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                 if (_ocrdma_dealloc_pd(dev, pd))
 748                         pr_err("%s: _ocrdma_dealloc_pd() failed\n", __func__);
 749         }
 750 exit:
 751         return ERR_PTR(status);
 752 }
 753
 754 int ocrdma_dealloc_pd(struct ib_pd *ibpd)
 755 {
 756         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
 757         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
 758         struct ocrdma_ucontext *uctx = NULL;
 759         int status = 0;
 760         u64 usr_db;
 761
 762         uctx = pd->uctx;
 763         if (uctx) {
 764                 u64 dpp_db = dev->nic_info.dpp_unmapped_addr +
 765                         (pd->id * PAGE_SIZE);
 766                 if (pd->dpp_enabled)
 767                         ocrdma_del_mmap(pd->uctx, dpp_db, PAGE_SIZE);
 768                 usr_db = ocrdma_get_db_addr(dev, pd->id);
 769                 ocrdma_del_mmap(pd->uctx, usr_db, dev->nic_info.db_page_size);
 770
 771                 if (is_ucontext_pd(uctx, pd)) {
 772                         ocrdma_release_ucontext_pd(uctx);
 773                         return status;
 774                 }
 775         }
 776         status = _ocrdma_dealloc_pd(dev, pd);
 777         return status;
 778 }
 779
 780 static int ocrdma_alloc_lkey(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
 781                             u32 pdid, int acc, u32 num_pbls, u32 addr_check)
 782 {
 783         int status;
 784
 785         mr->hwmr.fr_mr = 0;
 786         mr->hwmr.local_rd = 1;
 787         mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
 788         mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
 789         mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
 790         mr->hwmr.mw_bind = (acc & IB_ACCESS_MW_BIND) ? 1 : 0;
 791         mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
 792         mr->hwmr.num_pbls = num_pbls;
 793
 794         status = ocrdma_mbx_alloc_lkey(dev, &mr->hwmr, pdid, addr_check);
 795         if (status)
 796                 return status;
 797
 798         mr->ibmr.lkey = mr->hwmr.lkey;
 799         if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
 800                 mr->ibmr.rkey = mr->hwmr.lkey;
 801         return 0;
 802 }
 803
 804 struct ib_mr *ocrdma_get_dma_mr(struct ib_pd *ibpd, int acc)
 805 {
 806         int status;
 807         struct ocrdma_mr *mr;
 808         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
 809         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
 810
 811         if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE)) {
 812                 pr_err("%s err, invalid access rights\n", __func__);
 813                 return ERR_PTR(-EINVAL);
 814         }
 815
 816         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
 817         if (!mr)
 818                 return ERR_PTR(-ENOMEM);
 819
 820         status = ocrdma_alloc_lkey(dev, mr, pd->id, acc, 0,
 821                                    OCRDMA_ADDR_CHECK_DISABLE);
 822         if (status) {
 823                 kfree(mr);
 824                 return ERR_PTR(status);
 825         }
 826
 827         return &mr->ibmr;
 828 }
 829
 830 static void ocrdma_free_mr_pbl_tbl(struct ocrdma_dev *dev,
 831                                    struct ocrdma_hw_mr *mr)
 832 {
 833         struct pci_dev *pdev = dev->nic_info.pdev;
 834         int i = 0;
 835
 836         if (mr->pbl_table) {
 837                 for (i = 0; i < mr->num_pbls; i++) {
 838                         if (!mr->pbl_table[i].va)
 839                                 continue;
 840                         dma_free_coherent(&pdev->dev, mr->pbl_size,
 841                                           mr->pbl_table[i].va,
 842                                           mr->pbl_table[i].pa);
 843                 }
 844                 kfree(mr->pbl_table);
 845                 mr->pbl_table = NULL;
 846         }
 847 }
 848
 849 static int ocrdma_get_pbl_info(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
 850                               u32 num_pbes)
 851 {
 852         u32 num_pbls = 0;
 853         u32 idx = 0;
 854         int status = 0;
 855         u32 pbl_size;
 856
 857         do {
 858                 pbl_size = OCRDMA_MIN_HPAGE_SIZE * (1 << idx);
 859                 if (pbl_size > MAX_OCRDMA_PBL_SIZE) {
 860                         status = -EFAULT;
 861                         break;
 862                 }
 863                 num_pbls = roundup(num_pbes, (pbl_size / sizeof(u64)));
 864                 num_pbls = num_pbls / (pbl_size / sizeof(u64));
 865                 idx++;
 866         } while (num_pbls >= dev->attr.max_num_mr_pbl);
 867
 868         mr->hwmr.num_pbes = num_pbes;
 869         mr->hwmr.num_pbls = num_pbls;
 870         mr->hwmr.pbl_size = pbl_size;
 871         return status;
 872 }
 873
 874 static int ocrdma_build_pbl_tbl(struct ocrdma_dev *dev, struct ocrdma_hw_mr *mr)
 875 {
 876         int status = 0;
 877         int i;
 878         u32 dma_len = mr->pbl_size;
 879         struct pci_dev *pdev = dev->nic_info.pdev;
 880         void *va;
 881         dma_addr_t pa;
 882
 883         mr->pbl_table = kzalloc(sizeof(struct ocrdma_pbl) *
 884                                 mr->num_pbls, GFP_KERNEL);
 885
 886         if (!mr->pbl_table)
 887                 return -ENOMEM;
 888
 889         for (i = 0; i < mr->num_pbls; i++) {
 890                 va = dma_alloc_coherent(&pdev->dev, dma_len, &pa, GFP_KERNEL);
 891                 if (!va) {
 892                         ocrdma_free_mr_pbl_tbl(dev, mr);
 893                         status = -ENOMEM;
 894                         break;
 895                 }
 896                 memset(va, 0, dma_len);
 897                 mr->pbl_table[i].va = va;
 898                 mr->pbl_table[i].pa = pa;
 899         }
 900         return status;
 901 }
 902
 903 static void build_user_pbes(struct ocrdma_dev *dev, struct ocrdma_mr *mr,
 904                             u32 num_pbes)
 905 {
 906         struct ocrdma_pbe *pbe;
 907         struct scatterlist *sg;
 908         struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
 909         struct ib_umem *umem = mr->umem;
 910         int shift, pg_cnt, pages, pbe_cnt, entry, total_num_pbes = 0;
 911
 912         if (!mr->hwmr.num_pbes)
 913                 return;
 914
 915         pbe = (struct ocrdma_pbe *)pbl_tbl->va;
 916         pbe_cnt = 0;
 917
 918         shift = umem->page_shift;
 919
 920         for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
 921                 pages = sg_dma_len(sg) >> shift;
 922                 for (pg_cnt = 0; pg_cnt < pages; pg_cnt++) {
 923                         /* store the page address in pbe */
 924                         pbe->pa_lo =
 925                             cpu_to_le32(sg_dma_address(sg) +
 926                                         (pg_cnt << shift));
 927                         pbe->pa_hi =
 928                             cpu_to_le32(upper_32_bits(sg_dma_address(sg) +
 929                                          (pg_cnt << shift)));
 930                         pbe_cnt += 1;
 931                         total_num_pbes += 1;
 932                         pbe++;
 933
 934                         /* if done building pbes, issue the mbx cmd. */
 935                         if (total_num_pbes == num_pbes)
 936                                 return;
 937
 938                         /* if the given pbl is full storing the pbes,
 939                          * move to next pbl.
 940                          */
 941                         if (pbe_cnt ==
 942                                 (mr->hwmr.pbl_size / sizeof(u64))) {
 943                                 pbl_tbl++;
 944                                 pbe = (struct ocrdma_pbe *)pbl_tbl->va;
 945                                 pbe_cnt = 0;
 946                         }
 947
 948                 }
 949         }
 950 }
 951
 952 struct ib_mr *ocrdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
 953                                  u64 usr_addr, int acc, struct ib_udata *udata)
 954 {
 955         int status = -ENOMEM;
 956         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
 957         struct ocrdma_mr *mr;
 958         struct ocrdma_pd *pd;
 959         u32 num_pbes;
 960
 961         pd = get_ocrdma_pd(ibpd);
 962
 963         if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
 964                 return ERR_PTR(-EINVAL);
 965
 966         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
 967         if (!mr)
 968                 return ERR_PTR(status);
 969         mr->umem = ib_umem_get(ibpd->uobject->context, start, len, acc, 0);
 970         if (IS_ERR(mr->umem)) {
 971                 status = -EFAULT;
 972                 goto umem_err;
 973         }
 974         num_pbes = ib_umem_page_count(mr->umem);
 975         status = ocrdma_get_pbl_info(dev, mr, num_pbes);
 976         if (status)
 977                 goto umem_err;
 978
 979         mr->hwmr.pbe_size = BIT(mr->umem->page_shift);
 980         mr->hwmr.fbo = ib_umem_offset(mr->umem);
 981         mr->hwmr.va = usr_addr;
 982         mr->hwmr.len = len;
 983         mr->hwmr.remote_wr = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
 984         mr->hwmr.remote_rd = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
 985         mr->hwmr.local_wr = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
 986         mr->hwmr.local_rd = 1;
 987         mr->hwmr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
 988         status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
 989         if (status)
 990                 goto umem_err;
 991         build_user_pbes(dev, mr, num_pbes);
 992         status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, acc);
 993         if (status)
 994                 goto mbx_err;
 995         mr->ibmr.lkey = mr->hwmr.lkey;
 996         if (mr->hwmr.remote_wr || mr->hwmr.remote_rd)
 997                 mr->ibmr.rkey = mr->hwmr.lkey;
 998
 999         return &mr->ibmr;
1000
1001 mbx_err:
1002         ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
1003 umem_err:
1004         kfree(mr);
1005         return ERR_PTR(status);
1006 }
1007
1008 int ocrdma_dereg_mr(struct ib_mr *ib_mr)
1009 {
1010         struct ocrdma_mr *mr = get_ocrdma_mr(ib_mr);
1011         struct ocrdma_dev *dev = get_ocrdma_dev(ib_mr->device);
1012
1013         (void) ocrdma_mbx_dealloc_lkey(dev, mr->hwmr.fr_mr, mr->hwmr.lkey);
1014
1015         kfree(mr->pages);
1016         ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
1017
1018         /* it could be user registered memory. */
1019         if (mr->umem)
1020                 ib_umem_release(mr->umem);
1021         kfree(mr);
1022
1023         /* Don't stop cleanup, in case FW is unresponsive */
1024         if (dev->mqe_ctx.fw_error_state) {
1025                 pr_err("%s(%d) fw not responding.\n",
1026                        __func__, dev->id);
1027         }
1028         return 0;
1029 }
1030
1031 static int ocrdma_copy_cq_uresp(struct ocrdma_dev *dev, struct ocrdma_cq *cq,
1032                                 struct ib_udata *udata,
1033                                 struct ib_ucontext *ib_ctx)
1034 {
1035         int status;
1036         struct ocrdma_ucontext *uctx = get_ocrdma_ucontext(ib_ctx);
1037         struct ocrdma_create_cq_uresp uresp;
1038
1039         memset(&uresp, 0, sizeof(uresp));
1040         uresp.cq_id = cq->id;
1041         uresp.page_size = PAGE_ALIGN(cq->len);
1042         uresp.num_pages = 1;
1043         uresp.max_hw_cqe = cq->max_hw_cqe;
1044         uresp.page_addr[0] = virt_to_phys(cq->va);
1045         uresp.db_page_addr =  ocrdma_get_db_addr(dev, uctx->cntxt_pd->id);
1046         uresp.db_page_size = dev->nic_info.db_page_size;
1047         uresp.phase_change = cq->phase_change ? 1 : 0;
1048         status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1049         if (status) {
1050                 pr_err("%s(%d) copy error cqid=0x%x.\n",
1051                        __func__, dev->id, cq->id);
1052                 goto err;
1053         }
1054         status = ocrdma_add_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
1055         if (status)
1056                 goto err;
1057         status = ocrdma_add_mmap(uctx, uresp.page_addr[0], uresp.page_size);
1058         if (status) {
1059                 ocrdma_del_mmap(uctx, uresp.db_page_addr, uresp.db_page_size);
1060                 goto err;
1061         }
1062         cq->ucontext = uctx;
1063 err:
1064         return status;
1065 }
1066
1067 struct ib_cq *ocrdma_create_cq(struct ib_device *ibdev,
1068                                const struct ib_cq_init_attr *attr,
1069                                struct ib_ucontext *ib_ctx,
1070                                struct ib_udata *udata)
1071 {
1072         int entries = attr->cqe;
1073         struct ocrdma_cq *cq;
1074         struct ocrdma_dev *dev = get_ocrdma_dev(ibdev);
1075         struct ocrdma_ucontext *uctx = NULL;
1076         u16 pd_id = 0;
1077         int status;
1078         struct ocrdma_create_cq_ureq ureq;
1079
1080         if (attr->flags)
1081                 return ERR_PTR(-EINVAL);
1082
1083         if (udata) {
1084                 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1085                         return ERR_PTR(-EFAULT);
1086         } else
1087                 ureq.dpp_cq = 0;
1088         cq = kzalloc(sizeof(*cq), GFP_KERNEL);
1089         if (!cq)
1090                 return ERR_PTR(-ENOMEM);
1091
1092         spin_lock_init(&cq->cq_lock);
1093         spin_lock_init(&cq->comp_handler_lock);
1094         INIT_LIST_HEAD(&cq->sq_head);
1095         INIT_LIST_HEAD(&cq->rq_head);
1096
1097         if (ib_ctx) {
1098                 uctx = get_ocrdma_ucontext(ib_ctx);
1099                 pd_id = uctx->cntxt_pd->id;
1100         }
1101
1102         status = ocrdma_mbx_create_cq(dev, cq, entries, ureq.dpp_cq, pd_id);
1103         if (status) {
1104                 kfree(cq);
1105                 return ERR_PTR(status);
1106         }
1107         if (ib_ctx) {
1108                 status = ocrdma_copy_cq_uresp(dev, cq, udata, ib_ctx);
1109                 if (status)
1110                         goto ctx_err;
1111         }
1112         cq->phase = OCRDMA_CQE_VALID;
1113         dev->cq_tbl[cq->id] = cq;
1114         return &cq->ibcq;
1115
1116 ctx_err:
1117         ocrdma_mbx_destroy_cq(dev, cq);
1118         kfree(cq);
1119         return ERR_PTR(status);
1120 }
1121
1122 int ocrdma_resize_cq(struct ib_cq *ibcq, int new_cnt,
1123                      struct ib_udata *udata)
1124 {
1125         int status = 0;
1126         struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1127
1128         if (new_cnt < 1 || new_cnt > cq->max_hw_cqe) {
1129                 status = -EINVAL;
1130                 return status;
1131         }
1132         ibcq->cqe = new_cnt;
1133         return status;
1134 }
1135
1136 static void ocrdma_flush_cq(struct ocrdma_cq *cq)
1137 {
1138         int cqe_cnt;
1139         int valid_count = 0;
1140         unsigned long flags;
1141
1142         struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
1143         struct ocrdma_cqe *cqe = NULL;
1144
1145         cqe = cq->va;
1146         cqe_cnt = cq->cqe_cnt;
1147
1148         /* Last irq might have scheduled a polling thread
1149          * sync-up with it before hard flushing.
1150          */
1151         spin_lock_irqsave(&cq->cq_lock, flags);
1152         while (cqe_cnt) {
1153                 if (is_cqe_valid(cq, cqe))
1154                         valid_count++;
1155                 cqe++;
1156                 cqe_cnt--;
1157         }
1158         ocrdma_ring_cq_db(dev, cq->id, false, false, valid_count);
1159         spin_unlock_irqrestore(&cq->cq_lock, flags);
1160 }
1161
1162 int ocrdma_destroy_cq(struct ib_cq *ibcq)
1163 {
1164         struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
1165         struct ocrdma_eq *eq = NULL;
1166         struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
1167         int pdid = 0;
1168         u32 irq, indx;
1169
1170         dev->cq_tbl[cq->id] = NULL;
1171         indx = ocrdma_get_eq_table_index(dev, cq->eqn);
1172         BUG_ON(indx == -EINVAL);
1173
1174         eq = &dev->eq_tbl[indx];
1175         irq = ocrdma_get_irq(dev, eq);
1176         synchronize_irq(irq);
1177         ocrdma_flush_cq(cq);
1178
1179         (void)ocrdma_mbx_destroy_cq(dev, cq);
1180         if (cq->ucontext) {
1181                 pdid = cq->ucontext->cntxt_pd->id;
1182                 ocrdma_del_mmap(cq->ucontext, (u64) cq->pa,
1183                                 PAGE_ALIGN(cq->len));
1184                 ocrdma_del_mmap(cq->ucontext,
1185                                 ocrdma_get_db_addr(dev, pdid),
1186                                 dev->nic_info.db_page_size);
1187         }
1188
1189         kfree(cq);
1190         return 0;
1191 }
1192
1193 static int ocrdma_add_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1194 {
1195         int status = -EINVAL;
1196
1197         if (qp->id < OCRDMA_MAX_QP && dev->qp_tbl[qp->id] == NULL) {
1198                 dev->qp_tbl[qp->id] = qp;
1199                 status = 0;
1200         }
1201         return status;
1202 }
1203
1204 static void ocrdma_del_qpn_map(struct ocrdma_dev *dev, struct ocrdma_qp *qp)
1205 {
1206         dev->qp_tbl[qp->id] = NULL;
1207 }
1208
1209 static int ocrdma_check_qp_params(struct ib_pd *ibpd, struct ocrdma_dev *dev,
1210                                   struct ib_qp_init_attr *attrs)
1211 {
1212         if ((attrs->qp_type != IB_QPT_GSI) &&
1213             (attrs->qp_type != IB_QPT_RC) &&
1214             (attrs->qp_type != IB_QPT_UC) &&
1215             (attrs->qp_type != IB_QPT_UD)) {
1216                 pr_err("%s(%d) unsupported qp type=0x%x requested\n",
1217                        __func__, dev->id, attrs->qp_type);
1218                 return -EINVAL;
1219         }
1220         /* Skip the check for QP1 to support CM size of 128 */
1221         if ((attrs->qp_type != IB_QPT_GSI) &&
1222             (attrs->cap.max_send_wr > dev->attr.max_wqe)) {
1223                 pr_err("%s(%d) unsupported send_wr=0x%x requested\n",
1224                        __func__, dev->id, attrs->cap.max_send_wr);
1225                 pr_err("%s(%d) supported send_wr=0x%x\n",
1226                        __func__, dev->id, dev->attr.max_wqe);
1227                 return -EINVAL;
1228         }
1229         if (!attrs->srq && (attrs->cap.max_recv_wr > dev->attr.max_rqe)) {
1230                 pr_err("%s(%d) unsupported recv_wr=0x%x requested\n",
1231                        __func__, dev->id, attrs->cap.max_recv_wr);
1232                 pr_err("%s(%d) supported recv_wr=0x%x\n",
1233                        __func__, dev->id, dev->attr.max_rqe);
1234                 return -EINVAL;
1235         }
1236         if (attrs->cap.max_inline_data > dev->attr.max_inline_data) {
1237                 pr_err("%s(%d) unsupported inline data size=0x%x requested\n",
1238                        __func__, dev->id, attrs->cap.max_inline_data);
1239                 pr_err("%s(%d) supported inline data size=0x%x\n",
1240                        __func__, dev->id, dev->attr.max_inline_data);
1241                 return -EINVAL;
1242         }
1243         if (attrs->cap.max_send_sge > dev->attr.max_send_sge) {
1244                 pr_err("%s(%d) unsupported send_sge=0x%x requested\n",
1245                        __func__, dev->id, attrs->cap.max_send_sge);
1246                 pr_err("%s(%d) supported send_sge=0x%x\n",
1247                        __func__, dev->id, dev->attr.max_send_sge);
1248                 return -EINVAL;
1249         }
1250         if (attrs->cap.max_recv_sge > dev->attr.max_recv_sge) {
1251                 pr_err("%s(%d) unsupported recv_sge=0x%x requested\n",
1252                        __func__, dev->id, attrs->cap.max_recv_sge);
1253                 pr_err("%s(%d) supported recv_sge=0x%x\n",
1254                        __func__, dev->id, dev->attr.max_recv_sge);
1255                 return -EINVAL;
1256         }
1257         /* unprivileged user space cannot create special QP */
1258         if (ibpd->uobject && attrs->qp_type == IB_QPT_GSI) {
1259                 pr_err
1260                     ("%s(%d) Userspace can't create special QPs of type=0x%x\n",
1261                      __func__, dev->id, attrs->qp_type);
1262                 return -EINVAL;
1263         }
1264         /* allow creating only one GSI type of QP */
1265         if (attrs->qp_type == IB_QPT_GSI && dev->gsi_qp_created) {
1266                 pr_err("%s(%d) GSI special QPs already created.\n",
1267                        __func__, dev->id);
1268                 return -EINVAL;
1269         }
1270         /* verify consumer QPs are not trying to use GSI QP's CQ */
1271         if ((attrs->qp_type != IB_QPT_GSI) && (dev->gsi_qp_created)) {
1272                 if ((dev->gsi_sqcq == get_ocrdma_cq(attrs->send_cq)) ||
1273                         (dev->gsi_rqcq == get_ocrdma_cq(attrs->recv_cq))) {
1274                         pr_err("%s(%d) Consumer QP cannot use GSI CQs.\n",
1275                                 __func__, dev->id);
1276                         return -EINVAL;
1277                 }
1278         }
1279         return 0;
1280 }
1281
1282 static int ocrdma_copy_qp_uresp(struct ocrdma_qp *qp,
1283                                 struct ib_udata *udata, int dpp_offset,
1284                                 int dpp_credit_lmt, int srq)
1285 {
1286         int status;
1287         u64 usr_db;
1288         struct ocrdma_create_qp_uresp uresp;
1289         struct ocrdma_pd *pd = qp->pd;
1290         struct ocrdma_dev *dev = get_ocrdma_dev(pd->ibpd.device);
1291
1292         memset(&uresp, 0, sizeof(uresp));
1293         usr_db = dev->nic_info.unmapped_db +
1294                         (pd->id * dev->nic_info.db_page_size);
1295         uresp.qp_id = qp->id;
1296         uresp.sq_dbid = qp->sq.dbid;
1297         uresp.num_sq_pages = 1;
1298         uresp.sq_page_size = PAGE_ALIGN(qp->sq.len);
1299         uresp.sq_page_addr[0] = virt_to_phys(qp->sq.va);
1300         uresp.num_wqe_allocated = qp->sq.max_cnt;
1301         if (!srq) {
1302                 uresp.rq_dbid = qp->rq.dbid;
1303                 uresp.num_rq_pages = 1;
1304                 uresp.rq_page_size = PAGE_ALIGN(qp->rq.len);
1305                 uresp.rq_page_addr[0] = virt_to_phys(qp->rq.va);
1306                 uresp.num_rqe_allocated = qp->rq.max_cnt;
1307         }
1308         uresp.db_page_addr = usr_db;
1309         uresp.db_page_size = dev->nic_info.db_page_size;
1310         uresp.db_sq_offset = OCRDMA_DB_GEN2_SQ_OFFSET;
1311         uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1312         uresp.db_shift = OCRDMA_DB_RQ_SHIFT;
1313
1314         if (qp->dpp_enabled) {
1315                 uresp.dpp_credit = dpp_credit_lmt;
1316                 uresp.dpp_offset = dpp_offset;
1317         }
1318         status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1319         if (status) {
1320                 pr_err("%s(%d) user copy error.\n", __func__, dev->id);
1321                 goto err;
1322         }
1323         status = ocrdma_add_mmap(pd->uctx, uresp.sq_page_addr[0],
1324                                  uresp.sq_page_size);
1325         if (status)
1326                 goto err;
1327
1328         if (!srq) {
1329                 status = ocrdma_add_mmap(pd->uctx, uresp.rq_page_addr[0],
1330                                          uresp.rq_page_size);
1331                 if (status)
1332                         goto rq_map_err;
1333         }
1334         return status;
1335 rq_map_err:
1336         ocrdma_del_mmap(pd->uctx, uresp.sq_page_addr[0], uresp.sq_page_size);
1337 err:
1338         return status;
1339 }
1340
1341 static void ocrdma_set_qp_db(struct ocrdma_dev *dev, struct ocrdma_qp *qp,
1342                              struct ocrdma_pd *pd)
1343 {
1344         if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1345                 qp->sq_db = dev->nic_info.db +
1346                         (pd->id * dev->nic_info.db_page_size) +
1347                         OCRDMA_DB_GEN2_SQ_OFFSET;
1348                 qp->rq_db = dev->nic_info.db +
1349                         (pd->id * dev->nic_info.db_page_size) +
1350                         OCRDMA_DB_GEN2_RQ_OFFSET;
1351         } else {
1352                 qp->sq_db = dev->nic_info.db +
1353                         (pd->id * dev->nic_info.db_page_size) +
1354                         OCRDMA_DB_SQ_OFFSET;
1355                 qp->rq_db = dev->nic_info.db +
1356                         (pd->id * dev->nic_info.db_page_size) +
1357                         OCRDMA_DB_RQ_OFFSET;
1358         }
1359 }
1360
1361 static int ocrdma_alloc_wr_id_tbl(struct ocrdma_qp *qp)
1362 {
1363         qp->wqe_wr_id_tbl =
1364             kzalloc(sizeof(*(qp->wqe_wr_id_tbl)) * qp->sq.max_cnt,
1365                     GFP_KERNEL);
1366         if (qp->wqe_wr_id_tbl == NULL)
1367                 return -ENOMEM;
1368         qp->rqe_wr_id_tbl =
1369             kzalloc(sizeof(u64) * qp->rq.max_cnt, GFP_KERNEL);
1370         if (qp->rqe_wr_id_tbl == NULL)
1371                 return -ENOMEM;
1372
1373         return 0;
1374 }
1375
1376 static void ocrdma_set_qp_init_params(struct ocrdma_qp *qp,
1377                                       struct ocrdma_pd *pd,
1378                                       struct ib_qp_init_attr *attrs)
1379 {
1380         qp->pd = pd;
1381         spin_lock_init(&qp->q_lock);
1382         INIT_LIST_HEAD(&qp->sq_entry);
1383         INIT_LIST_HEAD(&qp->rq_entry);
1384
1385         qp->qp_type = attrs->qp_type;
1386         qp->cap_flags = OCRDMA_QP_INB_RD | OCRDMA_QP_INB_WR;
1387         qp->max_inline_data = attrs->cap.max_inline_data;
1388         qp->sq.max_sges = attrs->cap.max_send_sge;
1389         qp->rq.max_sges = attrs->cap.max_recv_sge;
1390         qp->state = OCRDMA_QPS_RST;
1391         qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
1392 }
1393
1394 static void ocrdma_store_gsi_qp_cq(struct ocrdma_dev *dev,
1395                                    struct ib_qp_init_attr *attrs)
1396 {
1397         if (attrs->qp_type == IB_QPT_GSI) {
1398                 dev->gsi_qp_created = 1;
1399                 dev->gsi_sqcq = get_ocrdma_cq(attrs->send_cq);
1400                 dev->gsi_rqcq = get_ocrdma_cq(attrs->recv_cq);
1401         }
1402 }
1403
1404 struct ib_qp *ocrdma_create_qp(struct ib_pd *ibpd,
1405                                struct ib_qp_init_attr *attrs,
1406                                struct ib_udata *udata)
1407 {
1408         int status;
1409         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1410         struct ocrdma_qp *qp;
1411         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1412         struct ocrdma_create_qp_ureq ureq;
1413         u16 dpp_credit_lmt, dpp_offset;
1414
1415         status = ocrdma_check_qp_params(ibpd, dev, attrs);
1416         if (status)
1417                 goto gen_err;
1418
1419         memset(&ureq, 0, sizeof(ureq));
1420         if (udata) {
1421                 if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1422                         return ERR_PTR(-EFAULT);
1423         }
1424         qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1425         if (!qp) {
1426                 status = -ENOMEM;
1427                 goto gen_err;
1428         }
1429         ocrdma_set_qp_init_params(qp, pd, attrs);
1430         if (udata == NULL)
1431                 qp->cap_flags |= (OCRDMA_QP_MW_BIND | OCRDMA_QP_LKEY0 |
1432                                         OCRDMA_QP_FAST_REG);
1433
1434         mutex_lock(&dev->dev_lock);
1435         status = ocrdma_mbx_create_qp(qp, attrs, ureq.enable_dpp_cq,
1436                                         ureq.dpp_cq_id,
1437                                         &dpp_offset, &dpp_credit_lmt);
1438         if (status)
1439                 goto mbx_err;
1440
1441         /* user space QP's wr_id table are managed in library */
1442         if (udata == NULL) {
1443                 status = ocrdma_alloc_wr_id_tbl(qp);
1444                 if (status)
1445                         goto map_err;
1446         }
1447
1448         status = ocrdma_add_qpn_map(dev, qp);
1449         if (status)
1450                 goto map_err;
1451         ocrdma_set_qp_db(dev, qp, pd);
1452         if (udata) {
1453                 status = ocrdma_copy_qp_uresp(qp, udata, dpp_offset,
1454                                               dpp_credit_lmt,
1455                                               (attrs->srq != NULL));
1456                 if (status)
1457                         goto cpy_err;
1458         }
1459         ocrdma_store_gsi_qp_cq(dev, attrs);
1460         qp->ibqp.qp_num = qp->id;
1461         mutex_unlock(&dev->dev_lock);
1462         return &qp->ibqp;
1463
1464 cpy_err:
1465         ocrdma_del_qpn_map(dev, qp);
1466 map_err:
1467         ocrdma_mbx_destroy_qp(dev, qp);
1468 mbx_err:
1469         mutex_unlock(&dev->dev_lock);
1470         kfree(qp->wqe_wr_id_tbl);
1471         kfree(qp->rqe_wr_id_tbl);
1472         kfree(qp);
1473         pr_err("%s(%d) error=%d\n", __func__, dev->id, status);
1474 gen_err:
1475         return ERR_PTR(status);
1476 }
1477
1478 int _ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1479                       int attr_mask)
1480 {
1481         int status = 0;
1482         struct ocrdma_qp *qp;
1483         struct ocrdma_dev *dev;
1484         enum ib_qp_state old_qps;
1485
1486         qp = get_ocrdma_qp(ibqp);
1487         dev = get_ocrdma_dev(ibqp->device);
1488         if (attr_mask & IB_QP_STATE)
1489                 status = ocrdma_qp_state_change(qp, attr->qp_state, &old_qps);
1490         /* if new and previous states are same hw doesn't need to
1491          * know about it.
1492          */
1493         if (status < 0)
1494                 return status;
1495         return ocrdma_mbx_modify_qp(dev, qp, attr, attr_mask);
1496 }
1497
1498 int ocrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1499                      int attr_mask, struct ib_udata *udata)
1500 {
1501         unsigned long flags;
1502         int status = -EINVAL;
1503         struct ocrdma_qp *qp;
1504         struct ocrdma_dev *dev;
1505         enum ib_qp_state old_qps, new_qps;
1506
1507         qp = get_ocrdma_qp(ibqp);
1508         dev = get_ocrdma_dev(ibqp->device);
1509
1510         /* syncronize with multiple context trying to change, retrive qps */
1511         mutex_lock(&dev->dev_lock);
1512         /* syncronize with wqe, rqe posting and cqe processing contexts */
1513         spin_lock_irqsave(&qp->q_lock, flags);
1514         old_qps = get_ibqp_state(qp->state);
1515         if (attr_mask & IB_QP_STATE)
1516                 new_qps = attr->qp_state;
1517         else
1518                 new_qps = old_qps;
1519         spin_unlock_irqrestore(&qp->q_lock, flags);
1520
1521         if (!ib_modify_qp_is_ok(old_qps, new_qps, ibqp->qp_type, attr_mask,
1522                                 IB_LINK_LAYER_ETHERNET)) {
1523                 pr_err("%s(%d) invalid attribute mask=0x%x specified for\n"
1524                        "qpn=0x%x of type=0x%x old_qps=0x%x, new_qps=0x%x\n",
1525                        __func__, dev->id, attr_mask, qp->id, ibqp->qp_type,
1526                        old_qps, new_qps);
1527                 goto param_err;
1528         }
1529
1530         status = _ocrdma_modify_qp(ibqp, attr, attr_mask);
1531         if (status > 0)
1532                 status = 0;
1533 param_err:
1534         mutex_unlock(&dev->dev_lock);
1535         return status;
1536 }
1537
1538 static enum ib_mtu ocrdma_mtu_int_to_enum(u16 mtu)
1539 {
1540         switch (mtu) {
1541         case 256:
1542                 return IB_MTU_256;
1543         case 512:
1544                 return IB_MTU_512;
1545         case 1024:
1546                 return IB_MTU_1024;
1547         case 2048:
1548                 return IB_MTU_2048;
1549         case 4096:
1550                 return IB_MTU_4096;
1551         default:
1552                 return IB_MTU_1024;
1553         }
1554 }
1555
1556 static int ocrdma_to_ib_qp_acc_flags(int qp_cap_flags)
1557 {
1558         int ib_qp_acc_flags = 0;
1559
1560         if (qp_cap_flags & OCRDMA_QP_INB_WR)
1561                 ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
1562         if (qp_cap_flags & OCRDMA_QP_INB_RD)
1563                 ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
1564         return ib_qp_acc_flags;
1565 }
1566
1567 int ocrdma_query_qp(struct ib_qp *ibqp,
1568                     struct ib_qp_attr *qp_attr,
1569                     int attr_mask, struct ib_qp_init_attr *qp_init_attr)
1570 {
1571         int status;
1572         u32 qp_state;
1573         struct ocrdma_qp_params params;
1574         struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
1575         struct ocrdma_dev *dev = get_ocrdma_dev(ibqp->device);
1576
1577         memset(&params, 0, sizeof(params));
1578         mutex_lock(&dev->dev_lock);
1579         status = ocrdma_mbx_query_qp(dev, qp, &params);
1580         mutex_unlock(&dev->dev_lock);
1581         if (status)
1582                 goto mbx_err;
1583         if (qp->qp_type == IB_QPT_UD)
1584                 qp_attr->qkey = params.qkey;
1585         qp_attr->path_mtu =
1586                 ocrdma_mtu_int_to_enum(params.path_mtu_pkey_indx &
1587                                 OCRDMA_QP_PARAMS_PATH_MTU_MASK) >>
1588                                 OCRDMA_QP_PARAMS_PATH_MTU_SHIFT;
1589         qp_attr->path_mig_state = IB_MIG_MIGRATED;
1590         qp_attr->rq_psn = params.hop_lmt_rq_psn & OCRDMA_QP_PARAMS_RQ_PSN_MASK;
1591         qp_attr->sq_psn = params.tclass_sq_psn & OCRDMA_QP_PARAMS_SQ_PSN_MASK;
1592         qp_attr->dest_qp_num =
1593             params.ack_to_rnr_rtc_dest_qpn & OCRDMA_QP_PARAMS_DEST_QPN_MASK;
1594
1595         qp_attr->qp_access_flags = ocrdma_to_ib_qp_acc_flags(qp->cap_flags);
1596         qp_attr->cap.max_send_wr = qp->sq.max_cnt - 1;
1597         qp_attr->cap.max_recv_wr = qp->rq.max_cnt - 1;
1598         qp_attr->cap.max_send_sge = qp->sq.max_sges;
1599         qp_attr->cap.max_recv_sge = qp->rq.max_sges;
1600         qp_attr->cap.max_inline_data = qp->max_inline_data;
1601         qp_init_attr->cap = qp_attr->cap;
1602         qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
1603
1604         rdma_ah_set_grh(&qp_attr->ah_attr, NULL,
1605                         params.rnt_rc_sl_fl &
1606                           OCRDMA_QP_PARAMS_FLOW_LABEL_MASK,
1607                         qp->sgid_idx,
1608                         (params.hop_lmt_rq_psn &
1609                          OCRDMA_QP_PARAMS_HOP_LMT_MASK) >>
1610                          OCRDMA_QP_PARAMS_HOP_LMT_SHIFT,
1611                         (params.tclass_sq_psn &
1612                          OCRDMA_QP_PARAMS_TCLASS_MASK) >>
1613                          OCRDMA_QP_PARAMS_TCLASS_SHIFT);
1614         rdma_ah_set_dgid_raw(&qp_attr->ah_attr, &params.dgid[0]);
1615
1616         rdma_ah_set_port_num(&qp_attr->ah_attr, 1);
1617         rdma_ah_set_sl(&qp_attr->ah_attr, (params.rnt_rc_sl_fl &
1618                                            OCRDMA_QP_PARAMS_SL_MASK) >>
1619                                            OCRDMA_QP_PARAMS_SL_SHIFT);
1620         qp_attr->timeout = (params.ack_to_rnr_rtc_dest_qpn &
1621                             OCRDMA_QP_PARAMS_ACK_TIMEOUT_MASK) >>
1622                                 OCRDMA_QP_PARAMS_ACK_TIMEOUT_SHIFT;
1623         qp_attr->rnr_retry = (params.ack_to_rnr_rtc_dest_qpn &
1624                               OCRDMA_QP_PARAMS_RNR_RETRY_CNT_MASK) >>
1625                                 OCRDMA_QP_PARAMS_RNR_RETRY_CNT_SHIFT;
1626         qp_attr->retry_cnt =
1627             (params.rnt_rc_sl_fl & OCRDMA_QP_PARAMS_RETRY_CNT_MASK) >>
1628                 OCRDMA_QP_PARAMS_RETRY_CNT_SHIFT;
1629         qp_attr->min_rnr_timer = 0;
1630         qp_attr->pkey_index = 0;
1631         qp_attr->port_num = 1;
1632         rdma_ah_set_path_bits(&qp_attr->ah_attr, 0);
1633         rdma_ah_set_static_rate(&qp_attr->ah_attr, 0);
1634         qp_attr->alt_pkey_index = 0;
1635         qp_attr->alt_port_num = 0;
1636         qp_attr->alt_timeout = 0;
1637         memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
1638         qp_state = (params.max_sge_recv_flags & OCRDMA_QP_PARAMS_STATE_MASK) >>
1639                     OCRDMA_QP_PARAMS_STATE_SHIFT;
1640         qp_attr->qp_state = get_ibqp_state(qp_state);
1641         qp_attr->cur_qp_state = qp_attr->qp_state;
1642         qp_attr->sq_draining = (qp_state == OCRDMA_QPS_SQ_DRAINING) ? 1 : 0;
1643         qp_attr->max_dest_rd_atomic =
1644             params.max_ord_ird >> OCRDMA_QP_PARAMS_MAX_ORD_SHIFT;
1645         qp_attr->max_rd_atomic =
1646             params.max_ord_ird & OCRDMA_QP_PARAMS_MAX_IRD_MASK;
1647         qp_attr->en_sqd_async_notify = (params.max_sge_recv_flags &
1648                                 OCRDMA_QP_PARAMS_FLAGS_SQD_ASYNC) ? 1 : 0;
1649         /* Sync driver QP state with FW */
1650         ocrdma_qp_state_change(qp, qp_attr->qp_state, NULL);
1651 mbx_err:
1652         return status;
1653 }
1654
1655 static void ocrdma_srq_toggle_bit(struct ocrdma_srq *srq, unsigned int idx)
1656 {
1657         unsigned int i = idx / 32;
1658         u32 mask = (1U << (idx % 32));
1659
1660         srq->idx_bit_fields[i] ^= mask;
1661 }
1662
1663 static int ocrdma_hwq_free_cnt(struct ocrdma_qp_hwq_info *q)
1664 {
1665         return ((q->max_wqe_idx - q->head) + q->tail) % q->max_cnt;
1666 }
1667
1668 static int is_hw_sq_empty(struct ocrdma_qp *qp)
1669 {
1670         return (qp->sq.tail == qp->sq.head);
1671 }
1672
1673 static int is_hw_rq_empty(struct ocrdma_qp *qp)
1674 {
1675         return (qp->rq.tail == qp->rq.head);
1676 }
1677
1678 static void *ocrdma_hwq_head(struct ocrdma_qp_hwq_info *q)
1679 {
1680         return q->va + (q->head * q->entry_size);
1681 }
1682
1683 static void *ocrdma_hwq_head_from_idx(struct ocrdma_qp_hwq_info *q,
1684                                       u32 idx)
1685 {
1686         return q->va + (idx * q->entry_size);
1687 }
1688
1689 static void ocrdma_hwq_inc_head(struct ocrdma_qp_hwq_info *q)
1690 {
1691         q->head = (q->head + 1) & q->max_wqe_idx;
1692 }
1693
1694 static void ocrdma_hwq_inc_tail(struct ocrdma_qp_hwq_info *q)
1695 {
1696         q->tail = (q->tail + 1) & q->max_wqe_idx;
1697 }
1698
1699 /* discard the cqe for a given QP */
1700 static void ocrdma_discard_cqes(struct ocrdma_qp *qp, struct ocrdma_cq *cq)
1701 {
1702         unsigned long cq_flags;
1703         unsigned long flags;
1704         int discard_cnt = 0;
1705         u32 cur_getp, stop_getp;
1706         struct ocrdma_cqe *cqe;
1707         u32 qpn = 0, wqe_idx = 0;
1708
1709         spin_lock_irqsave(&cq->cq_lock, cq_flags);
1710
1711         /* traverse through the CQEs in the hw CQ,
1712          * find the matching CQE for a given qp,
1713          * mark the matching one discarded by clearing qpn.
1714          * ring the doorbell in the poll_cq() as
1715          * we don't complete out of order cqe.
1716          */
1717
1718         cur_getp = cq->getp;
1719         /* find upto when do we reap the cq. */
1720         stop_getp = cur_getp;
1721         do {
1722                 if (is_hw_sq_empty(qp) && (!qp->srq && is_hw_rq_empty(qp)))
1723                         break;
1724
1725                 cqe = cq->va + cur_getp;
1726                 /* if (a) done reaping whole hw cq, or
1727                  *    (b) qp_xq becomes empty.
1728                  * then exit
1729                  */
1730                 qpn = cqe->cmn.qpn & OCRDMA_CQE_QPN_MASK;
1731                 /* if previously discarded cqe found, skip that too. */
1732                 /* check for matching qp */
1733                 if (qpn == 0 || qpn != qp->id)
1734                         goto skip_cqe;
1735
1736                 if (is_cqe_for_sq(cqe)) {
1737                         ocrdma_hwq_inc_tail(&qp->sq);
1738                 } else {
1739                         if (qp->srq) {
1740                                 wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
1741                                         OCRDMA_CQE_BUFTAG_SHIFT) &
1742                                         qp->srq->rq.max_wqe_idx;
1743                                 BUG_ON(wqe_idx < 1);
1744                                 spin_lock_irqsave(&qp->srq->q_lock, flags);
1745                                 ocrdma_hwq_inc_tail(&qp->srq->rq);
1746                                 ocrdma_srq_toggle_bit(qp->srq, wqe_idx - 1);
1747                                 spin_unlock_irqrestore(&qp->srq->q_lock, flags);
1748
1749                         } else {
1750                                 ocrdma_hwq_inc_tail(&qp->rq);
1751                         }
1752                 }
1753                 /* mark cqe discarded so that it is not picked up later
1754                  * in the poll_cq().
1755                  */
1756                 discard_cnt += 1;
1757                 cqe->cmn.qpn = 0;
1758 skip_cqe:
1759                 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
1760         } while (cur_getp != stop_getp);
1761         spin_unlock_irqrestore(&cq->cq_lock, cq_flags);
1762 }
1763
1764 void ocrdma_del_flush_qp(struct ocrdma_qp *qp)
1765 {
1766         int found = false;
1767         unsigned long flags;
1768         struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
1769         /* sync with any active CQ poll */
1770
1771         spin_lock_irqsave(&dev->flush_q_lock, flags);
1772         found = ocrdma_is_qp_in_sq_flushlist(qp->sq_cq, qp);
1773         if (found)
1774                 list_del(&qp->sq_entry);
1775         if (!qp->srq) {
1776                 found = ocrdma_is_qp_in_rq_flushlist(qp->rq_cq, qp);
1777                 if (found)
1778                         list_del(&qp->rq_entry);
1779         }
1780         spin_unlock_irqrestore(&dev->flush_q_lock, flags);
1781 }
1782
1783 int ocrdma_destroy_qp(struct ib_qp *ibqp)
1784 {
1785         struct ocrdma_pd *pd;
1786         struct ocrdma_qp *qp;
1787         struct ocrdma_dev *dev;
1788         struct ib_qp_attr attrs;
1789         int attr_mask;
1790         unsigned long flags;
1791
1792         qp = get_ocrdma_qp(ibqp);
1793         dev = get_ocrdma_dev(ibqp->device);
1794
1795         pd = qp->pd;
1796
1797         /* change the QP state to ERROR */
1798         if (qp->state != OCRDMA_QPS_RST) {
1799                 attrs.qp_state = IB_QPS_ERR;
1800                 attr_mask = IB_QP_STATE;
1801                 _ocrdma_modify_qp(ibqp, &attrs, attr_mask);
1802         }
1803         /* ensure that CQEs for newly created QP (whose id may be same with
1804          * one which just getting destroyed are same), dont get
1805          * discarded until the old CQEs are discarded.
1806          */
1807         mutex_lock(&dev->dev_lock);
1808         (void) ocrdma_mbx_destroy_qp(dev, qp);
1809
1810         /*
1811          * acquire CQ lock while destroy is in progress, in order to
1812          * protect against proessing in-flight CQEs for this QP.
1813          */
1814         spin_lock_irqsave(&qp->sq_cq->cq_lock, flags);
1815         if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1816                 spin_lock(&qp->rq_cq->cq_lock);
1817
1818         ocrdma_del_qpn_map(dev, qp);
1819
1820         if (qp->rq_cq && (qp->rq_cq != qp->sq_cq))
1821                 spin_unlock(&qp->rq_cq->cq_lock);
1822         spin_unlock_irqrestore(&qp->sq_cq->cq_lock, flags);
1823
1824         if (!pd->uctx) {
1825                 ocrdma_discard_cqes(qp, qp->sq_cq);
1826                 ocrdma_discard_cqes(qp, qp->rq_cq);
1827         }
1828         mutex_unlock(&dev->dev_lock);
1829
1830         if (pd->uctx) {
1831                 ocrdma_del_mmap(pd->uctx, (u64) qp->sq.pa,
1832                                 PAGE_ALIGN(qp->sq.len));
1833                 if (!qp->srq)
1834                         ocrdma_del_mmap(pd->uctx, (u64) qp->rq.pa,
1835                                         PAGE_ALIGN(qp->rq.len));
1836         }
1837
1838         ocrdma_del_flush_qp(qp);
1839
1840         kfree(qp->wqe_wr_id_tbl);
1841         kfree(qp->rqe_wr_id_tbl);
1842         kfree(qp);
1843         return 0;
1844 }
1845
1846 static int ocrdma_copy_srq_uresp(struct ocrdma_dev *dev, struct ocrdma_srq *srq,
1847                                 struct ib_udata *udata)
1848 {
1849         int status;
1850         struct ocrdma_create_srq_uresp uresp;
1851
1852         memset(&uresp, 0, sizeof(uresp));
1853         uresp.rq_dbid = srq->rq.dbid;
1854         uresp.num_rq_pages = 1;
1855         uresp.rq_page_addr[0] = virt_to_phys(srq->rq.va);
1856         uresp.rq_page_size = srq->rq.len;
1857         uresp.db_page_addr = dev->nic_info.unmapped_db +
1858             (srq->pd->id * dev->nic_info.db_page_size);
1859         uresp.db_page_size = dev->nic_info.db_page_size;
1860         uresp.num_rqe_allocated = srq->rq.max_cnt;
1861         if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
1862                 uresp.db_rq_offset = OCRDMA_DB_GEN2_RQ_OFFSET;
1863                 uresp.db_shift = 24;
1864         } else {
1865                 uresp.db_rq_offset = OCRDMA_DB_RQ_OFFSET;
1866                 uresp.db_shift = 16;
1867         }
1868
1869         status = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1870         if (status)
1871                 return status;
1872         status = ocrdma_add_mmap(srq->pd->uctx, uresp.rq_page_addr[0],
1873                                  uresp.rq_page_size);
1874         if (status)
1875                 return status;
1876         return status;
1877 }
1878
1879 struct ib_srq *ocrdma_create_srq(struct ib_pd *ibpd,
1880                                  struct ib_srq_init_attr *init_attr,
1881                                  struct ib_udata *udata)
1882 {
1883         int status = -ENOMEM;
1884         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
1885         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
1886         struct ocrdma_srq *srq;
1887
1888         if (init_attr->attr.max_sge > dev->attr.max_recv_sge)
1889                 return ERR_PTR(-EINVAL);
1890         if (init_attr->attr.max_wr > dev->attr.max_rqe)
1891                 return ERR_PTR(-EINVAL);
1892
1893         srq = kzalloc(sizeof(*srq), GFP_KERNEL);
1894         if (!srq)
1895                 return ERR_PTR(status);
1896
1897         spin_lock_init(&srq->q_lock);
1898         srq->pd = pd;
1899         srq->db = dev->nic_info.db + (pd->id * dev->nic_info.db_page_size);
1900         status = ocrdma_mbx_create_srq(dev, srq, init_attr, pd);
1901         if (status)
1902                 goto err;
1903
1904         if (udata == NULL) {
1905                 srq->rqe_wr_id_tbl = kzalloc(sizeof(u64) * srq->rq.max_cnt,
1906                             GFP_KERNEL);
1907                 if (srq->rqe_wr_id_tbl == NULL)
1908                         goto arm_err;
1909
1910                 srq->bit_fields_len = (srq->rq.max_cnt / 32) +
1911                     (srq->rq.max_cnt % 32 ? 1 : 0);
1912                 srq->idx_bit_fields =
1913                     kmalloc(srq->bit_fields_len * sizeof(u32), GFP_KERNEL);
1914                 if (srq->idx_bit_fields == NULL)
1915                         goto arm_err;
1916                 memset(srq->idx_bit_fields, 0xff,
1917                        srq->bit_fields_len * sizeof(u32));
1918         }
1919
1920         if (init_attr->attr.srq_limit) {
1921                 status = ocrdma_mbx_modify_srq(srq, &init_attr->attr);
1922                 if (status)
1923                         goto arm_err;
1924         }
1925
1926         if (udata) {
1927                 status = ocrdma_copy_srq_uresp(dev, srq, udata);
1928                 if (status)
1929                         goto arm_err;
1930         }
1931
1932         return &srq->ibsrq;
1933
1934 arm_err:
1935         ocrdma_mbx_destroy_srq(dev, srq);
1936 err:
1937         kfree(srq->rqe_wr_id_tbl);
1938         kfree(srq->idx_bit_fields);
1939         kfree(srq);
1940         return ERR_PTR(status);
1941 }
1942
1943 int ocrdma_modify_srq(struct ib_srq *ibsrq,
1944                       struct ib_srq_attr *srq_attr,
1945                       enum ib_srq_attr_mask srq_attr_mask,
1946                       struct ib_udata *udata)
1947 {
1948         int status;
1949         struct ocrdma_srq *srq;
1950
1951         srq = get_ocrdma_srq(ibsrq);
1952         if (srq_attr_mask & IB_SRQ_MAX_WR)
1953                 status = -EINVAL;
1954         else
1955                 status = ocrdma_mbx_modify_srq(srq, srq_attr);
1956         return status;
1957 }
1958
1959 int ocrdma_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
1960 {
1961         int status;
1962         struct ocrdma_srq *srq;
1963
1964         srq = get_ocrdma_srq(ibsrq);
1965         status = ocrdma_mbx_query_srq(srq, srq_attr);
1966         return status;
1967 }
1968
1969 int ocrdma_destroy_srq(struct ib_srq *ibsrq)
1970 {
1971         int status;
1972         struct ocrdma_srq *srq;
1973         struct ocrdma_dev *dev = get_ocrdma_dev(ibsrq->device);
1974
1975         srq = get_ocrdma_srq(ibsrq);
1976
1977         status = ocrdma_mbx_destroy_srq(dev, srq);
1978
1979         if (srq->pd->uctx)
1980                 ocrdma_del_mmap(srq->pd->uctx, (u64) srq->rq.pa,
1981                                 PAGE_ALIGN(srq->rq.len));
1982
1983         kfree(srq->idx_bit_fields);
1984         kfree(srq->rqe_wr_id_tbl);
1985         kfree(srq);
1986         return status;
1987 }
1988
1989 /* unprivileged verbs and their support functions. */
1990 static void ocrdma_build_ud_hdr(struct ocrdma_qp *qp,
1991                                 struct ocrdma_hdr_wqe *hdr,
1992                                 struct ib_send_wr *wr)
1993 {
1994         struct ocrdma_ewqe_ud_hdr *ud_hdr =
1995                 (struct ocrdma_ewqe_ud_hdr *)(hdr + 1);
1996         struct ocrdma_ah *ah = get_ocrdma_ah(ud_wr(wr)->ah);
1997
1998         ud_hdr->rsvd_dest_qpn = ud_wr(wr)->remote_qpn;
1999         if (qp->qp_type == IB_QPT_GSI)
2000                 ud_hdr->qkey = qp->qkey;
2001         else
2002                 ud_hdr->qkey = ud_wr(wr)->remote_qkey;
2003         ud_hdr->rsvd_ahid = ah->id;
2004         ud_hdr->hdr_type = ah->hdr_type;
2005         if (ah->av->valid & OCRDMA_AV_VLAN_VALID)
2006                 hdr->cw |= (OCRDMA_FLAG_AH_VLAN_PR << OCRDMA_WQE_FLAGS_SHIFT);
2007 }
2008
2009 static void ocrdma_build_sges(struct ocrdma_hdr_wqe *hdr,
2010                               struct ocrdma_sge *sge, int num_sge,
2011                               struct ib_sge *sg_list)
2012 {
2013         int i;
2014
2015         for (i = 0; i < num_sge; i++) {
2016                 sge[i].lrkey = sg_list[i].lkey;
2017                 sge[i].addr_lo = sg_list[i].addr;
2018                 sge[i].addr_hi = upper_32_bits(sg_list[i].addr);
2019                 sge[i].len = sg_list[i].length;
2020                 hdr->total_len += sg_list[i].length;
2021         }
2022         if (num_sge == 0)
2023                 memset(sge, 0, sizeof(*sge));
2024 }
2025
2026 static inline uint32_t ocrdma_sglist_len(struct ib_sge *sg_list, int num_sge)
2027 {
2028         uint32_t total_len = 0, i;
2029
2030         for (i = 0; i < num_sge; i++)
2031                 total_len += sg_list[i].length;
2032         return total_len;
2033 }
2034
2035
2036 static int ocrdma_build_inline_sges(struct ocrdma_qp *qp,
2037                                     struct ocrdma_hdr_wqe *hdr,
2038                                     struct ocrdma_sge *sge,
2039                                     struct ib_send_wr *wr, u32 wqe_size)
2040 {
2041         int i;
2042         char *dpp_addr;
2043
2044         if (wr->send_flags & IB_SEND_INLINE && qp->qp_type != IB_QPT_UD) {
2045                 hdr->total_len = ocrdma_sglist_len(wr->sg_list, wr->num_sge);
2046                 if (unlikely(hdr->total_len > qp->max_inline_data)) {
2047                         pr_err("%s() supported_len=0x%x,\n"
2048                                " unsupported len req=0x%x\n", __func__,
2049                                 qp->max_inline_data, hdr->total_len);
2050                         return -EINVAL;
2051                 }
2052                 dpp_addr = (char *)sge;
2053                 for (i = 0; i < wr->num_sge; i++) {
2054                         memcpy(dpp_addr,
2055                                (void *)(unsigned long)wr->sg_list[i].addr,
2056                                wr->sg_list[i].length);
2057                         dpp_addr += wr->sg_list[i].length;
2058                 }
2059
2060                 wqe_size += roundup(hdr->total_len, OCRDMA_WQE_ALIGN_BYTES);
2061                 if (0 == hdr->total_len)
2062                         wqe_size += sizeof(struct ocrdma_sge);
2063                 hdr->cw |= (OCRDMA_TYPE_INLINE << OCRDMA_WQE_TYPE_SHIFT);
2064         } else {
2065                 ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
2066                 if (wr->num_sge)
2067                         wqe_size += (wr->num_sge * sizeof(struct ocrdma_sge));
2068                 else
2069                         wqe_size += sizeof(struct ocrdma_sge);
2070                 hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2071         }
2072         hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2073         return 0;
2074 }
2075
2076 static int ocrdma_build_send(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2077                              struct ib_send_wr *wr)
2078 {
2079         int status;
2080         struct ocrdma_sge *sge;
2081         u32 wqe_size = sizeof(*hdr);
2082
2083         if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2084                 ocrdma_build_ud_hdr(qp, hdr, wr);
2085                 sge = (struct ocrdma_sge *)(hdr + 2);
2086                 wqe_size += sizeof(struct ocrdma_ewqe_ud_hdr);
2087         } else {
2088                 sge = (struct ocrdma_sge *)(hdr + 1);
2089         }
2090
2091         status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
2092         return status;
2093 }
2094
2095 static int ocrdma_build_write(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2096                               struct ib_send_wr *wr)
2097 {
2098         int status;
2099         struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
2100         struct ocrdma_sge *sge = ext_rw + 1;
2101         u32 wqe_size = sizeof(*hdr) + sizeof(*ext_rw);
2102
2103         status = ocrdma_build_inline_sges(qp, hdr, sge, wr, wqe_size);
2104         if (status)
2105                 return status;
2106         ext_rw->addr_lo = rdma_wr(wr)->remote_addr;
2107         ext_rw->addr_hi = upper_32_bits(rdma_wr(wr)->remote_addr);
2108         ext_rw->lrkey = rdma_wr(wr)->rkey;
2109         ext_rw->len = hdr->total_len;
2110         return 0;
2111 }
2112
2113 static void ocrdma_build_read(struct ocrdma_qp *qp, struct ocrdma_hdr_wqe *hdr,
2114                               struct ib_send_wr *wr)
2115 {
2116         struct ocrdma_sge *ext_rw = (struct ocrdma_sge *)(hdr + 1);
2117         struct ocrdma_sge *sge = ext_rw + 1;
2118         u32 wqe_size = ((wr->num_sge + 1) * sizeof(struct ocrdma_sge)) +
2119             sizeof(struct ocrdma_hdr_wqe);
2120
2121         ocrdma_build_sges(hdr, sge, wr->num_sge, wr->sg_list);
2122         hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2123         hdr->cw |= (OCRDMA_READ << OCRDMA_WQE_OPCODE_SHIFT);
2124         hdr->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2125
2126         ext_rw->addr_lo = rdma_wr(wr)->remote_addr;
2127         ext_rw->addr_hi = upper_32_bits(rdma_wr(wr)->remote_addr);
2128         ext_rw->lrkey = rdma_wr(wr)->rkey;
2129         ext_rw->len = hdr->total_len;
2130 }
2131
2132 static int get_encoded_page_size(int pg_sz)
2133 {
2134         /* Max size is 256M 4096 << 16 */
2135         int i = 0;
2136         for (; i < 17; i++)
2137                 if (pg_sz == (4096 << i))
2138                         break;
2139         return i;
2140 }
2141
2142 static int ocrdma_build_reg(struct ocrdma_qp *qp,
2143                             struct ocrdma_hdr_wqe *hdr,
2144                             struct ib_reg_wr *wr)
2145 {
2146         u64 fbo;
2147         struct ocrdma_ewqe_fr *fast_reg = (struct ocrdma_ewqe_fr *)(hdr + 1);
2148         struct ocrdma_mr *mr = get_ocrdma_mr(wr->mr);
2149         struct ocrdma_pbl *pbl_tbl = mr->hwmr.pbl_table;
2150         struct ocrdma_pbe *pbe;
2151         u32 wqe_size = sizeof(*fast_reg) + sizeof(*hdr);
2152         int num_pbes = 0, i;
2153
2154         wqe_size = roundup(wqe_size, OCRDMA_WQE_ALIGN_BYTES);
2155
2156         hdr->cw |= (OCRDMA_FR_MR << OCRDMA_WQE_OPCODE_SHIFT);
2157         hdr->cw |= ((wqe_size / OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT);
2158
2159         if (wr->access & IB_ACCESS_LOCAL_WRITE)
2160                 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_LOCAL_WR;
2161         if (wr->access & IB_ACCESS_REMOTE_WRITE)
2162                 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_WR;
2163         if (wr->access & IB_ACCESS_REMOTE_READ)
2164                 hdr->rsvd_lkey_flags |= OCRDMA_LKEY_FLAG_REMOTE_RD;
2165         hdr->lkey = wr->key;
2166         hdr->total_len = mr->ibmr.length;
2167
2168         fbo = mr->ibmr.iova - mr->pages[0];
2169
2170         fast_reg->va_hi = upper_32_bits(mr->ibmr.iova);
2171         fast_reg->va_lo = (u32) (mr->ibmr.iova & 0xffffffff);
2172         fast_reg->fbo_hi = upper_32_bits(fbo);
2173         fast_reg->fbo_lo = (u32) fbo & 0xffffffff;
2174         fast_reg->num_sges = mr->npages;
2175         fast_reg->size_sge = get_encoded_page_size(mr->ibmr.page_size);
2176
2177         pbe = pbl_tbl->va;
2178         for (i = 0; i < mr->npages; i++) {
2179                 u64 buf_addr = mr->pages[i];
2180
2181                 pbe->pa_lo = cpu_to_le32((u32) (buf_addr & PAGE_MASK));
2182                 pbe->pa_hi = cpu_to_le32((u32) upper_32_bits(buf_addr));
2183                 num_pbes += 1;
2184                 pbe++;
2185
2186                 /* if the pbl is full storing the pbes,
2187                  * move to next pbl.
2188                 */
2189                 if (num_pbes == (mr->hwmr.pbl_size/sizeof(u64))) {
2190                         pbl_tbl++;
2191                         pbe = (struct ocrdma_pbe *)pbl_tbl->va;
2192                 }
2193         }
2194
2195         return 0;
2196 }
2197
2198 static void ocrdma_ring_sq_db(struct ocrdma_qp *qp)
2199 {
2200         u32 val = qp->sq.dbid | (1 << OCRDMA_DB_SQ_SHIFT);
2201
2202         iowrite32(val, qp->sq_db);
2203 }
2204
2205 int ocrdma_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2206                      struct ib_send_wr **bad_wr)
2207 {
2208         int status = 0;
2209         struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2210         struct ocrdma_hdr_wqe *hdr;
2211         unsigned long flags;
2212
2213         spin_lock_irqsave(&qp->q_lock, flags);
2214         if (qp->state != OCRDMA_QPS_RTS && qp->state != OCRDMA_QPS_SQD) {
2215                 spin_unlock_irqrestore(&qp->q_lock, flags);
2216                 *bad_wr = wr;
2217                 return -EINVAL;
2218         }
2219
2220         while (wr) {
2221                 if (qp->qp_type == IB_QPT_UD &&
2222                     (wr->opcode != IB_WR_SEND &&
2223                      wr->opcode != IB_WR_SEND_WITH_IMM)) {
2224                         *bad_wr = wr;
2225                         status = -EINVAL;
2226                         break;
2227                 }
2228                 if (ocrdma_hwq_free_cnt(&qp->sq) == 0 ||
2229                     wr->num_sge > qp->sq.max_sges) {
2230                         *bad_wr = wr;
2231                         status = -ENOMEM;
2232                         break;
2233                 }
2234                 hdr = ocrdma_hwq_head(&qp->sq);
2235                 hdr->cw = 0;
2236                 if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2237                         hdr->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2238                 if (wr->send_flags & IB_SEND_FENCE)
2239                         hdr->cw |=
2240                             (OCRDMA_FLAG_FENCE_L << OCRDMA_WQE_FLAGS_SHIFT);
2241                 if (wr->send_flags & IB_SEND_SOLICITED)
2242                         hdr->cw |=
2243                             (OCRDMA_FLAG_SOLICIT << OCRDMA_WQE_FLAGS_SHIFT);
2244                 hdr->total_len = 0;
2245                 switch (wr->opcode) {
2246                 case IB_WR_SEND_WITH_IMM:
2247                         hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2248                         hdr->immdt = ntohl(wr->ex.imm_data);
2249                 case IB_WR_SEND:
2250                         hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2251                         ocrdma_build_send(qp, hdr, wr);
2252                         break;
2253                 case IB_WR_SEND_WITH_INV:
2254                         hdr->cw |= (OCRDMA_FLAG_INV << OCRDMA_WQE_FLAGS_SHIFT);
2255                         hdr->cw |= (OCRDMA_SEND << OCRDMA_WQE_OPCODE_SHIFT);
2256                         hdr->lkey = wr->ex.invalidate_rkey;
2257                         status = ocrdma_build_send(qp, hdr, wr);
2258                         break;
2259                 case IB_WR_RDMA_WRITE_WITH_IMM:
2260                         hdr->cw |= (OCRDMA_FLAG_IMM << OCRDMA_WQE_FLAGS_SHIFT);
2261                         hdr->immdt = ntohl(wr->ex.imm_data);
2262                 case IB_WR_RDMA_WRITE:
2263                         hdr->cw |= (OCRDMA_WRITE << OCRDMA_WQE_OPCODE_SHIFT);
2264                         status = ocrdma_build_write(qp, hdr, wr);
2265                         break;
2266                 case IB_WR_RDMA_READ:
2267                         ocrdma_build_read(qp, hdr, wr);
2268                         break;
2269                 case IB_WR_LOCAL_INV:
2270                         hdr->cw |=
2271                             (OCRDMA_LKEY_INV << OCRDMA_WQE_OPCODE_SHIFT);
2272                         hdr->cw |= ((sizeof(struct ocrdma_hdr_wqe) +
2273                                         sizeof(struct ocrdma_sge)) /
2274                                 OCRDMA_WQE_STRIDE) << OCRDMA_WQE_SIZE_SHIFT;
2275                         hdr->lkey = wr->ex.invalidate_rkey;
2276                         break;
2277                 case IB_WR_REG_MR:
2278                         status = ocrdma_build_reg(qp, hdr, reg_wr(wr));
2279                         break;
2280                 default:
2281                         status = -EINVAL;
2282                         break;
2283                 }
2284                 if (status) {
2285                         *bad_wr = wr;
2286                         break;
2287                 }
2288                 if (wr->send_flags & IB_SEND_SIGNALED || qp->signaled)
2289                         qp->wqe_wr_id_tbl[qp->sq.head].signaled = 1;
2290                 else
2291                         qp->wqe_wr_id_tbl[qp->sq.head].signaled = 0;
2292                 qp->wqe_wr_id_tbl[qp->sq.head].wrid = wr->wr_id;
2293                 ocrdma_cpu_to_le32(hdr, ((hdr->cw >> OCRDMA_WQE_SIZE_SHIFT) &
2294                                    OCRDMA_WQE_SIZE_MASK) * OCRDMA_WQE_STRIDE);
2295                 /* make sure wqe is written before adapter can access it */
2296                 wmb();
2297                 /* inform hw to start processing it */
2298                 ocrdma_ring_sq_db(qp);
2299
2300                 /* update pointer, counter for next wr */
2301                 ocrdma_hwq_inc_head(&qp->sq);
2302                 wr = wr->next;
2303         }
2304         spin_unlock_irqrestore(&qp->q_lock, flags);
2305         return status;
2306 }
2307
2308 static void ocrdma_ring_rq_db(struct ocrdma_qp *qp)
2309 {
2310         u32 val = qp->rq.dbid | (1 << OCRDMA_DB_RQ_SHIFT);
2311
2312         iowrite32(val, qp->rq_db);
2313 }
2314
2315 static void ocrdma_build_rqe(struct ocrdma_hdr_wqe *rqe, struct ib_recv_wr *wr,
2316                              u16 tag)
2317 {
2318         u32 wqe_size = 0;
2319         struct ocrdma_sge *sge;
2320         if (wr->num_sge)
2321                 wqe_size = (wr->num_sge * sizeof(*sge)) + sizeof(*rqe);
2322         else
2323                 wqe_size = sizeof(*sge) + sizeof(*rqe);
2324
2325         rqe->cw = ((wqe_size / OCRDMA_WQE_STRIDE) <<
2326                                 OCRDMA_WQE_SIZE_SHIFT);
2327         rqe->cw |= (OCRDMA_FLAG_SIG << OCRDMA_WQE_FLAGS_SHIFT);
2328         rqe->cw |= (OCRDMA_TYPE_LKEY << OCRDMA_WQE_TYPE_SHIFT);
2329         rqe->total_len = 0;
2330         rqe->rsvd_tag = tag;
2331         sge = (struct ocrdma_sge *)(rqe + 1);
2332         ocrdma_build_sges(rqe, sge, wr->num_sge, wr->sg_list);
2333         ocrdma_cpu_to_le32(rqe, wqe_size);
2334 }
2335
2336 int ocrdma_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2337                      struct ib_recv_wr **bad_wr)
2338 {
2339         int status = 0;
2340         unsigned long flags;
2341         struct ocrdma_qp *qp = get_ocrdma_qp(ibqp);
2342         struct ocrdma_hdr_wqe *rqe;
2343
2344         spin_lock_irqsave(&qp->q_lock, flags);
2345         if (qp->state == OCRDMA_QPS_RST || qp->state == OCRDMA_QPS_ERR) {
2346                 spin_unlock_irqrestore(&qp->q_lock, flags);
2347                 *bad_wr = wr;
2348                 return -EINVAL;
2349         }
2350         while (wr) {
2351                 if (ocrdma_hwq_free_cnt(&qp->rq) == 0 ||
2352                     wr->num_sge > qp->rq.max_sges) {
2353                         *bad_wr = wr;
2354                         status = -ENOMEM;
2355                         break;
2356                 }
2357                 rqe = ocrdma_hwq_head(&qp->rq);
2358                 ocrdma_build_rqe(rqe, wr, 0);
2359
2360                 qp->rqe_wr_id_tbl[qp->rq.head] = wr->wr_id;
2361                 /* make sure rqe is written before adapter can access it */
2362                 wmb();
2363
2364                 /* inform hw to start processing it */
2365                 ocrdma_ring_rq_db(qp);
2366
2367                 /* update pointer, counter for next wr */
2368                 ocrdma_hwq_inc_head(&qp->rq);
2369                 wr = wr->next;
2370         }
2371         spin_unlock_irqrestore(&qp->q_lock, flags);
2372         return status;
2373 }
2374
2375 /* cqe for srq's rqe can potentially arrive out of order.
2376  * index gives the entry in the shadow table where to store
2377  * the wr_id. tag/index is returned in cqe to reference back
2378  * for a given rqe.
2379  */
2380 static int ocrdma_srq_get_idx(struct ocrdma_srq *srq)
2381 {
2382         int row = 0;
2383         int indx = 0;
2384
2385         for (row = 0; row < srq->bit_fields_len; row++) {
2386                 if (srq->idx_bit_fields[row]) {
2387                         indx = ffs(srq->idx_bit_fields[row]);
2388                         indx = (row * 32) + (indx - 1);
2389                         BUG_ON(indx >= srq->rq.max_cnt);
2390                         ocrdma_srq_toggle_bit(srq, indx);
2391                         break;
2392                 }
2393         }
2394
2395         BUG_ON(row == srq->bit_fields_len);
2396         return indx + 1; /* Use from index 1 */
2397 }
2398
2399 static void ocrdma_ring_srq_db(struct ocrdma_srq *srq)
2400 {
2401         u32 val = srq->rq.dbid | (1 << 16);
2402
2403         iowrite32(val, srq->db + OCRDMA_DB_GEN2_SRQ_OFFSET);
2404 }
2405
2406 int ocrdma_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
2407                          struct ib_recv_wr **bad_wr)
2408 {
2409         int status = 0;
2410         unsigned long flags;
2411         struct ocrdma_srq *srq;
2412         struct ocrdma_hdr_wqe *rqe;
2413         u16 tag;
2414
2415         srq = get_ocrdma_srq(ibsrq);
2416
2417         spin_lock_irqsave(&srq->q_lock, flags);
2418         while (wr) {
2419                 if (ocrdma_hwq_free_cnt(&srq->rq) == 0 ||
2420                     wr->num_sge > srq->rq.max_sges) {
2421                         status = -ENOMEM;
2422                         *bad_wr = wr;
2423                         break;
2424                 }
2425                 tag = ocrdma_srq_get_idx(srq);
2426                 rqe = ocrdma_hwq_head(&srq->rq);
2427                 ocrdma_build_rqe(rqe, wr, tag);
2428
2429                 srq->rqe_wr_id_tbl[tag] = wr->wr_id;
2430                 /* make sure rqe is written before adapter can perform DMA */
2431                 wmb();
2432                 /* inform hw to start processing it */
2433                 ocrdma_ring_srq_db(srq);
2434                 /* update pointer, counter for next wr */
2435                 ocrdma_hwq_inc_head(&srq->rq);
2436                 wr = wr->next;
2437         }
2438         spin_unlock_irqrestore(&srq->q_lock, flags);
2439         return status;
2440 }
2441
2442 static enum ib_wc_status ocrdma_to_ibwc_err(u16 status)
2443 {
2444         enum ib_wc_status ibwc_status;
2445
2446         switch (status) {
2447         case OCRDMA_CQE_GENERAL_ERR:
2448                 ibwc_status = IB_WC_GENERAL_ERR;
2449                 break;
2450         case OCRDMA_CQE_LOC_LEN_ERR:
2451                 ibwc_status = IB_WC_LOC_LEN_ERR;
2452                 break;
2453         case OCRDMA_CQE_LOC_QP_OP_ERR:
2454                 ibwc_status = IB_WC_LOC_QP_OP_ERR;
2455                 break;
2456         case OCRDMA_CQE_LOC_EEC_OP_ERR:
2457                 ibwc_status = IB_WC_LOC_EEC_OP_ERR;
2458                 break;
2459         case OCRDMA_CQE_LOC_PROT_ERR:
2460                 ibwc_status = IB_WC_LOC_PROT_ERR;
2461                 break;
2462         case OCRDMA_CQE_WR_FLUSH_ERR:
2463                 ibwc_status = IB_WC_WR_FLUSH_ERR;
2464                 break;
2465         case OCRDMA_CQE_MW_BIND_ERR:
2466                 ibwc_status = IB_WC_MW_BIND_ERR;
2467                 break;
2468         case OCRDMA_CQE_BAD_RESP_ERR:
2469                 ibwc_status = IB_WC_BAD_RESP_ERR;
2470                 break;
2471         case OCRDMA_CQE_LOC_ACCESS_ERR:
2472                 ibwc_status = IB_WC_LOC_ACCESS_ERR;
2473                 break;
2474         case OCRDMA_CQE_REM_INV_REQ_ERR:
2475                 ibwc_status = IB_WC_REM_INV_REQ_ERR;
2476                 break;
2477         case OCRDMA_CQE_REM_ACCESS_ERR:
2478                 ibwc_status = IB_WC_REM_ACCESS_ERR;
2479                 break;
2480         case OCRDMA_CQE_REM_OP_ERR:
2481                 ibwc_status = IB_WC_REM_OP_ERR;
2482                 break;
2483         case OCRDMA_CQE_RETRY_EXC_ERR:
2484                 ibwc_status = IB_WC_RETRY_EXC_ERR;
2485                 break;
2486         case OCRDMA_CQE_RNR_RETRY_EXC_ERR:
2487                 ibwc_status = IB_WC_RNR_RETRY_EXC_ERR;
2488                 break;
2489         case OCRDMA_CQE_LOC_RDD_VIOL_ERR:
2490                 ibwc_status = IB_WC_LOC_RDD_VIOL_ERR;
2491                 break;
2492         case OCRDMA_CQE_REM_INV_RD_REQ_ERR:
2493                 ibwc_status = IB_WC_REM_INV_RD_REQ_ERR;
2494                 break;
2495         case OCRDMA_CQE_REM_ABORT_ERR:
2496                 ibwc_status = IB_WC_REM_ABORT_ERR;
2497                 break;
2498         case OCRDMA_CQE_INV_EECN_ERR:
2499                 ibwc_status = IB_WC_INV_EECN_ERR;
2500                 break;
2501         case OCRDMA_CQE_INV_EEC_STATE_ERR:
2502                 ibwc_status = IB_WC_INV_EEC_STATE_ERR;
2503                 break;
2504         case OCRDMA_CQE_FATAL_ERR:
2505                 ibwc_status = IB_WC_FATAL_ERR;
2506                 break;
2507         case OCRDMA_CQE_RESP_TIMEOUT_ERR:
2508                 ibwc_status = IB_WC_RESP_TIMEOUT_ERR;
2509                 break;
2510         default:
2511                 ibwc_status = IB_WC_GENERAL_ERR;
2512                 break;
2513         }
2514         return ibwc_status;
2515 }
2516
2517 static void ocrdma_update_wc(struct ocrdma_qp *qp, struct ib_wc *ibwc,
2518                       u32 wqe_idx)
2519 {
2520         struct ocrdma_hdr_wqe *hdr;
2521         struct ocrdma_sge *rw;
2522         int opcode;
2523
2524         hdr = ocrdma_hwq_head_from_idx(&qp->sq, wqe_idx);
2525
2526         ibwc->wr_id = qp->wqe_wr_id_tbl[wqe_idx].wrid;
2527         /* Undo the hdr->cw swap */
2528         opcode = le32_to_cpu(hdr->cw) & OCRDMA_WQE_OPCODE_MASK;
2529         switch (opcode) {
2530         case OCRDMA_WRITE:
2531                 ibwc->opcode = IB_WC_RDMA_WRITE;
2532                 break;
2533         case OCRDMA_READ:
2534                 rw = (struct ocrdma_sge *)(hdr + 1);
2535                 ibwc->opcode = IB_WC_RDMA_READ;
2536                 ibwc->byte_len = rw->len;
2537                 break;
2538         case OCRDMA_SEND:
2539                 ibwc->opcode = IB_WC_SEND;
2540                 break;
2541         case OCRDMA_FR_MR:
2542                 ibwc->opcode = IB_WC_REG_MR;
2543                 break;
2544         case OCRDMA_LKEY_INV:
2545                 ibwc->opcode = IB_WC_LOCAL_INV;
2546                 break;
2547         default:
2548                 ibwc->status = IB_WC_GENERAL_ERR;
2549                 pr_err("%s() invalid opcode received = 0x%x\n",
2550                        __func__, hdr->cw & OCRDMA_WQE_OPCODE_MASK);
2551                 break;
2552         }
2553 }
2554
2555 static void ocrdma_set_cqe_status_flushed(struct ocrdma_qp *qp,
2556                                                 struct ocrdma_cqe *cqe)
2557 {
2558         if (is_cqe_for_sq(cqe)) {
2559                 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2560                                 cqe->flags_status_srcqpn) &
2561                                         ~OCRDMA_CQE_STATUS_MASK);
2562                 cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2563                                 cqe->flags_status_srcqpn) |
2564                                 (OCRDMA_CQE_WR_FLUSH_ERR <<
2565                                         OCRDMA_CQE_STATUS_SHIFT));
2566         } else {
2567                 if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2568                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2569                                         cqe->flags_status_srcqpn) &
2570                                                 ~OCRDMA_CQE_UD_STATUS_MASK);
2571                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2572                                         cqe->flags_status_srcqpn) |
2573                                         (OCRDMA_CQE_WR_FLUSH_ERR <<
2574                                                 OCRDMA_CQE_UD_STATUS_SHIFT));
2575                 } else {
2576                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2577                                         cqe->flags_status_srcqpn) &
2578                                                 ~OCRDMA_CQE_STATUS_MASK);
2579                         cqe->flags_status_srcqpn = cpu_to_le32(le32_to_cpu(
2580                                         cqe->flags_status_srcqpn) |
2581                                         (OCRDMA_CQE_WR_FLUSH_ERR <<
2582                                                 OCRDMA_CQE_STATUS_SHIFT));
2583                 }
2584         }
2585 }
2586
2587 static bool ocrdma_update_err_cqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2588                                   struct ocrdma_qp *qp, int status)
2589 {
2590         bool expand = false;
2591
2592         ibwc->byte_len = 0;
2593         ibwc->qp = &qp->ibqp;
2594         ibwc->status = ocrdma_to_ibwc_err(status);
2595
2596         ocrdma_flush_qp(qp);
2597         ocrdma_qp_state_change(qp, IB_QPS_ERR, NULL);
2598
2599         /* if wqe/rqe pending for which cqe needs to be returned,
2600          * trigger inflating it.
2601          */
2602         if (!is_hw_rq_empty(qp) || !is_hw_sq_empty(qp)) {
2603                 expand = true;
2604                 ocrdma_set_cqe_status_flushed(qp, cqe);
2605         }
2606         return expand;
2607 }
2608
2609 static int ocrdma_update_err_rcqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2610                                   struct ocrdma_qp *qp, int status)
2611 {
2612         ibwc->opcode = IB_WC_RECV;
2613         ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2614         ocrdma_hwq_inc_tail(&qp->rq);
2615
2616         return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2617 }
2618
2619 static int ocrdma_update_err_scqe(struct ib_wc *ibwc, struct ocrdma_cqe *cqe,
2620                                   struct ocrdma_qp *qp, int status)
2621 {
2622         ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2623         ocrdma_hwq_inc_tail(&qp->sq);
2624
2625         return ocrdma_update_err_cqe(ibwc, cqe, qp, status);
2626 }
2627
2628
2629 static bool ocrdma_poll_err_scqe(struct ocrdma_qp *qp,
2630                                  struct ocrdma_cqe *cqe, struct ib_wc *ibwc,
2631                                  bool *polled, bool *stop)
2632 {
2633         bool expand;
2634         struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2635         int status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2636                 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2637         if (status < OCRDMA_MAX_CQE_ERR)
2638                 atomic_inc(&dev->cqe_err_stats[status]);
2639
2640         /* when hw sq is empty, but rq is not empty, so we continue
2641          * to keep the cqe in order to get the cq event again.
2642          */
2643         if (is_hw_sq_empty(qp) && !is_hw_rq_empty(qp)) {
2644                 /* when cq for rq and sq is same, it is safe to return
2645                  * flush cqe for RQEs.
2646                  */
2647                 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2648                         *polled = true;
2649                         status = OCRDMA_CQE_WR_FLUSH_ERR;
2650                         expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2651                 } else {
2652                         /* stop processing further cqe as this cqe is used for
2653                          * triggering cq event on buddy cq of RQ.
2654                          * When QP is destroyed, this cqe will be removed
2655                          * from the cq's hardware q.
2656                          */
2657                         *polled = false;
2658                         *stop = true;
2659                         expand = false;
2660                 }
2661         } else if (is_hw_sq_empty(qp)) {
2662                 /* Do nothing */
2663                 expand = false;
2664                 *polled = false;
2665                 *stop = false;
2666         } else {
2667                 *polled = true;
2668                 expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2669         }
2670         return expand;
2671 }
2672
2673 static bool ocrdma_poll_success_scqe(struct ocrdma_qp *qp,
2674                                      struct ocrdma_cqe *cqe,
2675                                      struct ib_wc *ibwc, bool *polled)
2676 {
2677         bool expand = false;
2678         int tail = qp->sq.tail;
2679         u32 wqe_idx;
2680
2681         if (!qp->wqe_wr_id_tbl[tail].signaled) {
2682                 *polled = false;    /* WC cannot be consumed yet */
2683         } else {
2684                 ibwc->status = IB_WC_SUCCESS;
2685                 ibwc->wc_flags = 0;
2686                 ibwc->qp = &qp->ibqp;
2687                 ocrdma_update_wc(qp, ibwc, tail);
2688                 *polled = true;
2689         }
2690         wqe_idx = (le32_to_cpu(cqe->wq.wqeidx) &
2691                         OCRDMA_CQE_WQEIDX_MASK) & qp->sq.max_wqe_idx;
2692         if (tail != wqe_idx)
2693                 expand = true; /* Coalesced CQE can't be consumed yet */
2694
2695         ocrdma_hwq_inc_tail(&qp->sq);
2696         return expand;
2697 }
2698
2699 static bool ocrdma_poll_scqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2700                              struct ib_wc *ibwc, bool *polled, bool *stop)
2701 {
2702         int status;
2703         bool expand;
2704
2705         status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2706                 OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2707
2708         if (status == OCRDMA_CQE_SUCCESS)
2709                 expand = ocrdma_poll_success_scqe(qp, cqe, ibwc, polled);
2710         else
2711                 expand = ocrdma_poll_err_scqe(qp, cqe, ibwc, polled, stop);
2712         return expand;
2713 }
2714
2715 static int ocrdma_update_ud_rcqe(struct ocrdma_dev *dev, struct ib_wc *ibwc,
2716                                  struct ocrdma_cqe *cqe)
2717 {
2718         int status;
2719         u16 hdr_type = 0;
2720
2721         status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2722                 OCRDMA_CQE_UD_STATUS_MASK) >> OCRDMA_CQE_UD_STATUS_SHIFT;
2723         ibwc->src_qp = le32_to_cpu(cqe->flags_status_srcqpn) &
2724                                                 OCRDMA_CQE_SRCQP_MASK;
2725         ibwc->pkey_index = 0;
2726         ibwc->wc_flags = IB_WC_GRH;
2727         ibwc->byte_len = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2728                           OCRDMA_CQE_UD_XFER_LEN_SHIFT) &
2729                           OCRDMA_CQE_UD_XFER_LEN_MASK;
2730
2731         if (ocrdma_is_udp_encap_supported(dev)) {
2732                 hdr_type = (le32_to_cpu(cqe->ud.rxlen_pkey) >>
2733                             OCRDMA_CQE_UD_L3TYPE_SHIFT) &
2734                             OCRDMA_CQE_UD_L3TYPE_MASK;
2735                 ibwc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
2736                 ibwc->network_hdr_type = hdr_type;
2737         }
2738
2739         return status;
2740 }
2741
2742 static void ocrdma_update_free_srq_cqe(struct ib_wc *ibwc,
2743                                        struct ocrdma_cqe *cqe,
2744                                        struct ocrdma_qp *qp)
2745 {
2746         unsigned long flags;
2747         struct ocrdma_srq *srq;
2748         u32 wqe_idx;
2749
2750         srq = get_ocrdma_srq(qp->ibqp.srq);
2751         wqe_idx = (le32_to_cpu(cqe->rq.buftag_qpn) >>
2752                 OCRDMA_CQE_BUFTAG_SHIFT) & srq->rq.max_wqe_idx;
2753         BUG_ON(wqe_idx < 1);
2754
2755         ibwc->wr_id = srq->rqe_wr_id_tbl[wqe_idx];
2756         spin_lock_irqsave(&srq->q_lock, flags);
2757         ocrdma_srq_toggle_bit(srq, wqe_idx - 1);
2758         spin_unlock_irqrestore(&srq->q_lock, flags);
2759         ocrdma_hwq_inc_tail(&srq->rq);
2760 }
2761
2762 static bool ocrdma_poll_err_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2763                                 struct ib_wc *ibwc, bool *polled, bool *stop,
2764                                 int status)
2765 {
2766         bool expand;
2767         struct ocrdma_dev *dev = get_ocrdma_dev(qp->ibqp.device);
2768
2769         if (status < OCRDMA_MAX_CQE_ERR)
2770                 atomic_inc(&dev->cqe_err_stats[status]);
2771
2772         /* when hw_rq is empty, but wq is not empty, so continue
2773          * to keep the cqe to get the cq event again.
2774          */
2775         if (is_hw_rq_empty(qp) && !is_hw_sq_empty(qp)) {
2776                 if (!qp->srq && (qp->sq_cq == qp->rq_cq)) {
2777                         *polled = true;
2778                         status = OCRDMA_CQE_WR_FLUSH_ERR;
2779                         expand = ocrdma_update_err_scqe(ibwc, cqe, qp, status);
2780                 } else {
2781                         *polled = false;
2782                         *stop = true;
2783                         expand = false;
2784                 }
2785         } else if (is_hw_rq_empty(qp)) {
2786                 /* Do nothing */
2787                 expand = false;
2788                 *polled = false;
2789                 *stop = false;
2790         } else {
2791                 *polled = true;
2792                 expand = ocrdma_update_err_rcqe(ibwc, cqe, qp, status);
2793         }
2794         return expand;
2795 }
2796
2797 static void ocrdma_poll_success_rcqe(struct ocrdma_qp *qp,
2798                                      struct ocrdma_cqe *cqe, struct ib_wc *ibwc)
2799 {
2800         struct ocrdma_dev *dev;
2801
2802         dev = get_ocrdma_dev(qp->ibqp.device);
2803         ibwc->opcode = IB_WC_RECV;
2804         ibwc->qp = &qp->ibqp;
2805         ibwc->status = IB_WC_SUCCESS;
2806
2807         if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI)
2808                 ocrdma_update_ud_rcqe(dev, ibwc, cqe);
2809         else
2810                 ibwc->byte_len = le32_to_cpu(cqe->rq.rxlen);
2811
2812         if (is_cqe_imm(cqe)) {
2813                 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2814                 ibwc->wc_flags |= IB_WC_WITH_IMM;
2815         } else if (is_cqe_wr_imm(cqe)) {
2816                 ibwc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2817                 ibwc->ex.imm_data = htonl(le32_to_cpu(cqe->rq.lkey_immdt));
2818                 ibwc->wc_flags |= IB_WC_WITH_IMM;
2819         } else if (is_cqe_invalidated(cqe)) {
2820                 ibwc->ex.invalidate_rkey = le32_to_cpu(cqe->rq.lkey_immdt);
2821                 ibwc->wc_flags |= IB_WC_WITH_INVALIDATE;
2822         }
2823         if (qp->ibqp.srq) {
2824                 ocrdma_update_free_srq_cqe(ibwc, cqe, qp);
2825         } else {
2826                 ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2827                 ocrdma_hwq_inc_tail(&qp->rq);
2828         }
2829 }
2830
2831 static bool ocrdma_poll_rcqe(struct ocrdma_qp *qp, struct ocrdma_cqe *cqe,
2832                              struct ib_wc *ibwc, bool *polled, bool *stop)
2833 {
2834         int status;
2835         bool expand = false;
2836
2837         ibwc->wc_flags = 0;
2838         if (qp->qp_type == IB_QPT_UD || qp->qp_type == IB_QPT_GSI) {
2839                 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2840                                         OCRDMA_CQE_UD_STATUS_MASK) >>
2841                                         OCRDMA_CQE_UD_STATUS_SHIFT;
2842         } else {
2843                 status = (le32_to_cpu(cqe->flags_status_srcqpn) &
2844                              OCRDMA_CQE_STATUS_MASK) >> OCRDMA_CQE_STATUS_SHIFT;
2845         }
2846
2847         if (status == OCRDMA_CQE_SUCCESS) {
2848                 *polled = true;
2849                 ocrdma_poll_success_rcqe(qp, cqe, ibwc);
2850         } else {
2851                 expand = ocrdma_poll_err_rcqe(qp, cqe, ibwc, polled, stop,
2852                                               status);
2853         }
2854         return expand;
2855 }
2856
2857 static void ocrdma_change_cq_phase(struct ocrdma_cq *cq, struct ocrdma_cqe *cqe,
2858                                    u16 cur_getp)
2859 {
2860         if (cq->phase_change) {
2861                 if (cur_getp == 0)
2862                         cq->phase = (~cq->phase & OCRDMA_CQE_VALID);
2863         } else {
2864                 /* clear valid bit */
2865                 cqe->flags_status_srcqpn = 0;
2866         }
2867 }
2868
2869 static int ocrdma_poll_hwcq(struct ocrdma_cq *cq, int num_entries,
2870                             struct ib_wc *ibwc)
2871 {
2872         u16 qpn = 0;
2873         int i = 0;
2874         bool expand = false;
2875         int polled_hw_cqes = 0;
2876         struct ocrdma_qp *qp = NULL;
2877         struct ocrdma_dev *dev = get_ocrdma_dev(cq->ibcq.device);
2878         struct ocrdma_cqe *cqe;
2879         u16 cur_getp; bool polled = false; bool stop = false;
2880
2881         cur_getp = cq->getp;
2882         while (num_entries) {
2883                 cqe = cq->va + cur_getp;
2884                 /* check whether valid cqe or not */
2885                 if (!is_cqe_valid(cq, cqe))
2886                         break;
2887                 qpn = (le32_to_cpu(cqe->cmn.qpn) & OCRDMA_CQE_QPN_MASK);
2888                 /* ignore discarded cqe */
2889                 if (qpn == 0)
2890                         goto skip_cqe;
2891                 qp = dev->qp_tbl[qpn];
2892                 BUG_ON(qp == NULL);
2893
2894                 if (is_cqe_for_sq(cqe)) {
2895                         expand = ocrdma_poll_scqe(qp, cqe, ibwc, &polled,
2896                                                   &stop);
2897                 } else {
2898                         expand = ocrdma_poll_rcqe(qp, cqe, ibwc, &polled,
2899                                                   &stop);
2900                 }
2901                 if (expand)
2902                         goto expand_cqe;
2903                 if (stop)
2904                         goto stop_cqe;
2905                 /* clear qpn to avoid duplicate processing by discard_cqe() */
2906                 cqe->cmn.qpn = 0;
2907 skip_cqe:
2908                 polled_hw_cqes += 1;
2909                 cur_getp = (cur_getp + 1) % cq->max_hw_cqe;
2910                 ocrdma_change_cq_phase(cq, cqe, cur_getp);
2911 expand_cqe:
2912                 if (polled) {
2913                         num_entries -= 1;
2914                         i += 1;
2915                         ibwc = ibwc + 1;
2916                         polled = false;
2917                 }
2918         }
2919 stop_cqe:
2920         cq->getp = cur_getp;
2921
2922         if (polled_hw_cqes)
2923                 ocrdma_ring_cq_db(dev, cq->id, false, false, polled_hw_cqes);
2924
2925         return i;
2926 }
2927
2928 /* insert error cqe if the QP's SQ or RQ's CQ matches the CQ under poll. */
2929 static int ocrdma_add_err_cqe(struct ocrdma_cq *cq, int num_entries,
2930                               struct ocrdma_qp *qp, struct ib_wc *ibwc)
2931 {
2932         int err_cqes = 0;
2933
2934         while (num_entries) {
2935                 if (is_hw_sq_empty(qp) && is_hw_rq_empty(qp))
2936                         break;
2937                 if (!is_hw_sq_empty(qp) && qp->sq_cq == cq) {
2938                         ocrdma_update_wc(qp, ibwc, qp->sq.tail);
2939                         ocrdma_hwq_inc_tail(&qp->sq);
2940                 } else if (!is_hw_rq_empty(qp) && qp->rq_cq == cq) {
2941                         ibwc->wr_id = qp->rqe_wr_id_tbl[qp->rq.tail];
2942                         ocrdma_hwq_inc_tail(&qp->rq);
2943                 } else {
2944                         return err_cqes;
2945                 }
2946                 ibwc->byte_len = 0;
2947                 ibwc->status = IB_WC_WR_FLUSH_ERR;
2948                 ibwc = ibwc + 1;
2949                 err_cqes += 1;
2950                 num_entries -= 1;
2951         }
2952         return err_cqes;
2953 }
2954
2955 int ocrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
2956 {
2957         int cqes_to_poll = num_entries;
2958         struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2959         struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2960         int num_os_cqe = 0, err_cqes = 0;
2961         struct ocrdma_qp *qp;
2962         unsigned long flags;
2963
2964         /* poll cqes from adapter CQ */
2965         spin_lock_irqsave(&cq->cq_lock, flags);
2966         num_os_cqe = ocrdma_poll_hwcq(cq, cqes_to_poll, wc);
2967         spin_unlock_irqrestore(&cq->cq_lock, flags);
2968         cqes_to_poll -= num_os_cqe;
2969
2970         if (cqes_to_poll) {
2971                 wc = wc + num_os_cqe;
2972                 /* adapter returns single error cqe when qp moves to
2973                  * error state. So insert error cqes with wc_status as
2974                  * FLUSHED for pending WQEs and RQEs of QP's SQ and RQ
2975                  * respectively which uses this CQ.
2976                  */
2977                 spin_lock_irqsave(&dev->flush_q_lock, flags);
2978                 list_for_each_entry(qp, &cq->sq_head, sq_entry) {
2979                         if (cqes_to_poll == 0)
2980                                 break;
2981                         err_cqes = ocrdma_add_err_cqe(cq, cqes_to_poll, qp, wc);
2982                         cqes_to_poll -= err_cqes;
2983                         num_os_cqe += err_cqes;
2984                         wc = wc + err_cqes;
2985                 }
2986                 spin_unlock_irqrestore(&dev->flush_q_lock, flags);
2987         }
2988         return num_os_cqe;
2989 }
2990
2991 int ocrdma_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags cq_flags)
2992 {
2993         struct ocrdma_cq *cq = get_ocrdma_cq(ibcq);
2994         struct ocrdma_dev *dev = get_ocrdma_dev(ibcq->device);
2995         u16 cq_id;
2996         unsigned long flags;
2997         bool arm_needed = false, sol_needed = false;
2998
2999         cq_id = cq->id;
3000
3001         spin_lock_irqsave(&cq->cq_lock, flags);
3002         if (cq_flags & IB_CQ_NEXT_COMP || cq_flags & IB_CQ_SOLICITED)
3003                 arm_needed = true;
3004         if (cq_flags & IB_CQ_SOLICITED)
3005                 sol_needed = true;
3006
3007         ocrdma_ring_cq_db(dev, cq_id, arm_needed, sol_needed, 0);
3008         spin_unlock_irqrestore(&cq->cq_lock, flags);
3009
3010         return 0;
3011 }
3012
3013 struct ib_mr *ocrdma_alloc_mr(struct ib_pd *ibpd,
3014                               enum ib_mr_type mr_type,
3015                               u32 max_num_sg)
3016 {
3017         int status;
3018         struct ocrdma_mr *mr;
3019         struct ocrdma_pd *pd = get_ocrdma_pd(ibpd);
3020         struct ocrdma_dev *dev = get_ocrdma_dev(ibpd->device);
3021
3022         if (mr_type != IB_MR_TYPE_MEM_REG)
3023                 return ERR_PTR(-EINVAL);
3024
3025         if (max_num_sg > dev->attr.max_pages_per_frmr)
3026                 return ERR_PTR(-EINVAL);
3027
3028         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3029         if (!mr)
3030                 return ERR_PTR(-ENOMEM);
3031
3032         mr->pages = kcalloc(max_num_sg, sizeof(u64), GFP_KERNEL);
3033         if (!mr->pages) {
3034                 status = -ENOMEM;
3035                 goto pl_err;
3036         }
3037
3038         status = ocrdma_get_pbl_info(dev, mr, max_num_sg);
3039         if (status)
3040                 goto pbl_err;
3041         mr->hwmr.fr_mr = 1;
3042         mr->hwmr.remote_rd = 0;
3043         mr->hwmr.remote_wr = 0;
3044         mr->hwmr.local_rd = 0;
3045         mr->hwmr.local_wr = 0;
3046         mr->hwmr.mw_bind = 0;
3047         status = ocrdma_build_pbl_tbl(dev, &mr->hwmr);
3048         if (status)
3049                 goto pbl_err;
3050         status = ocrdma_reg_mr(dev, &mr->hwmr, pd->id, 0);
3051         if (status)
3052                 goto mbx_err;
3053         mr->ibmr.rkey = mr->hwmr.lkey;
3054         mr->ibmr.lkey = mr->hwmr.lkey;
3055         dev->stag_arr[(mr->hwmr.lkey >> 8) & (OCRDMA_MAX_STAG - 1)] =
3056                 (unsigned long) mr;
3057         return &mr->ibmr;
3058 mbx_err:
3059         ocrdma_free_mr_pbl_tbl(dev, &mr->hwmr);
3060 pbl_err:
3061         kfree(mr->pages);
3062 pl_err:
3063         kfree(mr);
3064         return ERR_PTR(-ENOMEM);
3065 }
3066
3067 static int ocrdma_set_page(struct ib_mr *ibmr, u64 addr)
3068 {
3069         struct ocrdma_mr *mr = get_ocrdma_mr(ibmr);
3070
3071         if (unlikely(mr->npages == mr->hwmr.num_pbes))
3072                 return -ENOMEM;
3073
3074         mr->pages[mr->npages++] = addr;
3075
3076         return 0;
3077 }
3078
3079 int ocrdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
3080                      unsigned int *sg_offset)
3081 {
3082         struct ocrdma_mr *mr = get_ocrdma_mr(ibmr);
3083
3084         mr->npages = 0;
3085
3086         return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, ocrdma_set_page);
3087 }