->nic_info.pdev;
 605
 606         status = ocrdma_dealloc_ucontext_pd(uctx);
 607
 608         ocrdma_del_mmap(uctx, uctx->ah_tbl.pa, uctx->ah_tbl.len);
 609         dma_free_coherent(&pdev->dev, uctx->ah_tbl.len, uctx->ah_tbl.va,
 610                           uctx->ah_tbl.pa);
 611
 612         list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
 613                 list_del(&mm->entry);
 614                 kfree(mm);
 615         }
 616         kfree(uctx);
 617         return status;
 618 }
 619
 620 int ocrdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
 621 {
 622         struct ocrdma_ucontext *ucontext = get_ocrdma_ucontext(context);
 623         struct ocrdma_dev *dev = get_ocrdma_dev(context->device);
 624         unsigned long vm_page = vma->vm_pgoff << PAGE_SHIFT;
 625         u64 unmapped_db = (u64) dev->nic_info.unmapped_db;
 626         unsigned long len = (vma->vm_end - vma->vm_start);
 627         int status;
 628         bool found;
 629
 630         if (vma->vm_start & (PAGE_SIZE - 1))
 631                 return -EINVAL;
 632         found = ocrdma_search_mmap(ucontext, vma->vm_pgoff << PAGE_SHIFT, len);
 633         if (!found)
 634                 return -EINVAL;
 635
 636         if ((vm_page >= unmapped_db) && (vm_page <= (unmapped_db +
 637                 dev->nic_info.db_total_size)) &&
 638                 (len <= dev->nic_info.db_page_size)) {
 639                 if (vma->vm_flags & VM_READ)
 640                         return -EPERM;
 641
 642                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 643                 status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
 644                                             len, vma->vm_page_prot);
 645         } else if (dev->nic_info.dpp_unmapped_len &&
 646                 (vm_page >= (u64) dev->nic_info.dpp_unmapped_addr) &&
 647                 (vm_page <= (u64) (dev->nic_info.dpp_unmapped_addr +
 648                         dev->nic_info.dpp_unmapped_len)) &&
 649                 (len <= dev->nic_info.dpp_unmapped_len)) {
 650                 if (vma->vm_flags & VM_READ)
 651                         return -EPERM;
 652
 653                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 654                 status = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
 655                                             len, vma->vm_page_prot);
 656         } else {
 657                 status = remap_pfn_range(vma, vma->vm_start,
 658                                          vma->vm_pgoff, len, vma->vm_page_prot);
 659         }
 660         return status;
 661 }
 662
 663 static int ocrdma_copy_pd_uresp(struct ocrdma_dev *dev, struct ocrdma_pd *pd,
 664                                 struct ib_ucontext *ib_ctx,
 665                                 struct ib_udata *udata)
 666 {
 667         int status;
 668         u64 db_page_addr;
 669         u64 dpp_page_addr = 0;
 670         u32 db_page_size;
 671         struct ocrdma_alloc_pd_uresp rsp;
 672         struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
 673
 674         memset(&rsp, 0, sizeof(rsp));
 675         rsp.id = pd->id;
 676         rsp.dpp_enabled = pd->dpp_enabled;
 677         db_page_addr = ocrdma_get_db_addr(dev, pd->id);
 678         db_page_size = dev->nic_info.db_page_size;
 679
 680         status = ocrdma_add_mmap(uctx, db_page_addr, db_page_size);
 681         if (status)
 682                 return status;
 683
 684         if (pd->dpp_enabled) {
 685                 dpp_page_addr = dev->nic_info.dpp_unmapped_addr +
 686                                 (pd->id * PAGE_SIZE);
 687                 status = ocrdma_add_mmap(uctx, dpp_page_addr,
 688                                  PAGE_SIZE);
 689                 if (status)
 690                         goto dpp_map_err;
 691                 rsp.dpp_page_addr_hi = upper_32_bits(dpp_page_addr);
 692                 rsp.dpp_page_addr_lo = dpp_page_addr;
 693         }
 694
 695         status = ib_copy_to_udata(udata, &rsp, sizeof(rsp));
 696         if (status)
 697                 goto ucopy_err;
 698
 699         pd->uctx = uctx;
 700         return 0;
 701
 702 ucopy_err:
 703         if (pd->dpp_enabled)
 704                 ocrdma_del_mmap(pd->uctx, dpp_page_addr, PAGE_SIZE);
 705 dpp_map_err:
 706         ocrdma_del_mmap(pd->uctx, db_page_addr, db_page_size);
 707         return status;
 708 }
 709
 710 struct ib_pd *ocrdma_alloc_pd(struct ib_device *ibdev,
 711                               struct ib_ucontext *context,
 712                               struct ib_udata *udata)
 713 {
 714         struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
 715         struct ocrdma_pd *pd;
 716         struct ocrdma_ucontext *uctx = NULL;
 717         int status;
 718         u8 is_uctx_pd = false;
 719
 720         if (udata && context) {
 721                 uctx = get_ocrdma_ucontext(context);
 722                 pd = ocrdma_get_ucontext_pd(uctx);
 723                 if (pd) {
 724                         is_uctx_pd = true;
 725                         goto pd_mapping;
 726                 }
 727         }
 728
 729         pd = _ocrdma_alloc_pd(dev, uctx, udata);
 730         if (IS_ERR(pd)) {
 731                 status = PTR_ERR(pd);
 732                 goto exit;
 733         }
 734
 735 pd_mapping:
 736         if (udata && context) {
 737                 status = ocrdma_copy_pd_uresp(dev, pd, context, udata);
 738                 if (status)
 739                         goto err;
 740         }
 741         return &pd->ibpd;
 742
 743 err:
 744         if (is_uctx_pd) {
 745                 ocrdma_release_ucontext_pd(uctx);
 746         } else {
 747                 status = _ocrdma_dealloc_pd(dev, pd);
 748         }
 749 exit:
 750         return ERR_PTR(status);
 751 }
 752
 753 int ocrdma_dealloc_pd(struct ib_pd *ibpd)
 754 {
 755         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
 756         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
 757         struct ocrdma_ucontext *uctx = NULL;
 758         int status = 0;
 759         u64 usr_db;
 760
 761         uctx = pd->uctx;
 762         if (uctx) {
 763                 u64 dpp_db = dev->nic_info.dpp_unmapped_addr +
 764                         (pd->id * PAGE_SIZE);
 765                 if (pd->dpp_enabled)
 766                         ocrdma_del_mmap(pd->uctx, dpp_db, PAGE_SIZE);
 767                 usr_db = ocrdma_get_db_addr(dev, pd->id);
 768                 ocrdma_del_mmap(pd->uctx, usr_db, dev->nic_info.db_page_size);
 769
 770                 if (is_ucontext_pd(uctx, pd)) {
 771                         ocrdma_release_ucontext_pd(uctx);
 772                         return status;
 773                 }
 774         }
 775         status = _ocrdma_dealloc_pd(dev, pd);
 776         return status;
 777 }
 778
 779 static int ocrdma_alloc_lkey(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
 780                             u32 pdid, int acc, u32 num_pbls, u32 addr_check)
 781 {
 782         int status;
 783
 784         mr->hwmr.fr_mr = 0;
 785         mr->hwmr.local_rd = 1;
 786         mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
 787         mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
 788         mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
 789         mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
 790         mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
 791         mr->hwmr.num_pbls = num_pbls;
 792
 793         status = ocrdma_mbx_alloc_lkey(dev, &mr->hwmr, pdid, addr_check);
 794         if (status)
 795                 return status;
 796
 797         mr->ibmr.lkey = mr->hwmr.lkey;
 798         if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
 799                 mr->ibmr.rkey = mr->hwmr.lkey;
 800         return 0;
 801 }
 802
 803 struct ib_mr *ocrdma_get_dma_mr(struct ib_pd *ibpd, int acc)
 804 {
 805         int status;
 806         struct ocrdma_mr *mr;
 807         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
 808         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
 809
 810         if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE)) {
 811                 pr_err("%s err, invalid access rights\n", __func__);
 812                 return ERR_PTR(-EINVAL);
 813         }
 814
 815         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
 816         if (!mr)
 817                 return ERR_PTR(-ENOMEM);
 818
 819         status = ocrdma_alloc_lkey(dev, mr, pd->id, acc, 0,
 820                                    OCRDMA_ADDR_CHECK_DISABLE);
 821         if (status) {
 822                 kfree(mr);
 823                 return ERR_PTR(status);
 824         }
 825
 826         return &mr->ibmr;
 827 }
 828
 829 static void ocrdma_free_mr_pbl_tbl(struct ocrdma_dev *dev,
 830                                    struct ocrdma_hw_mr *mr)
 831 {
 832         struct pci_dev *pdev = dev->nic_info.pdev;
 833         int i = 0;
 834
 835         if (mr->pbl_table) {
 836                 for (i = 0; i < mr->num_pbls; i++) {
 837                         if (!mr->pbl_table[i].va)
 838                                 continue;
 839                         dma_free_coherent(&pdev->dev, mr->pbl_size,
 840                                           mr->pbl_table[i].va,
 841                                           mr->pbl_table[i].pa);
 842                 }
 843                 kfree(mr->pbl_table);
 844                 mr->pbl_table = NULL;
 845         }
 846 }
 847
 848 static int ocrdma_get_pbl_info(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
 849                               u32 num_pbes)
 850 {
 851         u32 num_pbls = 0;
 852         u32 idx = 0;
 853         int status = 0;
 854         u32 pbl_size;
 855
 856         do {
 857                 pbl_size = OCRDMA_MIN_HPAGE_SIZE * (1 << idx);
 858                 if (pbl_size > MAX_OCRDMA_PBL_SIZE) {
 859                         status = -EFAULT;
 860                         break;
 861                 }
 862                 num_pbls = roundup(num_pbes, (pbl_size / sizeof(u64)));
 863                 num_pbls = num_pbls / (pbl_size / sizeof(u64));
 864                 idx++;
 865         } while (num_pbls >= dev->attr.max_num_mr_pbl);
 866
 867         mr->hwmr.num_pbes = num_pbes;
 868         mr->hwmr.num_pbls = num_pbls;
 869         mr->hwmr.pbl_size = pbl_size;
 870         return status;
 871 }
 872
 873 static int ocrdma_build_pbl_tbl(struct ocrdma_dev *dev, struct ocrdma_hw_mr *mr)
 874 {
 875         int status = 0;
 876         int i;
 877         u32 dma_len = mr->pbl_size;
 878         struct pci_dev *pdev = dev->nic_info.pdev;
 879         void *va;
 880         dma_addr_t pa;
 881
 882         mr->pbl_table = kzalloc(sizeof(struct ocrdma_pbl) *
 883                                 mr->num_pbls, GFP_KERNEL);
 884
 885         if (!mr->pbl_table)
 886                 return -ENOMEM;
 887
 888         for (i = 0; i < mr->num_pbls; i++) {
 889                 va = dma_alloc_coherent(&pdev->dev, dma_len, &pa, GFP_KERNEL);
 890                 if (!va) {
 891                         ocrdma_free_mr_pbl_tbl(dev, mr);
 892                         status = -ENOMEM;
 893                         break;
 894                 }
 895                 memset(va, 0, dma_len);
 896                 mr->pbl_table[i].va = va;
 897                 mr->pbl_table[i].pa = pa;
 898         }
 899         return status;
 900 }
 901
 902 static void build_user_pbes(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
 903                             u32 num_pbes)
 904 {
 905         struct ocrdma_pbe *pbe;
 906         struct scatterlist *sg;
 907         struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
 908         struct ib_umem *umem = mr->umem;
 909         int shift, pg_cnt, pages, pbe_cnt, entry, total_num_pbes = 0;
 910
 911         if (!mr->hwmr.num_pbes)
 912                 return;
 913
 914         pbe = (struct ocrdma_pbe *)pbl_tbl->va;
 915         pbe_cnt = 0;
 916
 917         shift = umem->page_shift;
 918
 919         for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
 920                 pages = sg_dma_len(sg) >> shift;
 921                 for (pg_cnt = 0; pg_cnt < pages; pg_cnt++) {
 922                         /* store the page address in pbe */
 923                         pbe->pa_lo =
 924                             cpu_to_le32(sg_dma_address(sg) +
 925                                         (pg_cnt << shift));
 926                         pbe->pa_hi =
 927                             cpu_to_le32(upper_32_bits(sg_dma_address(sg) +
 928                                          (pg_cnt << shift)));
 929                         pbe_cnt += 1;
 930                         total_num_pbes += 1;
 931                         pbe++;
 932
 933                         /* if done building pbes, issue the mbx cmd. */
 934                         if (total_num_pbes == num_pbes)
 935                                 return;
 936
 937                         /* if the given pbl is full storing the pbes,
 938                          * move to next pbl.
 939                          */
 940                         if (pbe_cnt ==
 941                                 (mr->hwmr.pbl_size / sizeof(u64))) {
 942                                 pbl_tbl++;
 943                                 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
 944                                 pbe_cnt = 0;
 945                         }
 946
 947                 }
 948         }
 949 }
 950
 951 struct ib_mr *ocrdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
 952                                  u64 usr_addr, int acc, struct ib_udata *udata)
 953 {
 954         int status = -ENOMEM;
 955         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
 956         struct ocrdma_mr *mr;
 957         struct ocrdma_pd *pd;
 958         u32 num_pbes;
 959
 960         pd = get_ocrdma_pd(ibpd);
 961
 962         if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
 963                 return ERR_PTR(-EINVAL);
 964
 965         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
 966         if (!mr)
 967                 return ERR_PTR(status);
 968         mr->umem = ib_umem_get(ibpd->uobject->context, start, len, acc, 0);
 969         if (IS_ERR(mr->umem)) {
 970                 status = -EFAULT;
 971                 goto umem_err;
 972         }
 973         num_pbes = ib_umem_page_count(mr->umem);
 974         status = ocrdma_get_pbl_info(dev, mr, num_pbes);
 975         if (status)
 976                 goto umem_err;
 977
 978         mr->hwmr.pbe_size = BIT(mr->umem->page_shift);
 979         mr->hwmr.fbo = ib_umem_offset(mr->umem);
 980         mr->hwmr.va = usr_addr;
 981         mr->hwmr.len = len;
 982         mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
 983         mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
 984         mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
 985         mr->hwmr.local_rd = 1;
 986         mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
 987         status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
 988         if (status)
 989                 goto umem_err;
 990         build_user_pbes(dev, mr, num_pbes);
 991         status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
 992         if (status)
 993                 goto mbx_err;
 994         mr->ibmr.lkey = mr->hwmr.lkey;
 995         if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
 996                 mr->ibmr.rkey = mr->hwmr.lkey;
 997
 998         return &mr->ibmr;
 999
1000 mbx_err:
1001         ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
1002 umem_err:
1003         kfree(mr);
1004         return ERR_PTR(status);
1005 }
1006
1007 int ocrdma_dereg_mr(struct ib_mr *ib_mr)
1008 {
1009         struct ocrdma_mr *mr = get_ocrdma_mr(ib_mr);
1010         struct ocrdma_dev *dev = get_ocrdma_dev(ib_mr->device);
1011
1012         (void) ocrdma_mbx_dealloc_lkey(dev, mr->hwmr.fr_mr, mr->hwmr.lkey);
1013
1014         kfree(mr->pages);
1015         ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
1016
1017         /* it could be user registered memory. */
1018         if (mr->umem)
1019                 ib_umem_release(mr->umem);
1020         kfree(mr);
1021
1022         /* Don't stop cleanup, in case FW is unresponsive */
1023         if (dev->mqe_ctx.fw_error_state) {
1024                 pr_err("%s(%d) fw not responding.\n",
1025                        __func__, dev->id);
1026         }
1027         return 0;
1028 }
1029
1030 static int ocrdma_copy_cq_uresp(struct ocrdma_dev *dev, struct ocrdma_cq *cq,
1031                                 struct ib_udata *udata,
1032                                 struct ib_ucontext *ib_ctx)
1033 {
1034         int status;
1035         struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
1036         struct ocrdma_create_cq_uresp uresp;
1037
1038         memset(&uresp, 0, sizeof(uresp));
1039         uresp.cq_id = cq->id;
1040         uresp.page_size = PAGE_ALIGN(cq->len);
1041         uresp.num_pages = 1;
1042         uresp.max_hw_cqe = cq->max_hw_cqe;
1043         uresp.page_addr[0] = virt_to_phys(cq->va);
1044         uresp.db_page_addr =  ocrdma_get_db_addr(dev, uctx->cntxt_pd->id);
1045         uresp.db_page_size = dev->nic_info.db_page_size;
1046         uresp.phase_change = cq->phase_change ? 1 : 0;
1047         status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1048         if (status) {
1049                 pr_err("%s(%d) copy error cqid=0x%x.\n",
1050                        __func__, dev->id, cq->id);
1051                 goto err;
1052         }
1053         status = ocrdma_add_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
1054         if (status)
1055                 goto err;
1056         status = ocrdma_add_mmap(uctx, uresp.page_addr[0], uresp.page_size);
1057         if (status) {
1058                 ocrdma_del_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
1059                 goto err;
1060         }
1061         cq->ucontext = uctx;
1062 err:
1063         return status;
1064 }
1065
1066 struct ib_cq *ocrdma_create_cq(struct ib_device *ibdev,
1067                                const struct ib_cq_init_attr *attr,
1068                                struct ib_ucontext *ib_ctx,
1069                                struct ib_udata *udata)
1070 {
1071         int entries = attr->cqe;
1072         struct ocrdma_cq *cq;
1073         struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
1074         struct ocrdma_ucontext *uctx = NULL;
1075         u16 pd_id = 0;
1076         int status;
1077         struct ocrdma_create_cq_ureq ureq;
1078
1079         if (attr->flags)
1080                 return ERR_PTR(-EINVAL);
1081
1082         if (udata) {
1083                 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1084                         return ERR_PTR(-EFAULT);
1085         } else
1086                 ureq.dpp_cq = 0;
1087         cq = kzalloc(sizeof(*cq), GFP_KERNEL);
1088         if (!cq)
1089                 return ERR_PTR(-ENOMEM);
1090
1091         spin_lock_init(&cq->cq_lock);
1092         spin_lock_init(&cq->comp_handler_lock);
1093         INIT_LIST_HEAD(&cq->sq_head);
1094         INIT_LIST_HEAD(&cq->rq_head);
1095
1096         if (ib_ctx) {
1097                 uctx = get_ocrdma_ucontext(ib_ctx);
1098                 pd_id = uctx->cntxt_pd->id;
1099         }
1100
1101         status = ocrdma_mbx_create_cq(dev, cq, entries, ureq.dpp_cq, pd_id);
1102         if (status) {
1103                 kfree(cq);
1104                 return ERR_PTR(status);
1105         }
1106         if (ib_ctx) {
1107                 status = ocrdma_copy_cq_uresp(dev, cq, udata, ib_ctx);
1108                 if (status)
1109                         goto ctx_err;
1110         }
1111         cq->phase = OCRDMA_CQE_VALID;
1112         dev->cq_tbl[cq->id] = cq;
1113         return &cq->ibcq;
1114
1115 ctx_err:
1116         ocrdma_mbx_destroy_cq(dev, cq);
1117         kfree(cq);
1118         return ERR_PTR(status);
1119 }
1120
1121 int ocrdma_resize_cq(struct ib_cq *ibcq, int new_cnt,
1122                      struct ib_udata *udata)
1123 {
1124         int status = 0;
1125         struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1126
1127         if (new_cnt < 1 || new_cnt > cq->max_hw_cqe) {
1128                 status = -EINVAL;
1129                 return status;
1130         }
1131         ibcq->cqe = new_cnt;
1132         return status;
1133 }
1134
1135 static void ocrdma_flush_cq(struct ocrdma_cq *cq)
1136 {
1137         int cqe_cnt;
1138         int valid_count = 0;
1139         unsigned long flags;
1140
1141         struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
1142         struct ocrdma_cqe *cqe = NULL;
1143
1144         cqe = cq->va;
1145         cqe_cnt = cq->cqe_cnt;
1146
1147         /* Last irq might have scheduled a polling thread
1148          * sync-up with it before hard flushing.
1149          */
1150         spin_lock_irqsave(&cq->cq_lock, flags);
1151         while (cqe_cnt) {
1152                 if (is_cqe_valid(cq, cqe))
1153                         valid_count++;
1154                 cqe++;
1155                 cqe_cnt--;
1156         }
1157         ocrdma_ring_cq_db(dev, cq->id, false, false, valid_count);
1158         spin_unlock_irqrestore(&cq->cq_lock, flags);
1159 }
1160
1161 int ocrdma_destroy_cq(struct ib_cq *ibcq)
1162 {
1163         struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1164         struct ocrdma_eq *eq = NULL;
1165         struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
1166         int pdid = 0;
1167         u32 irq, indx;
1168
1169         dev->cq_tbl[cq->id] = NULL;
1170         indx = ocrdma_get_eq_table_index(dev, cq->eqn);
1171         BUG_ON(indx == -EINVAL);
1172
1173         eq = &dev->eq_tbl[indx];
1174         irq = ocrdma_get_irq(dev, eq);
1175         synchronize_irq(irq);
1176         ocrdma_flush_cq(cq);
1177
1178         (void)ocrdma_mbx_destroy_cq(dev, cq);
1179         if (cq->ucontext) {
1180                 pdid = cq->ucontext->cntxt_pd->id;
1181                 ocrdma_del_mmap(cq->ucontext, (u64) cq->pa,
1182                                 PAGE_ALIGN(cq->len));
1183                 ocrdma_del_mmap(cq->ucontext,
1184                                 ocrdma_get_db_addr(dev, pdid),
1185                                 dev->nic_info.db_page_size);
1186         }
1187
1188         kfree(cq);
1189         return 0;
1190 }
1191
1192 static int ocrdma_add_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1193 {
1194         int status = -EINVAL;
1195
1196         if (qp->id < OCRDMA_MAX_QP && dev->qp_tbl[qp->id] == NULL) {
1197                 dev->qp_tbl[qp->id] = qp;
1198                 status = 0;
1199         }
1200         return status;
1201 }
1202
1203 static void ocrdma_del_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1204 {
1205         dev->qp_tbl[qp->id] = NULL;
1206 }
1207
1208 static int ocrdma_check_qp_params(struct ib_pd *ibpd, struct ocrdma_dev *dev,
1209                                   struct ib_qp_init_attr *attrs)
1210 {
1211         if ((attrs->qp_type != IB_QPT_GSI) &&
1212             (attrs->qp_type != IB_QPT_RC) &&
1213             (attrs->qp_type != IB_QPT_UC) &&
1214             (attrs->qp_type != IB_QPT_UD)) {
1215                 pr_err("%s(%d) unsupported qp type=0x%x requested\n",
1216                        __func__, dev->id, attrs->qp_type);
1217                 return -EINVAL;
1218         }
1219         /* Skip the check for QP1 to support CM size of 128 */
1220         if ((attrs->qp_type != IB_QPT_GSI) &&
1221             (attrs->cap.max_send_wr > dev->attr.max_wqe)) {
1222                 pr_err("%s(%d) unsupported send_wr=0x%x requested\n",
1223                        __func__, dev->id, attrs->cap.max_send_wr);
1224                 pr_err("%s(%d) supported send_wr=0x%x\n",
1225                        __func__, dev->id, dev->attr.max_wqe);
1226                 return -EINVAL;
1227         }
1228         if (!attrs->srq && (attrs->cap.max_recv_wr > dev->attr.max_rqe)) {
1229                 pr_err("%s(%d) unsupported recv_wr=0x%x requested\n",
1230                        __func__, dev->id, attrs->cap.max_recv_wr);
1231                 pr_err("%s(%d) supported recv_wr=0x%x\n",
1232                        __func__, dev->id, dev->attr.max_rqe);
1233                 return -EINVAL;
1234         }
1235         if (attrs->cap.max_inline_data > dev->attr.max_inline_data) {
1236                 pr_err("%s(%d) unsupported inline data size=0x%x requested\n",
1237                        __func__, dev->id, attrs->cap.max_inline_data);
1238                 pr_err("%s(%d) supported inline data size=0x%x\n",
1239                        __func__, dev->id, dev->attr.max_inline_data);
1240                 return -EINVAL;
1241         }
1242         if (attrs->cap.max_send_sge > dev->attr.max_send_sge) {
1243                 pr_err("%s(%d) unsupported send_sge=0x%x requested\n",
1244                        __func__, dev->id, attrs->cap.max_send_sge);
1245                 pr_err("%s(%d) supported send_sge=0x%x\n",
1246                        __func__, dev->id, dev->attr.max_send_sge);
1247                 return -EINVAL;
1248         }
1249         if (attrs->cap.max_recv_sge > dev->attr.max_recv_sge) {
1250                 pr_err("%s(%d) unsupported recv_sge=0x%x requested\n",
1251                        __func__, dev->id, attrs->cap.max_recv_sge);
1252                 pr_err("%s(%d) supported recv_sge=0x%x\n",
1253                        __func__, dev->id, dev->attr.max_recv_sge);
1254                 return -EINVAL;
1255         }
1256         /* unprivileged user space cannot create special QP */
1257         if (ibpd->uobject && attrs->qp_type == IB_QPT_GSI) {
1258                 pr_err
1259                     ("%s(%d) Userspace can't create special QPs of type=0x%x\n",
1260                      __func__, dev->id, attrs->qp_type);
1261                 return -EINVAL;
1262         }
1263         /* allow creating only one GSI type of QP */
1264         if (attrs->qp_type == IB_QPT_GSI && dev->gsi_qp_created) {
1265                 pr_err("%s(%d) GSI special QPs already created.\n",
1266                        __func__, dev->id);
1267                 return -EINVAL;
1268         }
1269         /* verify consumer QPs are not trying to use GSI QP's CQ */
1270         if ((attrs->qp_type != IB_QPT_GSI) && (dev->gsi_qp_created)) {
1271                 if ((dev->gsi_sqcq == get_ocrdma_cq(attrs->send_cq)) ||
1272                         (dev->gsi_rqcq == get_ocrdma_cq(attrs->recv_cq))) {
1273                         pr_err("%s(%d) Consumer QP cannot use GSI CQs.\n",
1274                                 __func__, dev->id);
1275                         return -EINVAL;
1276                 }
1277         }
1278         return 0;
1279 }
1280
1281 static int ocrdma_copy_qp_uresp(struct ocrdma_qp *qp,
1282                                 struct ib_udata *udata, int dpp_offset,
1283                                 int dpp_credit_lmt, int srq)
1284 {
1285         int status;
1286         u64 usr_db;
1287         struct ocrdma_create_qp_uresp uresp;
1288         struct ocrdma_pd *pd = qp->pd;
1289         struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
1290
1291         memset(&uresp, 0, sizeof(uresp));
1292         usr_db = dev->nic_info.unmapped_db +
1293                         (pd->id * dev->nic_info.db_page_size);
1294         uresp.qp_id = qp->id;
1295         uresp.sq_dbid = qp->sq.dbid;
1296         uresp.num_sq_pages = 1;
1297         uresp.sq_page_size = PAGE_ALIGN(qp->sq.len);
1298         uresp.sq_page_addr[0] = virt_to_phys(qp->sq.va);
1299         uresp.num_wqe_allocated = qp->sq.max_cnt;
1300         if (!srq) {
1301                 uresp.rq_dbid = qp->rq.dbid;
1302                 uresp.num_rq_pages = 1;
1303                 uresp.rq_page_size = PAGE_ALIGN(qp->rq.len);
1304                 uresp.rq_page_addr[0] = virt_to_phys(qp->rq.va);
1305                 uresp.num_rqe_allocated = qp->rq.max_cnt;
1306         }
1307         uresp.db_page_addr = usr_db;
1308         uresp.db_page_size = dev->nic_info.db_page_size;
1309         uresp.db_sq_offset = OCRDMA_DB_GEN2_SQ_OFFSET;
1310         uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1311         uresp.db_shift = OCRDMA_DB_RQ_SHIFT;
1312
1313         if (qp->dpp_enabled) {
1314                 uresp.dpp_credit = dpp_credit_lmt;
1315                 uresp.dpp_offset = dpp_offset;
1316         }
1317         status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1318         if (status) {
1319                 pr_err("%s(%d) user copy error.\n", __func__, dev->id);
1320                 goto err;
1321         }
1322         status = ocrdma_add_mmap(pd->uctx, uresp.sq_page_addr[0],
1323                                  uresp.sq_page_size);
1324         if (status)
1325                 goto err;
1326
1327         if (!srq) {
1328                 status = ocrdma_add_mmap(pd->uctx, uresp.rq_page_addr[0],
1329                                          uresp.rq_page_size);
1330                 if (status)
1331                         goto rq_map_err;
1332         }
1333         return status;
1334 rq_map_err:
1335         ocrdma_del_mmap(pd->uctx, uresp.sq_page_addr[0], uresp.sq_page_size);
1336 err:
1337         return status;
1338 }
1339
1340 static void ocrdma_set_qp_db(struct ocrdma_dev *dev, struct ocrdma_qp *qp,
1341                              struct ocrdma_pd *pd)
1342 {
1343         if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1344                 qp->sq_db = dev->nic_info.db +
1345                         (pd->id * dev->nic_info.db_page_size) +
1346                         OCRDMA_DB_GEN2_SQ_OFFSET;
1347                 qp->rq_db = dev->nic_info.db +
1348                         (pd->id * dev->nic_info.db_page_size) +
1349                         OCRDMA_DB_GEN2_RQ_OFFSET;
1350         } else {
1351                 qp->sq_db = dev->nic_info.db +
1352                         (pd->id * dev->nic_info.db_page_size) +
1353                         OCRDMA_DB_SQ_OFFSET;
1354                 qp->rq_db = dev->nic_info.db +
1355                         (pd->id * dev->nic_info.db_page_size) +
1356                         OCRDMA_DB_RQ_OFFSET;
1357         }
1358 }
1359
1360 static int ocrdma_alloc_wr_id_tbl(struct ocrdma_qp *qp)
1361 {
1362         qp->wqe_wr_id_tbl =
1363             kzalloc(sizeof(*(qp->wqe_wr_id_tbl)) * qp->sq.max_cnt,
1364                     GFP_KERNEL);
1365         if (qp->wqe_wr_id_tbl == NULL)
1366                 return -ENOMEM;
1367         qp->rqe_wr_id_tbl =
1368             kzalloc(sizeof(u64) * qp->rq.max_cnt, GFP_KERNEL);
1369         if (qp->rqe_wr_id_tbl == NULL)
1370                 return -ENOMEM;
1371
1372         return 0;
1373 }
1374
1375 static void ocrdma_set_qp_init_params(struct ocrdma_qp *qp,
1376                                       struct ocrdma_pd *pd,
1377                                       struct ib_qp_init_attr *attrs)
1378 {
1379         qp->pd = pd;
1380         spin_lock_init(&qp->q_lock);
1381         INIT_LIST_HEAD(&qp->sq_entry);
1382         INIT_LIST_HEAD(&qp->rq_entry);
1383
1384         qp->qp_type = attrs->qp_type;
1385         qp->cap_flags = OCRDMA_QP_INB_RD | OCRDMA_QP_INB_WR;
1386         qp->max_inline_data = attrs->cap.max_inline_data;
1387         qp->sq.max_sges = attrs->cap.max_send_sge;
1388         qp->rq.max_sges = attrs->cap.max_recv_sge;
1389         qp->state = OCRDMA_QPS_RST;
1390         qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
1391 }
1392
1393 static void ocrdma_store_gsi_qp_cq(struct ocrdma_dev *dev,
1394                                    struct ib_qp_init_attr *attrs)
1395 {
1396         if (attrs->qp_type == IB_QPT_GSI) {
1397                 dev->gsi_qp_created = 1;
1398                 dev->gsi_sqcq = get_ocrdma_cq(attrs->send_cq);
1399                 dev->gsi_rqcq = get_ocrdma_cq(attrs->recv_cq);
1400         }
1401 }
1402
1403 struct ib_qp *ocrdma_create_qp(struct ib_pd *ibpd,
1404                                struct ib_qp_init_attr *attrs,
1405                                struct ib_udata *udata)
1406 {
1407         int status;
1408         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1409         struct ocrdma_qp *qp;
1410         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1411         struct ocrdma_create_qp_ureq ureq;
1412         u16 dpp_credit_lmt, dpp_offset;
1413
1414         status = ocrdma_check_qp_params(ibpd, dev, attrs);
1415         if (status)
1416                 goto gen_err;
1417
1418         memset(&ureq, 0, sizeof(ureq));
1419         if (udata) {
1420                 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1421                         return ERR_PTR(-EFAULT);
1422         }
1423         qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1424         if (!qp) {
1425                 status = -ENOMEM;
1426                 goto gen_err;
1427         }
1428         ocrdma_set_qp_init_params(qp, pd, attrs);
1429         if (udata == NULL)
1430                 qp->cap_flags |= (OCRDMA_QP_MW_BIND | OCRDMA_QP_LKEY0 |
1431                                         OCRDMA_QP_FAST_REG);
1432
1433         mutex_lock(&dev->dev_lock);
1434         status = ocrdma_mbx_create_qp(qp, attrs, ureq.enable_dpp_cq,
1435                                         ureq.dpp_cq_id,
1436                                         &dpp_offset, &dpp_credit_lmt);
1437         if (status)
1438                 goto mbx_err;
1439
1440         /* user space QP's wr_id table are managed in library */
1441         if (udata == NULL) {
1442                 status = ocrdma_alloc_wr_id_tbl(qp);
1443                 if (status)
1444                         goto map_err;
1445         }
1446
1447         status = ocrdma_add_qpn_map(dev, qp);
1448         if (status)
1449                 goto map_err;
1450         ocrdma_set_qp_db(dev, qp, pd);
1451         if (udata) {
1452                 status = ocrdma_copy_qp_uresp(qp, udata, dpp_offset,
1453                                               dpp_credit_lmt,
1454                                               (attrs->srq != NULL));
1455                 if (status)
1456                         goto cpy_err;
1457         }
1458         ocrdma_store_gsi_qp_cq(dev, attrs);
1459         qp->ibqp.qp_num = qp->id;
1460         mutex_unlock(&dev->dev_lock);
1461         return &qp->ibqp;
1462
1463 cpy_err:
1464         ocrdma_del_qpn_map(dev, qp);
1465 map_err:
1466         ocrdma_mbx_destroy_qp(dev, qp);
1467 mbx_err:
1468         mutex_unlock(&dev->dev_lock);
1469         kfree(qp->wqe_wr_id_tbl);
1470         kfree(qp->rqe_wr_id_tbl);
1471         kfree(qp);
1472         pr_err("%s(%d) error=%d\n", __func__, dev->id, status);
1473 gen_err:
1474         return ERR_PTR(status);
1475 }
1476
1477 int _ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1478                       int attr_mask)
1479 {
1480         int status = 0;
1481         struct ocrdma_qp *qp;
1482         struct ocrdma_dev *dev;
1483         enum ib_qp_state old_qps;
1484
1485         qp = get_ocrdma_qp(ibqp);
1486         dev = get_ocrdma_dev(ibqp->device);
1487         if (attr_mask & IB_QP_STATE)
1488                 status = ocrdma_qp_state_change(qp, attr->qp_state, &old_qps);
1489         /* if new and previous states are same hw doesn't need to
1490          * know about it.
1491          */
1492         if (status < 0)
1493                 return status;
1494         return ocrdma_mbx_modify_qp(dev, qp, attr, attr_mask);
1495 }
1496
1497 int ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1498                      int attr_mask, struct ib_udata *udata)
1499 {
1500         unsigned long flags;
1501         int status = -EINVAL;
1502         struct ocrdma_qp *qp;
1503         struct ocrdma_dev *dev;
1504         enum ib_qp_state old_qps, new_qps;
1505
1506         qp = get_ocrdma_qp(ibqp);
1507         dev = get_ocrdma_dev(ibqp->device);
1508
1509         /* syncronize with multiple context trying to change, retrive qps */
1510         mutex_lock(&dev->dev_lock);
1511         /* syncronize with wqe, rqe posting and cqe processing contexts */
1512         spin_lock_irqsave(&qp->q_lock, flags);
1513         old_qps = get_ibqp_state(qp->state);
1514         if (attr_mask & IB_QP_STATE)
1515                 new_qps = attr->qp_state;
1516         else
1517                 new_qps = old_qps;
1518         spin_unlock_irqrestore(&qp->q_lock, flags);
1519
1520         if (!ib_modify_qp_is_ok(old_qps, new_qps, ibqp->qp_type, attr_mask,
1521                                 IB_LINK_LAYER_ETHERNET)) {
1522                 pr_err("%s(%d) invalid attribute mask=0x%x specified for\n"
1523                        "qpn=0x%x of type=0x%x old_qps=0x%x, new_qps=0x%x\n",
1524                        __func__, dev->id, attr_mask, qp->id, ibqp->qp_type,
1525                        old_qps, new_qps);
1526                 goto param_err;
1527         }
1528
1529         status = _ocrdma_modify_qp(ibqp, attr, attr_mask);
1530         if (status > 0)
1531                 status = 0;
1532 param_err:
1533         mutex_unlock(&dev->dev_lock);
1534         return status;
1535 }
1536
1537 static enum ib_mtu ocrdma_mtu_int_to_enum(u16 mtu)
1538 {
1539         switch (mtu) {
1540         case 256:
1541                 return IB_MTU_256;
1542         case 512:
1543                 return IB_MTU_512;
1544         case 1024:
1545                 return IB_MTU_1024;
1546         case 2048:
1547                 return IB_MTU_2048;
1548         case 4096:
1549                 return IB_MTU_4096;
1550         default:
1551                 return IB_MTU_1024;
1552         }
1553 }
1554
1555 static int ocrdma_to_ib_qp_acc_flags(int qp_cap_flags)
1556 {
1557         int ib_qp_acc_flags = 0;
1558
1559         if (qp_cap_flags & OCRDMA_QP_INB_WR)
1560                 ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
1561         if (qp_cap_flags & OCRDMA_QP_INB_RD)
1562                 ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
1563         return ib_qp_acc_flags;
1564 }
1565
1566 int ocrdma_query_qp(struct ib_qp *ibqp,
1567                     struct ib_qp_attr *qp_attr,
1568                     int attr_mask, struct ib_qp_init_attr *qp_init_attr)
1569 {
1570         int status;
1571         u32 qp_state;
1572         struct ocrdma_qp_params params;
1573         struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1574         struct ocrdma_dev *dev = get_ocrdma_dev(ibqp->device);
1575
1576         memset(&params, 0, sizeof(params));
1577         mutex_lock(&dev->dev_lock);
1578         status = ocrdma_mbx_query_qp(dev, qp, &params);
1579         mutex_unlock(&dev->dev_lock);
1580         if (status)
1581                 goto mbx_err;
1582         if (qp->qp_type == IB_QPT_UD)
1583                 qp_attr->qkey = params.qkey;
1584         qp_attr->path_mtu =
1585                 ocrdma_mtu_int_to_enum(params.path_mtu_pkey_indx &
1586                                 OCRDMA_QP_PARAMS_PATH_MTU_MASK) >>
1587                                 OCRDMA_QP_PARAMS_PATH_MTU_SHIFT;
1588         qp_attr->path_mig_state = IB_MIG_MIGRATED;
1589         qp_attr->rq_psn = params.hop_lmt_rq_psn & OCRDMA_QP_PARAMS_RQ_PSN_MASK;
1590         qp_attr->sq_psn = params.tclass_sq_psn & OCRDMA_QP_PARAMS_SQ_PSN_MASK;
1591         qp_attr->dest_qp_num =
1592             params.ack_to_rnr_rtc_dest_qpn & OCRDMA_QP_PARAMS_DEST_QPN_MASK;
1593
1594         qp_attr->qp_access_flags = ocrdma_to_ib_qp_acc_flags(qp->cap_flags);
1595         qp_attr->cap.max_send_wr = qp->sq.max_cnt - 1;
1596         qp_attr->cap.max_recv_wr = qp->rq.max_cnt - 1;
1597         qp_attr->cap.max_send_sge = qp->sq.max_sges;
1598         qp_attr->cap.max_recv_sge = qp->rq.max_sges;
1599         qp_attr->cap.max_inline_data = qp->max_inline_data;
1600         qp_init_attr->cap = qp_attr->cap;
1601         qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
1602
1603         rdma_ah_set_grh(&qp_attr->ah_attr, NULL,
1604                         params.rnt_rc_sl_fl &
1605                           OCRDMA_QP_PARAMS_FLOW_LABEL_MASK,
1606                         qp->sgid_idx,
1607                         (params.hop_lmt_rq_psn &
1608                          OCRDMA_QP_PARAMS_HOP_LMT_MASK) >>
1609                          OCRDMA_QP_PARAMS_HOP_LMT_SHIFT,
1610                         (params.tclass_sq_psn &
1611                          OCRDMA_QP_PARAMS_TCLASS_MASK) >>
1612                          OCRDMA_QP_PARAMS_TCLASS_SHIFT);
1613         rdma_ah_set_dgid_raw(&qp_attr->ah_attr, &params.dgid[0]);
1614
1615         rdma_ah_set_port_num(&qp_attr->ah_attr, 1);
1616         rdma_ah_set_sl(&qp_attr->ah_attr, (params.rnt_rc_sl_fl &
1617                                            OCRDMA_QP_PARAMS_SL_MASK) >>
1618                                            OCRDMA_QP_PARAMS_SL_SHIFT);
1619         qp_attr->timeout = (params.ack_to_rnr_rtc_dest_qpn &
1620                             OCRDMA_QP_PARAMS_ACK_TIMEOUT_MASK) >>
1621                                 OCRDMA_QP_PARAMS_ACK_TIMEOUT_SHIFT;
1622         qp_attr->rnr_retry = (params.ack_to_rnr_rtc_dest_qpn &
1623                               OCRDMA_QP_PARAMS_RNR_RETRY_CNT_MASK) >>
1624                                 OCRDMA_QP_PARAMS_RNR_RETRY_CNT_SHIFT;
1625         qp_attr->retry_cnt =
1626             (params.rnt_rc_sl_fl & OCRDMA_QP_PARAMS_RETRY_CNT_MASK) >>
1627                 OCRDMA_QP_PARAMS_RETRY_CNT_SHIFT;
1628         qp_attr->min_rnr_timer = 0;
1629         qp_attr->pkey_index = 0;
1630         qp_attr->port_num = 1;
1631         rdma_ah_set_path_bits(&qp_attr->ah_attr, 0);
1632         rdma_ah_set_static_rate(&qp_attr->ah_attr, 0);
1633         qp_attr->alt_pkey_index = 0;
1634         qp_attr->alt_port_num = 0;
1635         qp_attr->alt_timeout = 0;
1636         memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
1637         qp_state = (params.max_sge_recv_flags & OCRDMA_QP_PARAMS_STATE_MASK) >>
1638                     OCRDMA_QP_PARAMS_STATE_SHIFT;
1639         qp_attr->qp_state = get_ibqp_state(qp_state);
1640         qp_attr->cur_qp_state = qp_attr->qp_state;
1641         qp_attr->sq_draining = (qp_state == OCRDMA_QPS_SQ_DRAINING) ? 1 : 0;
1642         qp_attr->max_dest_rd_atomic =
1643             params.max_ord_ird >> OCRDMA_QP_PARAMS_MAX_ORD_SHIFT;
1644         qp_attr->max_rd_atomic =
1645             params.max_ord_ird & OCRDMA_QP_PARAMS_MAX_IRD_MASK;
1646         qp_attr->en_sqd_async_notify = (params.max_sge_recv_flags &
1647                                 OCRDMA_QP_PARAMS_FLAGS_SQD_ASYNC) ? 1 : 0;
1648         /* Sync driver QP state with FW */
1649         ocrdma_qp_state_change(qp, qp_attr->qp_state, NULL);
1650 mbx_err:
1651         return status;
1652 }
1653
1654 static void ocrdma_srq_toggle_bit(struct ocrdma_srq *srq, unsigned int idx)
1655 {
1656         unsigned int i = idx / 32;
1657         u32 mask = (1U << (idx % 32));
1658
1659         srq->idx_bit_fields[i] ^= mask;
1660 }
1661
1662 static int ocrdma_hwq_free_cnt(struct ocrdma_qp_hwq_info *q)
1663 {
1664         return ((q->max_wqe_idx - q->head) + q->tail) % q->max_cnt;
1665 }
1666
1667 static int is_hw_sq_empty(struct ocrdma_qp *qp)
1668 {
1669         return (qp->sq.tail == qp->sq.head);
1670 }
1671
1672 static int is_hw_rq_empty(struct ocrdma_qp *qp)
1673 {
1674         return (qp->rq.tail == qp->rq.head);
1675 }
1676
1677 static void *ocrdma_hwq_head(struct ocrdma_qp_hwq_info *q)
1678 {
1679         return q->va + (q->head * q->entry_size);
1680 }
1681
1682 static void *ocrdma_hwq_head_from_idx(struct ocrdma_qp_hwq_info *q,
1683                                       u32 idx)
1684 {
1685         return q->va + (idx * q->entry_size);
1686 }
1687
1688 static void ocrdma_hwq_inc_head(struct ocrdma_qp_hwq_info *q)
1689 {
1690         q->head = (q->head + 1) & q->max_wqe_idx;
1691 }
1692
1693 static void ocrdma_hwq_inc_tail(struct ocrdma_qp_hwq_info *q)
1694 {
1695         q->tail = (q->tail + 1) & q->max_wqe_idx;
1696 }
1697
1698 /* discard the cqe for a given QP */
1699 static void ocrdma_discard_cqes(struct ocrdma_qp *qp, struct ocrdma_cq *cq)
1700 {
1701         unsigned long cq_flags;
1702         unsigned long flags;
1703         int discard_cnt = 0;
1704         u32 cur_getp, stop_getp;
1705         struct ocrdma_cqe *cqe;
1706         u32 qpn = 0, wqe_idx = 0;
1707
1708         spin_lock_irqsave(&cq->cq_lock, cq_flags);
1709
1710         /* traverse through the CQEs in the hw CQ,
1711          * find the matching CQE for a given qp,
1712          * mark the matching one discarded by clearing qpn.
1713          * ring the doorbell in the poll_cq() as
1714          * we don't complete out of order cqe.
1715          */
1716
1717         cur_getp = cq->getp;
1718         /* find upto when do we reap the cq. */
1719         stop_getp = cur_getp;
1720         do {
1721                 if (is_hw_sq_empty(qp) && (!qp->srq && is_hw_rq_empty(qp)))
1722                         break;
1723
1724                 cqe = cq->va + cur_getp;
1725                 /* if (a) done reaping whole hw cq, or
1726                  *    (b) qp_xq becomes empty.
1727                  * then exit
1728                  */
1729                 qpn = cqe->cmn.qpn & OCRDMA_CQE_QPN_MASK;
1730                 /* if previously discarded cqe found, skip that too. */
1731                 /* check for matching qp */
1732                 if (qpn == 0 || qpn != qp->id)
1733                         goto skip_cqe;
1734
1735                 if (is_cqe_for_sq(cqe)) {
1736                         ocrdma_hwq_inc_tail(&qp->sq);
1737                 } else {
1738                         if (qp->srq) {
1739                                 wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
1740                                         OCRDMA_CQE_BUFTAG_SHIFT) &
1741                                         qp->srq->rq.max_wqe_idx;
1742                                 BUG_ON(wqe_idx < 1);
1743                                 spin_lock_irqsave(&qp->srq->q_lock, flags);
1744                                 ocrdma_hwq_inc_tail(&qp->srq->rq);
1745                                 ocrdma_srq_toggle_bit(qp->srq, wqe_idx - 1);
1746                                 spin_unlock_irqrestore(&qp->srq->q_lock, flags);
1747
1748                         } else {
1749                                 ocrdma_hwq_inc_tail(&qp->rq);
1750                         }
1751                 }
1752                 /* mark cqe discarded so that it is not picked up later
1753                  * in the poll_cq().
1754                  */
1755                 discard_cnt += 1;
1756                 cqe->cmn.qpn = 0;
1757 skip_cqe:
1758                 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
1759         } while (cur_getp != stop_getp);
1760         spin_unlock_irqrestore(&cq->cq_lock, cq_flags);
1761 }
1762
1763 void ocrdma_del_flush_qp(struct ocrdma_qp *qp)
1764 {
1765         int found = false;
1766         unsigned long flags;
1767         struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
1768         /* sync with any active CQ poll */
1769
1770         spin_lock_irqsave(&dev->flush_q_lock, flags);
1771         found = ocrdma_is_qp_in_sq_flushlist(qp->sq_cq, qp);
1772         if (found)
1773                 list_del(&qp->sq_entry);
1774         if (!qp->srq) {
1775                 found = ocrdma_is_qp_in_rq_flushlist(qp->rq_cq, qp);
1776                 if (found)
1777                         list_del(&qp->rq_entry);
1778         }
1779         spin_unlock_irqrestore(&dev->flush_q_lock, flags);
1780 }
1781
1782 int ocrdma_destroy_qp(struct ib_qp *ibqp)
1783 {
1784         struct ocrdma_pd *pd;
1785         struct ocrdma_qp *qp;
1786         struct ocrdma_dev *dev;
1787         struct ib_qp_attr attrs;
1788         int attr_mask;
1789         unsigned long flags;
1790
1791         qp = get_ocrdma_qp(ibqp);
1792         dev = get_ocrdma_dev(ibqp->device);
1793
1794         pd = qp->pd;
1795
1796         /* change the QP state to ERROR */
1797         if (qp->state != OCRDMA_QPS_RST) {
1798                 attrs.qp_state = IB_QPS_ERR;
1799                 attr_mask = IB_QP_STATE;
1800                 _ocrdma_modify_qp(ibqp, &attrs, attr_mask);
1801         }
1802         /* ensure that CQEs for newly created QP (whose id may be same with
1803          * one which just getting destroyed are same), dont get
1804          * discarded until the old CQEs are discarded.
1805          */
1806         mutex_lock(&dev->dev_lock);
1807         (void) ocrdma_mbx_destroy_qp(dev, qp);
1808
1809         /*
1810          * acquire CQ lock while destroy is in progress, in order to
1811          * protect against proessing in-flight CQEs for this QP.
1812          */
1813         spin_lock_irqsave(&qp->sq_cq->cq_lock, flags);
1814         if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1815                 spin_lock(&qp->rq_cq->cq_lock);
1816
1817         ocrdma_del_qpn_map(dev, qp);
1818
1819         if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1820                 spin_unlock(&qp->rq_cq->cq_lock);
1821         spin_unlock_irqrestore(&qp->sq_cq->cq_lock, flags);
1822
1823         if (!pd->uctx) {
1824                 ocrdma_discard_cqes(qp, qp->sq_cq);
1825                 ocrdma_discard_cqes(qp, qp->rq_cq);
1826         }
1827         mutex_unlock(&dev->dev_lock);
1828
1829         if (pd->uctx) {
1830                 ocrdma_del_mmap(pd->uctx, (u64) qp->sq.pa,
1831                                 PAGE_ALIGN(qp->sq.len));
1832                 if (!qp->srq)
1833                         ocrdma_del_mmap(pd->uctx, (u64) qp->rq.pa,
1834                                         PAGE_ALIGN(qp->rq.len));
1835         }
1836
1837         ocrdma_del_flush_qp(qp);
1838
1839         kfree(qp->wqe_wr_id_tbl);
1840         kfree(qp->rqe_wr_id_tbl);
1841         kfree(qp);
1842         return 0;
1843 }
1844
1845 static int ocrdma_copy_srq_uresp(struct ocrdma_dev *dev, struct ocrdma_srq *srq,
1846                                 struct ib_udata *udata)
1847 {
1848         int status;
1849         struct ocrdma_create_srq_uresp uresp;
1850
1851         memset(&uresp, 0, sizeof(uresp));
1852         uresp.rq_dbid = srq->rq.dbid;
1853         uresp.num_rq_pages = 1;
1854         uresp.rq_page_addr[0] = virt_to_phys(srq->rq.va);
1855         uresp.rq_page_size = srq->rq.len;
1856         uresp.db_page_addr = dev->nic_info.unmapped_db +
1857             (srq->pd->id * dev->nic_info.db_page_size);
1858         uresp.db_page_size = dev->nic_info.db_page_size;
1859         uresp.num_rqe_allocated = srq->rq.max_cnt;
1860         if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1861                 uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1862                 uresp.db_shift = 24;
1863         } else {
1864                 uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
1865                 uresp.db_shift = 16;
1866         }
1867
1868         status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1869         if (status)
1870                 return status;
1871         status = ocrdma_add_mmap(srq->pd->uctx, uresp.rq_page_addr[0],
1872                                  uresp.rq_page_size);
1873         if (status)
1874                 return status;
1875         return status;
1876 }
1877
1878 struct ib_srq *ocrdma_create_srq(struct ib_pd *ibpd,
1879                                  struct ib_srq_init_attr *init_attr,
1880                                  struct ib_udata *udata)
1881 {
1882         int status = -ENOMEM;
1883         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1884         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1885         struct ocrdma_srq *srq;
1886
1887         if (init_attr->attr.max_sge > dev->attr.max_recv_sge)
1888                 return ERR_PTR(-EINVAL);
1889         if (init_attr->attr.max_wr > dev->attr.max_rqe)
1890                 return ERR_PTR(-EINVAL);
1891
1892         srq = kzalloc(sizeof(*srq), GFP_KERNEL);
1893         if (!srq)
1894                 return ERR_PTR(status);
1895
1896         spin_lock_init(&srq->q_lock);
1897         srq->pd = pd;
1898         srq->db = dev->nic_info.db + (pd->id * dev->nic_info.db_page_size);
1899         status = ocrdma_mbx_create_srq(dev, srq, init_attr, pd);
1900         if (status)
1901                 goto err;
1902
1903         if (udata == NULL) {
1904                 srq->rqe_wr_id_tbl = kzalloc(sizeof(u64) * srq->rq.max_cnt,
1905                             GFP_KERNEL);
1906                 if (srq->rqe_wr_id_tbl == NULL)
1907                         goto arm_err;
1908
1909                 srq->bit_fields_len = (srq->rq.max_cnt / 32) +
1910                     (srq->rq.max_cnt % 32 ? 1 : 0);
1911                 srq->idx_bit_fields =
1912                     kmalloc(srq->bit_fields_len * sizeof(u32), GFP_KERNEL);
1913                 if (srq->idx_bit_fields == NULL)
1914                         goto arm_err;
1915                 memset(srq->idx_bit_fields, 0xff,
1916                        srq->bit_fields_len * sizeof(u32));
1917         }
1918
1919         if (init_attr->attr.srq_limit) {
1920                 status = ocrdma_mbx_modify_srq(srq, &init_attr->attr);
1921                 if (status)
1922                         goto arm_err;
1923         }
1924
1925         if (udata) {
1926                 status = ocrdma_copy_srq_uresp(dev, srq, udata);
1927                 if (status)
1928                         goto arm_err;
1929         }
1930
1931         return &srq->ibsrq;
1932
1933 arm_err:
1934         ocrdma_mbx_destroy_srq(dev, srq);
1935 err:
1936         kfree(srq->rqe_wr_id_tbl);
1937         kfree(srq->idx_bit_fields);
1938         kfree(srq);
1939         return ERR_PTR(status);
1940 }
1941
1942 int ocrdma_modify_srq(struct ib_srq *ibsrq,
1943                       struct ib_srq_attr *srq_attr,
1944                       enum ib_srq_attr_mask srq_attr_mask,
1945                       struct ib_udata *udata)
1946 {
1947         int status;
1948         struct ocrdma_srq *srq;
1949
1950         srq = get_ocrdma_srq(ibsrq);
1951         if (srq_attr_mask & IB_SRQ_MAX_WR)
1952                 status = -EINVAL;
1953         else
1954                 status = ocrdma_mbx_modify_srq(srq, srq_attr);
1955         return status;
1956 }
1957
1958 int ocrdma_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
1959 {
1960         int status;
1961         struct ocrdma_srq *srq;
1962
1963         srq = get_ocrdma_srq(ibsrq);
1964         status = ocrdma_mbx_query_srq(srq, srq_attr);
1965         return status;
1966 }
1967
1968 int ocrdma_destroy_srq(struct ib_srq *ibsrq)
1969 {
1970         int status;
1971         struct ocrdma_srq *srq;
1972         struct ocrdma_dev *dev = get_ocrdma_dev(ibsrq->device);
1973
1974         srq = get_ocrdma_srq(ibsrq);
1975
1976         status = ocrdma_mbx_destroy_srq(dev, srq);
1977
1978         if (srq->pd->uctx)
1979                 ocrdma_del_mmap(srq->pd->uctx, (u64) srq->rq.pa,
1980                                 PAGE_ALIGN(srq->rq.len));
1981
1982         kfree(srq->idx_bit_fields);
1983         kfree(srq->rqe_wr_id_tbl);
1984         kfree(srq);
1985         return status;
1986 }
1987
1988 /* unprivileged verbs and their support functions. */
1989 static void ocrdma_build_ud_hdr(struct ocrdma_qp *qp,
1990                                 struct ocrdma_hdr_wqe *hdr,
1991                                 struct ib_send_wr *wr)
1992 {
1993         struct ocrdma_ewqe_ud_hdr *ud_hdr =
1994                 (struct ocrdma_ewqe_ud_hdr *)(hdr + 1);
1995         struct ocrdma_ah *ah = get_ocrdma_ah(ud_wr(wr)->ah);
1996
1997         ud_hdr->rsvd_dest_qpn = ud_wr(wr)->remote_qpn;
1998         if (qp->qp_type == IB_QPT_GSI)
1999                 ud_hdr->qkey = qp->qkey;
2000         else
2001                 ud_hdr->qkey = ud_wr(wr)->remote_qkey;
2002         ud_hdr->rsvd_ahid = ah->id;
2003         ud_hdr->hdr_type = ah->hdr_type;
2004         if (ah->av->valid & OCRDMA_AV_VLAN_VALID)
2005                 hdr->cw |= (OCRDMA_FLAG_AH_VLAN_PR << OCRDMA_WQE_FLAGS_SHIFT);
2006 }
2007
2008 static void ocrdma_build_sges(struct ocrdma_hdr_wqe *hdr,
2009                               struct ocrdma_sge *sge, int num_sge,
2010                               struct ib_sge *sg_list)
2011 {
2012         int i;
2013
2014         for (i = 0; i < num_sge; i++) {
2015                 sge[i].lrkey = sg_list[i].lkey;
2016                 sge[i].addr_lo = sg_list[i].addr;
2017                 sge[i].addr_hi = upper_32_bits(sg_list[i].addr);
2018                 sge[i].len = sg_list[i].length;
2019                 hdr->total_len += sg_list[i].length;
2020         }
2021         if (num_sge == 0)
2022                 memset(sge, 0, sizeof(*sge));
2023 }
2024
2025 static inline uint32_t ocrdma_sglist_len(struct ib_sge *sg_list, int num_sge)
2026 {
2027         uint32_t total_len = 0, i;
2028
2029         for (i = 0; i < num_sge; i++)
2030                 total_len += sg_list[i].length;
2031         return total_len;
2032 }
2033
2034
2035 static int ocrdma_build_inline_sges(struct ocrdma_qp *qp,
2036                                     struct ocrdma_hdr_wqe *hdr,
2037                                     struct ocrdma_sge *sge,
2038                                     struct ib_send_wr *wr, u32 wqe_size)
2039 {
2040         int i;
2041         char *dpp_addr;
2042
2043         if (wr->send_flags & IB_SEND_INLINE && qp->qp_type != IB_QPT_UD) {
2044                 hdr->total_len = ocrdma_sglist_len(wr->sg_list, wr->num_sge);
2045                 if (unlikely(hdr->total_len > qp->max_inline_data)) {
2046                         pr_err("%s() supported_len=0x%x,\n"
2047                                " unsupported len req=0x%x\n", __func__,
2048                                 qp->max_inline_data, hdr->total_len);
2049                         return -EINVAL;
2050                 }
2051                 dpp_addr = (char *)sge;
2052                 for (i = 0; i < wr->num_sge; i++) {
2053                         memcpy(dpp_addr,
2054                                (void *)(unsigned long)wr->sg_list[i].addr,
2055                                wr->sg_list[i].length);
2056                         dpp_addr += wr->sg_list[i].length;
2057                 }
2058
2059                 wqe_size += roundup(hdr->total_len, OCRDMA_WQE_ALIGN_BYTES);
2060                 if (0 == hdr->total_len)
2061                         wqe_size += sizeof(struct ocrdma_sge);
2062                 hdr->cw |= (OCRDMA_TYPE_INLINE << OCRDMA_WQE_TYPE_SHIFT);
2063         } else {
2064                 ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
2065                 if (wr->num_sge)
2066                         wqe_size += (wr->num_sge * sizeof(struct ocrdma_sge));
2067                 else
2068                         wqe_size += sizeof(struct ocrdma_sge);
2069                 hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2070         }
2071         hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2072         return 0;
2073 }
2074
2075 static int ocrdma_build_send(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2076                              struct ib_send_wr *wr)
2077 {
2078         int status;
2079         struct ocrdma_sge *sge;
2080         u32 wqe_size = sizeof(*hdr);
2081
2082         if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2083                 ocrdma_build_ud_hdr(qp, hdr, wr);
2084                 sge = (struct ocrdma_sge *)(hdr + 2);
2085                 wqe_size += sizeof(struct ocrdma_ewqe_ud_hdr);
2086         } else {
2087                 sge = (struct ocrdma_sge *)(hdr + 1);
2088         }
2089
2090         status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
2091         return status;
2092 }
2093
2094 static int ocrdma_build_write(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2095                               struct ib_send_wr *wr)
2096 {
2097         int status;
2098         struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
2099         struct ocrdma_sge *sge = ext_rw + 1;
2100         u32 wqe_size = sizeof(*hdr) + sizeof(*ext_rw);
2101
2102         status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
2103         if (status)
2104                 return status;
2105         ext_rw->addr_lo = rdma_wr(wr)->remote_addr;
2106         ext_rw->addr_hi = upper_32_bits(rdma_wr(wr)->remote_addr);
2107         ext_rw->lrkey = rdma_wr(wr)->rkey;
2108         ext_rw->len = hdr->total_len;
2109         return 0;
2110 }
2111
2112 static void ocrdma_build_read(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2113                               struct ib_send_wr *wr)
2114 {
2115         struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
2116         struct ocrdma_sge *sge = ext_rw + 1;
2117         u32 wqe_size = ((wr->num_sge + 1) * sizeof(struct ocrdma_sge)) +
2118             sizeof(struct ocrdma_hdr_wqe);
2119
2120         ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
2121         hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2122         hdr->cw |= (OCRDMA_READ << OCRDMA_WQE_OPCODE_SHIFT);
2123         hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2124
2125         ext_rw->addr_lo = rdma_wr(wr)->remote_addr;
2126         ext_rw->addr_hi = upper_32_bits(rdma_wr(wr)->remote_addr);
2127         ext_rw->lrkey = rdma_wr(wr)->rkey;
2128         ext_rw->len = hdr->total_len;
2129 }
2130
2131 static int get_encoded_page_size(int pg_sz)
2132 {
2133         /* Max size is 256M 4096 << 16 */
2134         int i = 0;
2135         for (; i < 17; i++)
2136                 if (pg_sz == (4096 << i))
2137                         break;
2138         return i;
2139 }
2140
2141 static int ocrdma_build_reg(struct ocrdma_qp *qp,
2142                             struct ocrdma_hdr_wqe *hdr,
2143                             struct ib_reg_wr *wr)
2144 {
2145         u64 fbo;
2146         struct ocrdma_ewqe_fr *fast_reg = (struct ocrdma_ewqe_fr *)(hdr + 1);
2147         struct ocrdma_mr *mr = get_ocrdma_mr(wr->mr);
2148         struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
2149         struct ocrdma_pbe *pbe;
2150         u32 wqe_size = sizeof(*fast_reg) + sizeof(*hdr);
2151         int num_pbes = 0, i;
2152
2153         wqe_size = roundup(wqe_size, OCRDMA_WQE_ALIGN_BYTES);
2154
2155         hdr->cw |= (OCRDMA_FR_MR << OCRDMA_WQE_OPCODE_SHIFT);
2156         hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2157
2158         if (wr->access & IB_ACCESS_LOCAL_WRITE)
2159                 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_LOCAL_WR;
2160         if (wr->access & IB_ACCESS_REMOTE_WRITE)
2161                 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_WR;
2162         if (wr->access & IB_ACCESS_REMOTE_READ)
2163                 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_RD;
2164         hdr->lkey = wr->key;
2165         hdr->total_len = mr->ibmr.length;
2166
2167         fbo = mr->ibmr.iova - mr->pages[0];
2168
2169         fast_reg->va_hi = upper_32_bits(mr->ibmr.iova);
2170         fast_reg->va_lo = (u32) (mr->ibmr.iova & 0xffffffff);
2171         fast_reg->fbo_hi = upper_32_bits(fbo);
2172         fast_reg->fbo_lo = (u32) fbo & 0xffffffff;
2173         fast_reg->num_sges = mr->npages;
2174         fast_reg->size_sge = get_encoded_page_size(mr->ibmr.page_size);
2175
2176         pbe = pbl_tbl->va;
2177         for (i = 0; i < mr->npages; i++) {
2178                 u64 buf_addr = mr->pages[i];
2179
2180                 pbe->pa_lo = cpu_to_le32((u32) (buf_addr & PAGE_MASK));
2181                 pbe->pa_hi = cpu_to_le32((u32) upper_32_bits(buf_addr));
2182                 num_pbes += 1;
2183                 pbe++;
2184
2185                 /* if the pbl is full storing the pbes,
2186                  * move to next pbl.
2187                 */
2188                 if (num_pbes == (mr->hwmr.pbl_size/sizeof(u64))) {
2189                         pbl_tbl++;
2190                         pbe = (struct ocrdma_pbe *)pbl_tbl->va;
2191                 }
2192         }
2193
2194         return 0;
2195 }
2196
2197 static void ocrdma_ring_sq_db(struct ocrdma_qp *qp)
2198 {
2199         u32 val = qp->sq.dbid | (1 << OCRDMA_DB_SQ_SHIFT);
2200
2201         iowrite32(val, qp->sq_db);
2202 }
2203
2204 int ocrdma_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2205                      struct ib_send_wr **bad_wr)
2206 {
2207         int status = 0;
2208         struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2209         struct ocrdma_hdr_wqe *hdr;
2210         unsigned long flags;
2211
2212         spin_lock_irqsave(&qp->q_lock, flags);
2213         if (qp->state != OCRDMA_QPS_RTS && qp->state != OCRDMA_QPS_SQD) {
2214                 spin_unlock_irqrestore(&qp->q_lock, flags);
2215                 *bad_wr = wr;
2216                 return -EINVAL;
2217         }
2218
2219         while (wr) {
2220                 if (qp->qp_type == IB_QPT_UD &&
2221                     (wr->opcode != IB_WR_SEND &&
2222                      wr->opcode != IB_WR_SEND_WITH_IMM)) {
2223                         *bad_wr = wr;
2224                         status = -EINVAL;
2225                         break;
2226                 }
2227                 if (ocrdma_hwq_free_cnt(&qp->sq) == 0 ||
2228                     wr->num_sge > qp->sq.max_sges) {
2229                         *bad_wr = wr;
2230                         status = -ENOMEM;
2231                         break;
2232                 }
2233                 hdr = ocrdma_hwq_head(&qp->sq);
2234                 hdr->cw = 0;
2235                 if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2236                         hdr->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2237                 if (wr->send_flags & IB_SEND_FENCE)
2238                         hdr->cw |=
2239                             (OCRDMA_FLAG_FENCE_L << OCRDMA_WQE_FLAGS_SHIFT);
2240                 if (wr->send_flags & IB_SEND_SOLICITED)
2241                         hdr->cw |=
2242                             (OCRDMA_FLAG_SOLICIT << OCRDMA_WQE_FLAGS_SHIFT);
2243                 hdr->total_len = 0;
2244                 switch (wr->opcode) {
2245                 case IB_WR_SEND_WITH_IMM:
2246                         hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2247                         hdr->immdt = ntohl(wr->ex.imm_data);
2248                 case IB_WR_SEND:
2249                         hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2250                         ocrdma_build_send(qp, hdr, wr);
2251                         break;
2252                 case IB_WR_SEND_WITH_INV:
2253                         hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
2254                         hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2255                         hdr->lkey = wr->ex.invalidate_rkey;
2256                         status = ocrdma_build_send(qp, hdr, wr);
2257                         break;
2258                 case IB_WR_RDMA_WRITE_WITH_IMM:
2259                         hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2260                         hdr->immdt = ntohl(wr->ex.imm_data);
2261                 case IB_WR_RDMA_WRITE:
2262                         hdr->cw |= (OCRDMA_WRITE << OCRDMA_WQE_OPCODE_SHIFT);
2263                         status = ocrdma_build_write(qp, hdr, wr);
2264                         break;
2265                 case IB_WR_RDMA_READ:
2266                         ocrdma_build_read(qp, hdr, wr);
2267                         break;
2268                 case IB_WR_LOCAL_INV:
2269                         hdr->cw |=
2270                             (OCRDMA_LKEY_INV << OCRDMA_WQE_OPCODE_SHIFT);
2271                         hdr->cw |= ((sizeof(struct ocrdma_hdr_wqe) +
2272                                         sizeof(struct ocrdma_sge)) /
2273                                 OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT;
2274                         hdr->lkey = wr->ex.invalidate_rkey;
2275                         break;
2276                 case IB_WR_REG_MR:
2277                         status = ocrdma_build_reg(qp, hdr, reg_wr(wr));
2278                         break;
2279                 default:
2280                         status = -EINVAL;
2281                         break;
2282                 }
2283                 if (status) {
2284                         *bad_wr = wr;
2285                         break;
2286                 }
2287                 if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2288                         qp->wqe_wr_id_tbl[qp->sq.head].signaled = 1;
2289                 else
2290                         qp->wqe_wr_id_tbl[qp->sq.head].signaled = 0;
2291                 qp->wqe_wr_id_tbl[qp->sq.head].wrid = wr->wr_id;
2292                 ocrdma_cpu_to_le32(hdr, ((hdr->cw >> OCRDMA_WQE_SIZE_SHIFT) &
2293                                    OCRDMA_WQE_SIZE_MASK) * OCRDMA_WQE_STRIDE);
2294                 /* make sure wqe is written before adapter can access it */
2295                 wmb();
2296                 /* inform hw to start processing it */
2297                 ocrdma_ring_sq_db(qp);
2298
2299                 /* update pointer, counter for next wr */
2300                 ocrdma_hwq_inc_head(&qp->sq);
2301                 wr = wr->next;
2302         }
2303         spin_unlock_irqrestore(&qp->q_lock, flags);
2304         return status;
2305 }
2306
2307 static void ocrdma_ring_rq_db(struct ocrdma_qp *qp)
2308 {
2309         u32 val = qp->rq.dbid | (1 << OCRDMA_DB_RQ_SHIFT);
2310
2311         iowrite32(val, qp->rq_db);
2312 }
2313
2314 static void ocrdma_build_rqe(struct ocrdma_hdr_wqe *rqe, struct ib_recv_wr *wr,
2315                              u16 tag)
2316 {
2317         u32 wqe_size = 0;
2318         struct ocrdma_sge *sge;
2319         if (wr->num_sge)
2320                 wqe_size = (wr->num_sge * sizeof(*sge)) + sizeof(*rqe);
2321         else
2322                 wqe_size = sizeof(*sge) + sizeof(*rqe);
2323
2324         rqe->cw = ((wqe_size / OCRDMA_WQE_STRIDE) <<
2325                                 OCRDMA_WQE_SIZE_SHIFT);
2326         rqe->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2327         rqe->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2328         rqe->total_len = 0;
2329         rqe->rsvd_tag = tag;
2330         sge = (struct ocrdma_sge *)(rqe + 1);
2331         ocrdma_build_sges(rqe, sge, wr->num_sge, wr->sg_list);
2332         ocrdma_cpu_to_le32(rqe, wqe_size);
2333 }
2334
2335 int ocrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2336                      struct ib_recv_wr **bad_wr)
2337 {
2338         int status = 0;
2339         unsigned long flags;
2340         struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2341         struct ocrdma_hdr_wqe *rqe;
2342
2343         spin_lock_irqsave(&qp->q_lock, flags);
2344         if (qp->state == OCRDMA_QPS_RST || qp->state == OCRDMA_QPS_ERR) {
2345                 spin_unlock_irqrestore(&qp->q_lock, flags);
2346                 *bad_wr = wr;
2347                 return -EINVAL;
2348         }
2349         while (wr) {
2350                 if (ocrdma_hwq_free_cnt(&qp->rq) == 0 ||
2351                     wr->num_sge > qp->rq.max_sges) {
2352                         *bad_wr = wr;
2353                         status = -ENOMEM;
2354                         break;
2355                 }
2356                 rqe = ocrdma_hwq_head(&qp->rq);
2357                 ocrdma_build_rqe(rqe, wr, 0);
2358
2359                 qp->rqe_wr_id_tbl[qp->rq.head] = wr->wr_id;
2360                 /* make sure rqe is written before adapter can access it */
2361                 wmb();
2362
2363                 /* inform hw to start processing it */
2364                 ocrdma_ring_rq_db(qp);
2365
2366                 /* update pointer, counter for next wr */
2367                 ocrdma_hwq_inc_head(&qp->rq);
2368                 wr = wr->next;
2369         }
2370         spin_unlock_irqrestore(&qp->q_lock, flags);
2371         return status;
2372 }
2373
2374 /* cqe for srq's rqe can potentially arrive out of order.
2375  * index gives the entry in the shadow table where to store
2376  * the wr_id. tag/index is returned in cqe to reference back
2377  * for a given rqe.
2378  */
2379 static int ocrdma_srq_get_idx(struct ocrdma_srq *srq)
2380 {
2381         int row = 0;
2382         int indx = 0;
2383
2384         for (row = 0; row < srq->bit_fields_len; row++) {
2385                 if (srq->idx_bit_fields[row]) {
2386                         indx = ffs(srq->idx_bit_fields[row]);
2387                         indx = (row * 32) + (indx - 1);
2388                         BUG_ON(indx >= srq->rq.max_cnt);
2389                         ocrdma_srq_toggle_bit(srq, indx);
2390                         break;
2391                 }
2392         }
2393
2394         BUG_ON(row == srq->bit_fields_len);
2395         return indx + 1; /* Use from index 1 */
2396 }
2397
2398 static void ocrdma_ring_srq_db(struct ocrdma_srq *srq)
2399 {
2400         u32 val = srq->rq.dbid | (1 << 16);
2401
2402         iowrite32(val, srq->db + OCRDMA_DB_GEN2_SRQ_OFFSET);
2403 }
2404
2405 int ocrdma_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
2406                          struct ib_recv_wr **bad_wr)
2407 {
2408         int status = 0;
2409         unsigned long flags;
2410         struct ocrdma_srq *srq;
2411         struct ocrdma_hdr_wqe *rqe;
2412         u16 tag;
2413
2414         srq = get_ocrdma_srq(ibsrq);
2415
2416         spin_lock_irqsave(&srq->q_lock, flags);
2417         while (wr) {
2418                 if (ocrdma_hwq_free_cnt(&srq->rq) == 0 ||
2419                     wr->num_sge > srq->rq.max_sges) {
2420                         status = -ENOMEM;
2421                         *bad_wr = wr;
2422                         break;
2423                 }
2424                 tag = ocrdma_srq_get_idx(srq);
2425                 rqe = ocrdma_hwq_head(&srq->rq);
2426                 ocrdma_build_rqe(rqe, wr, tag);
2427
2428                 srq->rqe_wr_id_tbl[tag] = wr->wr_id;
2429                 /* make sure rqe is written before adapter can perform DMA */
2430                 wmb();
2431                 /* inform hw to start processing it */
2432                 ocrdma_ring_srq_db(srq);
2433                 /* update pointer, counter for next wr */
2434                 ocrdma_hwq_inc_head(&srq->rq);
2435                 wr = wr->next;
2436         }
2437         spin_unlock_irqrestore(&srq->q_lock, flags);
2438         return status;
2439 }
2440
2441 static enum ib_wc_status ocrdma_to_ibwc_err(u16 status)
2442 {
2443         enum ib_wc_status ibwc_status;
2444
2445         switch (status) {
2446         case OCRDMA_CQE_GENERAL_ERR:
2447                 ibwc_status = IB_WC_GENERAL_ERR;
2448                 break;
2449         case OCRDMA_CQE_LOC_LEN_ERR:
2450                 ibwc_status = IB_WC_LOC_LEN_ERR;
2451                 break;
2452         case OCRDMA_CQE_LOC_QP_OP_ERR:
2453                 ibwc_status = IB_WC_LOC_QP_OP_ERR;
2454                 break;
2455         case OCRDMA_CQE_LOC_EEC_OP_ERR:
2456                 ibwc_status = IB_WC_LOC_EEC_OP_ERR;
2457                 break;
2458         case OCRDMA_CQE_LOC_PROT_ERR:
2459                 ibwc_status = IB_WC_LOC_PROT_ERR;
2460                 break;
2461         case OCRDMA_CQE_WR_FLUSH_ERR:
2462                 ibwc_status = IB_WC_WR_FLUSH_ERR;
2463                 break;
2464         case OCRDMA_CQE_MW_BIND_ERR:
2465                 ibwc_status = IB_WC_MW_BIND_ERR;
2466                 break;
2467         case OCRDMA_CQE_BAD_RESP_ERR:
2468                 ibwc_status = IB_WC_BAD_RESP_ERR;
2469                 break;
2470         case OCRDMA_CQE_LOC_ACCESS_ERR:
2471                 ibwc_status = IB_WC_LOC_ACCESS_ERR;
2472                 break;
2473         case OCRDMA_CQE_REM_INV_REQ_ERR:
2474                 ibwc_status = IB_WC_REM_INV_REQ_ERR;
2475                 break;
2476         case OCRDMA_CQE_REM_ACCESS_ERR:
2477                 ibwc_status = IB_WC_REM_ACCESS_ERR;
2478                 break;
2479         case OCRDMA_CQE_REM_OP_ERR:
2480                 ibwc_status = IB_WC_REM_OP_ERR;
2481                 break;
2482         case OCRDMA_CQE_RETRY_EXC_ERR:
2483                 ibwc_status = IB_WC_RETRY_EXC_ERR;
2484                 break;
2485         case OCRDMA_CQE_RNR_RETRY_EXC_ERR:
2486                 ibwc_status = IB_WC_RNR_RETRY_EXC_ERR;
2487                 break;
2488         case OCRDMA_CQE_LOC_RDD_VIOL_ERR:
2489                 ibwc_status = IB_WC_LOC_RDD_VIOL_ERR;
2490                 break;
2491         case OCRDMA_CQE_REM_INV_RD_REQ_ERR:
2492                 ibwc_status = IB_WC_REM_INV_RD_REQ_ERR;
2493                 break;
2494         case OCRDMA_CQE_REM_ABORT_ERR:
2495                 ibwc_status = IB_WC_REM_ABORT_ERR;
2496                 break;
2497         case OCRDMA_CQE_INV_EECN_ERR:
2498                 ibwc_status = IB_WC_INV_EECN_ERR;
2499                 break;
2500         case OCRDMA_CQE_INV_EEC_STATE_ERR:
2501                 ibwc_status = IB_WC_INV_EEC_STATE_ERR;
2502                 break;
2503         case OCRDMA_CQE_FATAL_ERR:
2504                 ibwc_status = IB_WC_FATAL_ERR;
2505                 break;
2506         case OCRDMA_CQE_RESP_TIMEOUT_ERR:
2507                 ibwc_status = IB_WC_RESP_TIMEOUT_ERR;
2508                 break;
2509         default:
2510                 ibwc_status = IB_WC_GENERAL_ERR;
2511                 break;
2512         }
2513         return ibwc_status;
2514 }
2515
2516 static void ocrdma_update_wc(struct ocrdma_qp *qp, struct ib_wc *ibwc,
2517                       u32 wqe_idx)
2518 {
2519         struct ocrdma_hdr_wqe *hdr;
2520         struct ocrdma_sge *rw;
2521         int opcode;
2522
2523         hdr = ocrdma_hwq_head_from_idx(&qp->sq, wqe_idx);
2524
2525         ibwc->wr_id = qp->wqe_wr_id_tbl[wqe_idx].wrid;
2526         /* Undo the hdr->cw swap */
2527         opcode = le32_to_cpu(hdr->cw) & OCRDMA_WQE_OPCODE_MASK;
2528         switch (opcode) {
2529         case OCRDMA_WRITE:
2530                 ibwc->opcode = IB_WC_RDMA_WRITE;
2531                 break;
2532         case OCRDMA_READ:
2533                 rw = (struct ocrdma_sge *)(hdr + 1);
2534                 ibwc->opcode = IB_WC_RDMA_READ;
2535                 ibwc->byte_len = rw->len;
2536                 break;
2537         case OCRDMA_SEND:
2538                 ibwc->opcode = IB_WC_SEND;
2539                 break;
2540         case OCRDMA_FR_MR:
2541                 ibwc->opcode = IB_WC_REG_MR;
2542                 break;
2543         case OCRDMA_LKEY_INV:
2544                 ibwc->opcode = IB_WC_LOCAL_INV;
2545                 break;
2546         default:
2547                 ibwc->status = IB_WC_GENERAL_ERR;
2548                 pr_err("%s() invalid opcode received = 0x%x\n",
2549                        __func__, hdr->cw & OCRDMA_WQE_OPCODE_MASK);
2550                 break;
2551         }
2552 }
2553
2554 static void ocrdma_set_cqe_status_flushed(struct ocrdma_qp *qp,
2555                                                 struct ocrdma_cqe *cqe)
2556 {
2557         if (is_cqe_for_sq(cqe)) {
2558                 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2559                                 cqe->flags_status_srcqpn) &
2560                                         ~OCRDMA_CQE_STATUS_MASK);
2561                 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2562                                 cqe->flags_status_srcqpn) |
2563                                 (OCRDMA_CQE_WR_FLUSH_ERR <<
2564                                         OCRDMA_CQE_STATUS_SHIFT));
2565         } else {
2566                 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2567                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2568                                         cqe->flags_status_srcqpn) &
2569                                                 ~OCRDMA_CQE_UD_STATUS_MASK);
2570                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2571                                         cqe->flags_status_srcqpn) |
2572                                         (OCRDMA_CQE_WR_FLUSH_ERR <<
2573                                                 OCRDMA_CQE_UD_STATUS_SHIFT));
2574                 } else {
2575                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2576                                         cqe->flags_status_srcqpn) &
2577                                                 ~OCRDMA_CQE_STATUS_MASK);
2578                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2579                                         cqe->flags_status_srcqpn) |
2580                                         (OCRDMA_CQE_WR_FLUSH_ERR <<
2581                                                 OCRDMA_CQE_STATUS_SHIFT));
2582                 }
2583         }
2584 }
2585
2586 static bool ocrdma_update_err_cqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2587                                   struct ocrdma_qp *qp, int status)
2588 {
2589         bool expand = false;
2590
2591         ibwc->byte_len = 0;
2592         ibwc->qp = &qp->ibqp;
2593         ibwc->status = ocrdma_to_ibwc_err(status);
2594
2595         ocrdma_flush_qp(qp);
2596         ocrdma_qp_state_change(qp, IB_QPS_ERR, NULL);
2597
2598         /* if wqe/rqe pending for which cqe needs to be returned,
2599          * trigger inflating it.
2600          */
2601         if (!is_hw_rq_empty(qp) || !is_hw_sq_empty(qp)) {
2602                 expand = true;
2603                 ocrdma_set_cqe_status_flushed(qp, cqe);
2604         }
2605         return expand;
2606 }
2607
2608 static int ocrdma_update_err_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2609                                   struct ocrdma_qp *qp, int status)
2610 {
2611         ibwc->opcode = IB_WC_RECV;
2612         ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2613         ocrdma_hwq_inc_tail(&qp->rq);
2614
2615         return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2616 }
2617
2618 static int ocrdma_update_err_scqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2619                                   struct ocrdma_qp *qp, int status)
2620 {
2621         ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2622         ocrdma_hwq_inc_tail(&qp->sq);
2623
2624         return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2625 }
2626
2627
2628 static bool ocrdma_poll_err_scqe(struct ocrdma_qp *qp,
2629                                  struct ocrdma_cqe *cqe, struct ib_wc *ibwc,
2630                                  bool *polled, bool *stop)
2631 {
2632         bool expand;
2633         struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2634         int status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2635                 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2636         if (status < OCRDMA_MAX_CQE_ERR)
2637                 atomic_inc(&dev->cqe_err_stats[status]);
2638
2639         /* when hw sq is empty, but rq is not empty, so we continue
2640          * to keep the cqe in order to get the cq event again.
2641          */
2642         if (is_hw_sq_empty(qp) && !is_hw_rq_empty(qp)) {
2643                 /* when cq for rq and sq is same, it is safe to return
2644                  * flush cqe for RQEs.
2645                  */
2646                 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2647                         *polled = true;
2648                         status = OCRDMA_CQE_WR_FLUSH_ERR;
2649                         expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2650                 } else {
2651                         /* stop processing further cqe as this cqe is used for
2652                          * triggering cq event on buddy cq of RQ.
2653                          * When QP is destroyed, this cqe will be removed
2654                          * from the cq's hardware q.
2655                          */
2656                         *polled = false;
2657                         *stop = true;
2658                         expand = false;
2659                 }
2660         } else if (is_hw_sq_empty(qp)) {
2661                 /* Do nothing */
2662                 expand = false;
2663                 *polled = false;
2664                 *stop = false;
2665         } else {
2666                 *polled = true;
2667                 expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2668         }
2669         return expand;
2670 }
2671
2672 static bool ocrdma_poll_success_scqe(struct ocrdma_qp *qp,
2673                                      struct ocrdma_cqe *cqe,
2674                                      struct ib_wc *ibwc, bool *polled)
2675 {
2676         bool expand = false;
2677         int tail = qp->sq.tail;
2678         u32 wqe_idx;
2679
2680         if (!qp->wqe_wr_id_tbl[tail].signaled) {
2681                 *polled = false;    /* WC cannot be consumed yet */
2682         } else {
2683                 ibwc->status = IB_WC_SUCCESS;
2684                 ibwc->wc_flags = 0;
2685                 ibwc->qp = &qp->ibqp;
2686                 ocrdma_update_wc(qp, ibwc, tail);
2687                 *polled = true;
2688         }
2689         wqe_idx = (le32_to_cpu(cqe->wq.wqeidx) &
2690                         OCRDMA_CQE_WQEIDX_MASK) & qp->sq.max_wqe_idx;
2691         if (tail != wqe_idx)
2692                 expand = true; /* Coalesced CQE can't be consumed yet */
2693
2694         ocrdma_hwq_inc_tail(&qp->sq);
2695         return expand;
2696 }
2697
2698 static bool ocrdma_poll_scqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2699                              struct ib_wc *ibwc, bool *polled, bool *stop)
2700 {
2701         int status;
2702         bool expand;
2703
2704         status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2705                 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2706
2707         if (status == OCRDMA_CQE_SUCCESS)
2708                 expand = ocrdma_poll_success_scqe(qp, cqe, ibwc, polled);
2709         else
2710                 expand = ocrdma_poll_err_scqe(qp, cqe, ibwc, polled, stop);
2711         return expand;
2712 }
2713
2714 static int ocrdma_update_ud_rcqe(struct ocrdma_dev *dev, struct ib_wc *ibwc,
2715                                  struct ocrdma_cqe *cqe)
2716 {
2717         int status;
2718         u16 hdr_type = 0;
2719
2720         status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2721                 OCRDMA_CQE_UD_STATUS_MASK) >> OCRDMA_CQE_UD_STATUS_SHIFT;
2722         ibwc->src_qp = le32_to_cpu(cqe->flags_status_srcqpn) &
2723                                                 OCRDMA_CQE_SRCQP_MASK;
2724         ibwc->pkey_index = 0;
2725         ibwc->wc_flags = IB_WC_GRH;
2726         ibwc->byte_len = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2727                           OCRDMA_CQE_UD_XFER_LEN_SHIFT) &
2728                           OCRDMA_CQE_UD_XFER_LEN_MASK;
2729
2730         if (ocrdma_is_udp_encap_supported(dev)) {
2731                 hdr_type = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2732                             OCRDMA_CQE_UD_L3TYPE_SHIFT) &
2733                             OCRDMA_CQE_UD_L3TYPE_MASK;
2734                 ibwc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
2735                 ibwc->network_hdr_type = hdr_type;
2736         }
2737
2738         return status;
2739 }
2740
2741 static void ocrdma_update_free_srq_cqe(struct ib_wc *ibwc,
2742                                        struct ocrdma_cqe *cqe,
2743                                        struct ocrdma_qp *qp)
2744 {
2745         unsigned long flags;
2746         struct ocrdma_srq *srq;
2747         u32 wqe_idx;
2748
2749         srq = get_ocrdma_srq(qp->ibqp.srq);
2750         wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
2751                 OCRDMA_CQE_BUFTAG_SHIFT) & srq->rq.max_wqe_idx;
2752         BUG_ON(wqe_idx < 1);
2753
2754         ibwc->wr_id = srq->rqe_wr_id_tbl[wqe_idx];
2755         spin_lock_irqsave(&srq->q_lock, flags);
2756         ocrdma_srq_toggle_bit(srq, wqe_idx - 1);
2757         spin_unlock_irqrestore(&srq->q_lock, flags);
2758         ocrdma_hwq_inc_tail(&srq->rq);
2759 }
2760
2761 static bool ocrdma_poll_err_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2762                                 struct ib_wc *ibwc, bool *polled, bool *stop,
2763                                 int status)
2764 {
2765         bool expand;
2766         struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2767
2768         if (status < OCRDMA_MAX_CQE_ERR)
2769                 atomic_inc(&dev->cqe_err_stats[status]);
2770
2771         /* when hw_rq is empty, but wq is not empty, so continue
2772          * to keep the cqe to get the cq event again.
2773          */
2774         if (is_hw_rq_empty(qp) && !is_hw_sq_empty(qp)) {
2775                 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2776                         *polled = true;
2777                         status = OCRDMA_CQE_WR_FLUSH_ERR;
2778                         expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2779                 } else {
2780                         *polled = false;
2781                         *stop = true;
2782                         expand = false;
2783                 }
2784         } else if (is_hw_rq_empty(qp)) {
2785                 /* Do nothing */
2786                 expand = false;
2787                 *polled = false;
2788                 *stop = false;
2789         } else {
2790                 *polled = true;
2791                 expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2792         }
2793         return expand;
2794 }
2795
2796 static void ocrdma_poll_success_rcqe(struct ocrdma_qp *qp,
2797                                      struct ocrdma_cqe *cqe, struct ib_wc *ibwc)
2798 {
2799         struct ocrdma_dev *dev;
2800
2801         dev = get_ocrdma_dev(qp->ibqp.device);
2802         ibwc->opcode = IB_WC_RECV;
2803         ibwc->qp = &qp->ibqp;
2804         ibwc->status = IB_WC_SUCCESS;
2805
2806         if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI)
2807                 ocrdma_update_ud_rcqe(dev, ibwc, cqe);
2808         else
2809                 ibwc->byte_len = le32_to_cpu(cqe->rq.rxlen);
2810
2811         if (is_cqe_imm(cqe)) {
2812                 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2813                 ibwc->wc_flags |= IB_WC_WITH_IMM;
2814         } else if (is_cqe_wr_imm(cqe)) {
2815                 ibwc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2816                 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2817                 ibwc->wc_flags |= IB_WC_WITH_IMM;
2818         } else if (is_cqe_invalidated(cqe)) {
2819                 ibwc->ex.invalidate_rkey = le32_to_cpu(cqe->rq.lkey_immdt);
2820                 ibwc->wc_flags |= IB_WC_WITH_INVALIDATE;
2821         }
2822         if (qp->ibqp.srq) {
2823                 ocrdma_update_free_srq_cqe(ibwc, cqe, qp);
2824         } else {
2825                 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2826                 ocrdma_hwq_inc_tail(&qp->rq);
2827         }
2828 }
2829
2830 static bool ocrdma_poll_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2831                              struct ib_wc *ibwc, bool *polled, bool *stop)
2832 {
2833         int status;
2834         bool expand = false;
2835
2836         ibwc->wc_flags = 0;
2837         if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2838                 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2839                                         OCRDMA_CQE_UD_STATUS_MASK) >>
2840                                         OCRDMA_CQE_UD_STATUS_SHIFT;
2841         } else {
2842                 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2843                              OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2844         }
2845
2846         if (status == OCRDMA_CQE_SUCCESS) {
2847                 *polled = true;
2848                 ocrdma_poll_success_rcqe(qp, cqe, ibwc);
2849         } else {
2850                 expand = ocrdma_poll_err_rcqe(qp, cqe, ibwc, polled, stop,
2851                                               status);
2852         }
2853         return expand;
2854 }
2855
2856 static void ocrdma_change_cq_phase(struct ocrdma_cq *cq, struct ocrdma_cqe *cqe,
2857                                    u16 cur_getp)
2858 {
2859         if (cq->phase_change) {
2860                 if (cur_getp == 0)
2861                         cq->phase = (~cq->phase & OCRDMA_CQE_VALID);
2862         } else {
2863                 /* clear valid bit */
2864                 cqe->flags_status_srcqpn = 0;
2865         }
2866 }
2867
2868 static int ocrdma_poll_hwcq(struct ocrdma_cq *cq, int num_entries,
2869                             struct ib_wc *ibwc)
2870 {
2871         u16 qpn = 0;
2872         int i = 0;
2873         bool expand = false;
2874         int polled_hw_cqes = 0;
2875         struct ocrdma_qp *qp = NULL;
2876         struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
2877         struct ocrdma_cqe *cqe;
2878         u16 cur_getp; bool polled = false; bool stop = false;
2879
2880         cur_getp = cq->getp;
2881         while (num_entries) {
2882                 cqe = cq->va + cur_getp;
2883                 /* check whether valid cqe or not */
2884                 if (!is_cqe_valid(cq, cqe))
2885                         break;
2886                 qpn = (le32_to_cpu(cqe->cmn.qpn) & OCRDMA_CQE_QPN_MASK);
2887                 /* ignore discarded cqe */
2888                 if (qpn == 0)
2889                         goto skip_cqe;
2890                 qp = dev->qp_tbl[qpn];
2891                 BUG_ON(qp == NULL);
2892
2893                 if (is_cqe_for_sq(cqe)) {
2894                         expand = ocrdma_poll_scqe(qp, cqe, ibwc, &polled,
2895                                                   &stop);
2896                 } else {
2897                         expand = ocrdma_poll_rcqe(qp, cqe, ibwc, &polled,
2898                                                   &stop);
2899                 }
2900                 if (expand)
2901                         goto expand_cqe;
2902                 if (stop)
2903                         goto stop_cqe;
2904                 /* clear qpn to avoid duplicate processing by discard_cqe() */
2905                 cqe->cmn.qpn = 0;
2906 skip_cqe:
2907                 polled_hw_cqes += 1;
2908                 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
2909                 ocrdma_change_cq_phase(cq, cqe, cur_getp);
2910 expand_cqe:
2911                 if (polled) {
2912                         num_entries -= 1;
2913                         i += 1;
2914                         ibwc = ibwc + 1;
2915                         polled = false;
2916                 }
2917         }
2918 stop_cqe:
2919         cq->getp = cur_getp;
2920
2921         if (polled_hw_cqes)
2922                 ocrdma_ring_cq_db(dev, cq->id, false, false, polled_hw_cqes);
2923
2924         return i;
2925 }
2926
2927 /* insert error cqe if the QP's SQ or RQ's CQ matches the CQ under poll. */
2928 static int ocrdma_add_err_cqe(struct ocrdma_cq *cq, int num_entries,
2929                               struct ocrdma_qp *qp, struct ib_wc *ibwc)
2930 {
2931         int err_cqes = 0;
2932
2933         while (num_entries) {
2934                 if (is_hw_sq_empty(qp) && is_hw_rq_empty(qp))
2935                         break;
2936                 if (!is_hw_sq_empty(qp) && qp->sq_cq == cq) {
2937                         ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2938                         ocrdma_hwq_inc_tail(&qp->sq);
2939                 } else if (!is_hw_rq_empty(qp) && qp->rq_cq == cq) {
2940                         ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2941                         ocrdma_hwq_inc_tail(&qp->rq);
2942                 } else {
2943                         return err_cqes;
2944                 }
2945                 ibwc->byte_len = 0;
2946                 ibwc->status = IB_WC_WR_FLUSH_ERR;
2947                 ibwc = ibwc + 1;
2948                 err_cqes += 1;
2949                 num_entries -= 1;
2950         }
2951         return err_cqes;
2952 }
2953
2954 int ocrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
2955 {
2956         int cqes_to_poll = num_entries;
2957         struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2958         struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2959         int num_os_cqe = 0, err_cqes = 0;
2960         struct ocrdma_qp *qp;
2961         unsigned long flags;
2962
2963         /* poll cqes from adapter CQ */
2964         spin_lock_irqsave(&cq->cq_lock, flags);
2965         num_os_cqe = ocrdma_poll_hwcq(cq, cqes_to_poll, wc);
2966         spin_unlock_irqrestore(&cq->cq_lock, flags);
2967         cqes_to_poll -= num_os_cqe;
2968
2969         if (cqes_to_poll) {
2970                 wc = wc + num_os_cqe;
2971                 /* adapter returns single error cqe when qp moves to
2972                  * error state. So insert error cqes with wc_status as
2973                  * FLUSHED for pending WQEs and RQEs of QP's SQ and RQ
2974                  * respectively which uses this CQ.
2975                  */
2976                 spin_lock_irqsave(&dev->flush_q_lock, flags);
2977                 list_for_each_entry(qp, &cq->sq_head, sq_entry) {
2978                         if (cqes_to_poll == 0)
2979                                 break;
2980                         err_cqes = ocrdma_add_err_cqe(cq, cqes_to_poll, qp, wc);
2981                         cqes_to_poll -= err_cqes;
2982                         num_os_cqe += err_cqes;
2983                         wc = wc + err_cqes;
2984                 }
2985                 spin_unlock_irqrestore(&dev->flush_q_lock, flags);
2986         }
2987         return num_os_cqe;
2988 }
2989
2990 int ocrdma_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags cq_flags)
2991 {
2992         struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2993         struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2994         u16 cq_id;
2995         unsigned long flags;
2996         bool arm_needed = false, sol_needed = false;
2997
2998         cq_id = cq->id;
2999
3000         spin_lock_irqsave(&cq->cq_lock, flags);
3001         if (cq_flags & IB_CQ_NEXT_COMP || cq_flags & IB_CQ_SOLICITED)
3002                 arm_needed = true;
3003         if (cq_flags & IB_CQ_SOLICITED)
3004                 sol_needed = true;
3005
3006         ocrdma_ring_cq_db(dev, cq_id, arm_needed, sol_needed, 0);
3007         spin_unlock_irqrestore(&cq->cq_lock, flags);
3008
3009         return 0;
3010 }
3011
3012 struct ib_mr *ocrdma_alloc_mr(struct ib_pd *ibpd,
3013                               enum ib_mr_type mr_type,
3014                               u32 max_num_sg)
3015 {
3016         int status;
3017         struct ocrdma_mr *mr;
3018         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
3019         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
3020
3021         if (mr_type != IB_MR_TYPE_MEM_REG)
3022                 return ERR_PTR(-EINVAL);
3023
3024         if (max_num_sg > dev->attr.max_pages_per_frmr)
3025                 return ERR_PTR(-EINVAL);
3026
3027         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3028         if (!mr)
3029                 return ERR_PTR(-ENOMEM);
3030
3031         mr->pages = kcalloc(max_num_sg, sizeof(u64), GFP_KERNEL);
3032         if (!mr->pages) {
3033                 status = -ENOMEM;
3034                 goto pl_err;
3035         }
3036
3037         status = ocrdma_get_pbl_info(dev, mr, max_num_sg);
3038         if (status)
3039                 goto pbl_err;
3040         mr->hwmr.fr_mr = 1;
3041         mr->hwmr.remote_rd = 0;
3042         mr->hwmr.remote_wr = 0;
3043         mr->hwmr.local_rd = 0;
3044         mr->hwmr.local_wr = 0;
3045         mr->hwmr.mw_bind = 0;
3046         status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
3047         if (status)
3048                 goto pbl_err;
3049         status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, 0);
3050         if (status)
3051                 goto mbx_err;
3052         mr->ibmr.rkey = mr->hwmr.lkey;
3053         mr->ibmr.lkey = mr->hwmr.lkey;
3054         dev->stag_arr[(mr->hwmr.lkey >> 8) & (OCRDMA_MAX_STAG - 1)] =
3055                 (unsigned long) mr;
3056         return &mr->ibmr;
3057 mbx_err:
3058         ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
3059 pbl_err:
3060         kfree(mr->pages);
3061 pl_err:
3062         kfree(mr);
3063         return ERR_PTR(-ENOMEM);
3064 }
3065
3066 static int ocrdma_set_page(struct ib_mr *ibmr, u64 addr)
3067 {
3068         struct ocrdma_mr *mr = get_ocrdma_mr(ibmr);
3069
3070         if (unlikely(mr->npages == mr->hwmr.num_pbes))
3071                 return -ENOMEM;
3072
3073         mr->pages[mr->npages++] = addr;
3074
3075         return 0;
3076 }
3077
3078 int ocrdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
3079                      unsigned int *sg_offset)
3080 {
3081         struct ocrdma_mr *mr = get_ocrdma_mr(ibmr);
3082
3083         mr->npages = 0;
3084
3085         return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, ocrdma_set_page);
3086 }