drivers/infiniband/hw/irdma/verbs.c

   1 // SPDX-License-Identifier: GPL-2.0 or Linux-OpenIB
   2 /* Copyright (c) 2015 - 2021 Intel Corporation */
   3 #include "main.h"
   4
   5 /**
   6  * irdma_query_device - get device attributes
   7  * @ibdev: device pointer from stack
   8  * @props: returning device attributes
   9  * @udata: user data
  10  */
  11 static int irdma_query_device(struct ib_device *ibdev,
  12                               struct ib_device_attr *props,
  13                               struct ib_udata *udata)
  14 {
  15         struct irdma_device *iwdev = to_iwdev(ibdev);
  16         struct irdma_pci_f *rf = iwdev->rf;
  17         struct pci_dev *pcidev = iwdev->rf->pcidev;
  18         struct irdma_hw_attrs *hw_attrs = &rf->sc_dev.hw_attrs;
  19
  20         if (udata->inlen || udata->outlen)
  21                 return -EINVAL;
  22
  23         memset(props, 0, sizeof(*props));
  24         addrconf_addr_eui48((u8 *)&props->sys_image_guid,
  25                             iwdev->netdev->dev_addr);
  26         props->fw_ver = (u64)irdma_fw_major_ver(&rf->sc_dev) << 32 |
  27                         irdma_fw_minor_ver(&rf->sc_dev);
  28         props->device_cap_flags = IB_DEVICE_MEM_WINDOW |
  29                                   IB_DEVICE_MEM_MGT_EXTENSIONS;
  30         props->kernel_cap_flags = IBK_LOCAL_DMA_LKEY;
  31         props->vendor_id = pcidev->vendor;
  32         props->vendor_part_id = pcidev->device;
  33
  34         props->hw_ver = rf->pcidev->revision;
  35         props->page_size_cap = hw_attrs->page_size_cap;
  36         props->max_mr_size = hw_attrs->max_mr_size;
  37         props->max_qp = rf->max_qp - rf->used_qps;
  38         props->max_qp_wr = hw_attrs->max_qp_wr;
  39         props->max_send_sge = hw_attrs->uk_attrs.max_hw_wq_frags;
  40         props->max_recv_sge = hw_attrs->uk_attrs.max_hw_wq_frags;
  41         props->max_cq = rf->max_cq - rf->used_cqs;
  42         props->max_cqe = rf->max_cqe;
  43         props->max_mr = rf->max_mr - rf->used_mrs;
  44         props->max_mw = props->max_mr;
  45         props->max_pd = rf->max_pd - rf->used_pds;
  46         props->max_sge_rd = hw_attrs->uk_attrs.max_hw_read_sges;
  47         props->max_qp_rd_atom = hw_attrs->max_hw_ird;
  48         props->max_qp_init_rd_atom = hw_attrs->max_hw_ord;
  49         if (rdma_protocol_roce(ibdev, 1))
  50                 props->max_pkeys = IRDMA_PKEY_TBL_SZ;
  51         props->max_ah = rf->max_ah;
  52         props->max_mcast_grp = rf->max_mcg;
  53         props->max_mcast_qp_attach = IRDMA_MAX_MGS_PER_CTX;
  54         props->max_total_mcast_qp_attach = rf->max_qp * IRDMA_MAX_MGS_PER_CTX;
  55         props->max_fast_reg_page_list_len = IRDMA_MAX_PAGES_PER_FMR;
  56 #define HCA_CLOCK_TIMESTAMP_MASK 0x1ffff
  57         if (hw_attrs->uk_attrs.hw_rev >= IRDMA_GEN_2)
  58                 props->timestamp_mask = HCA_CLOCK_TIMESTAMP_MASK;
  59
  60         return 0;
  61 }
  62
  63 /**
  64  * irdma_get_eth_speed_and_width - Get IB port speed and width from netdev speed
  65  * @link_speed: netdev phy link speed
  66  * @active_speed: IB port speed
  67  * @active_width: IB port width
  68  */
  69 static void irdma_get_eth_speed_and_width(u32 link_speed, u16 *active_speed,
  70                                           u8 *active_width)
  71 {
  72         if (link_speed <= SPEED_1000) {
  73                 *active_width = IB_WIDTH_1X;
  74                 *active_speed = IB_SPEED_SDR;
  75         } else if (link_speed <= SPEED_10000) {
  76                 *active_width = IB_WIDTH_1X;
  77                 *active_speed = IB_SPEED_FDR10;
  78         } else if (link_speed <= SPEED_20000) {
  79                 *active_width = IB_WIDTH_4X;
  80                 *active_speed = IB_SPEED_DDR;
  81         } else if (link_speed <= SPEED_25000) {
  82                 *active_width = IB_WIDTH_1X;
  83                 *active_speed = IB_SPEED_EDR;
  84         } else if (link_speed <= SPEED_40000) {
  85                 *active_width = IB_WIDTH_4X;
  86                 *active_speed = IB_SPEED_FDR10;
  87         } else {
  88                 *active_width = IB_WIDTH_4X;
  89                 *active_speed = IB_SPEED_EDR;
  90         }
  91 }
  92
  93 /**
  94  * irdma_query_port - get port attributes
  95  * @ibdev: device pointer from stack
  96  * @port: port number for query
  97  * @props: returning device attributes
  98  */
  99 static int irdma_query_port(struct ib_device *ibdev, u32 port,
 100                             struct ib_port_attr *props)
 101 {
 102         struct irdma_device *iwdev = to_iwdev(ibdev);
 103         struct net_device *netdev = iwdev->netdev;
 104
 105         /* no need to zero out pros here. done by caller */
 106
 107         props->max_mtu = IB_MTU_4096;
 108         props->active_mtu = ib_mtu_int_to_enum(netdev->mtu);
 109         props->lid = 1;
 110         props->lmc = 0;
 111         props->sm_lid = 0;
 112         props->sm_sl = 0;
 113         if (netif_carrier_ok(netdev) && netif_running(netdev)) {
 114                 props->state = IB_PORT_ACTIVE;
 115                 props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
 116         } else {
 117                 props->state = IB_PORT_DOWN;
 118                 props->phys_state = IB_PORT_PHYS_STATE_DISABLED;
 119         }
 120         irdma_get_eth_speed_and_width(SPEED_100000, &props->active_speed,
 121                                       &props->active_width);
 122
 123         if (rdma_protocol_roce(ibdev, 1)) {
 124                 props->gid_tbl_len = 32;
 125                 props->ip_gids = true;
 126                 props->pkey_tbl_len = IRDMA_PKEY_TBL_SZ;
 127         } else {
 128                 props->gid_tbl_len = 1;
 129         }
 130         props->qkey_viol_cntr = 0;
 131         props->port_cap_flags |= IB_PORT_CM_SUP | IB_PORT_REINIT_SUP;
 132         props->max_msg_sz = iwdev->rf->sc_dev.hw_attrs.max_hw_outbound_msg_size;
 133
 134         return 0;
 135 }
 136
 137 /**
 138  * irdma_disassociate_ucontext - Disassociate user context
 139  * @context: ib user context
 140  */
 141 static void irdma_disassociate_ucontext(struct ib_ucontext *context)
 142 {
 143 }
 144
 145 static int irdma_mmap_legacy(struct irdma_ucontext *ucontext,
 146                              struct vm_area_struct *vma)
 147 {
 148         u64 pfn;
 149
 150         if (vma->vm_pgoff || vma->vm_end - vma->vm_start != PAGE_SIZE)
 151                 return -EINVAL;
 152
 153         vma->vm_private_data = ucontext;
 154         pfn = ((uintptr_t)ucontext->iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET] +
 155                pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;
 156
 157         return rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn, PAGE_SIZE,
 158                                  pgprot_noncached(vma->vm_page_prot), NULL);
 159 }
 160
 161 static void irdma_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
 162 {
 163         struct irdma_user_mmap_entry *entry = to_irdma_mmap_entry(rdma_entry);
 164
 165         kfree(entry);
 166 }
 167
 168 static struct rdma_user_mmap_entry*
 169 irdma_user_mmap_entry_insert(struct irdma_ucontext *ucontext, u64 bar_offset,
 170                              enum irdma_mmap_flag mmap_flag, u64 *mmap_offset)
 171 {
 172         struct irdma_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 173         int ret;
 174
 175         if (!entry)
 176                 return NULL;
 177
 178         entry->bar_offset = bar_offset;
 179         entry->mmap_flag = mmap_flag;
 180
 181         ret = rdma_user_mmap_entry_insert(&ucontext->ibucontext,
 182                                           &entry->rdma_entry, PAGE_SIZE);
 183         if (ret) {
 184                 kfree(entry);
 185                 return NULL;
 186         }
 187         *mmap_offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
 188
 189         return &entry->rdma_entry;
 190 }
 191
 192 /**
 193  * irdma_mmap - user memory map
 194  * @context: context created during alloc
 195  * @vma: kernel info for user memory map
 196  */
 197 static int irdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
 198 {
 199         struct rdma_user_mmap_entry *rdma_entry;
 200         struct irdma_user_mmap_entry *entry;
 201         struct irdma_ucontext *ucontext;
 202         u64 pfn;
 203         int ret;
 204
 205         ucontext = to_ucontext(context);
 206
 207         /* Legacy support for libi40iw with hard-coded mmap key */
 208         if (ucontext->legacy_mode)
 209                 return irdma_mmap_legacy(ucontext, vma);
 210
 211         rdma_entry = rdma_user_mmap_entry_get(&ucontext->ibucontext, vma);
 212         if (!rdma_entry) {
 213                 ibdev_dbg(&ucontext->iwdev->ibdev,
 214                           "VERBS: pgoff[0x%lx] does not have valid entry\n",
 215                           vma->vm_pgoff);
 216                 return -EINVAL;
 217         }
 218
 219         entry = to_irdma_mmap_entry(rdma_entry);
 220         ibdev_dbg(&ucontext->iwdev->ibdev,
 221                   "VERBS: bar_offset [0x%llx] mmap_flag [%d]\n",
 222                   entry->bar_offset, entry->mmap_flag);
 223
 224         pfn = (entry->bar_offset +
 225               pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;
 226
 227         switch (entry->mmap_flag) {
 228         case IRDMA_MMAP_IO_NC:
 229                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
 230                                         pgprot_noncached(vma->vm_page_prot),
 231                                         rdma_entry);
 232                 break;
 233         case IRDMA_MMAP_IO_WC:
 234                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
 235                                         pgprot_writecombine(vma->vm_page_prot),
 236                                         rdma_entry);
 237                 break;
 238         default:
 239                 ret = -EINVAL;
 240         }
 241
 242         if (ret)
 243                 ibdev_dbg(&ucontext->iwdev->ibdev,
 244                           "VERBS: bar_offset [0x%llx] mmap_flag[%d] err[%d]\n",
 245                           entry->bar_offset, entry->mmap_flag, ret);
 246         rdma_user_mmap_entry_put(rdma_entry);
 247
 248         return ret;
 249 }
 250
 251 /**
 252  * irdma_alloc_push_page - allocate a push page for qp
 253  * @iwqp: qp pointer
 254  */
 255 static void irdma_alloc_push_page(struct irdma_qp *iwqp)
 256 {
 257         struct irdma_cqp_request *cqp_request;
 258         struct cqp_cmds_info *cqp_info;
 259         struct irdma_device *iwdev = iwqp->iwdev;
 260         struct irdma_sc_qp *qp = &iwqp->sc_qp;
 261         int status;
 262
 263         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
 264         if (!cqp_request)
 265                 return;
 266
 267         cqp_info = &cqp_request->info;
 268         cqp_info->cqp_cmd = IRDMA_OP_MANAGE_PUSH_PAGE;
 269         cqp_info->post_sq = 1;
 270         cqp_info->in.u.manage_push_page.info.push_idx = 0;
 271         cqp_info->in.u.manage_push_page.info.qs_handle =
 272                 qp->vsi->qos[qp->user_pri].qs_handle;
 273         cqp_info->in.u.manage_push_page.info.free_page = 0;
 274         cqp_info->in.u.manage_push_page.info.push_page_type = 0;
 275         cqp_info->in.u.manage_push_page.cqp = &iwdev->rf->cqp.sc_cqp;
 276         cqp_info->in.u.manage_push_page.scratch = (uintptr_t)cqp_request;
 277
 278         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
 279         if (!status && cqp_request->compl_info.op_ret_val <
 280             iwdev->rf->sc_dev.hw_attrs.max_hw_device_pages) {
 281                 qp->push_idx = cqp_request->compl_info.op_ret_val;
 282                 qp->push_offset = 0;
 283         }
 284
 285         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
 286 }
 287
 288 /**
 289  * irdma_alloc_ucontext - Allocate the user context data structure
 290  * @uctx: uverbs context pointer
 291  * @udata: user data
 292  *
 293  * This keeps track of all objects associated with a particular
 294  * user-mode client.
 295  */
 296 static int irdma_alloc_ucontext(struct ib_ucontext *uctx,
 297                                 struct ib_udata *udata)
 298 {
 299         struct ib_device *ibdev = uctx->device;
 300         struct irdma_device *iwdev = to_iwdev(ibdev);
 301         struct irdma_alloc_ucontext_req req;
 302         struct irdma_alloc_ucontext_resp uresp = {};
 303         struct irdma_ucontext *ucontext = to_ucontext(uctx);
 304         struct irdma_uk_attrs *uk_attrs;
 305
 306         if (ib_copy_from_udata(&req, udata, min(sizeof(req), udata->inlen)))
 307                 return -EINVAL;
 308
 309         if (req.userspace_ver < 4 || req.userspace_ver > IRDMA_ABI_VER)
 310                 goto ver_error;
 311
 312         ucontext->iwdev = iwdev;
 313         ucontext->abi_ver = req.userspace_ver;
 314
 315         uk_attrs = &iwdev->rf->sc_dev.hw_attrs.uk_attrs;
 316         /* GEN_1 legacy support with libi40iw */
 317         if (udata->outlen < sizeof(uresp)) {
 318                 if (uk_attrs->hw_rev != IRDMA_GEN_1)
 319                         return -EOPNOTSUPP;
 320
 321                 ucontext->legacy_mode = true;
 322                 uresp.max_qps = iwdev->rf->max_qp;
 323                 uresp.max_pds = iwdev->rf->sc_dev.hw_attrs.max_hw_pds;
 324                 uresp.wq_size = iwdev->rf->sc_dev.hw_attrs.max_qp_wr * 2;
 325                 uresp.kernel_ver = req.userspace_ver;
 326                 if (ib_copy_to_udata(udata, &uresp,
 327                                      min(sizeof(uresp), udata->outlen)))
 328                         return -EFAULT;
 329         } else {
 330                 u64 bar_off = (uintptr_t)iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET];
 331
 332                 ucontext->db_mmap_entry =
 333                         irdma_user_mmap_entry_insert(ucontext, bar_off,
 334                                                      IRDMA_MMAP_IO_NC,
 335                                                      &uresp.db_mmap_key);
 336                 if (!ucontext->db_mmap_entry)
 337                         return -ENOMEM;
 338
 339                 uresp.kernel_ver = IRDMA_ABI_VER;
 340                 uresp.feature_flags = uk_attrs->feature_flags;
 341                 uresp.max_hw_wq_frags = uk_attrs->max_hw_wq_frags;
 342                 uresp.max_hw_read_sges = uk_attrs->max_hw_read_sges;
 343                 uresp.max_hw_inline = uk_attrs->max_hw_inline;
 344                 uresp.max_hw_rq_quanta = uk_attrs->max_hw_rq_quanta;
 345                 uresp.max_hw_wq_quanta = uk_attrs->max_hw_wq_quanta;
 346                 uresp.max_hw_sq_chunk = uk_attrs->max_hw_sq_chunk;
 347                 uresp.max_hw_cq_size = uk_attrs->max_hw_cq_size;
 348                 uresp.min_hw_cq_size = uk_attrs->min_hw_cq_size;
 349                 uresp.hw_rev = uk_attrs->hw_rev;
 350                 if (ib_copy_to_udata(udata, &uresp,
 351                                      min(sizeof(uresp), udata->outlen))) {
 352                         rdma_user_mmap_entry_remove(ucontext->db_mmap_entry);
 353                         return -EFAULT;
 354                 }
 355         }
 356
 357         INIT_LIST_HEAD(&ucontext->cq_reg_mem_list);
 358         spin_lock_init(&ucontext->cq_reg_mem_list_lock);
 359         INIT_LIST_HEAD(&ucontext->qp_reg_mem_list);
 360         spin_lock_init(&ucontext->qp_reg_mem_list_lock);
 361
 362         return 0;
 363
 364 ver_error:
 365         ibdev_err(&iwdev->ibdev,
 366                   "Invalid userspace driver version detected. Detected version %d, should be %d\n",
 367                   req.userspace_ver, IRDMA_ABI_VER);
 368         return -EINVAL;
 369 }
 370
 371 /**
 372  * irdma_dealloc_ucontext - deallocate the user context data structure
 373  * @context: user context created during alloc
 374  */
 375 static void irdma_dealloc_ucontext(struct ib_ucontext *context)
 376 {
 377         struct irdma_ucontext *ucontext = to_ucontext(context);
 378
 379         rdma_user_mmap_entry_remove(ucontext->db_mmap_entry);
 380 }
 381
 382 /**
 383  * irdma_alloc_pd - allocate protection domain
 384  * @pd: PD pointer
 385  * @udata: user data
 386  */
 387 static int irdma_alloc_pd(struct ib_pd *pd, struct ib_udata *udata)
 388 {
 389         struct irdma_pd *iwpd = to_iwpd(pd);
 390         struct irdma_device *iwdev = to_iwdev(pd->device);
 391         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 392         struct irdma_pci_f *rf = iwdev->rf;
 393         struct irdma_alloc_pd_resp uresp = {};
 394         struct irdma_sc_pd *sc_pd;
 395         u32 pd_id = 0;
 396         int err;
 397
 398         err = irdma_alloc_rsrc(rf, rf->allocated_pds, rf->max_pd, &pd_id,
 399                                &rf->next_pd);
 400         if (err)
 401                 return err;
 402
 403         sc_pd = &iwpd->sc_pd;
 404         if (udata) {
 405                 struct irdma_ucontext *ucontext =
 406                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
 407                                                   ibucontext);
 408                 irdma_sc_pd_init(dev, sc_pd, pd_id, ucontext->abi_ver);
 409                 uresp.pd_id = pd_id;
 410                 if (ib_copy_to_udata(udata, &uresp,
 411                                      min(sizeof(uresp), udata->outlen))) {
 412                         err = -EFAULT;
 413                         goto error;
 414                 }
 415         } else {
 416                 irdma_sc_pd_init(dev, sc_pd, pd_id, IRDMA_ABI_VER);
 417         }
 418
 419         return 0;
 420 error:
 421         irdma_free_rsrc(rf, rf->allocated_pds, pd_id);
 422
 423         return err;
 424 }
 425
 426 /**
 427  * irdma_dealloc_pd - deallocate pd
 428  * @ibpd: ptr of pd to be deallocated
 429  * @udata: user data
 430  */
 431 static int irdma_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 432 {
 433         struct irdma_pd *iwpd = to_iwpd(ibpd);
 434         struct irdma_device *iwdev = to_iwdev(ibpd->device);
 435
 436         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_pds, iwpd->sc_pd.pd_id);
 437
 438         return 0;
 439 }
 440
 441 /**
 442  * irdma_get_pbl - Retrieve pbl from a list given a virtual
 443  * address
 444  * @va: user virtual address
 445  * @pbl_list: pbl list to search in (QP's or CQ's)
 446  */
 447 static struct irdma_pbl *irdma_get_pbl(unsigned long va,
 448                                        struct list_head *pbl_list)
 449 {
 450         struct irdma_pbl *iwpbl;
 451
 452         list_for_each_entry (iwpbl, pbl_list, list) {
 453                 if (iwpbl->user_base == va) {
 454                         list_del(&iwpbl->list);
 455                         iwpbl->on_list = false;
 456                         return iwpbl;
 457                 }
 458         }
 459
 460         return NULL;
 461 }
 462
 463 /**
 464  * irdma_clean_cqes - clean cq entries for qp
 465  * @iwqp: qp ptr (user or kernel)
 466  * @iwcq: cq ptr
 467  */
 468 static void irdma_clean_cqes(struct irdma_qp *iwqp, struct irdma_cq *iwcq)
 469 {
 470         struct irdma_cq_uk *ukcq = &iwcq->sc_cq.cq_uk;
 471         unsigned long flags;
 472
 473         spin_lock_irqsave(&iwcq->lock, flags);
 474         irdma_uk_clean_cq(&iwqp->sc_qp.qp_uk, ukcq);
 475         spin_unlock_irqrestore(&iwcq->lock, flags);
 476 }
 477
 478 static void irdma_remove_push_mmap_entries(struct irdma_qp *iwqp)
 479 {
 480         if (iwqp->push_db_mmap_entry) {
 481                 rdma_user_mmap_entry_remove(iwqp->push_db_mmap_entry);
 482                 iwqp->push_db_mmap_entry = NULL;
 483         }
 484         if (iwqp->push_wqe_mmap_entry) {
 485                 rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
 486                 iwqp->push_wqe_mmap_entry = NULL;
 487         }
 488 }
 489
 490 static int irdma_setup_push_mmap_entries(struct irdma_ucontext *ucontext,
 491                                          struct irdma_qp *iwqp,
 492                                          u64 *push_wqe_mmap_key,
 493                                          u64 *push_db_mmap_key)
 494 {
 495         struct irdma_device *iwdev = ucontext->iwdev;
 496         u64 rsvd, bar_off;
 497
 498         rsvd = IRDMA_PF_BAR_RSVD;
 499         bar_off = (uintptr_t)iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET];
 500         /* skip over db page */
 501         bar_off += IRDMA_HW_PAGE_SIZE;
 502         /* push wqe page */
 503         bar_off += rsvd + iwqp->sc_qp.push_idx * IRDMA_HW_PAGE_SIZE;
 504         iwqp->push_wqe_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
 505                                         bar_off, IRDMA_MMAP_IO_WC,
 506                                         push_wqe_mmap_key);
 507         if (!iwqp->push_wqe_mmap_entry)
 508                 return -ENOMEM;
 509
 510         /* push doorbell page */
 511         bar_off += IRDMA_HW_PAGE_SIZE;
 512         iwqp->push_db_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
 513                                         bar_off, IRDMA_MMAP_IO_NC,
 514                                         push_db_mmap_key);
 515         if (!iwqp->push_db_mmap_entry) {
 516                 rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
 517                 return -ENOMEM;
 518         }
 519
 520         return 0;
 521 }
 522
 523 /**
 524  * irdma_destroy_qp - destroy qp
 525  * @ibqp: qp's ib pointer also to get to device's qp address
 526  * @udata: user data
 527  */
 528 static int irdma_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
 529 {
 530         struct irdma_qp *iwqp = to_iwqp(ibqp);
 531         struct irdma_device *iwdev = iwqp->iwdev;
 532
 533         iwqp->sc_qp.qp_uk.destroy_pending = true;
 534
 535         if (iwqp->iwarp_state == IRDMA_QP_STATE_RTS)
 536                 irdma_modify_qp_to_err(&iwqp->sc_qp);
 537
 538         if (!iwqp->user_mode)
 539                 cancel_delayed_work_sync(&iwqp->dwork_flush);
 540
 541         irdma_qp_rem_ref(&iwqp->ibqp);
 542         wait_for_completion(&iwqp->free_qp);
 543         irdma_free_lsmm_rsrc(iwqp);
 544         irdma_cqp_qp_destroy_cmd(&iwdev->rf->sc_dev, &iwqp->sc_qp);
 545
 546         if (!iwqp->user_mode) {
 547                 if (iwqp->iwscq) {
 548                         irdma_clean_cqes(iwqp, iwqp->iwscq);
 549                         if (iwqp->iwrcq != iwqp->iwscq)
 550                                 irdma_clean_cqes(iwqp, iwqp->iwrcq);
 551                 }
 552         }
 553         irdma_remove_push_mmap_entries(iwqp);
 554         irdma_free_qp_rsrc(iwqp);
 555
 556         return 0;
 557 }
 558
 559 /**
 560  * irdma_setup_virt_qp - setup for allocation of virtual qp
 561  * @iwdev: irdma device
 562  * @iwqp: qp ptr
 563  * @init_info: initialize info to return
 564  */
 565 static void irdma_setup_virt_qp(struct irdma_device *iwdev,
 566                                struct irdma_qp *iwqp,
 567                                struct irdma_qp_init_info *init_info)
 568 {
 569         struct irdma_pbl *iwpbl = iwqp->iwpbl;
 570         struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;
 571
 572         iwqp->page = qpmr->sq_page;
 573         init_info->shadow_area_pa = qpmr->shadow;
 574         if (iwpbl->pbl_allocated) {
 575                 init_info->virtual_map = true;
 576                 init_info->sq_pa = qpmr->sq_pbl.idx;
 577                 init_info->rq_pa = qpmr->rq_pbl.idx;
 578         } else {
 579                 init_info->sq_pa = qpmr->sq_pbl.addr;
 580                 init_info->rq_pa = qpmr->rq_pbl.addr;
 581         }
 582 }
 583
 584 /**
 585  * irdma_setup_kmode_qp - setup initialization for kernel mode qp
 586  * @iwdev: iwarp device
 587  * @iwqp: qp ptr (user or kernel)
 588  * @info: initialize info to return
 589  * @init_attr: Initial QP create attributes
 590  */
 591 static int irdma_setup_kmode_qp(struct irdma_device *iwdev,
 592                                 struct irdma_qp *iwqp,
 593                                 struct irdma_qp_init_info *info,
 594                                 struct ib_qp_init_attr *init_attr)
 595 {
 596         struct irdma_dma_mem *mem = &iwqp->kqp.dma_mem;
 597         u32 sqdepth, rqdepth;
 598         u8 sqshift, rqshift;
 599         u32 size;
 600         int status;
 601         struct irdma_qp_uk_init_info *ukinfo = &info->qp_uk_init_info;
 602         struct irdma_uk_attrs *uk_attrs = &iwdev->rf->sc_dev.hw_attrs.uk_attrs;
 603
 604         irdma_get_wqe_shift(uk_attrs,
 605                 uk_attrs->hw_rev >= IRDMA_GEN_2 ? ukinfo->max_sq_frag_cnt + 1 :
 606                                                   ukinfo->max_sq_frag_cnt,
 607                 ukinfo->max_inline_data, &sqshift);
 608         status = irdma_get_sqdepth(uk_attrs, ukinfo->sq_size, sqshift,
 609                                    &sqdepth);
 610         if (status)
 611                 return status;
 612
 613         if (uk_attrs->hw_rev == IRDMA_GEN_1)
 614                 rqshift = IRDMA_MAX_RQ_WQE_SHIFT_GEN1;
 615         else
 616                 irdma_get_wqe_shift(uk_attrs, ukinfo->max_rq_frag_cnt, 0,
 617                                     &rqshift);
 618
 619         status = irdma_get_rqdepth(uk_attrs, ukinfo->rq_size, rqshift,
 620                                    &rqdepth);
 621         if (status)
 622                 return status;
 623
 624         iwqp->kqp.sq_wrid_mem =
 625                 kcalloc(sqdepth, sizeof(*iwqp->kqp.sq_wrid_mem), GFP_KERNEL);
 626         if (!iwqp->kqp.sq_wrid_mem)
 627                 return -ENOMEM;
 628
 629         iwqp->kqp.rq_wrid_mem =
 630                 kcalloc(rqdepth, sizeof(*iwqp->kqp.rq_wrid_mem), GFP_KERNEL);
 631         if (!iwqp->kqp.rq_wrid_mem) {
 632                 kfree(iwqp->kqp.sq_wrid_mem);
 633                 iwqp->kqp.sq_wrid_mem = NULL;
 634                 return -ENOMEM;
 635         }
 636
 637         ukinfo->sq_wrtrk_array = iwqp->kqp.sq_wrid_mem;
 638         ukinfo->rq_wrid_array = iwqp->kqp.rq_wrid_mem;
 639
 640         size = (sqdepth + rqdepth) * IRDMA_QP_WQE_MIN_SIZE;
 641         size += (IRDMA_SHADOW_AREA_SIZE << 3);
 642
 643         mem->size = ALIGN(size, 256);
 644         mem->va = dma_alloc_coherent(iwdev->rf->hw.device, mem->size,
 645                                      &mem->pa, GFP_KERNEL);
 646         if (!mem->va) {
 647                 kfree(iwqp->kqp.sq_wrid_mem);
 648                 iwqp->kqp.sq_wrid_mem = NULL;
 649                 kfree(iwqp->kqp.rq_wrid_mem);
 650                 iwqp->kqp.rq_wrid_mem = NULL;
 651                 return -ENOMEM;
 652         }
 653
 654         ukinfo->sq = mem->va;
 655         info->sq_pa = mem->pa;
 656         ukinfo->rq = &ukinfo->sq[sqdepth];
 657         info->rq_pa = info->sq_pa + (sqdepth * IRDMA_QP_WQE_MIN_SIZE);
 658         ukinfo->shadow_area = ukinfo->rq[rqdepth].elem;
 659         info->shadow_area_pa = info->rq_pa + (rqdepth * IRDMA_QP_WQE_MIN_SIZE);
 660         ukinfo->sq_size = sqdepth >> sqshift;
 661         ukinfo->rq_size = rqdepth >> rqshift;
 662         ukinfo->qp_id = iwqp->ibqp.qp_num;
 663
 664         init_attr->cap.max_send_wr = (sqdepth - IRDMA_SQ_RSVD) >> sqshift;
 665         init_attr->cap.max_recv_wr = (rqdepth - IRDMA_RQ_RSVD) >> rqshift;
 666
 667         return 0;
 668 }
 669
 670 static int irdma_cqp_create_qp_cmd(struct irdma_qp *iwqp)
 671 {
 672         struct irdma_pci_f *rf = iwqp->iwdev->rf;
 673         struct irdma_cqp_request *cqp_request;
 674         struct cqp_cmds_info *cqp_info;
 675         struct irdma_create_qp_info *qp_info;
 676         int status;
 677
 678         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
 679         if (!cqp_request)
 680                 return -ENOMEM;
 681
 682         cqp_info = &cqp_request->info;
 683         qp_info = &cqp_request->info.in.u.qp_create.info;
 684         memset(qp_info, 0, sizeof(*qp_info));
 685         qp_info->mac_valid = true;
 686         qp_info->cq_num_valid = true;
 687         qp_info->next_iwarp_state = IRDMA_QP_STATE_IDLE;
 688
 689         cqp_info->cqp_cmd = IRDMA_OP_QP_CREATE;
 690         cqp_info->post_sq = 1;
 691         cqp_info->in.u.qp_create.qp = &iwqp->sc_qp;
 692         cqp_info->in.u.qp_create.scratch = (uintptr_t)cqp_request;
 693         status = irdma_handle_cqp_op(rf, cqp_request);
 694         irdma_put_cqp_request(&rf->cqp, cqp_request);
 695
 696         return status;
 697 }
 698
 699 static void irdma_roce_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
 700                                                struct irdma_qp_host_ctx_info *ctx_info)
 701 {
 702         struct irdma_device *iwdev = iwqp->iwdev;
 703         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 704         struct irdma_roce_offload_info *roce_info;
 705         struct irdma_udp_offload_info *udp_info;
 706
 707         udp_info = &iwqp->udp_info;
 708         udp_info->snd_mss = ib_mtu_enum_to_int(ib_mtu_int_to_enum(iwdev->vsi.mtu));
 709         udp_info->cwnd = iwdev->roce_cwnd;
 710         udp_info->rexmit_thresh = 2;
 711         udp_info->rnr_nak_thresh = 2;
 712         udp_info->src_port = 0xc000;
 713         udp_info->dst_port = ROCE_V2_UDP_DPORT;
 714         roce_info = &iwqp->roce_info;
 715         ether_addr_copy(roce_info->mac_addr, iwdev->netdev->dev_addr);
 716
 717         roce_info->rd_en = true;
 718         roce_info->wr_rdresp_en = true;
 719         roce_info->bind_en = true;
 720         roce_info->dcqcn_en = false;
 721         roce_info->rtomin = 5;
 722
 723         roce_info->ack_credits = iwdev->roce_ackcreds;
 724         roce_info->ird_size = dev->hw_attrs.max_hw_ird;
 725         roce_info->ord_size = dev->hw_attrs.max_hw_ord;
 726
 727         if (!iwqp->user_mode) {
 728                 roce_info->priv_mode_en = true;
 729                 roce_info->fast_reg_en = true;
 730                 roce_info->udprivcq_en = true;
 731         }
 732         roce_info->roce_tver = 0;
 733
 734         ctx_info->roce_info = &iwqp->roce_info;
 735         ctx_info->udp_info = &iwqp->udp_info;
 736         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
 737 }
 738
 739 static void irdma_iw_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
 740                                              struct irdma_qp_host_ctx_info *ctx_info)
 741 {
 742         struct irdma_device *iwdev = iwqp->iwdev;
 743         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 744         struct irdma_iwarp_offload_info *iwarp_info;
 745
 746         iwarp_info = &iwqp->iwarp_info;
 747         ether_addr_copy(iwarp_info->mac_addr, iwdev->netdev->dev_addr);
 748         iwarp_info->rd_en = true;
 749         iwarp_info->wr_rdresp_en = true;
 750         iwarp_info->bind_en = true;
 751         iwarp_info->ecn_en = true;
 752         iwarp_info->rtomin = 5;
 753
 754         if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
 755                 iwarp_info->ib_rd_en = true;
 756         if (!iwqp->user_mode) {
 757                 iwarp_info->priv_mode_en = true;
 758                 iwarp_info->fast_reg_en = true;
 759         }
 760         iwarp_info->ddp_ver = 1;
 761         iwarp_info->rdmap_ver = 1;
 762
 763         ctx_info->iwarp_info = &iwqp->iwarp_info;
 764         ctx_info->iwarp_info_valid = true;
 765         irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
 766         ctx_info->iwarp_info_valid = false;
 767 }
 768
 769 static int irdma_validate_qp_attrs(struct ib_qp_init_attr *init_attr,
 770                                    struct irdma_device *iwdev)
 771 {
 772         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 773         struct irdma_uk_attrs *uk_attrs = &dev->hw_attrs.uk_attrs;
 774
 775         if (init_attr->create_flags)
 776                 return -EOPNOTSUPP;
 777
 778         if (init_attr->cap.max_inline_data > uk_attrs->max_hw_inline ||
 779             init_attr->cap.max_send_sge > uk_attrs->max_hw_wq_frags ||
 780             init_attr->cap.max_recv_sge > uk_attrs->max_hw_wq_frags)
 781                 return -EINVAL;
 782
 783         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 784                 if (init_attr->qp_type != IB_QPT_RC &&
 785                     init_attr->qp_type != IB_QPT_UD &&
 786                     init_attr->qp_type != IB_QPT_GSI)
 787                         return -EOPNOTSUPP;
 788         } else {
 789                 if (init_attr->qp_type != IB_QPT_RC)
 790                         return -EOPNOTSUPP;
 791         }
 792
 793         return 0;
 794 }
 795
 796 static void irdma_flush_worker(struct work_struct *work)
 797 {
 798         struct delayed_work *dwork = to_delayed_work(work);
 799         struct irdma_qp *iwqp = container_of(dwork, struct irdma_qp, dwork_flush);
 800
 801         irdma_generate_flush_completions(iwqp);
 802 }
 803
 804 /**
 805  * irdma_create_qp - create qp
 806  * @ibqp: ptr of qp
 807  * @init_attr: attributes for qp
 808  * @udata: user data for create qp
 809  */
 810 static int irdma_create_qp(struct ib_qp *ibqp,
 811                            struct ib_qp_init_attr *init_attr,
 812                            struct ib_udata *udata)
 813 {
 814         struct ib_pd *ibpd = ibqp->pd;
 815         struct irdma_pd *iwpd = to_iwpd(ibpd);
 816         struct irdma_device *iwdev = to_iwdev(ibpd->device);
 817         struct irdma_pci_f *rf = iwdev->rf;
 818         struct irdma_qp *iwqp = to_iwqp(ibqp);
 819         struct irdma_create_qp_req req;
 820         struct irdma_create_qp_resp uresp = {};
 821         u32 qp_num = 0;
 822         int err_code;
 823         int sq_size;
 824         int rq_size;
 825         struct irdma_sc_qp *qp;
 826         struct irdma_sc_dev *dev = &rf->sc_dev;
 827         struct irdma_uk_attrs *uk_attrs = &dev->hw_attrs.uk_attrs;
 828         struct irdma_qp_init_info init_info = {};
 829         struct irdma_qp_host_ctx_info *ctx_info;
 830         unsigned long flags;
 831
 832         err_code = irdma_validate_qp_attrs(init_attr, iwdev);
 833         if (err_code)
 834                 return err_code;
 835
 836         sq_size = init_attr->cap.max_send_wr;
 837         rq_size = init_attr->cap.max_recv_wr;
 838
 839         init_info.vsi = &iwdev->vsi;
 840         init_info.qp_uk_init_info.uk_attrs = uk_attrs;
 841         init_info.qp_uk_init_info.sq_size = sq_size;
 842         init_info.qp_uk_init_info.rq_size = rq_size;
 843         init_info.qp_uk_init_info.max_sq_frag_cnt = init_attr->cap.max_send_sge;
 844         init_info.qp_uk_init_info.max_rq_frag_cnt = init_attr->cap.max_recv_sge;
 845         init_info.qp_uk_init_info.max_inline_data = init_attr->cap.max_inline_data;
 846
 847         qp = &iwqp->sc_qp;
 848         qp->qp_uk.back_qp = iwqp;
 849         qp->push_idx = IRDMA_INVALID_PUSH_PAGE_INDEX;
 850
 851         iwqp->iwdev = iwdev;
 852         iwqp->q2_ctx_mem.size = ALIGN(IRDMA_Q2_BUF_SIZE + IRDMA_QP_CTX_SIZE,
 853                                       256);
 854         iwqp->q2_ctx_mem.va = dma_alloc_coherent(dev->hw->device,
 855                                                  iwqp->q2_ctx_mem.size,
 856                                                  &iwqp->q2_ctx_mem.pa,
 857                                                  GFP_KERNEL);
 858         if (!iwqp->q2_ctx_mem.va)
 859                 return -ENOMEM;
 860
 861         init_info.q2 = iwqp->q2_ctx_mem.va;
 862         init_info.q2_pa = iwqp->q2_ctx_mem.pa;
 863         init_info.host_ctx = (__le64 *)(init_info.q2 + IRDMA_Q2_BUF_SIZE);
 864         init_info.host_ctx_pa = init_info.q2_pa + IRDMA_Q2_BUF_SIZE;
 865
 866         if (init_attr->qp_type == IB_QPT_GSI)
 867                 qp_num = 1;
 868         else
 869                 err_code = irdma_alloc_rsrc(rf, rf->allocated_qps, rf->max_qp,
 870                                             &qp_num, &rf->next_qp);
 871         if (err_code)
 872                 goto error;
 873
 874         iwqp->iwpd = iwpd;
 875         iwqp->ibqp.qp_num = qp_num;
 876         qp = &iwqp->sc_qp;
 877         iwqp->iwscq = to_iwcq(init_attr->send_cq);
 878         iwqp->iwrcq = to_iwcq(init_attr->recv_cq);
 879         iwqp->host_ctx.va = init_info.host_ctx;
 880         iwqp->host_ctx.pa = init_info.host_ctx_pa;
 881         iwqp->host_ctx.size = IRDMA_QP_CTX_SIZE;
 882
 883         init_info.pd = &iwpd->sc_pd;
 884         init_info.qp_uk_init_info.qp_id = iwqp->ibqp.qp_num;
 885         if (!rdma_protocol_roce(&iwdev->ibdev, 1))
 886                 init_info.qp_uk_init_info.first_sq_wq = 1;
 887         iwqp->ctx_info.qp_compl_ctx = (uintptr_t)qp;
 888         init_waitqueue_head(&iwqp->waitq);
 889         init_waitqueue_head(&iwqp->mod_qp_waitq);
 890
 891         if (udata) {
 892                 err_code = ib_copy_from_udata(&req, udata,
 893                                               min(sizeof(req), udata->inlen));
 894                 if (err_code) {
 895                         ibdev_dbg(&iwdev->ibdev,
 896                                   "VERBS: ib_copy_from_data fail\n");
 897                         goto error;
 898                 }
 899
 900                 iwqp->ctx_info.qp_compl_ctx = req.user_compl_ctx;
 901                 iwqp->user_mode = 1;
 902                 if (req.user_wqe_bufs) {
 903                         struct irdma_ucontext *ucontext =
 904                                 rdma_udata_to_drv_context(udata,
 905                                                           struct irdma_ucontext,
 906                                                           ibucontext);
 907
 908                         init_info.qp_uk_init_info.legacy_mode = ucontext->legacy_mode;
 909                         spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
 910                         iwqp->iwpbl = irdma_get_pbl((unsigned long)req.user_wqe_bufs,
 911                                                     &ucontext->qp_reg_mem_list);
 912                         spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
 913
 914                         if (!iwqp->iwpbl) {
 915                                 err_code = -ENODATA;
 916                                 ibdev_dbg(&iwdev->ibdev, "VERBS: no pbl info\n");
 917                                 goto error;
 918                         }
 919                 }
 920                 init_info.qp_uk_init_info.abi_ver = iwpd->sc_pd.abi_ver;
 921                 irdma_setup_virt_qp(iwdev, iwqp, &init_info);
 922         } else {
 923                 INIT_DELAYED_WORK(&iwqp->dwork_flush, irdma_flush_worker);
 924                 init_info.qp_uk_init_info.abi_ver = IRDMA_ABI_VER;
 925                 err_code = irdma_setup_kmode_qp(iwdev, iwqp, &init_info, init_attr);
 926         }
 927
 928         if (err_code) {
 929                 ibdev_dbg(&iwdev->ibdev, "VERBS: setup qp failed\n");
 930                 goto error;
 931         }
 932
 933         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 934                 if (init_attr->qp_type == IB_QPT_RC) {
 935                         init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_ROCE_RC;
 936                         init_info.qp_uk_init_info.qp_caps = IRDMA_SEND_WITH_IMM |
 937                                                             IRDMA_WRITE_WITH_IMM |
 938                                                             IRDMA_ROCE;
 939                 } else {
 940                         init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_ROCE_UD;
 941                         init_info.qp_uk_init_info.qp_caps = IRDMA_SEND_WITH_IMM |
 942                                                             IRDMA_ROCE;
 943                 }
 944         } else {
 945                 init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_IWARP;
 946                 init_info.qp_uk_init_info.qp_caps = IRDMA_WRITE_WITH_IMM;
 947         }
 948
 949         if (dev->hw_attrs.uk_attrs.hw_rev > IRDMA_GEN_1)
 950                 init_info.qp_uk_init_info.qp_caps |= IRDMA_PUSH_MODE;
 951
 952         err_code = irdma_sc_qp_init(qp, &init_info);
 953         if (err_code) {
 954                 ibdev_dbg(&iwdev->ibdev, "VERBS: qp_init fail\n");
 955                 goto error;
 956         }
 957
 958         ctx_info = &iwqp->ctx_info;
 959         ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
 960         ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
 961
 962         if (rdma_protocol_roce(&iwdev->ibdev, 1))
 963                 irdma_roce_fill_and_set_qpctx_info(iwqp, ctx_info);
 964         else
 965                 irdma_iw_fill_and_set_qpctx_info(iwqp, ctx_info);
 966
 967         err_code = irdma_cqp_create_qp_cmd(iwqp);
 968         if (err_code)
 969                 goto error;
 970
 971         refcount_set(&iwqp->refcnt, 1);
 972         spin_lock_init(&iwqp->lock);
 973         spin_lock_init(&iwqp->sc_qp.pfpdu.lock);
 974         iwqp->sig_all = (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ? 1 : 0;
 975         rf->qp_table[qp_num] = iwqp;
 976         iwqp->max_send_wr = sq_size;
 977         iwqp->max_recv_wr = rq_size;
 978
 979         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 980                 if (dev->ws_add(&iwdev->vsi, 0)) {
 981                         irdma_cqp_qp_destroy_cmd(&rf->sc_dev, &iwqp->sc_qp);
 982                         err_code = -EINVAL;
 983                         goto error;
 984                 }
 985
 986                 irdma_qp_add_qos(&iwqp->sc_qp);
 987         }
 988
 989         if (udata) {
 990                 /* GEN_1 legacy support with libi40iw does not have expanded uresp struct */
 991                 if (udata->outlen < sizeof(uresp)) {
 992                         uresp.lsmm = 1;
 993                         uresp.push_idx = IRDMA_INVALID_PUSH_PAGE_INDEX_GEN_1;
 994                 } else {
 995                         if (rdma_protocol_iwarp(&iwdev->ibdev, 1))
 996                                 uresp.lsmm = 1;
 997                 }
 998                 uresp.actual_sq_size = sq_size;
 999                 uresp.actual_rq_size = rq_size;
1000                 uresp.qp_id = qp_num;
1001                 uresp.qp_caps = qp->qp_uk.qp_caps;
1002
1003                 err_code = ib_copy_to_udata(udata, &uresp,
1004                                             min(sizeof(uresp), udata->outlen));
1005                 if (err_code) {
1006                         ibdev_dbg(&iwdev->ibdev, "VERBS: copy_to_udata failed\n");
1007                         irdma_destroy_qp(&iwqp->ibqp, udata);
1008                         return err_code;
1009                 }
1010         }
1011
1012         init_completion(&iwqp->free_qp);
1013         return 0;
1014
1015 error:
1016         irdma_free_qp_rsrc(iwqp);
1017         return err_code;
1018 }
1019
1020 static int irdma_get_ib_acc_flags(struct irdma_qp *iwqp)
1021 {
1022         int acc_flags = 0;
1023
1024         if (rdma_protocol_roce(iwqp->ibqp.device, 1)) {
1025                 if (iwqp->roce_info.wr_rdresp_en) {
1026                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
1027                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
1028                 }
1029                 if (iwqp->roce_info.rd_en)
1030                         acc_flags |= IB_ACCESS_REMOTE_READ;
1031                 if (iwqp->roce_info.bind_en)
1032                         acc_flags |= IB_ACCESS_MW_BIND;
1033         } else {
1034                 if (iwqp->iwarp_info.wr_rdresp_en) {
1035                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
1036                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
1037                 }
1038                 if (iwqp->iwarp_info.rd_en)
1039                         acc_flags |= IB_ACCESS_REMOTE_READ;
1040                 if (iwqp->iwarp_info.bind_en)
1041                         acc_flags |= IB_ACCESS_MW_BIND;
1042         }
1043         return acc_flags;
1044 }
1045
1046 /**
1047  * irdma_query_qp - query qp attributes
1048  * @ibqp: qp pointer
1049  * @attr: attributes pointer
1050  * @attr_mask: Not used
1051  * @init_attr: qp attributes to return
1052  */
1053 static int irdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1054                           int attr_mask, struct ib_qp_init_attr *init_attr)
1055 {
1056         struct irdma_qp *iwqp = to_iwqp(ibqp);
1057         struct irdma_sc_qp *qp = &iwqp->sc_qp;
1058
1059         memset(attr, 0, sizeof(*attr));
1060         memset(init_attr, 0, sizeof(*init_attr));
1061
1062         attr->qp_state = iwqp->ibqp_state;
1063         attr->cur_qp_state = iwqp->ibqp_state;
1064         attr->cap.max_send_wr = iwqp->max_send_wr;
1065         attr->cap.max_recv_wr = iwqp->max_recv_wr;
1066         attr->cap.max_inline_data = qp->qp_uk.max_inline_data;
1067         attr->cap.max_send_sge = qp->qp_uk.max_sq_frag_cnt;
1068         attr->cap.max_recv_sge = qp->qp_uk.max_rq_frag_cnt;
1069         attr->qp_access_flags = irdma_get_ib_acc_flags(iwqp);
1070         attr->port_num = 1;
1071         if (rdma_protocol_roce(ibqp->device, 1)) {
1072                 attr->path_mtu = ib_mtu_int_to_enum(iwqp->udp_info.snd_mss);
1073                 attr->qkey = iwqp->roce_info.qkey;
1074                 attr->rq_psn = iwqp->udp_info.epsn;
1075                 attr->sq_psn = iwqp->udp_info.psn_nxt;
1076                 attr->dest_qp_num = iwqp->roce_info.dest_qp;
1077                 attr->pkey_index = iwqp->roce_info.p_key;
1078                 attr->retry_cnt = iwqp->udp_info.rexmit_thresh;
1079                 attr->rnr_retry = iwqp->udp_info.rnr_nak_thresh;
1080                 attr->max_rd_atomic = iwqp->roce_info.ord_size;
1081                 attr->max_dest_rd_atomic = iwqp->roce_info.ird_size;
1082         }
1083
1084         init_attr->event_handler = iwqp->ibqp.event_handler;
1085         init_attr->qp_context = iwqp->ibqp.qp_context;
1086         init_attr->send_cq = iwqp->ibqp.send_cq;
1087         init_attr->recv_cq = iwqp->ibqp.recv_cq;
1088         init_attr->cap = attr->cap;
1089
1090         return 0;
1091 }
1092
1093 /**
1094  * irdma_query_pkey - Query partition key
1095  * @ibdev: device pointer from stack
1096  * @port: port number
1097  * @index: index of pkey
1098  * @pkey: pointer to store the pkey
1099  */
1100 static int irdma_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1101                             u16 *pkey)
1102 {
1103         if (index >= IRDMA_PKEY_TBL_SZ)
1104                 return -EINVAL;
1105
1106         *pkey = IRDMA_DEFAULT_PKEY;
1107         return 0;
1108 }
1109
1110 /**
1111  * irdma_modify_qp_roce - modify qp request
1112  * @ibqp: qp's pointer for modify
1113  * @attr: access attributes
1114  * @attr_mask: state mask
1115  * @udata: user data
1116  */
1117 int irdma_modify_qp_roce(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1118                          int attr_mask, struct ib_udata *udata)
1119 {
1120         struct irdma_pd *iwpd = to_iwpd(ibqp->pd);
1121         struct irdma_qp *iwqp = to_iwqp(ibqp);
1122         struct irdma_device *iwdev = iwqp->iwdev;
1123         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
1124         struct irdma_qp_host_ctx_info *ctx_info;
1125         struct irdma_roce_offload_info *roce_info;
1126         struct irdma_udp_offload_info *udp_info;
1127         struct irdma_modify_qp_info info = {};
1128         struct irdma_modify_qp_resp uresp = {};
1129         struct irdma_modify_qp_req ureq = {};
1130         unsigned long flags;
1131         u8 issue_modify_qp = 0;
1132         int ret = 0;
1133
1134         ctx_info = &iwqp->ctx_info;
1135         roce_info = &iwqp->roce_info;
1136         udp_info = &iwqp->udp_info;
1137
1138         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1139                 return -EOPNOTSUPP;
1140
1141         if (attr_mask & IB_QP_DEST_QPN)
1142                 roce_info->dest_qp = attr->dest_qp_num;
1143
1144         if (attr_mask & IB_QP_PKEY_INDEX) {
1145                 ret = irdma_query_pkey(ibqp->device, 0, attr->pkey_index,
1146                                        &roce_info->p_key);
1147                 if (ret)
1148                         return ret;
1149         }
1150
1151         if (attr_mask & IB_QP_QKEY)
1152                 roce_info->qkey = attr->qkey;
1153
1154         if (attr_mask & IB_QP_PATH_MTU)
1155                 udp_info->snd_mss = ib_mtu_enum_to_int(attr->path_mtu);
1156
1157         if (attr_mask & IB_QP_SQ_PSN) {
1158                 udp_info->psn_nxt = attr->sq_psn;
1159                 udp_info->lsn =  0xffff;
1160                 udp_info->psn_una = attr->sq_psn;
1161                 udp_info->psn_max = attr->sq_psn;
1162         }
1163
1164         if (attr_mask & IB_QP_RQ_PSN)
1165                 udp_info->epsn = attr->rq_psn;
1166
1167         if (attr_mask & IB_QP_RNR_RETRY)
1168                 udp_info->rnr_nak_thresh = attr->rnr_retry;
1169
1170         if (attr_mask & IB_QP_RETRY_CNT)
1171                 udp_info->rexmit_thresh = attr->retry_cnt;
1172
1173         ctx_info->roce_info->pd_id = iwpd->sc_pd.pd_id;
1174
1175         if (attr_mask & IB_QP_AV) {
1176                 struct irdma_av *av = &iwqp->roce_ah.av;
1177                 const struct ib_gid_attr *sgid_attr;
1178                 u16 vlan_id = VLAN_N_VID;
1179                 u32 local_ip[4];
1180
1181                 memset(&iwqp->roce_ah, 0, sizeof(iwqp->roce_ah));
1182                 if (attr->ah_attr.ah_flags & IB_AH_GRH) {
1183                         udp_info->ttl = attr->ah_attr.grh.hop_limit;
1184                         udp_info->flow_label = attr->ah_attr.grh.flow_label;
1185                         udp_info->tos = attr->ah_attr.grh.traffic_class;
1186                         udp_info->src_port =
1187                                 rdma_get_udp_sport(udp_info->flow_label,
1188                                                    ibqp->qp_num,
1189                                                    roce_info->dest_qp);
1190                         irdma_qp_rem_qos(&iwqp->sc_qp);
1191                         dev->ws_remove(iwqp->sc_qp.vsi, ctx_info->user_pri);
1192                         ctx_info->user_pri = rt_tos2priority(udp_info->tos);
1193                         iwqp->sc_qp.user_pri = ctx_info->user_pri;
1194                         if (dev->ws_add(iwqp->sc_qp.vsi, ctx_info->user_pri))
1195                                 return -ENOMEM;
1196                         irdma_qp_add_qos(&iwqp->sc_qp);
1197                 }
1198                 sgid_attr = attr->ah_attr.grh.sgid_attr;
1199                 ret = rdma_read_gid_l2_fields(sgid_attr, &vlan_id,
1200                                               ctx_info->roce_info->mac_addr);
1201                 if (ret)
1202                         return ret;
1203
1204                 if (vlan_id >= VLAN_N_VID && iwdev->dcb_vlan_mode)
1205                         vlan_id = 0;
1206                 if (vlan_id < VLAN_N_VID) {
1207                         udp_info->insert_vlan_tag = true;
1208                         udp_info->vlan_tag = vlan_id |
1209                                 ctx_info->user_pri << VLAN_PRIO_SHIFT;
1210                 } else {
1211                         udp_info->insert_vlan_tag = false;
1212                 }
1213
1214                 av->attrs = attr->ah_attr;
1215                 rdma_gid2ip((struct sockaddr *)&av->sgid_addr, &sgid_attr->gid);
1216                 rdma_gid2ip((struct sockaddr *)&av->dgid_addr, &attr->ah_attr.grh.dgid);
1217                 if (av->net_type == RDMA_NETWORK_IPV6) {
1218                         __be32 *daddr =
1219                                 av->dgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32;
1220                         __be32 *saddr =
1221                                 av->sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32;
1222
1223                         irdma_copy_ip_ntohl(&udp_info->dest_ip_addr[0], daddr);
1224                         irdma_copy_ip_ntohl(&udp_info->local_ipaddr[0], saddr);
1225
1226                         udp_info->ipv4 = false;
1227                         irdma_copy_ip_ntohl(local_ip, daddr);
1228
1229                         udp_info->arp_idx = irdma_arp_table(iwdev->rf,
1230                                                             &local_ip[0],
1231                                                             false, NULL,
1232                                                             IRDMA_ARP_RESOLVE);
1233                 } else if (av->net_type == RDMA_NETWORK_IPV4) {
1234                         __be32 saddr = av->sgid_addr.saddr_in.sin_addr.s_addr;
1235                         __be32 daddr = av->dgid_addr.saddr_in.sin_addr.s_addr;
1236
1237                         local_ip[0] = ntohl(daddr);
1238
1239                         udp_info->ipv4 = true;
1240                         udp_info->dest_ip_addr[0] = 0;
1241                         udp_info->dest_ip_addr[1] = 0;
1242                         udp_info->dest_ip_addr[2] = 0;
1243                         udp_info->dest_ip_addr[3] = local_ip[0];
1244
1245                         udp_info->local_ipaddr[0] = 0;
1246                         udp_info->local_ipaddr[1] = 0;
1247                         udp_info->local_ipaddr[2] = 0;
1248                         udp_info->local_ipaddr[3] = ntohl(saddr);
1249                 }
1250                 udp_info->arp_idx =
1251                         irdma_add_arp(iwdev->rf, local_ip, udp_info->ipv4,
1252                                       attr->ah_attr.roce.dmac);
1253         }
1254
1255         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1256                 if (attr->max_rd_atomic > dev->hw_attrs.max_hw_ord) {
1257                         ibdev_err(&iwdev->ibdev,
1258                                   "rd_atomic = %d, above max_hw_ord=%d\n",
1259                                   attr->max_rd_atomic,
1260                                   dev->hw_attrs.max_hw_ord);
1261                         return -EINVAL;
1262                 }
1263                 if (attr->max_rd_atomic)
1264                         roce_info->ord_size = attr->max_rd_atomic;
1265                 info.ord_valid = true;
1266         }
1267
1268         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1269                 if (attr->max_dest_rd_atomic > dev->hw_attrs.max_hw_ird) {
1270                         ibdev_err(&iwdev->ibdev,
1271                                   "rd_atomic = %d, above max_hw_ird=%d\n",
1272                                    attr->max_rd_atomic,
1273                                    dev->hw_attrs.max_hw_ird);
1274                         return -EINVAL;
1275                 }
1276                 if (attr->max_dest_rd_atomic)
1277                         roce_info->ird_size = attr->max_dest_rd_atomic;
1278         }
1279
1280         if (attr_mask & IB_QP_ACCESS_FLAGS) {
1281                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
1282                         roce_info->wr_rdresp_en = true;
1283                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
1284                         roce_info->wr_rdresp_en = true;
1285                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
1286                         roce_info->rd_en = true;
1287         }
1288
1289         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
1290
1291         ibdev_dbg(&iwdev->ibdev,
1292                   "VERBS: caller: %pS qp_id=%d to_ibqpstate=%d ibqpstate=%d irdma_qpstate=%d attr_mask=0x%x\n",
1293                   __builtin_return_address(0), ibqp->qp_num, attr->qp_state,
1294                   iwqp->ibqp_state, iwqp->iwarp_state, attr_mask);
1295
1296         spin_lock_irqsave(&iwqp->lock, flags);
1297         if (attr_mask & IB_QP_STATE) {
1298                 if (!ib_modify_qp_is_ok(iwqp->ibqp_state, attr->qp_state,
1299                                         iwqp->ibqp.qp_type, attr_mask)) {
1300                         ibdev_warn(&iwdev->ibdev, "modify_qp invalid for qp_id=%d, old_state=0x%x, new_state=0x%x\n",
1301                                    iwqp->ibqp.qp_num, iwqp->ibqp_state,
1302                                    attr->qp_state);
1303                         ret = -EINVAL;
1304                         goto exit;
1305                 }
1306                 info.curr_iwarp_state = iwqp->iwarp_state;
1307
1308                 switch (attr->qp_state) {
1309                 case IB_QPS_INIT:
1310                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1311                                 ret = -EINVAL;
1312                                 goto exit;
1313                         }
1314
1315                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
1316                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
1317                                 issue_modify_qp = 1;
1318                         }
1319                         break;
1320                 case IB_QPS_RTR:
1321                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1322                                 ret = -EINVAL;
1323                                 goto exit;
1324                         }
1325                         info.arp_cache_idx_valid = true;
1326                         info.cq_num_valid = true;
1327                         info.next_iwarp_state = IRDMA_QP_STATE_RTR;
1328                         issue_modify_qp = 1;
1329                         break;
1330                 case IB_QPS_RTS:
1331                         if (iwqp->ibqp_state < IB_QPS_RTR ||
1332                             iwqp->ibqp_state == IB_QPS_ERR) {
1333                                 ret = -EINVAL;
1334                                 goto exit;
1335                         }
1336
1337                         info.arp_cache_idx_valid = true;
1338                         info.cq_num_valid = true;
1339                         info.ord_valid = true;
1340                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
1341                         issue_modify_qp = 1;
1342                         if (iwdev->push_mode && udata &&
1343                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1344                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1345                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1346                                 irdma_alloc_push_page(iwqp);
1347                                 spin_lock_irqsave(&iwqp->lock, flags);
1348                         }
1349                         break;
1350                 case IB_QPS_SQD:
1351                         if (iwqp->iwarp_state == IRDMA_QP_STATE_SQD)
1352                                 goto exit;
1353
1354                         if (iwqp->iwarp_state != IRDMA_QP_STATE_RTS) {
1355                                 ret = -EINVAL;
1356                                 goto exit;
1357                         }
1358
1359                         info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1360                         issue_modify_qp = 1;
1361                         break;
1362                 case IB_QPS_SQE:
1363                 case IB_QPS_ERR:
1364                 case IB_QPS_RESET:
1365                         if (iwqp->iwarp_state == IRDMA_QP_STATE_RTS) {
1366                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1367                                 info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1368                                 irdma_hw_modify_qp(iwdev, iwqp, &info, true);
1369                                 spin_lock_irqsave(&iwqp->lock, flags);
1370                         }
1371
1372                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1373                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1374                                 if (udata) {
1375                                         if (ib_copy_from_udata(&ureq, udata,
1376                                             min(sizeof(ureq), udata->inlen)))
1377                                                 return -EINVAL;
1378
1379                                         irdma_flush_wqes(iwqp,
1380                                             (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1381                                             (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1382                                             IRDMA_REFLUSH);
1383                                 }
1384                                 return 0;
1385                         }
1386
1387                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1388                         issue_modify_qp = 1;
1389                         break;
1390                 default:
1391                         ret = -EINVAL;
1392                         goto exit;
1393                 }
1394
1395                 iwqp->ibqp_state = attr->qp_state;
1396         }
1397
1398         ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1399         ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1400         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1401         spin_unlock_irqrestore(&iwqp->lock, flags);
1402
1403         if (attr_mask & IB_QP_STATE) {
1404                 if (issue_modify_qp) {
1405                         ctx_info->rem_endpoint_idx = udp_info->arp_idx;
1406                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1407                                 return -EINVAL;
1408                         spin_lock_irqsave(&iwqp->lock, flags);
1409                         if (iwqp->iwarp_state == info.curr_iwarp_state) {
1410                                 iwqp->iwarp_state = info.next_iwarp_state;
1411                                 iwqp->ibqp_state = attr->qp_state;
1412                         }
1413                         if (iwqp->ibqp_state > IB_QPS_RTS &&
1414                             !iwqp->flush_issued) {
1415                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1416                                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_SQ |
1417                                                        IRDMA_FLUSH_RQ |
1418                                                        IRDMA_FLUSH_WAIT);
1419                                 iwqp->flush_issued = 1;
1420                         } else {
1421                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1422                         }
1423                 } else {
1424                         iwqp->ibqp_state = attr->qp_state;
1425                 }
1426                 if (udata && dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1427                         struct irdma_ucontext *ucontext;
1428
1429                         ucontext = rdma_udata_to_drv_context(udata,
1430                                         struct irdma_ucontext, ibucontext);
1431                         if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1432                             !iwqp->push_wqe_mmap_entry &&
1433                             !irdma_setup_push_mmap_entries(ucontext, iwqp,
1434                                 &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1435                                 uresp.push_valid = 1;
1436                                 uresp.push_offset = iwqp->sc_qp.push_offset;
1437                         }
1438                         ret = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1439                                                udata->outlen));
1440                         if (ret) {
1441                                 irdma_remove_push_mmap_entries(iwqp);
1442                                 ibdev_dbg(&iwdev->ibdev,
1443                                           "VERBS: copy_to_udata failed\n");
1444                                 return ret;
1445                         }
1446                 }
1447         }
1448
1449         return 0;
1450 exit:
1451         spin_unlock_irqrestore(&iwqp->lock, flags);
1452
1453         return ret;
1454 }
1455
1456 /**
1457  * irdma_modify_qp - modify qp request
1458  * @ibqp: qp's pointer for modify
1459  * @attr: access attributes
1460  * @attr_mask: state mask
1461  * @udata: user data
1462  */
1463 int irdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
1464                     struct ib_udata *udata)
1465 {
1466         struct irdma_qp *iwqp = to_iwqp(ibqp);
1467         struct irdma_device *iwdev = iwqp->iwdev;
1468         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
1469         struct irdma_qp_host_ctx_info *ctx_info;
1470         struct irdma_tcp_offload_info *tcp_info;
1471         struct irdma_iwarp_offload_info *offload_info;
1472         struct irdma_modify_qp_info info = {};
1473         struct irdma_modify_qp_resp uresp = {};
1474         struct irdma_modify_qp_req ureq = {};
1475         u8 issue_modify_qp = 0;
1476         u8 dont_wait = 0;
1477         int err;
1478         unsigned long flags;
1479
1480         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1481                 return -EOPNOTSUPP;
1482
1483         ctx_info = &iwqp->ctx_info;
1484         offload_info = &iwqp->iwarp_info;
1485         tcp_info = &iwqp->tcp_info;
1486         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
1487         ibdev_dbg(&iwdev->ibdev,
1488                   "VERBS: caller: %pS qp_id=%d to_ibqpstate=%d ibqpstate=%d irdma_qpstate=%d last_aeq=%d hw_tcp_state=%d hw_iwarp_state=%d attr_mask=0x%x\n",
1489                   __builtin_return_address(0), ibqp->qp_num, attr->qp_state,
1490                   iwqp->ibqp_state, iwqp->iwarp_state, iwqp->last_aeq,
1491                   iwqp->hw_tcp_state, iwqp->hw_iwarp_state, attr_mask);
1492
1493         spin_lock_irqsave(&iwqp->lock, flags);
1494         if (attr_mask & IB_QP_STATE) {
1495                 info.curr_iwarp_state = iwqp->iwarp_state;
1496                 switch (attr->qp_state) {
1497                 case IB_QPS_INIT:
1498                 case IB_QPS_RTR:
1499                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1500                                 err = -EINVAL;
1501                                 goto exit;
1502                         }
1503
1504                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
1505                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
1506                                 issue_modify_qp = 1;
1507                         }
1508                         if (iwdev->push_mode && udata &&
1509                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1510                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1511                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1512                                 irdma_alloc_push_page(iwqp);
1513                                 spin_lock_irqsave(&iwqp->lock, flags);
1514                         }
1515                         break;
1516                 case IB_QPS_RTS:
1517                         if (iwqp->iwarp_state > IRDMA_QP_STATE_RTS ||
1518                             !iwqp->cm_id) {
1519                                 err = -EINVAL;
1520                                 goto exit;
1521                         }
1522
1523                         issue_modify_qp = 1;
1524                         iwqp->hw_tcp_state = IRDMA_TCP_STATE_ESTABLISHED;
1525                         iwqp->hte_added = 1;
1526                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
1527                         info.tcp_ctx_valid = true;
1528                         info.ord_valid = true;
1529                         info.arp_cache_idx_valid = true;
1530                         info.cq_num_valid = true;
1531                         break;
1532                 case IB_QPS_SQD:
1533                         if (iwqp->hw_iwarp_state > IRDMA_QP_STATE_RTS) {
1534                                 err = 0;
1535                                 goto exit;
1536                         }
1537
1538                         if (iwqp->iwarp_state == IRDMA_QP_STATE_CLOSING ||
1539                             iwqp->iwarp_state < IRDMA_QP_STATE_RTS) {
1540                                 err = 0;
1541                                 goto exit;
1542                         }
1543
1544                         if (iwqp->iwarp_state > IRDMA_QP_STATE_CLOSING) {
1545                                 err = -EINVAL;
1546                                 goto exit;
1547                         }
1548
1549                         info.next_iwarp_state = IRDMA_QP_STATE_CLOSING;
1550                         issue_modify_qp = 1;
1551                         break;
1552                 case IB_QPS_SQE:
1553                         if (iwqp->iwarp_state >= IRDMA_QP_STATE_TERMINATE) {
1554                                 err = -EINVAL;
1555                                 goto exit;
1556                         }
1557
1558                         info.next_iwarp_state = IRDMA_QP_STATE_TERMINATE;
1559                         issue_modify_qp = 1;
1560                         break;
1561                 case IB_QPS_ERR:
1562                 case IB_QPS_RESET:
1563                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1564                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1565                                 if (udata) {
1566                                         if (ib_copy_from_udata(&ureq, udata,
1567                                             min(sizeof(ureq), udata->inlen)))
1568                                                 return -EINVAL;
1569
1570                                         irdma_flush_wqes(iwqp,
1571                                             (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1572                                             (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1573                                             IRDMA_REFLUSH);
1574                                 }
1575                                 return 0;
1576                         }
1577
1578                         if (iwqp->sc_qp.term_flags) {
1579                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1580                                 irdma_terminate_del_timer(&iwqp->sc_qp);
1581                                 spin_lock_irqsave(&iwqp->lock, flags);
1582                         }
1583                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1584                         if (iwqp->hw_tcp_state > IRDMA_TCP_STATE_CLOSED &&
1585                             iwdev->iw_status &&
1586                             iwqp->hw_tcp_state != IRDMA_TCP_STATE_TIME_WAIT)
1587                                 info.reset_tcp_conn = true;
1588                         else
1589                                 dont_wait = 1;
1590
1591                         issue_modify_qp = 1;
1592                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1593                         break;
1594                 default:
1595                         err = -EINVAL;
1596                         goto exit;
1597                 }
1598
1599                 iwqp->ibqp_state = attr->qp_state;
1600         }
1601         if (attr_mask & IB_QP_ACCESS_FLAGS) {
1602                 ctx_info->iwarp_info_valid = true;
1603                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
1604                         offload_info->wr_rdresp_en = true;
1605                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
1606                         offload_info->wr_rdresp_en = true;
1607                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
1608                         offload_info->rd_en = true;
1609         }
1610
1611         if (ctx_info->iwarp_info_valid) {
1612                 ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1613                 ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1614                 irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1615         }
1616         spin_unlock_irqrestore(&iwqp->lock, flags);
1617
1618         if (attr_mask & IB_QP_STATE) {
1619                 if (issue_modify_qp) {
1620                         ctx_info->rem_endpoint_idx = tcp_info->arp_idx;
1621                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1622                                 return -EINVAL;
1623                 }
1624
1625                 spin_lock_irqsave(&iwqp->lock, flags);
1626                 if (iwqp->iwarp_state == info.curr_iwarp_state) {
1627                         iwqp->iwarp_state = info.next_iwarp_state;
1628                         iwqp->ibqp_state = attr->qp_state;
1629                 }
1630                 spin_unlock_irqrestore(&iwqp->lock, flags);
1631         }
1632
1633         if (issue_modify_qp && iwqp->ibqp_state > IB_QPS_RTS) {
1634                 if (dont_wait) {
1635                         if (iwqp->hw_tcp_state) {
1636                                 spin_lock_irqsave(&iwqp->lock, flags);
1637                                 iwqp->hw_tcp_state = IRDMA_TCP_STATE_CLOSED;
1638                                 iwqp->last_aeq = IRDMA_AE_RESET_SENT;
1639                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1640                         }
1641                         irdma_cm_disconn(iwqp);
1642                 } else {
1643                         int close_timer_started;
1644
1645                         spin_lock_irqsave(&iwdev->cm_core.ht_lock, flags);
1646
1647                         if (iwqp->cm_node) {
1648                                 refcount_inc(&iwqp->cm_node->refcnt);
1649                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1650                                 close_timer_started = atomic_inc_return(&iwqp->close_timer_started);
1651                                 if (iwqp->cm_id && close_timer_started == 1)
1652                                         irdma_schedule_cm_timer(iwqp->cm_node,
1653                                                 (struct irdma_puda_buf *)iwqp,
1654                                                 IRDMA_TIMER_TYPE_CLOSE, 1, 0);
1655
1656                                 irdma_rem_ref_cm_node(iwqp->cm_node);
1657                         } else {
1658                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1659                         }
1660                 }
1661         }
1662         if (attr_mask & IB_QP_STATE && udata &&
1663             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1664                 struct irdma_ucontext *ucontext;
1665
1666                 ucontext = rdma_udata_to_drv_context(udata,
1667                                         struct irdma_ucontext, ibucontext);
1668                 if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1669                     !iwqp->push_wqe_mmap_entry &&
1670                     !irdma_setup_push_mmap_entries(ucontext, iwqp,
1671                         &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1672                         uresp.push_valid = 1;
1673                         uresp.push_offset = iwqp->sc_qp.push_offset;
1674                 }
1675
1676                 err = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1677                                        udata->outlen));
1678                 if (err) {
1679                         irdma_remove_push_mmap_entries(iwqp);
1680                         ibdev_dbg(&iwdev->ibdev,
1681                                   "VERBS: copy_to_udata failed\n");
1682                         return err;
1683                 }
1684         }
1685
1686         return 0;
1687 exit:
1688         spin_unlock_irqrestore(&iwqp->lock, flags);
1689
1690         return err;
1691 }
1692
1693 /**
1694  * irdma_cq_free_rsrc - free up resources for cq
1695  * @rf: RDMA PCI function
1696  * @iwcq: cq ptr
1697  */
1698 static void irdma_cq_free_rsrc(struct irdma_pci_f *rf, struct irdma_cq *iwcq)
1699 {
1700         struct irdma_sc_cq *cq = &iwcq->sc_cq;
1701
1702         if (!iwcq->user_mode) {
1703                 dma_free_coherent(rf->sc_dev.hw->device, iwcq->kmem.size,
1704                                   iwcq->kmem.va, iwcq->kmem.pa);
1705                 iwcq->kmem.va = NULL;
1706                 dma_free_coherent(rf->sc_dev.hw->device,
1707                                   iwcq->kmem_shadow.size,
1708                                   iwcq->kmem_shadow.va, iwcq->kmem_shadow.pa);
1709                 iwcq->kmem_shadow.va = NULL;
1710         }
1711
1712         irdma_free_rsrc(rf, rf->allocated_cqs, cq->cq_uk.cq_id);
1713 }
1714
1715 /**
1716  * irdma_free_cqbuf - worker to free a cq buffer
1717  * @work: provides access to the cq buffer to free
1718  */
1719 static void irdma_free_cqbuf(struct work_struct *work)
1720 {
1721         struct irdma_cq_buf *cq_buf = container_of(work, struct irdma_cq_buf, work);
1722
1723         dma_free_coherent(cq_buf->hw->device, cq_buf->kmem_buf.size,
1724                           cq_buf->kmem_buf.va, cq_buf->kmem_buf.pa);
1725         cq_buf->kmem_buf.va = NULL;
1726         kfree(cq_buf);
1727 }
1728
1729 /**
1730  * irdma_process_resize_list - remove resized cq buffers from the resize_list
1731  * @iwcq: cq which owns the resize_list
1732  * @iwdev: irdma device
1733  * @lcqe_buf: the buffer where the last cqe is received
1734  */
1735 static int irdma_process_resize_list(struct irdma_cq *iwcq,
1736                                      struct irdma_device *iwdev,
1737                                      struct irdma_cq_buf *lcqe_buf)
1738 {
1739         struct list_head *tmp_node, *list_node;
1740         struct irdma_cq_buf *cq_buf;
1741         int cnt = 0;
1742
1743         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
1744                 cq_buf = list_entry(list_node, struct irdma_cq_buf, list);
1745                 if (cq_buf == lcqe_buf)
1746                         return cnt;
1747
1748                 list_del(&cq_buf->list);
1749                 queue_work(iwdev->cleanup_wq, &cq_buf->work);
1750                 cnt++;
1751         }
1752
1753         return cnt;
1754 }
1755
1756 /**
1757  * irdma_destroy_cq - destroy cq
1758  * @ib_cq: cq pointer
1759  * @udata: user data
1760  */
1761 static int irdma_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata)
1762 {
1763         struct irdma_device *iwdev = to_iwdev(ib_cq->device);
1764         struct irdma_cq *iwcq = to_iwcq(ib_cq);
1765         struct irdma_sc_cq *cq = &iwcq->sc_cq;
1766         struct irdma_sc_dev *dev = cq->dev;
1767         struct irdma_sc_ceq *ceq = dev->ceq[cq->ceq_id];
1768         struct irdma_ceq *iwceq = container_of(ceq, struct irdma_ceq, sc_ceq);
1769         unsigned long flags;
1770
1771         spin_lock_irqsave(&iwcq->lock, flags);
1772         if (!list_empty(&iwcq->cmpl_generated))
1773                 irdma_remove_cmpls_list(iwcq);
1774         if (!list_empty(&iwcq->resize_list))
1775                 irdma_process_resize_list(iwcq, iwdev, NULL);
1776         spin_unlock_irqrestore(&iwcq->lock, flags);
1777
1778         irdma_cq_wq_destroy(iwdev->rf, cq);
1779
1780         spin_lock_irqsave(&iwceq->ce_lock, flags);
1781         irdma_sc_cleanup_ceqes(cq, ceq);
1782         spin_unlock_irqrestore(&iwceq->ce_lock, flags);
1783         irdma_cq_free_rsrc(iwdev->rf, iwcq);
1784
1785         return 0;
1786 }
1787
1788 /**
1789  * irdma_resize_cq - resize cq
1790  * @ibcq: cq to be resized
1791  * @entries: desired cq size
1792  * @udata: user data
1793  */
1794 static int irdma_resize_cq(struct ib_cq *ibcq, int entries,
1795                            struct ib_udata *udata)
1796 {
1797         struct irdma_cq *iwcq = to_iwcq(ibcq);
1798         struct irdma_sc_dev *dev = iwcq->sc_cq.dev;
1799         struct irdma_cqp_request *cqp_request;
1800         struct cqp_cmds_info *cqp_info;
1801         struct irdma_modify_cq_info *m_info;
1802         struct irdma_modify_cq_info info = {};
1803         struct irdma_dma_mem kmem_buf;
1804         struct irdma_cq_mr *cqmr_buf;
1805         struct irdma_pbl *iwpbl_buf;
1806         struct irdma_device *iwdev;
1807         struct irdma_pci_f *rf;
1808         struct irdma_cq_buf *cq_buf = NULL;
1809         unsigned long flags;
1810         int ret;
1811
1812         iwdev = to_iwdev(ibcq->device);
1813         rf = iwdev->rf;
1814
1815         if (!(rf->sc_dev.hw_attrs.uk_attrs.feature_flags &
1816             IRDMA_FEATURE_CQ_RESIZE))
1817                 return -EOPNOTSUPP;
1818
1819         if (entries > rf->max_cqe)
1820                 return -EINVAL;
1821
1822         if (!iwcq->user_mode) {
1823                 entries++;
1824                 if (rf->sc_dev.hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
1825                         entries *= 2;
1826         }
1827
1828         info.cq_size = max(entries, 4);
1829
1830         if (info.cq_size == iwcq->sc_cq.cq_uk.cq_size - 1)
1831                 return 0;
1832
1833         if (udata) {
1834                 struct irdma_resize_cq_req req = {};
1835                 struct irdma_ucontext *ucontext =
1836                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
1837                                                   ibucontext);
1838
1839                 /* CQ resize not supported with legacy GEN_1 libi40iw */
1840                 if (ucontext->legacy_mode)
1841                         return -EOPNOTSUPP;
1842
1843                 if (ib_copy_from_udata(&req, udata,
1844                                        min(sizeof(req), udata->inlen)))
1845                         return -EINVAL;
1846
1847                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
1848                 iwpbl_buf = irdma_get_pbl((unsigned long)req.user_cq_buffer,
1849                                           &ucontext->cq_reg_mem_list);
1850                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
1851
1852                 if (!iwpbl_buf)
1853                         return -ENOMEM;
1854
1855                 cqmr_buf = &iwpbl_buf->cq_mr;
1856                 if (iwpbl_buf->pbl_allocated) {
1857                         info.virtual_map = true;
1858                         info.pbl_chunk_size = 1;
1859                         info.first_pm_pbl_idx = cqmr_buf->cq_pbl.idx;
1860                 } else {
1861                         info.cq_pa = cqmr_buf->cq_pbl.addr;
1862                 }
1863         } else {
1864                 /* Kmode CQ resize */
1865                 int rsize;
1866
1867                 rsize = info.cq_size * sizeof(struct irdma_cqe);
1868                 kmem_buf.size = ALIGN(round_up(rsize, 256), 256);
1869                 kmem_buf.va = dma_alloc_coherent(dev->hw->device,
1870                                                  kmem_buf.size, &kmem_buf.pa,
1871                                                  GFP_KERNEL);
1872                 if (!kmem_buf.va)
1873                         return -ENOMEM;
1874
1875                 info.cq_base = kmem_buf.va;
1876                 info.cq_pa = kmem_buf.pa;
1877                 cq_buf = kzalloc(sizeof(*cq_buf), GFP_KERNEL);
1878                 if (!cq_buf) {
1879                         ret = -ENOMEM;
1880                         goto error;
1881                 }
1882         }
1883
1884         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
1885         if (!cqp_request) {
1886                 ret = -ENOMEM;
1887                 goto error;
1888         }
1889
1890         info.shadow_read_threshold = iwcq->sc_cq.shadow_read_threshold;
1891         info.cq_resize = true;
1892
1893         cqp_info = &cqp_request->info;
1894         m_info = &cqp_info->in.u.cq_modify.info;
1895         memcpy(m_info, &info, sizeof(*m_info));
1896
1897         cqp_info->cqp_cmd = IRDMA_OP_CQ_MODIFY;
1898         cqp_info->in.u.cq_modify.cq = &iwcq->sc_cq;
1899         cqp_info->in.u.cq_modify.scratch = (uintptr_t)cqp_request;
1900         cqp_info->post_sq = 1;
1901         ret = irdma_handle_cqp_op(rf, cqp_request);
1902         irdma_put_cqp_request(&rf->cqp, cqp_request);
1903         if (ret)
1904                 goto error;
1905
1906         spin_lock_irqsave(&iwcq->lock, flags);
1907         if (cq_buf) {
1908                 cq_buf->kmem_buf = iwcq->kmem;
1909                 cq_buf->hw = dev->hw;
1910                 memcpy(&cq_buf->cq_uk, &iwcq->sc_cq.cq_uk, sizeof(cq_buf->cq_uk));
1911                 INIT_WORK(&cq_buf->work, irdma_free_cqbuf);
1912                 list_add_tail(&cq_buf->list, &iwcq->resize_list);
1913                 iwcq->kmem = kmem_buf;
1914         }
1915
1916         irdma_sc_cq_resize(&iwcq->sc_cq, &info);
1917         ibcq->cqe = info.cq_size - 1;
1918         spin_unlock_irqrestore(&iwcq->lock, flags);
1919
1920         return 0;
1921 error:
1922         if (!udata) {
1923                 dma_free_coherent(dev->hw->device, kmem_buf.size, kmem_buf.va,
1924                                   kmem_buf.pa);
1925                 kmem_buf.va = NULL;
1926         }
1927         kfree(cq_buf);
1928
1929         return ret;
1930 }
1931
1932 static inline int cq_validate_flags(u32 flags, u8 hw_rev)
1933 {
1934         /* GEN1 does not support CQ create flags */
1935         if (hw_rev == IRDMA_GEN_1)
1936                 return flags ? -EOPNOTSUPP : 0;
1937
1938         return flags & ~IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION ? -EOPNOTSUPP : 0;
1939 }
1940
1941 /**
1942  * irdma_create_cq - create cq
1943  * @ibcq: CQ allocated
1944  * @attr: attributes for cq
1945  * @udata: user data
1946  */
1947 static int irdma_create_cq(struct ib_cq *ibcq,
1948                            const struct ib_cq_init_attr *attr,
1949                            struct ib_udata *udata)
1950 {
1951         struct ib_device *ibdev = ibcq->device;
1952         struct irdma_device *iwdev = to_iwdev(ibdev);
1953         struct irdma_pci_f *rf = iwdev->rf;
1954         struct irdma_cq *iwcq = to_iwcq(ibcq);
1955         u32 cq_num = 0;
1956         struct irdma_sc_cq *cq;
1957         struct irdma_sc_dev *dev = &rf->sc_dev;
1958         struct irdma_cq_init_info info = {};
1959         struct irdma_cqp_request *cqp_request;
1960         struct cqp_cmds_info *cqp_info;
1961         struct irdma_cq_uk_init_info *ukinfo = &info.cq_uk_init_info;
1962         unsigned long flags;
1963         int err_code;
1964         int entries = attr->cqe;
1965
1966         err_code = cq_validate_flags(attr->flags, dev->hw_attrs.uk_attrs.hw_rev);
1967         if (err_code)
1968                 return err_code;
1969         err_code = irdma_alloc_rsrc(rf, rf->allocated_cqs, rf->max_cq, &cq_num,
1970                                     &rf->next_cq);
1971         if (err_code)
1972                 return err_code;
1973
1974         cq = &iwcq->sc_cq;
1975         cq->back_cq = iwcq;
1976         spin_lock_init(&iwcq->lock);
1977         INIT_LIST_HEAD(&iwcq->resize_list);
1978         INIT_LIST_HEAD(&iwcq->cmpl_generated);
1979         info.dev = dev;
1980         ukinfo->cq_size = max(entries, 4);
1981         ukinfo->cq_id = cq_num;
1982         iwcq->ibcq.cqe = info.cq_uk_init_info.cq_size;
1983         if (attr->comp_vector < rf->ceqs_count)
1984                 info.ceq_id = attr->comp_vector;
1985         info.ceq_id_valid = true;
1986         info.ceqe_mask = 1;
1987         info.type = IRDMA_CQ_TYPE_IWARP;
1988         info.vsi = &iwdev->vsi;
1989
1990         if (udata) {
1991                 struct irdma_ucontext *ucontext;
1992                 struct irdma_create_cq_req req = {};
1993                 struct irdma_cq_mr *cqmr;
1994                 struct irdma_pbl *iwpbl;
1995                 struct irdma_pbl *iwpbl_shadow;
1996                 struct irdma_cq_mr *cqmr_shadow;
1997
1998                 iwcq->user_mode = true;
1999                 ucontext =
2000                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
2001                                                   ibucontext);
2002                 if (ib_copy_from_udata(&req, udata,
2003                                        min(sizeof(req), udata->inlen))) {
2004                         err_code = -EFAULT;
2005                         goto cq_free_rsrc;
2006                 }
2007
2008                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2009                 iwpbl = irdma_get_pbl((unsigned long)req.user_cq_buf,
2010                                       &ucontext->cq_reg_mem_list);
2011                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2012                 if (!iwpbl) {
2013                         err_code = -EPROTO;
2014                         goto cq_free_rsrc;
2015                 }
2016
2017                 iwcq->iwpbl = iwpbl;
2018                 iwcq->cq_mem_size = 0;
2019                 cqmr = &iwpbl->cq_mr;
2020
2021                 if (rf->sc_dev.hw_attrs.uk_attrs.feature_flags &
2022                     IRDMA_FEATURE_CQ_RESIZE && !ucontext->legacy_mode) {
2023                         spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2024                         iwpbl_shadow = irdma_get_pbl(
2025                                         (unsigned long)req.user_shadow_area,
2026                                         &ucontext->cq_reg_mem_list);
2027                         spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2028
2029                         if (!iwpbl_shadow) {
2030                                 err_code = -EPROTO;
2031                                 goto cq_free_rsrc;
2032                         }
2033                         iwcq->iwpbl_shadow = iwpbl_shadow;
2034                         cqmr_shadow = &iwpbl_shadow->cq_mr;
2035                         info.shadow_area_pa = cqmr_shadow->cq_pbl.addr;
2036                         cqmr->split = true;
2037                 } else {
2038                         info.shadow_area_pa = cqmr->shadow;
2039                 }
2040                 if (iwpbl->pbl_allocated) {
2041                         info.virtual_map = true;
2042                         info.pbl_chunk_size = 1;
2043                         info.first_pm_pbl_idx = cqmr->cq_pbl.idx;
2044                 } else {
2045                         info.cq_base_pa = cqmr->cq_pbl.addr;
2046                 }
2047         } else {
2048                 /* Kmode allocations */
2049                 int rsize;
2050
2051                 if (entries < 1 || entries > rf->max_cqe) {
2052                         err_code = -EINVAL;
2053                         goto cq_free_rsrc;
2054                 }
2055
2056                 entries++;
2057                 if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
2058                         entries *= 2;
2059                 ukinfo->cq_size = entries;
2060
2061                 rsize = info.cq_uk_init_info.cq_size * sizeof(struct irdma_cqe);
2062                 iwcq->kmem.size = ALIGN(round_up(rsize, 256), 256);
2063                 iwcq->kmem.va = dma_alloc_coherent(dev->hw->device,
2064                                                    iwcq->kmem.size,
2065                                                    &iwcq->kmem.pa, GFP_KERNEL);
2066                 if (!iwcq->kmem.va) {
2067                         err_code = -ENOMEM;
2068                         goto cq_free_rsrc;
2069                 }
2070
2071                 iwcq->kmem_shadow.size = ALIGN(IRDMA_SHADOW_AREA_SIZE << 3,
2072                                                64);
2073                 iwcq->kmem_shadow.va = dma_alloc_coherent(dev->hw->device,
2074                                                           iwcq->kmem_shadow.size,
2075                                                           &iwcq->kmem_shadow.pa,
2076                                                           GFP_KERNEL);
2077                 if (!iwcq->kmem_shadow.va) {
2078                         err_code = -ENOMEM;
2079                         goto cq_free_rsrc;
2080                 }
2081                 info.shadow_area_pa = iwcq->kmem_shadow.pa;
2082                 ukinfo->shadow_area = iwcq->kmem_shadow.va;
2083                 ukinfo->cq_base = iwcq->kmem.va;
2084                 info.cq_base_pa = iwcq->kmem.pa;
2085         }
2086
2087         if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
2088                 info.shadow_read_threshold = min(info.cq_uk_init_info.cq_size / 2,
2089                                                  (u32)IRDMA_MAX_CQ_READ_THRESH);
2090
2091         if (irdma_sc_cq_init(cq, &info)) {
2092                 ibdev_dbg(&iwdev->ibdev, "VERBS: init cq fail\n");
2093                 err_code = -EPROTO;
2094                 goto cq_free_rsrc;
2095         }
2096
2097         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
2098         if (!cqp_request) {
2099                 err_code = -ENOMEM;
2100                 goto cq_free_rsrc;
2101         }
2102
2103         cqp_info = &cqp_request->info;
2104         cqp_info->cqp_cmd = IRDMA_OP_CQ_CREATE;
2105         cqp_info->post_sq = 1;
2106         cqp_info->in.u.cq_create.cq = cq;
2107         cqp_info->in.u.cq_create.check_overflow = true;
2108         cqp_info->in.u.cq_create.scratch = (uintptr_t)cqp_request;
2109         err_code = irdma_handle_cqp_op(rf, cqp_request);
2110         irdma_put_cqp_request(&rf->cqp, cqp_request);
2111         if (err_code)
2112                 goto cq_free_rsrc;
2113
2114         if (udata) {
2115                 struct irdma_create_cq_resp resp = {};
2116
2117                 resp.cq_id = info.cq_uk_init_info.cq_id;
2118                 resp.cq_size = info.cq_uk_init_info.cq_size;
2119                 if (ib_copy_to_udata(udata, &resp,
2120                                      min(sizeof(resp), udata->outlen))) {
2121                         ibdev_dbg(&iwdev->ibdev,
2122                                   "VERBS: copy to user data\n");
2123                         err_code = -EPROTO;
2124                         goto cq_destroy;
2125                 }
2126         }
2127         return 0;
2128 cq_destroy:
2129         irdma_cq_wq_destroy(rf, cq);
2130 cq_free_rsrc:
2131         irdma_cq_free_rsrc(rf, iwcq);
2132
2133         return err_code;
2134 }
2135
2136 /**
2137  * irdma_get_mr_access - get hw MR access permissions from IB access flags
2138  * @access: IB access flags
2139  */
2140 static inline u16 irdma_get_mr_access(int access)
2141 {
2142         u16 hw_access = 0;
2143
2144         hw_access |= (access & IB_ACCESS_LOCAL_WRITE) ?
2145                      IRDMA_ACCESS_FLAGS_LOCALWRITE : 0;
2146         hw_access |= (access & IB_ACCESS_REMOTE_WRITE) ?
2147                      IRDMA_ACCESS_FLAGS_REMOTEWRITE : 0;
2148         hw_access |= (access & IB_ACCESS_REMOTE_READ) ?
2149                      IRDMA_ACCESS_FLAGS_REMOTEREAD : 0;
2150         hw_access |= (access & IB_ACCESS_MW_BIND) ?
2151                      IRDMA_ACCESS_FLAGS_BIND_WINDOW : 0;
2152         hw_access |= (access & IB_ZERO_BASED) ?
2153                      IRDMA_ACCESS_FLAGS_ZERO_BASED : 0;
2154         hw_access |= IRDMA_ACCESS_FLAGS_LOCALREAD;
2155
2156         return hw_access;
2157 }
2158
2159 /**
2160  * irdma_free_stag - free stag resource
2161  * @iwdev: irdma device
2162  * @stag: stag to free
2163  */
2164 static void irdma_free_stag(struct irdma_device *iwdev, u32 stag)
2165 {
2166         u32 stag_idx;
2167
2168         stag_idx = (stag & iwdev->rf->mr_stagmask) >> IRDMA_CQPSQ_STAG_IDX_S;
2169         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_mrs, stag_idx);
2170 }
2171
2172 /**
2173  * irdma_create_stag - create random stag
2174  * @iwdev: irdma device
2175  */
2176 static u32 irdma_create_stag(struct irdma_device *iwdev)
2177 {
2178         u32 stag = 0;
2179         u32 stag_index = 0;
2180         u32 next_stag_index;
2181         u32 driver_key;
2182         u32 random;
2183         u8 consumer_key;
2184         int ret;
2185
2186         get_random_bytes(&random, sizeof(random));
2187         consumer_key = (u8)random;
2188
2189         driver_key = random & ~iwdev->rf->mr_stagmask;
2190         next_stag_index = (random & iwdev->rf->mr_stagmask) >> 8;
2191         next_stag_index %= iwdev->rf->max_mr;
2192
2193         ret = irdma_alloc_rsrc(iwdev->rf, iwdev->rf->allocated_mrs,
2194                                iwdev->rf->max_mr, &stag_index,
2195                                &next_stag_index);
2196         if (ret)
2197                 return stag;
2198         stag = stag_index << IRDMA_CQPSQ_STAG_IDX_S;
2199         stag |= driver_key;
2200         stag += (u32)consumer_key;
2201
2202         return stag;
2203 }
2204
2205 /**
2206  * irdma_next_pbl_addr - Get next pbl address
2207  * @pbl: pointer to a pble
2208  * @pinfo: info pointer
2209  * @idx: index
2210  */
2211 static inline u64 *irdma_next_pbl_addr(u64 *pbl, struct irdma_pble_info **pinfo,
2212                                        u32 *idx)
2213 {
2214         *idx += 1;
2215         if (!(*pinfo) || *idx != (*pinfo)->cnt)
2216                 return ++pbl;
2217         *idx = 0;
2218         (*pinfo)++;
2219
2220         return (*pinfo)->addr;
2221 }
2222
2223 /**
2224  * irdma_copy_user_pgaddrs - copy user page address to pble's os locally
2225  * @iwmr: iwmr for IB's user page addresses
2226  * @pbl: ple pointer to save 1 level or 0 level pble
2227  * @level: indicated level 0, 1 or 2
2228  */
2229 static void irdma_copy_user_pgaddrs(struct irdma_mr *iwmr, u64 *pbl,
2230                                     enum irdma_pble_level level)
2231 {
2232         struct ib_umem *region = iwmr->region;
2233         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2234         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2235         struct irdma_pble_info *pinfo;
2236         struct ib_block_iter biter;
2237         u32 idx = 0;
2238         u32 pbl_cnt = 0;
2239
2240         pinfo = (level == PBLE_LEVEL_1) ? NULL : palloc->level2.leaf;
2241
2242         if (iwmr->type == IRDMA_MEMREG_TYPE_QP)
2243                 iwpbl->qp_mr.sq_page = sg_page(region->sgt_append.sgt.sgl);
2244
2245         rdma_umem_for_each_dma_block(region, &biter, iwmr->page_size) {
2246                 *pbl = rdma_block_iter_dma_address(&biter);
2247                 if (++pbl_cnt == palloc->total_cnt)
2248                         break;
2249                 pbl = irdma_next_pbl_addr(pbl, &pinfo, &idx);
2250         }
2251 }
2252
2253 /**
2254  * irdma_check_mem_contiguous - check if pbls stored in arr are contiguous
2255  * @arr: lvl1 pbl array
2256  * @npages: page count
2257  * @pg_size: page size
2258  *
2259  */
2260 static bool irdma_check_mem_contiguous(u64 *arr, u32 npages, u32 pg_size)
2261 {
2262         u32 pg_idx;
2263
2264         for (pg_idx = 0; pg_idx < npages; pg_idx++) {
2265                 if ((*arr + (pg_size * pg_idx)) != arr[pg_idx])
2266                         return false;
2267         }
2268
2269         return true;
2270 }
2271
2272 /**
2273  * irdma_check_mr_contiguous - check if MR is physically contiguous
2274  * @palloc: pbl allocation struct
2275  * @pg_size: page size
2276  */
2277 static bool irdma_check_mr_contiguous(struct irdma_pble_alloc *palloc,
2278                                       u32 pg_size)
2279 {
2280         struct irdma_pble_level2 *lvl2 = &palloc->level2;
2281         struct irdma_pble_info *leaf = lvl2->leaf;
2282         u64 *arr = NULL;
2283         u64 *start_addr = NULL;
2284         int i;
2285         bool ret;
2286
2287         if (palloc->level == PBLE_LEVEL_1) {
2288                 arr = palloc->level1.addr;
2289                 ret = irdma_check_mem_contiguous(arr, palloc->total_cnt,
2290                                                  pg_size);
2291                 return ret;
2292         }
2293
2294         start_addr = leaf->addr;
2295
2296         for (i = 0; i < lvl2->leaf_cnt; i++, leaf++) {
2297                 arr = leaf->addr;
2298                 if ((*start_addr + (i * pg_size * PBLE_PER_PAGE)) != *arr)
2299                         return false;
2300                 ret = irdma_check_mem_contiguous(arr, leaf->cnt, pg_size);
2301                 if (!ret)
2302                         return false;
2303         }
2304
2305         return true;
2306 }
2307
2308 /**
2309  * irdma_setup_pbles - copy user pg address to pble's
2310  * @rf: RDMA PCI function
2311  * @iwmr: mr pointer for this memory registration
2312  * @use_pbles: flag if to use pble's
2313  */
2314 static int irdma_setup_pbles(struct irdma_pci_f *rf, struct irdma_mr *iwmr,
2315                              bool use_pbles)
2316 {
2317         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2318         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2319         struct irdma_pble_info *pinfo;
2320         u64 *pbl;
2321         int status;
2322         enum irdma_pble_level level = PBLE_LEVEL_1;
2323
2324         if (use_pbles) {
2325                 status = irdma_get_pble(rf->pble_rsrc, palloc, iwmr->page_cnt,
2326                                         false);
2327                 if (status)
2328                         return status;
2329
2330                 iwpbl->pbl_allocated = true;
2331                 level = palloc->level;
2332                 pinfo = (level == PBLE_LEVEL_1) ? &palloc->level1 :
2333                                                   palloc->level2.leaf;
2334                 pbl = pinfo->addr;
2335         } else {
2336                 pbl = iwmr->pgaddrmem;
2337         }
2338
2339         irdma_copy_user_pgaddrs(iwmr, pbl, level);
2340
2341         if (use_pbles)
2342                 iwmr->pgaddrmem[0] = *pbl;
2343
2344         return 0;
2345 }
2346
2347 /**
2348  * irdma_handle_q_mem - handle memory for qp and cq
2349  * @iwdev: irdma device
2350  * @req: information for q memory management
2351  * @iwpbl: pble struct
2352  * @use_pbles: flag to use pble
2353  */
2354 static int irdma_handle_q_mem(struct irdma_device *iwdev,
2355                               struct irdma_mem_reg_req *req,
2356                               struct irdma_pbl *iwpbl, bool use_pbles)
2357 {
2358         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2359         struct irdma_mr *iwmr = iwpbl->iwmr;
2360         struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;
2361         struct irdma_cq_mr *cqmr = &iwpbl->cq_mr;
2362         struct irdma_hmc_pble *hmc_p;
2363         u64 *arr = iwmr->pgaddrmem;
2364         u32 pg_size, total;
2365         int err = 0;
2366         bool ret = true;
2367
2368         pg_size = iwmr->page_size;
2369         err = irdma_setup_pbles(iwdev->rf, iwmr, use_pbles);
2370         if (err)
2371                 return err;
2372
2373         if (use_pbles && palloc->level != PBLE_LEVEL_1) {
2374                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2375                 iwpbl->pbl_allocated = false;
2376                 return -ENOMEM;
2377         }
2378
2379         if (use_pbles)
2380                 arr = palloc->level1.addr;
2381
2382         switch (iwmr->type) {
2383         case IRDMA_MEMREG_TYPE_QP:
2384                 total = req->sq_pages + req->rq_pages;
2385                 hmc_p = &qpmr->sq_pbl;
2386                 qpmr->shadow = (dma_addr_t)arr[total];
2387
2388                 if (use_pbles) {
2389                         ret = irdma_check_mem_contiguous(arr, req->sq_pages,
2390                                                          pg_size);
2391                         if (ret)
2392                                 ret = irdma_check_mem_contiguous(&arr[req->sq_pages],
2393                                                                  req->rq_pages,
2394                                                                  pg_size);
2395                 }
2396
2397                 if (!ret) {
2398                         hmc_p->idx = palloc->level1.idx;
2399                         hmc_p = &qpmr->rq_pbl;
2400                         hmc_p->idx = palloc->level1.idx + req->sq_pages;
2401                 } else {
2402                         hmc_p->addr = arr[0];
2403                         hmc_p = &qpmr->rq_pbl;
2404                         hmc_p->addr = arr[req->sq_pages];
2405                 }
2406                 break;
2407         case IRDMA_MEMREG_TYPE_CQ:
2408                 hmc_p = &cqmr->cq_pbl;
2409
2410                 if (!cqmr->split)
2411                         cqmr->shadow = (dma_addr_t)arr[req->cq_pages];
2412
2413                 if (use_pbles)
2414                         ret = irdma_check_mem_contiguous(arr, req->cq_pages,
2415                                                          pg_size);
2416
2417                 if (!ret)
2418                         hmc_p->idx = palloc->level1.idx;
2419                 else
2420                         hmc_p->addr = arr[0];
2421         break;
2422         default:
2423                 ibdev_dbg(&iwdev->ibdev, "VERBS: MR type error\n");
2424                 err = -EINVAL;
2425         }
2426
2427         if (use_pbles && ret) {
2428                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2429                 iwpbl->pbl_allocated = false;
2430         }
2431
2432         return err;
2433 }
2434
2435 /**
2436  * irdma_hw_alloc_mw - create the hw memory window
2437  * @iwdev: irdma device
2438  * @iwmr: pointer to memory window info
2439  */
2440 static int irdma_hw_alloc_mw(struct irdma_device *iwdev, struct irdma_mr *iwmr)
2441 {
2442         struct irdma_mw_alloc_info *info;
2443         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2444         struct irdma_cqp_request *cqp_request;
2445         struct cqp_cmds_info *cqp_info;
2446         int status;
2447
2448         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2449         if (!cqp_request)
2450                 return -ENOMEM;
2451
2452         cqp_info = &cqp_request->info;
2453         info = &cqp_info->in.u.mw_alloc.info;
2454         memset(info, 0, sizeof(*info));
2455         if (iwmr->ibmw.type == IB_MW_TYPE_1)
2456                 info->mw_wide = true;
2457
2458         info->page_size = PAGE_SIZE;
2459         info->mw_stag_index = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2460         info->pd_id = iwpd->sc_pd.pd_id;
2461         info->remote_access = true;
2462         cqp_info->cqp_cmd = IRDMA_OP_MW_ALLOC;
2463         cqp_info->post_sq = 1;
2464         cqp_info->in.u.mw_alloc.dev = &iwdev->rf->sc_dev;
2465         cqp_info->in.u.mw_alloc.scratch = (uintptr_t)cqp_request;
2466         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2467         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2468
2469         return status;
2470 }
2471
2472 /**
2473  * irdma_alloc_mw - Allocate memory window
2474  * @ibmw: Memory Window
2475  * @udata: user data pointer
2476  */
2477 static int irdma_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata)
2478 {
2479         struct irdma_device *iwdev = to_iwdev(ibmw->device);
2480         struct irdma_mr *iwmr = to_iwmw(ibmw);
2481         int err_code;
2482         u32 stag;
2483
2484         stag = irdma_create_stag(iwdev);
2485         if (!stag)
2486                 return -ENOMEM;
2487
2488         iwmr->stag = stag;
2489         ibmw->rkey = stag;
2490
2491         err_code = irdma_hw_alloc_mw(iwdev, iwmr);
2492         if (err_code) {
2493                 irdma_free_stag(iwdev, stag);
2494                 return err_code;
2495         }
2496
2497         return 0;
2498 }
2499
2500 /**
2501  * irdma_dealloc_mw - Dealloc memory window
2502  * @ibmw: memory window structure.
2503  */
2504 static int irdma_dealloc_mw(struct ib_mw *ibmw)
2505 {
2506         struct ib_pd *ibpd = ibmw->pd;
2507         struct irdma_pd *iwpd = to_iwpd(ibpd);
2508         struct irdma_mr *iwmr = to_iwmr((struct ib_mr *)ibmw);
2509         struct irdma_device *iwdev = to_iwdev(ibmw->device);
2510         struct irdma_cqp_request *cqp_request;
2511         struct cqp_cmds_info *cqp_info;
2512         struct irdma_dealloc_stag_info *info;
2513
2514         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2515         if (!cqp_request)
2516                 return -ENOMEM;
2517
2518         cqp_info = &cqp_request->info;
2519         info = &cqp_info->in.u.dealloc_stag.info;
2520         memset(info, 0, sizeof(*info));
2521         info->pd_id = iwpd->sc_pd.pd_id;
2522         info->stag_idx = ibmw->rkey >> IRDMA_CQPSQ_STAG_IDX_S;
2523         info->mr = false;
2524         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
2525         cqp_info->post_sq = 1;
2526         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
2527         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
2528         irdma_handle_cqp_op(iwdev->rf, cqp_request);
2529         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2530         irdma_free_stag(iwdev, iwmr->stag);
2531
2532         return 0;
2533 }
2534
2535 /**
2536  * irdma_hw_alloc_stag - cqp command to allocate stag
2537  * @iwdev: irdma device
2538  * @iwmr: irdma mr pointer
2539  */
2540 static int irdma_hw_alloc_stag(struct irdma_device *iwdev,
2541                                struct irdma_mr *iwmr)
2542 {
2543         struct irdma_allocate_stag_info *info;
2544         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2545         int status;
2546         struct irdma_cqp_request *cqp_request;
2547         struct cqp_cmds_info *cqp_info;
2548
2549         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2550         if (!cqp_request)
2551                 return -ENOMEM;
2552
2553         cqp_info = &cqp_request->info;
2554         info = &cqp_info->in.u.alloc_stag.info;
2555         memset(info, 0, sizeof(*info));
2556         info->page_size = PAGE_SIZE;
2557         info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2558         info->pd_id = iwpd->sc_pd.pd_id;
2559         info->total_len = iwmr->len;
2560         info->remote_access = true;
2561         cqp_info->cqp_cmd = IRDMA_OP_ALLOC_STAG;
2562         cqp_info->post_sq = 1;
2563         cqp_info->in.u.alloc_stag.dev = &iwdev->rf->sc_dev;
2564         cqp_info->in.u.alloc_stag.scratch = (uintptr_t)cqp_request;
2565         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2566         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2567
2568         return status;
2569 }
2570
2571 /**
2572  * irdma_alloc_mr - register stag for fast memory registration
2573  * @pd: ibpd pointer
2574  * @mr_type: memory for stag registrion
2575  * @max_num_sg: man number of pages
2576  */
2577 static struct ib_mr *irdma_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
2578                                     u32 max_num_sg)
2579 {
2580         struct irdma_device *iwdev = to_iwdev(pd->device);
2581         struct irdma_pble_alloc *palloc;
2582         struct irdma_pbl *iwpbl;
2583         struct irdma_mr *iwmr;
2584         u32 stag;
2585         int err_code;
2586
2587         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2588         if (!iwmr)
2589                 return ERR_PTR(-ENOMEM);
2590
2591         stag = irdma_create_stag(iwdev);
2592         if (!stag) {
2593                 err_code = -ENOMEM;
2594                 goto err;
2595         }
2596
2597         iwmr->stag = stag;
2598         iwmr->ibmr.rkey = stag;
2599         iwmr->ibmr.lkey = stag;
2600         iwmr->ibmr.pd = pd;
2601         iwmr->ibmr.device = pd->device;
2602         iwpbl = &iwmr->iwpbl;
2603         iwpbl->iwmr = iwmr;
2604         iwmr->type = IRDMA_MEMREG_TYPE_MEM;
2605         palloc = &iwpbl->pble_alloc;
2606         iwmr->page_cnt = max_num_sg;
2607         err_code = irdma_get_pble(iwdev->rf->pble_rsrc, palloc, iwmr->page_cnt,
2608                                   false);
2609         if (err_code)
2610                 goto err_get_pble;
2611
2612         err_code = irdma_hw_alloc_stag(iwdev, iwmr);
2613         if (err_code)
2614                 goto err_alloc_stag;
2615
2616         iwpbl->pbl_allocated = true;
2617
2618         return &iwmr->ibmr;
2619 err_alloc_stag:
2620         irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2621 err_get_pble:
2622         irdma_free_stag(iwdev, stag);
2623 err:
2624         kfree(iwmr);
2625
2626         return ERR_PTR(err_code);
2627 }
2628
2629 /**
2630  * irdma_set_page - populate pbl list for fmr
2631  * @ibmr: ib mem to access iwarp mr pointer
2632  * @addr: page dma address fro pbl list
2633  */
2634 static int irdma_set_page(struct ib_mr *ibmr, u64 addr)
2635 {
2636         struct irdma_mr *iwmr = to_iwmr(ibmr);
2637         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2638         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2639         u64 *pbl;
2640
2641         if (unlikely(iwmr->npages == iwmr->page_cnt))
2642                 return -ENOMEM;
2643
2644         if (palloc->level == PBLE_LEVEL_2) {
2645                 struct irdma_pble_info *palloc_info =
2646                         palloc->level2.leaf + (iwmr->npages >> PBLE_512_SHIFT);
2647
2648                 palloc_info->addr[iwmr->npages & (PBLE_PER_PAGE - 1)] = addr;
2649         } else {
2650                 pbl = palloc->level1.addr;
2651                 pbl[iwmr->npages] = addr;
2652         }
2653         iwmr->npages++;
2654
2655         return 0;
2656 }
2657
2658 /**
2659  * irdma_map_mr_sg - map of sg list for fmr
2660  * @ibmr: ib mem to access iwarp mr pointer
2661  * @sg: scatter gather list
2662  * @sg_nents: number of sg pages
2663  * @sg_offset: scatter gather list for fmr
2664  */
2665 static int irdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
2666                            int sg_nents, unsigned int *sg_offset)
2667 {
2668         struct irdma_mr *iwmr = to_iwmr(ibmr);
2669
2670         iwmr->npages = 0;
2671
2672         return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, irdma_set_page);
2673 }
2674
2675 /**
2676  * irdma_hwreg_mr - send cqp command for memory registration
2677  * @iwdev: irdma device
2678  * @iwmr: irdma mr pointer
2679  * @access: access for MR
2680  */
2681 static int irdma_hwreg_mr(struct irdma_device *iwdev, struct irdma_mr *iwmr,
2682                           u16 access)
2683 {
2684         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2685         struct irdma_reg_ns_stag_info *stag_info;
2686         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2687         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2688         struct irdma_cqp_request *cqp_request;
2689         struct cqp_cmds_info *cqp_info;
2690         int ret;
2691
2692         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2693         if (!cqp_request)
2694                 return -ENOMEM;
2695
2696         cqp_info = &cqp_request->info;
2697         stag_info = &cqp_info->in.u.mr_reg_non_shared.info;
2698         memset(stag_info, 0, sizeof(*stag_info));
2699         stag_info->va = iwpbl->user_base;
2700         stag_info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2701         stag_info->stag_key = (u8)iwmr->stag;
2702         stag_info->total_len = iwmr->len;
2703         stag_info->access_rights = irdma_get_mr_access(access);
2704         stag_info->pd_id = iwpd->sc_pd.pd_id;
2705         if (stag_info->access_rights & IRDMA_ACCESS_FLAGS_ZERO_BASED)
2706                 stag_info->addr_type = IRDMA_ADDR_TYPE_ZERO_BASED;
2707         else
2708                 stag_info->addr_type = IRDMA_ADDR_TYPE_VA_BASED;
2709         stag_info->page_size = iwmr->page_size;
2710
2711         if (iwpbl->pbl_allocated) {
2712                 if (palloc->level == PBLE_LEVEL_1) {
2713                         stag_info->first_pm_pbl_index = palloc->level1.idx;
2714                         stag_info->chunk_size = 1;
2715                 } else {
2716                         stag_info->first_pm_pbl_index = palloc->level2.root.idx;
2717                         stag_info->chunk_size = 3;
2718                 }
2719         } else {
2720                 stag_info->reg_addr_pa = iwmr->pgaddrmem[0];
2721         }
2722
2723         cqp_info->cqp_cmd = IRDMA_OP_MR_REG_NON_SHARED;
2724         cqp_info->post_sq = 1;
2725         cqp_info->in.u.mr_reg_non_shared.dev = &iwdev->rf->sc_dev;
2726         cqp_info->in.u.mr_reg_non_shared.scratch = (uintptr_t)cqp_request;
2727         ret = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2728         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2729
2730         return ret;
2731 }
2732
2733 /**
2734  * irdma_reg_user_mr - Register a user memory region
2735  * @pd: ptr of pd
2736  * @start: virtual start address
2737  * @len: length of mr
2738  * @virt: virtual address
2739  * @access: access of mr
2740  * @udata: user data
2741  */
2742 static struct ib_mr *irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,
2743                                        u64 virt, int access,
2744                                        struct ib_udata *udata)
2745 {
2746         struct irdma_device *iwdev = to_iwdev(pd->device);
2747         struct irdma_ucontext *ucontext;
2748         struct irdma_pble_alloc *palloc;
2749         struct irdma_pbl *iwpbl;
2750         struct irdma_mr *iwmr;
2751         struct ib_umem *region;
2752         struct irdma_mem_reg_req req;
2753         u32 total, stag = 0;
2754         u8 shadow_pgcnt = 1;
2755         bool use_pbles = false;
2756         unsigned long flags;
2757         int err = -EINVAL;
2758         int ret;
2759
2760         if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size)
2761                 return ERR_PTR(-EINVAL);
2762
2763         region = ib_umem_get(pd->device, start, len, access);
2764
2765         if (IS_ERR(region)) {
2766                 ibdev_dbg(&iwdev->ibdev,
2767                           "VERBS: Failed to create ib_umem region\n");
2768                 return (struct ib_mr *)region;
2769         }
2770
2771         if (ib_copy_from_udata(&req, udata, min(sizeof(req), udata->inlen))) {
2772                 ib_umem_release(region);
2773                 return ERR_PTR(-EFAULT);
2774         }
2775
2776         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2777         if (!iwmr) {
2778                 ib_umem_release(region);
2779                 return ERR_PTR(-ENOMEM);
2780         }
2781
2782         iwpbl = &iwmr->iwpbl;
2783         iwpbl->iwmr = iwmr;
2784         iwmr->region = region;
2785         iwmr->ibmr.pd = pd;
2786         iwmr->ibmr.device = pd->device;
2787         iwmr->ibmr.iova = virt;
2788         iwmr->page_size = PAGE_SIZE;
2789
2790         if (req.reg_type == IRDMA_MEMREG_TYPE_MEM) {
2791                 iwmr->page_size = ib_umem_find_best_pgsz(region,
2792                                                          iwdev->rf->sc_dev.hw_attrs.page_size_cap,
2793                                                          virt);
2794                 if (unlikely(!iwmr->page_size)) {
2795                         kfree(iwmr);
2796                         ib_umem_release(region);
2797                         return ERR_PTR(-EOPNOTSUPP);
2798                 }
2799         }
2800         iwmr->len = region->length;
2801         iwpbl->user_base = virt;
2802         palloc = &iwpbl->pble_alloc;
2803         iwmr->type = req.reg_type;
2804         iwmr->page_cnt = ib_umem_num_dma_blocks(region, iwmr->page_size);
2805
2806         switch (req.reg_type) {
2807         case IRDMA_MEMREG_TYPE_QP:
2808                 total = req.sq_pages + req.rq_pages + shadow_pgcnt;
2809                 if (total > iwmr->page_cnt) {
2810                         err = -EINVAL;
2811                         goto error;
2812                 }
2813                 total = req.sq_pages + req.rq_pages;
2814                 use_pbles = (total > 2);
2815                 err = irdma_handle_q_mem(iwdev, &req, iwpbl, use_pbles);
2816                 if (err)
2817                         goto error;
2818
2819                 ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext,
2820                                                      ibucontext);
2821                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
2822                 list_add_tail(&iwpbl->list, &ucontext->qp_reg_mem_list);
2823                 iwpbl->on_list = true;
2824                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
2825                 break;
2826         case IRDMA_MEMREG_TYPE_CQ:
2827                 if (iwdev->rf->sc_dev.hw_attrs.uk_attrs.feature_flags & IRDMA_FEATURE_CQ_RESIZE)
2828                         shadow_pgcnt = 0;
2829                 total = req.cq_pages + shadow_pgcnt;
2830                 if (total > iwmr->page_cnt) {
2831                         err = -EINVAL;
2832                         goto error;
2833                 }
2834
2835                 use_pbles = (req.cq_pages > 1);
2836                 err = irdma_handle_q_mem(iwdev, &req, iwpbl, use_pbles);
2837                 if (err)
2838                         goto error;
2839
2840                 ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext,
2841                                                      ibucontext);
2842                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2843                 list_add_tail(&iwpbl->list, &ucontext->cq_reg_mem_list);
2844                 iwpbl->on_list = true;
2845                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2846                 break;
2847         case IRDMA_MEMREG_TYPE_MEM:
2848                 use_pbles = (iwmr->page_cnt != 1);
2849
2850                 err = irdma_setup_pbles(iwdev->rf, iwmr, use_pbles);
2851                 if (err)
2852                         goto error;
2853
2854                 if (use_pbles) {
2855                         ret = irdma_check_mr_contiguous(palloc,
2856                                                         iwmr->page_size);
2857                         if (ret) {
2858                                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2859                                 iwpbl->pbl_allocated = false;
2860                         }
2861                 }
2862
2863                 stag = irdma_create_stag(iwdev);
2864                 if (!stag) {
2865                         err = -ENOMEM;
2866                         goto error;
2867                 }
2868
2869                 iwmr->stag = stag;
2870                 iwmr->ibmr.rkey = stag;
2871                 iwmr->ibmr.lkey = stag;
2872                 err = irdma_hwreg_mr(iwdev, iwmr, access);
2873                 if (err) {
2874                         irdma_free_stag(iwdev, stag);
2875                         goto error;
2876                 }
2877
2878                 break;
2879         default:
2880                 goto error;
2881         }
2882
2883         iwmr->type = req.reg_type;
2884
2885         return &iwmr->ibmr;
2886
2887 error:
2888         if (palloc->level != PBLE_LEVEL_0 && iwpbl->pbl_allocated)
2889                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2890         ib_umem_release(region);
2891         kfree(iwmr);
2892
2893         return ERR_PTR(err);
2894 }
2895
2896 /**
2897  * irdma_reg_phys_mr - register kernel physical memory
2898  * @pd: ibpd pointer
2899  * @addr: physical address of memory to register
2900  * @size: size of memory to register
2901  * @access: Access rights
2902  * @iova_start: start of virtual address for physical buffers
2903  */
2904 struct ib_mr *irdma_reg_phys_mr(struct ib_pd *pd, u64 addr, u64 size, int access,
2905                                 u64 *iova_start)
2906 {
2907         struct irdma_device *iwdev = to_iwdev(pd->device);
2908         struct irdma_pbl *iwpbl;
2909         struct irdma_mr *iwmr;
2910         u32 stag;
2911         int ret;
2912
2913         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2914         if (!iwmr)
2915                 return ERR_PTR(-ENOMEM);
2916
2917         iwmr->ibmr.pd = pd;
2918         iwmr->ibmr.device = pd->device;
2919         iwpbl = &iwmr->iwpbl;
2920         iwpbl->iwmr = iwmr;
2921         iwmr->type = IRDMA_MEMREG_TYPE_MEM;
2922         iwpbl->user_base = *iova_start;
2923         stag = irdma_create_stag(iwdev);
2924         if (!stag) {
2925                 ret = -ENOMEM;
2926                 goto err;
2927         }
2928
2929         iwmr->stag = stag;
2930         iwmr->ibmr.iova = *iova_start;
2931         iwmr->ibmr.rkey = stag;
2932         iwmr->ibmr.lkey = stag;
2933         iwmr->page_cnt = 1;
2934         iwmr->pgaddrmem[0] = addr;
2935         iwmr->len = size;
2936         iwmr->page_size = SZ_4K;
2937         ret = irdma_hwreg_mr(iwdev, iwmr, access);
2938         if (ret) {
2939                 irdma_free_stag(iwdev, stag);
2940                 goto err;
2941         }
2942
2943         return &iwmr->ibmr;
2944
2945 err:
2946         kfree(iwmr);
2947
2948         return ERR_PTR(ret);
2949 }
2950
2951 /**
2952  * irdma_get_dma_mr - register physical mem
2953  * @pd: ptr of pd
2954  * @acc: access for memory
2955  */
2956 static struct ib_mr *irdma_get_dma_mr(struct ib_pd *pd, int acc)
2957 {
2958         u64 kva = 0;
2959
2960         return irdma_reg_phys_mr(pd, 0, 0, acc, &kva);
2961 }
2962
2963 /**
2964  * irdma_del_memlist - Deleting pbl list entries for CQ/QP
2965  * @iwmr: iwmr for IB's user page addresses
2966  * @ucontext: ptr to user context
2967  */
2968 static void irdma_del_memlist(struct irdma_mr *iwmr,
2969                               struct irdma_ucontext *ucontext)
2970 {
2971         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2972         unsigned long flags;
2973
2974         switch (iwmr->type) {
2975         case IRDMA_MEMREG_TYPE_CQ:
2976                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2977                 if (iwpbl->on_list) {
2978                         iwpbl->on_list = false;
2979                         list_del(&iwpbl->list);
2980                 }
2981                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2982                 break;
2983         case IRDMA_MEMREG_TYPE_QP:
2984                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
2985                 if (iwpbl->on_list) {
2986                         iwpbl->on_list = false;
2987                         list_del(&iwpbl->list);
2988                 }
2989                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
2990                 break;
2991         default:
2992                 break;
2993         }
2994 }
2995
2996 /**
2997  * irdma_dereg_mr - deregister mr
2998  * @ib_mr: mr ptr for dereg
2999  * @udata: user data
3000  */
3001 static int irdma_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
3002 {
3003         struct ib_pd *ibpd = ib_mr->pd;
3004         struct irdma_pd *iwpd = to_iwpd(ibpd);
3005         struct irdma_mr *iwmr = to_iwmr(ib_mr);
3006         struct irdma_device *iwdev = to_iwdev(ib_mr->device);
3007         struct irdma_dealloc_stag_info *info;
3008         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
3009         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
3010         struct irdma_cqp_request *cqp_request;
3011         struct cqp_cmds_info *cqp_info;
3012
3013         if (iwmr->type != IRDMA_MEMREG_TYPE_MEM) {
3014                 if (iwmr->region) {
3015                         struct irdma_ucontext *ucontext;
3016
3017                         ucontext = rdma_udata_to_drv_context(udata,
3018                                                 struct irdma_ucontext,
3019                                                 ibucontext);
3020                         irdma_del_memlist(iwmr, ucontext);
3021                 }
3022                 goto done;
3023         }
3024
3025         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
3026         if (!cqp_request)
3027                 return -ENOMEM;
3028
3029         cqp_info = &cqp_request->info;
3030         info = &cqp_info->in.u.dealloc_stag.info;
3031         memset(info, 0, sizeof(*info));
3032         info->pd_id = iwpd->sc_pd.pd_id;
3033         info->stag_idx = ib_mr->rkey >> IRDMA_CQPSQ_STAG_IDX_S;
3034         info->mr = true;
3035         if (iwpbl->pbl_allocated)
3036                 info->dealloc_pbl = true;
3037
3038         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
3039         cqp_info->post_sq = 1;
3040         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
3041         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
3042         irdma_handle_cqp_op(iwdev->rf, cqp_request);
3043         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
3044         irdma_free_stag(iwdev, iwmr->stag);
3045 done:
3046         if (iwpbl->pbl_allocated)
3047                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
3048         ib_umem_release(iwmr->region);
3049         kfree(iwmr);
3050
3051         return 0;
3052 }
3053
3054 /**
3055  * irdma_post_send -  kernel application wr
3056  * @ibqp: qp ptr for wr
3057  * @ib_wr: work request ptr
3058  * @bad_wr: return of bad wr if err
3059  */
3060 static int irdma_post_send(struct ib_qp *ibqp,
3061                            const struct ib_send_wr *ib_wr,
3062                            const struct ib_send_wr **bad_wr)
3063 {
3064         struct irdma_qp *iwqp;
3065         struct irdma_qp_uk *ukqp;
3066         struct irdma_sc_dev *dev;
3067         struct irdma_post_sq_info info;
3068         int err = 0;
3069         unsigned long flags;
3070         bool inv_stag;
3071         struct irdma_ah *ah;
3072
3073         iwqp = to_iwqp(ibqp);
3074         ukqp = &iwqp->sc_qp.qp_uk;
3075         dev = &iwqp->iwdev->rf->sc_dev;
3076
3077         spin_lock_irqsave(&iwqp->lock, flags);
3078         while (ib_wr) {
3079                 memset(&info, 0, sizeof(info));
3080                 inv_stag = false;
3081                 info.wr_id = (ib_wr->wr_id);
3082                 if ((ib_wr->send_flags & IB_SEND_SIGNALED) || iwqp->sig_all)
3083                         info.signaled = true;
3084                 if (ib_wr->send_flags & IB_SEND_FENCE)
3085                         info.read_fence = true;
3086                 switch (ib_wr->opcode) {
3087                 case IB_WR_SEND_WITH_IMM:
3088                         if (ukqp->qp_caps & IRDMA_SEND_WITH_IMM) {
3089                                 info.imm_data_valid = true;
3090                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
3091                         } else {
3092                                 err = -EINVAL;
3093                                 break;
3094                         }
3095                         fallthrough;
3096                 case IB_WR_SEND:
3097                 case IB_WR_SEND_WITH_INV:
3098                         if (ib_wr->opcode == IB_WR_SEND ||
3099                             ib_wr->opcode == IB_WR_SEND_WITH_IMM) {
3100                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
3101                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL;
3102                                 else
3103                                         info.op_type = IRDMA_OP_TYPE_SEND;
3104                         } else {
3105                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
3106                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL_INV;
3107                                 else
3108                                         info.op_type = IRDMA_OP_TYPE_SEND_INV;
3109                                 info.stag_to_inv = ib_wr->ex.invalidate_rkey;
3110                         }
3111
3112                         if (ib_wr->send_flags & IB_SEND_INLINE) {
3113                                 info.op.inline_send.data = (void *)(unsigned long)
3114                                                            ib_wr->sg_list[0].addr;
3115                                 info.op.inline_send.len = ib_wr->sg_list[0].length;
3116                                 if (iwqp->ibqp.qp_type == IB_QPT_UD ||
3117                                     iwqp->ibqp.qp_type == IB_QPT_GSI) {
3118                                         ah = to_iwah(ud_wr(ib_wr)->ah);
3119                                         info.op.inline_send.ah_id = ah->sc_ah.ah_info.ah_idx;
3120                                         info.op.inline_send.qkey = ud_wr(ib_wr)->remote_qkey;
3121                                         info.op.inline_send.dest_qp = ud_wr(ib_wr)->remote_qpn;
3122                                 }
3123                                 err = irdma_uk_inline_send(ukqp, &info, false);
3124                         } else {
3125                                 info.op.send.num_sges = ib_wr->num_sge;
3126                                 info.op.send.sg_list = ib_wr->sg_list;
3127                                 if (iwqp->ibqp.qp_type == IB_QPT_UD ||
3128                                     iwqp->ibqp.qp_type == IB_QPT_GSI) {
3129                                         ah = to_iwah(ud_wr(ib_wr)->ah);
3130                                         info.op.send.ah_id = ah->sc_ah.ah_info.ah_idx;
3131                                         info.op.send.qkey = ud_wr(ib_wr)->remote_qkey;
3132                                         info.op.send.dest_qp = ud_wr(ib_wr)->remote_qpn;
3133                                 }
3134                                 err = irdma_uk_send(ukqp, &info, false);
3135                         }
3136                         break;
3137                 case IB_WR_RDMA_WRITE_WITH_IMM:
3138                         if (ukqp->qp_caps & IRDMA_WRITE_WITH_IMM) {
3139                                 info.imm_data_valid = true;
3140                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
3141                         } else {
3142                                 err = -EINVAL;
3143                                 break;
3144                         }
3145                         fallthrough;
3146                 case IB_WR_RDMA_WRITE:
3147                         if (ib_wr->send_flags & IB_SEND_SOLICITED)
3148                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE_SOL;
3149                         else
3150                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE;
3151
3152                         if (ib_wr->send_flags & IB_SEND_INLINE) {
3153                                 info.op.inline_rdma_write.data = (void *)(uintptr_t)ib_wr->sg_list[0].addr;
3154                                 info.op.inline_rdma_write.len =
3155                                                 ib_wr->sg_list[0].length;
3156                                 info.op.inline_rdma_write.rem_addr.addr =
3157                                                 rdma_wr(ib_wr)->remote_addr;
3158                                 info.op.inline_rdma_write.rem_addr.lkey =
3159                                                 rdma_wr(ib_wr)->rkey;
3160                                 err = irdma_uk_inline_rdma_write(ukqp, &info, false);
3161                         } else {
3162                                 info.op.rdma_write.lo_sg_list = (void *)ib_wr->sg_list;
3163                                 info.op.rdma_write.num_lo_sges = ib_wr->num_sge;
3164                                 info.op.rdma_write.rem_addr.addr = rdma_wr(ib_wr)->remote_addr;
3165                                 info.op.rdma_write.rem_addr.lkey = rdma_wr(ib_wr)->rkey;
3166                                 err = irdma_uk_rdma_write(ukqp, &info, false);
3167                         }
3168                         break;
3169                 case IB_WR_RDMA_READ_WITH_INV:
3170                         inv_stag = true;
3171                         fallthrough;
3172                 case IB_WR_RDMA_READ:
3173                         if (ib_wr->num_sge >
3174                             dev->hw_attrs.uk_attrs.max_hw_read_sges) {
3175                                 err = -EINVAL;
3176                                 break;
3177                         }
3178                         info.op_type = IRDMA_OP_TYPE_RDMA_READ;
3179                         info.op.rdma_read.rem_addr.addr = rdma_wr(ib_wr)->remote_addr;
3180                         info.op.rdma_read.rem_addr.lkey = rdma_wr(ib_wr)->rkey;
3181                         info.op.rdma_read.lo_sg_list = (void *)ib_wr->sg_list;
3182                         info.op.rdma_read.num_lo_sges = ib_wr->num_sge;
3183                         err = irdma_uk_rdma_read(ukqp, &info, inv_stag, false);
3184                         break;
3185                 case IB_WR_LOCAL_INV:
3186                         info.op_type = IRDMA_OP_TYPE_INV_STAG;
3187                         info.op.inv_local_stag.target_stag = ib_wr->ex.invalidate_rkey;
3188                         err = irdma_uk_stag_local_invalidate(ukqp, &info, true);
3189                         break;
3190                 case IB_WR_REG_MR: {
3191                         struct irdma_mr *iwmr = to_iwmr(reg_wr(ib_wr)->mr);
3192                         struct irdma_pble_alloc *palloc = &iwmr->iwpbl.pble_alloc;
3193                         struct irdma_fast_reg_stag_info stag_info = {};
3194
3195                         stag_info.signaled = info.signaled;
3196                         stag_info.read_fence = info.read_fence;
3197                         stag_info.access_rights = irdma_get_mr_access(reg_wr(ib_wr)->access);
3198                         stag_info.stag_key = reg_wr(ib_wr)->key & 0xff;
3199                         stag_info.stag_idx = reg_wr(ib_wr)->key >> 8;
3200                         stag_info.page_size = reg_wr(ib_wr)->mr->page_size;
3201                         stag_info.wr_id = ib_wr->wr_id;
3202                         stag_info.addr_type = IRDMA_ADDR_TYPE_VA_BASED;
3203                         stag_info.va = (void *)(uintptr_t)iwmr->ibmr.iova;
3204                         stag_info.total_len = iwmr->ibmr.length;
3205                         stag_info.reg_addr_pa = *palloc->level1.addr;
3206                         stag_info.first_pm_pbl_index = palloc->level1.idx;
3207                         stag_info.local_fence = ib_wr->send_flags & IB_SEND_FENCE;
3208                         if (iwmr->npages > IRDMA_MIN_PAGES_PER_FMR)
3209                                 stag_info.chunk_size = 1;
3210                         err = irdma_sc_mr_fast_register(&iwqp->sc_qp, &stag_info,
3211                                                         true);
3212                         break;
3213                 }
3214                 default:
3215                         err = -EINVAL;
3216                         ibdev_dbg(&iwqp->iwdev->ibdev,
3217                                   "VERBS: upost_send bad opcode = 0x%x\n",
3218                                   ib_wr->opcode);
3219                         break;
3220                 }
3221
3222                 if (err)
3223                         break;
3224                 ib_wr = ib_wr->next;
3225         }
3226
3227         if (!iwqp->flush_issued) {
3228                 if (iwqp->hw_iwarp_state <= IRDMA_QP_STATE_RTS)
3229                         irdma_uk_qp_post_wr(ukqp);
3230                 spin_unlock_irqrestore(&iwqp->lock, flags);
3231         } else {
3232                 spin_unlock_irqrestore(&iwqp->lock, flags);
3233                 mod_delayed_work(iwqp->iwdev->cleanup_wq, &iwqp->dwork_flush,
3234                                  msecs_to_jiffies(IRDMA_FLUSH_DELAY_MS));
3235         }
3236         if (err)
3237                 *bad_wr = ib_wr;
3238
3239         return err;
3240 }
3241
3242 /**
3243  * irdma_post_recv - post receive wr for kernel application
3244  * @ibqp: ib qp pointer
3245  * @ib_wr: work request for receive
3246  * @bad_wr: bad wr caused an error
3247  */
3248 static int irdma_post_recv(struct ib_qp *ibqp,
3249                            const struct ib_recv_wr *ib_wr,
3250                            const struct ib_recv_wr **bad_wr)
3251 {
3252         struct irdma_qp *iwqp;
3253         struct irdma_qp_uk *ukqp;
3254         struct irdma_post_rq_info post_recv = {};
3255         unsigned long flags;
3256         int err = 0;
3257
3258         iwqp = to_iwqp(ibqp);
3259         ukqp = &iwqp->sc_qp.qp_uk;
3260
3261         spin_lock_irqsave(&iwqp->lock, flags);
3262         while (ib_wr) {
3263                 post_recv.num_sges = ib_wr->num_sge;
3264                 post_recv.wr_id = ib_wr->wr_id;
3265                 post_recv.sg_list = ib_wr->sg_list;
3266                 err = irdma_uk_post_receive(ukqp, &post_recv);
3267                 if (err) {
3268                         ibdev_dbg(&iwqp->iwdev->ibdev,
3269                                   "VERBS: post_recv err %d\n", err);
3270                         goto out;
3271                 }
3272
3273                 ib_wr = ib_wr->next;
3274         }
3275
3276 out:
3277         spin_unlock_irqrestore(&iwqp->lock, flags);
3278         if (iwqp->flush_issued)
3279                 mod_delayed_work(iwqp->iwdev->cleanup_wq, &iwqp->dwork_flush,
3280                                  msecs_to_jiffies(IRDMA_FLUSH_DELAY_MS));
3281
3282         if (err)
3283                 *bad_wr = ib_wr;
3284
3285         return err;
3286 }
3287
3288 /**
3289  * irdma_flush_err_to_ib_wc_status - return change flush error code to IB status
3290  * @opcode: iwarp flush code
3291  */
3292 static enum ib_wc_status irdma_flush_err_to_ib_wc_status(enum irdma_flush_opcode opcode)
3293 {
3294         switch (opcode) {
3295         case FLUSH_PROT_ERR:
3296                 return IB_WC_LOC_PROT_ERR;
3297         case FLUSH_REM_ACCESS_ERR:
3298                 return IB_WC_REM_ACCESS_ERR;
3299         case FLUSH_LOC_QP_OP_ERR:
3300                 return IB_WC_LOC_QP_OP_ERR;
3301         case FLUSH_REM_OP_ERR:
3302                 return IB_WC_REM_OP_ERR;
3303         case FLUSH_LOC_LEN_ERR:
3304                 return IB_WC_LOC_LEN_ERR;
3305         case FLUSH_GENERAL_ERR:
3306                 return IB_WC_WR_FLUSH_ERR;
3307         case FLUSH_RETRY_EXC_ERR:
3308                 return IB_WC_RETRY_EXC_ERR;
3309         case FLUSH_MW_BIND_ERR:
3310                 return IB_WC_MW_BIND_ERR;
3311         case FLUSH_FATAL_ERR:
3312         default:
3313                 return IB_WC_FATAL_ERR;
3314         }
3315 }
3316
3317 /**
3318  * irdma_process_cqe - process cqe info
3319  * @entry: processed cqe
3320  * @cq_poll_info: cqe info
3321  */
3322 static void irdma_process_cqe(struct ib_wc *entry,
3323                               struct irdma_cq_poll_info *cq_poll_info)
3324 {
3325         struct irdma_qp *iwqp;
3326         struct irdma_sc_qp *qp;
3327
3328         entry->wc_flags = 0;
3329         entry->pkey_index = 0;
3330         entry->wr_id = cq_poll_info->wr_id;
3331
3332         qp = cq_poll_info->qp_handle;
3333         iwqp = qp->qp_uk.back_qp;
3334         entry->qp = qp->qp_uk.back_qp;
3335
3336         if (cq_poll_info->error) {
3337                 entry->status = (cq_poll_info->comp_status == IRDMA_COMPL_STATUS_FLUSHED) ?
3338                                 irdma_flush_err_to_ib_wc_status(cq_poll_info->minor_err) : IB_WC_GENERAL_ERR;
3339
3340                 entry->vendor_err = cq_poll_info->major_err << 16 |
3341                                     cq_poll_info->minor_err;
3342         } else {
3343                 entry->status = IB_WC_SUCCESS;
3344                 if (cq_poll_info->imm_valid) {
3345                         entry->ex.imm_data = htonl(cq_poll_info->imm_data);
3346                         entry->wc_flags |= IB_WC_WITH_IMM;
3347                 }
3348                 if (cq_poll_info->ud_smac_valid) {
3349                         ether_addr_copy(entry->smac, cq_poll_info->ud_smac);
3350                         entry->wc_flags |= IB_WC_WITH_SMAC;
3351                 }
3352
3353                 if (cq_poll_info->ud_vlan_valid) {
3354                         u16 vlan = cq_poll_info->ud_vlan & VLAN_VID_MASK;
3355
3356                         entry->sl = cq_poll_info->ud_vlan >> VLAN_PRIO_SHIFT;
3357                         if (vlan) {
3358                                 entry->vlan_id = vlan;
3359                                 entry->wc_flags |= IB_WC_WITH_VLAN;
3360                         }
3361                 } else {
3362                         entry->sl = 0;
3363                 }
3364         }
3365
3366         switch (cq_poll_info->op_type) {
3367         case IRDMA_OP_TYPE_RDMA_WRITE:
3368         case IRDMA_OP_TYPE_RDMA_WRITE_SOL:
3369                 entry->opcode = IB_WC_RDMA_WRITE;
3370                 break;
3371         case IRDMA_OP_TYPE_RDMA_READ_INV_STAG:
3372         case IRDMA_OP_TYPE_RDMA_READ:
3373                 entry->opcode = IB_WC_RDMA_READ;
3374                 break;
3375         case IRDMA_OP_TYPE_SEND_INV:
3376         case IRDMA_OP_TYPE_SEND_SOL:
3377         case IRDMA_OP_TYPE_SEND_SOL_INV:
3378         case IRDMA_OP_TYPE_SEND:
3379                 entry->opcode = IB_WC_SEND;
3380                 break;
3381         case IRDMA_OP_TYPE_FAST_REG_NSMR:
3382                 entry->opcode = IB_WC_REG_MR;
3383                 break;
3384         case IRDMA_OP_TYPE_INV_STAG:
3385                 entry->opcode = IB_WC_LOCAL_INV;
3386                 break;
3387         case IRDMA_OP_TYPE_REC_IMM:
3388         case IRDMA_OP_TYPE_REC:
3389                 entry->opcode = cq_poll_info->op_type == IRDMA_OP_TYPE_REC_IMM ?
3390                         IB_WC_RECV_RDMA_WITH_IMM : IB_WC_RECV;
3391                 if (qp->qp_uk.qp_type != IRDMA_QP_TYPE_ROCE_UD &&
3392                     cq_poll_info->stag_invalid_set) {
3393                         entry->ex.invalidate_rkey = cq_poll_info->inv_stag;
3394                         entry->wc_flags |= IB_WC_WITH_INVALIDATE;
3395                 }
3396                 break;
3397         default:
3398                 ibdev_err(&iwqp->iwdev->ibdev,
3399                           "Invalid opcode = %d in CQE\n", cq_poll_info->op_type);
3400                 entry->status = IB_WC_GENERAL_ERR;
3401                 return;
3402         }
3403
3404         if (qp->qp_uk.qp_type == IRDMA_QP_TYPE_ROCE_UD) {
3405                 entry->src_qp = cq_poll_info->ud_src_qpn;
3406                 entry->slid = 0;
3407                 entry->wc_flags |=
3408                         (IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE);
3409                 entry->network_hdr_type = cq_poll_info->ipv4 ?
3410                                                   RDMA_NETWORK_IPV4 :
3411                                                   RDMA_NETWORK_IPV6;
3412         } else {
3413                 entry->src_qp = cq_poll_info->qp_id;
3414         }
3415
3416         entry->byte_len = cq_poll_info->bytes_xfered;
3417 }
3418
3419 /**
3420  * irdma_poll_one - poll one entry of the CQ
3421  * @ukcq: ukcq to poll
3422  * @cur_cqe: current CQE info to be filled in
3423  * @entry: ibv_wc object to be filled for non-extended CQ or NULL for extended CQ
3424  *
3425  * Returns the internal irdma device error code or 0 on success
3426  */
3427 static inline int irdma_poll_one(struct irdma_cq_uk *ukcq,
3428                                  struct irdma_cq_poll_info *cur_cqe,
3429                                  struct ib_wc *entry)
3430 {
3431         int ret = irdma_uk_cq_poll_cmpl(ukcq, cur_cqe);
3432
3433         if (ret)
3434                 return ret;
3435
3436         irdma_process_cqe(entry, cur_cqe);
3437
3438         return 0;
3439 }
3440
3441 /**
3442  * __irdma_poll_cq - poll cq for completion (kernel apps)
3443  * @iwcq: cq to poll
3444  * @num_entries: number of entries to poll
3445  * @entry: wr of a completed entry
3446  */
3447 static int __irdma_poll_cq(struct irdma_cq *iwcq, int num_entries, struct ib_wc *entry)
3448 {
3449         struct list_head *tmp_node, *list_node;
3450         struct irdma_cq_buf *last_buf = NULL;
3451         struct irdma_cq_poll_info *cur_cqe = &iwcq->cur_cqe;
3452         struct irdma_cq_buf *cq_buf;
3453         int ret;
3454         struct irdma_device *iwdev;
3455         struct irdma_cq_uk *ukcq;
3456         bool cq_new_cqe = false;
3457         int resized_bufs = 0;
3458         int npolled = 0;
3459
3460         iwdev = to_iwdev(iwcq->ibcq.device);
3461         ukcq = &iwcq->sc_cq.cq_uk;
3462
3463         /* go through the list of previously resized CQ buffers */
3464         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
3465                 cq_buf = container_of(list_node, struct irdma_cq_buf, list);
3466                 while (npolled < num_entries) {
3467                         ret = irdma_poll_one(&cq_buf->cq_uk, cur_cqe, entry + npolled);
3468                         if (!ret) {
3469                                 ++npolled;
3470                                 cq_new_cqe = true;
3471                                 continue;
3472                         }
3473                         if (ret == -ENOENT)
3474                                 break;
3475                          /* QP using the CQ is destroyed. Skip reporting this CQE */
3476                         if (ret == -EFAULT) {
3477                                 cq_new_cqe = true;
3478                                 continue;
3479                         }
3480                         goto error;
3481                 }
3482
3483                 /* save the resized CQ buffer which received the last cqe */
3484                 if (cq_new_cqe)
3485                         last_buf = cq_buf;
3486                 cq_new_cqe = false;
3487         }
3488
3489         /* check the current CQ for new cqes */
3490         while (npolled < num_entries) {
3491                 ret = irdma_poll_one(ukcq, cur_cqe, entry + npolled);
3492                 if (ret == -ENOENT) {
3493                         ret = irdma_generated_cmpls(iwcq, cur_cqe);
3494                         if (!ret)
3495                                 irdma_process_cqe(entry + npolled, cur_cqe);
3496                 }
3497                 if (!ret) {
3498                         ++npolled;
3499                         cq_new_cqe = true;
3500                         continue;
3501                 }
3502
3503                 if (ret == -ENOENT)
3504                         break;
3505                 /* QP using the CQ is destroyed. Skip reporting this CQE */
3506                 if (ret == -EFAULT) {
3507                         cq_new_cqe = true;
3508                         continue;
3509                 }
3510                 goto error;
3511         }
3512
3513         if (cq_new_cqe)
3514                 /* all previous CQ resizes are complete */
3515                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, NULL);
3516         else if (last_buf)
3517                 /* only CQ resizes up to the last_buf are complete */
3518                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, last_buf);
3519         if (resized_bufs)
3520                 /* report to the HW the number of complete CQ resizes */
3521                 irdma_uk_cq_set_resized_cnt(ukcq, resized_bufs);
3522
3523         return npolled;
3524 error:
3525         ibdev_dbg(&iwdev->ibdev, "%s: Error polling CQ, irdma_err: %d\n",
3526                   __func__, ret);
3527
3528         return ret;
3529 }
3530
3531 /**
3532  * irdma_poll_cq - poll cq for completion (kernel apps)
3533  * @ibcq: cq to poll
3534  * @num_entries: number of entries to poll
3535  * @entry: wr of a completed entry
3536  */
3537 static int irdma_poll_cq(struct ib_cq *ibcq, int num_entries,
3538                          struct ib_wc *entry)
3539 {
3540         struct irdma_cq *iwcq;
3541         unsigned long flags;
3542         int ret;
3543
3544         iwcq = to_iwcq(ibcq);
3545
3546         spin_lock_irqsave(&iwcq->lock, flags);
3547         ret = __irdma_poll_cq(iwcq, num_entries, entry);
3548         spin_unlock_irqrestore(&iwcq->lock, flags);
3549
3550         return ret;
3551 }
3552
3553 /**
3554  * irdma_req_notify_cq - arm cq kernel application
3555  * @ibcq: cq to arm
3556  * @notify_flags: notofication flags
3557  */
3558 static int irdma_req_notify_cq(struct ib_cq *ibcq,
3559                                enum ib_cq_notify_flags notify_flags)
3560 {
3561         struct irdma_cq *iwcq;
3562         struct irdma_cq_uk *ukcq;
3563         unsigned long flags;
3564         enum irdma_cmpl_notify cq_notify;
3565         bool promo_event = false;
3566         int ret = 0;
3567
3568         cq_notify = notify_flags == IB_CQ_SOLICITED ?
3569                     IRDMA_CQ_COMPL_SOLICITED : IRDMA_CQ_COMPL_EVENT;
3570         iwcq = to_iwcq(ibcq);
3571         ukcq = &iwcq->sc_cq.cq_uk;
3572
3573         spin_lock_irqsave(&iwcq->lock, flags);
3574         /* Only promote to arm the CQ for any event if the last arm event was solicited. */
3575         if (iwcq->last_notify == IRDMA_CQ_COMPL_SOLICITED && notify_flags != IB_CQ_SOLICITED)
3576                 promo_event = true;
3577
3578         if (!atomic_cmpxchg(&iwcq->armed, 0, 1) || promo_event) {
3579                 iwcq->last_notify = cq_notify;
3580                 irdma_uk_cq_request_notification(ukcq, cq_notify);
3581         }
3582
3583         if ((notify_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
3584             (!irdma_cq_empty(iwcq) || !list_empty(&iwcq->cmpl_generated)))
3585                 ret = 1;
3586         spin_unlock_irqrestore(&iwcq->lock, flags);
3587
3588         return ret;
3589 }
3590
3591 static int irdma_roce_port_immutable(struct ib_device *ibdev, u32 port_num,
3592                                      struct ib_port_immutable *immutable)
3593 {
3594         struct ib_port_attr attr;
3595         int err;
3596
3597         immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
3598         err = ib_query_port(ibdev, port_num, &attr);
3599         if (err)
3600                 return err;
3601
3602         immutable->max_mad_size = IB_MGMT_MAD_SIZE;
3603         immutable->pkey_tbl_len = attr.pkey_tbl_len;
3604         immutable->gid_tbl_len = attr.gid_tbl_len;
3605
3606         return 0;
3607 }
3608
3609 static int irdma_iw_port_immutable(struct ib_device *ibdev, u32 port_num,
3610                                    struct ib_port_immutable *immutable)
3611 {
3612         struct ib_port_attr attr;
3613         int err;
3614
3615         immutable->core_cap_flags = RDMA_CORE_PORT_IWARP;
3616         err = ib_query_port(ibdev, port_num, &attr);
3617         if (err)
3618                 return err;
3619         immutable->gid_tbl_len = attr.gid_tbl_len;
3620
3621         return 0;
3622 }
3623
3624 static const struct rdma_stat_desc irdma_hw_stat_descs[] = {
3625         /* 32bit names */
3626         [IRDMA_HW_STAT_INDEX_RXVLANERR].name = "rxVlanErrors",
3627         [IRDMA_HW_STAT_INDEX_IP4RXDISCARD].name = "ip4InDiscards",
3628         [IRDMA_HW_STAT_INDEX_IP4RXTRUNC].name = "ip4InTruncatedPkts",
3629         [IRDMA_HW_STAT_INDEX_IP4TXNOROUTE].name = "ip4OutNoRoutes",
3630         [IRDMA_HW_STAT_INDEX_IP6RXDISCARD].name = "ip6InDiscards",
3631         [IRDMA_HW_STAT_INDEX_IP6RXTRUNC].name = "ip6InTruncatedPkts",
3632         [IRDMA_HW_STAT_INDEX_IP6TXNOROUTE].name = "ip6OutNoRoutes",
3633         [IRDMA_HW_STAT_INDEX_TCPRTXSEG].name = "tcpRetransSegs",
3634         [IRDMA_HW_STAT_INDEX_TCPRXOPTERR].name = "tcpInOptErrors",
3635         [IRDMA_HW_STAT_INDEX_TCPRXPROTOERR].name = "tcpInProtoErrors",
3636         [IRDMA_HW_STAT_INDEX_RXRPCNPHANDLED].name = "cnpHandled",
3637         [IRDMA_HW_STAT_INDEX_RXRPCNPIGNORED].name = "cnpIgnored",
3638         [IRDMA_HW_STAT_INDEX_TXNPCNPSENT].name = "cnpSent",
3639
3640         /* 64bit names */
3641         [IRDMA_HW_STAT_INDEX_IP4RXOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3642                 "ip4InOctets",
3643         [IRDMA_HW_STAT_INDEX_IP4RXPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3644                 "ip4InPkts",
3645         [IRDMA_HW_STAT_INDEX_IP4RXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3646                 "ip4InReasmRqd",
3647         [IRDMA_HW_STAT_INDEX_IP4RXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3648                 "ip4InMcastOctets",
3649         [IRDMA_HW_STAT_INDEX_IP4RXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3650                 "ip4InMcastPkts",
3651         [IRDMA_HW_STAT_INDEX_IP4TXOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3652                 "ip4OutOctets",
3653         [IRDMA_HW_STAT_INDEX_IP4TXPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3654                 "ip4OutPkts",
3655         [IRDMA_HW_STAT_INDEX_IP4TXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3656                 "ip4OutSegRqd",
3657         [IRDMA_HW_STAT_INDEX_IP4TXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3658                 "ip4OutMcastOctets",
3659         [IRDMA_HW_STAT_INDEX_IP4TXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3660                 "ip4OutMcastPkts",
3661         [IRDMA_HW_STAT_INDEX_IP6RXOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3662                 "ip6InOctets",
3663         [IRDMA_HW_STAT_INDEX_IP6RXPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3664                 "ip6InPkts",
3665         [IRDMA_HW_STAT_INDEX_IP6RXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3666                 "ip6InReasmRqd",
3667         [IRDMA_HW_STAT_INDEX_IP6RXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3668                 "ip6InMcastOctets",
3669         [IRDMA_HW_STAT_INDEX_IP6RXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3670                 "ip6InMcastPkts",
3671         [IRDMA_HW_STAT_INDEX_IP6TXOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3672                 "ip6OutOctets",
3673         [IRDMA_HW_STAT_INDEX_IP6TXPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3674                 "ip6OutPkts",
3675         [IRDMA_HW_STAT_INDEX_IP6TXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3676                 "ip6OutSegRqd",
3677         [IRDMA_HW_STAT_INDEX_IP6TXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3678                 "ip6OutMcastOctets",
3679         [IRDMA_HW_STAT_INDEX_IP6TXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3680                 "ip6OutMcastPkts",
3681         [IRDMA_HW_STAT_INDEX_TCPRXSEGS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3682                 "tcpInSegs",
3683         [IRDMA_HW_STAT_INDEX_TCPTXSEG + IRDMA_HW_STAT_INDEX_MAX_32].name =
3684                 "tcpOutSegs",
3685         [IRDMA_HW_STAT_INDEX_RDMARXRDS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3686                 "iwInRdmaReads",
3687         [IRDMA_HW_STAT_INDEX_RDMARXSNDS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3688                 "iwInRdmaSends",
3689         [IRDMA_HW_STAT_INDEX_RDMARXWRS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3690                 "iwInRdmaWrites",
3691         [IRDMA_HW_STAT_INDEX_RDMATXRDS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3692                 "iwOutRdmaReads",
3693         [IRDMA_HW_STAT_INDEX_RDMATXSNDS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3694                 "iwOutRdmaSends",
3695         [IRDMA_HW_STAT_INDEX_RDMATXWRS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3696                 "iwOutRdmaWrites",
3697         [IRDMA_HW_STAT_INDEX_RDMAVBND + IRDMA_HW_STAT_INDEX_MAX_32].name =
3698                 "iwRdmaBnd",
3699         [IRDMA_HW_STAT_INDEX_RDMAVINV + IRDMA_HW_STAT_INDEX_MAX_32].name =
3700                 "iwRdmaInv",
3701         [IRDMA_HW_STAT_INDEX_UDPRXPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3702                 "RxUDP",
3703         [IRDMA_HW_STAT_INDEX_UDPTXPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3704                 "TxUDP",
3705         [IRDMA_HW_STAT_INDEX_RXNPECNMARKEDPKTS + IRDMA_HW_STAT_INDEX_MAX_32]
3706                 .name = "RxECNMrkd",
3707 };
3708
3709 static void irdma_get_dev_fw_str(struct ib_device *dev, char *str)
3710 {
3711         struct irdma_device *iwdev = to_iwdev(dev);
3712
3713         snprintf(str, IB_FW_VERSION_NAME_MAX, "%u.%u",
3714                  irdma_fw_major_ver(&iwdev->rf->sc_dev),
3715                  irdma_fw_minor_ver(&iwdev->rf->sc_dev));
3716 }
3717
3718 /**
3719  * irdma_alloc_hw_port_stats - Allocate a hw stats structure
3720  * @ibdev: device pointer from stack
3721  * @port_num: port number
3722  */
3723 static struct rdma_hw_stats *irdma_alloc_hw_port_stats(struct ib_device *ibdev,
3724                                                        u32 port_num)
3725 {
3726         int num_counters = IRDMA_HW_STAT_INDEX_MAX_32 +
3727                            IRDMA_HW_STAT_INDEX_MAX_64;
3728         unsigned long lifespan = RDMA_HW_STATS_DEFAULT_LIFESPAN;
3729
3730         BUILD_BUG_ON(ARRAY_SIZE(irdma_hw_stat_descs) !=
3731                      (IRDMA_HW_STAT_INDEX_MAX_32 + IRDMA_HW_STAT_INDEX_MAX_64));
3732
3733         return rdma_alloc_hw_stats_struct(irdma_hw_stat_descs, num_counters,
3734                                           lifespan);
3735 }
3736
3737 /**
3738  * irdma_get_hw_stats - Populates the rdma_hw_stats structure
3739  * @ibdev: device pointer from stack
3740  * @stats: stats pointer from stack
3741  * @port_num: port number
3742  * @index: which hw counter the stack is requesting we update
3743  */
3744 static int irdma_get_hw_stats(struct ib_device *ibdev,
3745                               struct rdma_hw_stats *stats, u32 port_num,
3746                               int index)
3747 {
3748         struct irdma_device *iwdev = to_iwdev(ibdev);
3749         struct irdma_dev_hw_stats *hw_stats = &iwdev->vsi.pestat->hw_stats;
3750
3751         if (iwdev->rf->rdma_ver >= IRDMA_GEN_2)
3752                 irdma_cqp_gather_stats_cmd(&iwdev->rf->sc_dev, iwdev->vsi.pestat, true);
3753         else
3754                 irdma_cqp_gather_stats_gen1(&iwdev->rf->sc_dev, iwdev->vsi.pestat);
3755
3756         memcpy(&stats->value[0], hw_stats, sizeof(*hw_stats));
3757
3758         return stats->num_counters;
3759 }
3760
3761 /**
3762  * irdma_query_gid - Query port GID
3763  * @ibdev: device pointer from stack
3764  * @port: port number
3765  * @index: Entry index
3766  * @gid: Global ID
3767  */
3768 static int irdma_query_gid(struct ib_device *ibdev, u32 port, int index,
3769                            union ib_gid *gid)
3770 {
3771         struct irdma_device *iwdev = to_iwdev(ibdev);
3772
3773         memset(gid->raw, 0, sizeof(gid->raw));
3774         ether_addr_copy(gid->raw, iwdev->netdev->dev_addr);
3775
3776         return 0;
3777 }
3778
3779 /**
3780  * mcast_list_add -  Add a new mcast item to list
3781  * @rf: RDMA PCI function
3782  * @new_elem: pointer to element to add
3783  */
3784 static void mcast_list_add(struct irdma_pci_f *rf,
3785                            struct mc_table_list *new_elem)
3786 {
3787         list_add(&new_elem->list, &rf->mc_qht_list.list);
3788 }
3789
3790 /**
3791  * mcast_list_del - Remove an mcast item from list
3792  * @mc_qht_elem: pointer to mcast table list element
3793  */
3794 static void mcast_list_del(struct mc_table_list *mc_qht_elem)
3795 {
3796         if (mc_qht_elem)
3797                 list_del(&mc_qht_elem->list);
3798 }
3799
3800 /**
3801  * mcast_list_lookup_ip - Search mcast list for address
3802  * @rf: RDMA PCI function
3803  * @ip_mcast: pointer to mcast IP address
3804  */
3805 static struct mc_table_list *mcast_list_lookup_ip(struct irdma_pci_f *rf,
3806                                                   u32 *ip_mcast)
3807 {
3808         struct mc_table_list *mc_qht_el;
3809         struct list_head *pos, *q;
3810
3811         list_for_each_safe (pos, q, &rf->mc_qht_list.list) {
3812                 mc_qht_el = list_entry(pos, struct mc_table_list, list);
3813                 if (!memcmp(mc_qht_el->mc_info.dest_ip, ip_mcast,
3814                             sizeof(mc_qht_el->mc_info.dest_ip)))
3815                         return mc_qht_el;
3816         }
3817
3818         return NULL;
3819 }
3820
3821 /**
3822  * irdma_mcast_cqp_op - perform a mcast cqp operation
3823  * @iwdev: irdma device
3824  * @mc_grp_ctx: mcast group info
3825  * @op: operation
3826  *
3827  * returns error status
3828  */
3829 static int irdma_mcast_cqp_op(struct irdma_device *iwdev,
3830                               struct irdma_mcast_grp_info *mc_grp_ctx, u8 op)
3831 {
3832         struct cqp_cmds_info *cqp_info;
3833         struct irdma_cqp_request *cqp_request;
3834         int status;
3835
3836         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
3837         if (!cqp_request)
3838                 return -ENOMEM;
3839
3840         cqp_request->info.in.u.mc_create.info = *mc_grp_ctx;
3841         cqp_info = &cqp_request->info;
3842         cqp_info->cqp_cmd = op;
3843         cqp_info->post_sq = 1;
3844         cqp_info->in.u.mc_create.scratch = (uintptr_t)cqp_request;
3845         cqp_info->in.u.mc_create.cqp = &iwdev->rf->cqp.sc_cqp;
3846         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
3847         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
3848
3849         return status;
3850 }
3851
3852 /**
3853  * irdma_mcast_mac - Get the multicast MAC for an IP address
3854  * @ip_addr: IPv4 or IPv6 address
3855  * @mac: pointer to result MAC address
3856  * @ipv4: flag indicating IPv4 or IPv6
3857  *
3858  */
3859 void irdma_mcast_mac(u32 *ip_addr, u8 *mac, bool ipv4)
3860 {
3861         u8 *ip = (u8 *)ip_addr;
3862
3863         if (ipv4) {
3864                 unsigned char mac4[ETH_ALEN] = {0x01, 0x00, 0x5E, 0x00,
3865                                                 0x00, 0x00};
3866
3867                 mac4[3] = ip[2] & 0x7F;
3868                 mac4[4] = ip[1];
3869                 mac4[5] = ip[0];
3870                 ether_addr_copy(mac, mac4);
3871         } else {
3872                 unsigned char mac6[ETH_ALEN] = {0x33, 0x33, 0x00, 0x00,
3873                                                 0x00, 0x00};
3874
3875                 mac6[2] = ip[3];
3876                 mac6[3] = ip[2];
3877                 mac6[4] = ip[1];
3878                 mac6[5] = ip[0];
3879                 ether_addr_copy(mac, mac6);
3880         }
3881 }
3882
3883 /**
3884  * irdma_attach_mcast - attach a qp to a multicast group
3885  * @ibqp: ptr to qp
3886  * @ibgid: pointer to global ID
3887  * @lid: local ID
3888  *
3889  * returns error status
3890  */
3891 static int irdma_attach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
3892 {
3893         struct irdma_qp *iwqp = to_iwqp(ibqp);
3894         struct irdma_device *iwdev = iwqp->iwdev;
3895         struct irdma_pci_f *rf = iwdev->rf;
3896         struct mc_table_list *mc_qht_elem;
3897         struct irdma_mcast_grp_ctx_entry_info mcg_info = {};
3898         unsigned long flags;
3899         u32 ip_addr[4] = {};
3900         u32 mgn;
3901         u32 no_mgs;
3902         int ret = 0;
3903         bool ipv4;
3904         u16 vlan_id;
3905         union irdma_sockaddr sgid_addr;
3906         unsigned char dmac[ETH_ALEN];
3907
3908         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
3909
3910         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid)) {
3911                 irdma_copy_ip_ntohl(ip_addr,
3912                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
3913                 irdma_netdev_vlan_ipv6(ip_addr, &vlan_id, NULL);
3914                 ipv4 = false;
3915                 ibdev_dbg(&iwdev->ibdev,
3916                           "VERBS: qp_id=%d, IP6address=%pI6\n", ibqp->qp_num,
3917                           ip_addr);
3918                 irdma_mcast_mac(ip_addr, dmac, false);
3919         } else {
3920                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
3921                 ipv4 = true;
3922                 vlan_id = irdma_get_vlan_ipv4(ip_addr);
3923                 irdma_mcast_mac(ip_addr, dmac, true);
3924                 ibdev_dbg(&iwdev->ibdev,
3925                           "VERBS: qp_id=%d, IP4address=%pI4, MAC=%pM\n",
3926                           ibqp->qp_num, ip_addr, dmac);
3927         }
3928
3929         spin_lock_irqsave(&rf->qh_list_lock, flags);
3930         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
3931         if (!mc_qht_elem) {
3932                 struct irdma_dma_mem *dma_mem_mc;
3933
3934                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3935                 mc_qht_elem = kzalloc(sizeof(*mc_qht_elem), GFP_KERNEL);
3936                 if (!mc_qht_elem)
3937                         return -ENOMEM;
3938
3939                 mc_qht_elem->mc_info.ipv4_valid = ipv4;
3940                 memcpy(mc_qht_elem->mc_info.dest_ip, ip_addr,
3941                        sizeof(mc_qht_elem->mc_info.dest_ip));
3942                 ret = irdma_alloc_rsrc(rf, rf->allocated_mcgs, rf->max_mcg,
3943                                        &mgn, &rf->next_mcg);
3944                 if (ret) {
3945                         kfree(mc_qht_elem);
3946                         return -ENOMEM;
3947                 }
3948
3949                 mc_qht_elem->mc_info.mgn = mgn;
3950                 dma_mem_mc = &mc_qht_elem->mc_grp_ctx.dma_mem_mc;
3951                 dma_mem_mc->size = ALIGN(sizeof(u64) * IRDMA_MAX_MGS_PER_CTX,
3952                                          IRDMA_HW_PAGE_SIZE);
3953                 dma_mem_mc->va = dma_alloc_coherent(rf->hw.device,
3954                                                     dma_mem_mc->size,
3955                                                     &dma_mem_mc->pa,
3956                                                     GFP_KERNEL);
3957                 if (!dma_mem_mc->va) {
3958                         irdma_free_rsrc(rf, rf->allocated_mcgs, mgn);
3959                         kfree(mc_qht_elem);
3960                         return -ENOMEM;
3961                 }
3962
3963                 mc_qht_elem->mc_grp_ctx.mg_id = (u16)mgn;
3964                 memcpy(mc_qht_elem->mc_grp_ctx.dest_ip_addr, ip_addr,
3965                        sizeof(mc_qht_elem->mc_grp_ctx.dest_ip_addr));
3966                 mc_qht_elem->mc_grp_ctx.ipv4_valid = ipv4;
3967                 mc_qht_elem->mc_grp_ctx.vlan_id = vlan_id;
3968                 if (vlan_id < VLAN_N_VID)
3969                         mc_qht_elem->mc_grp_ctx.vlan_valid = true;
3970                 mc_qht_elem->mc_grp_ctx.hmc_fcn_id = iwdev->vsi.fcn_id;
3971                 mc_qht_elem->mc_grp_ctx.qs_handle =
3972                         iwqp->sc_qp.vsi->qos[iwqp->sc_qp.user_pri].qs_handle;
3973                 ether_addr_copy(mc_qht_elem->mc_grp_ctx.dest_mac_addr, dmac);
3974
3975                 spin_lock_irqsave(&rf->qh_list_lock, flags);
3976                 mcast_list_add(rf, mc_qht_elem);
3977         } else {
3978                 if (mc_qht_elem->mc_grp_ctx.no_of_mgs ==
3979                     IRDMA_MAX_MGS_PER_CTX) {
3980                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3981                         return -ENOMEM;
3982                 }
3983         }
3984
3985         mcg_info.qp_id = iwqp->ibqp.qp_num;
3986         no_mgs = mc_qht_elem->mc_grp_ctx.no_of_mgs;
3987         irdma_sc_add_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
3988         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3989
3990         /* Only if there is a change do we need to modify or create */
3991         if (!no_mgs) {
3992                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
3993                                          IRDMA_OP_MC_CREATE);
3994         } else if (no_mgs != mc_qht_elem->mc_grp_ctx.no_of_mgs) {
3995                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
3996                                          IRDMA_OP_MC_MODIFY);
3997         } else {
3998                 return 0;
3999         }
4000
4001         if (ret)
4002                 goto error;
4003
4004         return 0;
4005
4006 error:
4007         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
4008         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
4009                 mcast_list_del(mc_qht_elem);
4010                 dma_free_coherent(rf->hw.device,
4011                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.size,
4012                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.va,
4013                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.pa);
4014                 mc_qht_elem->mc_grp_ctx.dma_mem_mc.va = NULL;
4015                 irdma_free_rsrc(rf, rf->allocated_mcgs,
4016                                 mc_qht_elem->mc_grp_ctx.mg_id);
4017                 kfree(mc_qht_elem);
4018         }
4019
4020         return ret;
4021 }
4022
4023 /**
4024  * irdma_detach_mcast - detach a qp from a multicast group
4025  * @ibqp: ptr to qp
4026  * @ibgid: pointer to global ID
4027  * @lid: local ID
4028  *
4029  * returns error status
4030  */
4031 static int irdma_detach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
4032 {
4033         struct irdma_qp *iwqp = to_iwqp(ibqp);
4034         struct irdma_device *iwdev = iwqp->iwdev;
4035         struct irdma_pci_f *rf = iwdev->rf;
4036         u32 ip_addr[4] = {};
4037         struct mc_table_list *mc_qht_elem;
4038         struct irdma_mcast_grp_ctx_entry_info mcg_info = {};
4039         int ret;
4040         unsigned long flags;
4041         union irdma_sockaddr sgid_addr;
4042
4043         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
4044         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid))
4045                 irdma_copy_ip_ntohl(ip_addr,
4046                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4047         else
4048                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
4049
4050         spin_lock_irqsave(&rf->qh_list_lock, flags);
4051         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
4052         if (!mc_qht_elem) {
4053                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4054                 ibdev_dbg(&iwdev->ibdev,
4055                           "VERBS: address not found MCG\n");
4056                 return 0;
4057         }
4058
4059         mcg_info.qp_id = iwqp->ibqp.qp_num;
4060         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
4061         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
4062                 mcast_list_del(mc_qht_elem);
4063                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4064                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4065                                          IRDMA_OP_MC_DESTROY);
4066                 if (ret) {
4067                         ibdev_dbg(&iwdev->ibdev,
4068                                   "VERBS: failed MC_DESTROY MCG\n");
4069                         spin_lock_irqsave(&rf->qh_list_lock, flags);
4070                         mcast_list_add(rf, mc_qht_elem);
4071                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4072                         return -EAGAIN;
4073                 }
4074
4075                 dma_free_coherent(rf->hw.device,
4076                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.size,
4077                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.va,
4078                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.pa);
4079                 mc_qht_elem->mc_grp_ctx.dma_mem_mc.va = NULL;
4080                 irdma_free_rsrc(rf, rf->allocated_mcgs,
4081                                 mc_qht_elem->mc_grp_ctx.mg_id);
4082                 kfree(mc_qht_elem);
4083         } else {
4084                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4085                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4086                                          IRDMA_OP_MC_MODIFY);
4087                 if (ret) {
4088                         ibdev_dbg(&iwdev->ibdev,
4089                                   "VERBS: failed Modify MCG\n");
4090                         return ret;
4091                 }
4092         }
4093
4094         return 0;
4095 }
4096
4097 static int irdma_create_hw_ah(struct irdma_device *iwdev, struct irdma_ah *ah, bool sleep)
4098 {
4099         struct irdma_pci_f *rf = iwdev->rf;
4100         int err;
4101
4102         err = irdma_alloc_rsrc(rf, rf->allocated_ahs, rf->max_ah, &ah->sc_ah.ah_info.ah_idx,
4103                                &rf->next_ah);
4104         if (err)
4105                 return err;
4106
4107         err = irdma_ah_cqp_op(rf, &ah->sc_ah, IRDMA_OP_AH_CREATE, sleep,
4108                               irdma_gsi_ud_qp_ah_cb, &ah->sc_ah);
4109
4110         if (err) {
4111                 ibdev_dbg(&iwdev->ibdev, "VERBS: CQP-OP Create AH fail");
4112                 goto err_ah_create;
4113         }
4114
4115         if (!sleep) {
4116                 int cnt = CQP_COMPL_WAIT_TIME_MS * CQP_TIMEOUT_THRESHOLD;
4117
4118                 do {
4119                         irdma_cqp_ce_handler(rf, &rf->ccq.sc_cq);
4120                         mdelay(1);
4121                 } while (!ah->sc_ah.ah_info.ah_valid && --cnt);
4122
4123                 if (!cnt) {
4124                         ibdev_dbg(&iwdev->ibdev, "VERBS: CQP create AH timed out");
4125                         err = -ETIMEDOUT;
4126                         goto err_ah_create;
4127                 }
4128         }
4129         return 0;
4130
4131 err_ah_create:
4132         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_ahs, ah->sc_ah.ah_info.ah_idx);
4133
4134         return err;
4135 }
4136
4137 static int irdma_setup_ah(struct ib_ah *ibah, struct rdma_ah_init_attr *attr)
4138 {
4139         struct irdma_pd *pd = to_iwpd(ibah->pd);
4140         struct irdma_ah *ah = container_of(ibah, struct irdma_ah, ibah);
4141         struct rdma_ah_attr *ah_attr = attr->ah_attr;
4142         const struct ib_gid_attr *sgid_attr;
4143         struct irdma_device *iwdev = to_iwdev(ibah->pd->device);
4144         struct irdma_pci_f *rf = iwdev->rf;
4145         struct irdma_sc_ah *sc_ah;
4146         struct irdma_ah_info *ah_info;
4147         union irdma_sockaddr sgid_addr, dgid_addr;
4148         int err;
4149         u8 dmac[ETH_ALEN];
4150
4151         ah->pd = pd;
4152         sc_ah = &ah->sc_ah;
4153         sc_ah->ah_info.vsi = &iwdev->vsi;
4154         irdma_sc_init_ah(&rf->sc_dev, sc_ah);
4155         ah->sgid_index = ah_attr->grh.sgid_index;
4156         sgid_attr = ah_attr->grh.sgid_attr;
4157         memcpy(&ah->dgid, &ah_attr->grh.dgid, sizeof(ah->dgid));
4158         rdma_gid2ip((struct sockaddr *)&sgid_addr, &sgid_attr->gid);
4159         rdma_gid2ip((struct sockaddr *)&dgid_addr, &ah_attr->grh.dgid);
4160         ah->av.attrs = *ah_attr;
4161         ah->av.net_type = rdma_gid_attr_network_type(sgid_attr);
4162         ah_info = &sc_ah->ah_info;
4163         ah_info->pd_idx = pd->sc_pd.pd_id;
4164         if (ah_attr->ah_flags & IB_AH_GRH) {
4165                 ah_info->flow_label = ah_attr->grh.flow_label;
4166                 ah_info->hop_ttl = ah_attr->grh.hop_limit;
4167                 ah_info->tc_tos = ah_attr->grh.traffic_class;
4168         }
4169
4170         ether_addr_copy(dmac, ah_attr->roce.dmac);
4171         if (ah->av.net_type == RDMA_NETWORK_IPV4) {
4172                 ah_info->ipv4_valid = true;
4173                 ah_info->dest_ip_addr[0] =
4174                         ntohl(dgid_addr.saddr_in.sin_addr.s_addr);
4175                 ah_info->src_ip_addr[0] =
4176                         ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
4177                 ah_info->do_lpbk = irdma_ipv4_is_lpb(ah_info->src_ip_addr[0],
4178                                                      ah_info->dest_ip_addr[0]);
4179                 if (ipv4_is_multicast(dgid_addr.saddr_in.sin_addr.s_addr)) {
4180                         ah_info->do_lpbk = true;
4181                         irdma_mcast_mac(ah_info->dest_ip_addr, dmac, true);
4182                 }
4183         } else {
4184                 irdma_copy_ip_ntohl(ah_info->dest_ip_addr,
4185                                     dgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4186                 irdma_copy_ip_ntohl(ah_info->src_ip_addr,
4187                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4188                 ah_info->do_lpbk = irdma_ipv6_is_lpb(ah_info->src_ip_addr,
4189                                                      ah_info->dest_ip_addr);
4190                 if (rdma_is_multicast_addr(&dgid_addr.saddr_in6.sin6_addr)) {
4191                         ah_info->do_lpbk = true;
4192                         irdma_mcast_mac(ah_info->dest_ip_addr, dmac, false);
4193                 }
4194         }
4195
4196         err = rdma_read_gid_l2_fields(sgid_attr, &ah_info->vlan_tag,
4197                                       ah_info->mac_addr);
4198         if (err)
4199                 return err;
4200
4201         ah_info->dst_arpindex = irdma_add_arp(iwdev->rf, ah_info->dest_ip_addr,
4202                                               ah_info->ipv4_valid, dmac);
4203
4204         if (ah_info->dst_arpindex == -1)
4205                 return -EINVAL;
4206
4207         if (ah_info->vlan_tag >= VLAN_N_VID && iwdev->dcb_vlan_mode)
4208                 ah_info->vlan_tag = 0;
4209
4210         if (ah_info->vlan_tag < VLAN_N_VID) {
4211                 ah_info->insert_vlan_tag = true;
4212                 ah_info->vlan_tag |=
4213                         rt_tos2priority(ah_info->tc_tos) << VLAN_PRIO_SHIFT;
4214         }
4215
4216         return 0;
4217 }
4218
4219 /**
4220  * irdma_ah_exists - Check for existing identical AH
4221  * @iwdev: irdma device
4222  * @new_ah: AH to check for
4223  *
4224  * returns true if AH is found, false if not found.
4225  */
4226 static bool irdma_ah_exists(struct irdma_device *iwdev,
4227                             struct irdma_ah *new_ah)
4228 {
4229         struct irdma_ah *ah;
4230         u32 key = new_ah->sc_ah.ah_info.dest_ip_addr[0] ^
4231                   new_ah->sc_ah.ah_info.dest_ip_addr[1] ^
4232                   new_ah->sc_ah.ah_info.dest_ip_addr[2] ^
4233                   new_ah->sc_ah.ah_info.dest_ip_addr[3];
4234
4235         hash_for_each_possible(iwdev->ah_hash_tbl, ah, list, key) {
4236                 /* Set ah_valid and ah_id the same so memcmp can work */
4237                 new_ah->sc_ah.ah_info.ah_idx = ah->sc_ah.ah_info.ah_idx;
4238                 new_ah->sc_ah.ah_info.ah_valid = ah->sc_ah.ah_info.ah_valid;
4239                 if (!memcmp(&ah->sc_ah.ah_info, &new_ah->sc_ah.ah_info,
4240                             sizeof(ah->sc_ah.ah_info))) {
4241                         refcount_inc(&ah->refcnt);
4242                         new_ah->parent_ah = ah;
4243                         return true;
4244                 }
4245         }
4246
4247         return false;
4248 }
4249
4250 /**
4251  * irdma_destroy_ah - Destroy address handle
4252  * @ibah: pointer to address handle
4253  * @ah_flags: flags for sleepable
4254  */
4255 static int irdma_destroy_ah(struct ib_ah *ibah, u32 ah_flags)
4256 {
4257         struct irdma_device *iwdev = to_iwdev(ibah->device);
4258         struct irdma_ah *ah = to_iwah(ibah);
4259
4260         if ((ah_flags & RDMA_DESTROY_AH_SLEEPABLE) && ah->parent_ah) {
4261                 mutex_lock(&iwdev->ah_tbl_lock);
4262                 if (!refcount_dec_and_test(&ah->parent_ah->refcnt)) {
4263                         mutex_unlock(&iwdev->ah_tbl_lock);
4264                         return 0;
4265                 }
4266                 hash_del(&ah->parent_ah->list);
4267                 kfree(ah->parent_ah);
4268                 mutex_unlock(&iwdev->ah_tbl_lock);
4269         }
4270
4271         irdma_ah_cqp_op(iwdev->rf, &ah->sc_ah, IRDMA_OP_AH_DESTROY,
4272                         false, NULL, ah);
4273
4274         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_ahs,
4275                         ah->sc_ah.ah_info.ah_idx);
4276
4277         return 0;
4278 }
4279
4280 /**
4281  * irdma_create_user_ah - create user address handle
4282  * @ibah: address handle
4283  * @attr: address handle attributes
4284  * @udata: User data
4285  *
4286  * returns 0 on success, error otherwise
4287  */
4288 static int irdma_create_user_ah(struct ib_ah *ibah,
4289                                 struct rdma_ah_init_attr *attr,
4290                                 struct ib_udata *udata)
4291 {
4292         struct irdma_ah *ah = container_of(ibah, struct irdma_ah, ibah);
4293         struct irdma_device *iwdev = to_iwdev(ibah->pd->device);
4294         struct irdma_create_ah_resp uresp;
4295         struct irdma_ah *parent_ah;
4296         int err;
4297
4298         err = irdma_setup_ah(ibah, attr);
4299         if (err)
4300                 return err;
4301         mutex_lock(&iwdev->ah_tbl_lock);
4302         if (!irdma_ah_exists(iwdev, ah)) {
4303                 err = irdma_create_hw_ah(iwdev, ah, true);
4304                 if (err) {
4305                         mutex_unlock(&iwdev->ah_tbl_lock);
4306                         return err;
4307                 }
4308                 /* Add new AH to list */
4309                 parent_ah = kmemdup(ah, sizeof(*ah), GFP_KERNEL);
4310                 if (parent_ah) {
4311                         u32 key = parent_ah->sc_ah.ah_info.dest_ip_addr[0] ^
4312                                   parent_ah->sc_ah.ah_info.dest_ip_addr[1] ^
4313                                   parent_ah->sc_ah.ah_info.dest_ip_addr[2] ^
4314                                   parent_ah->sc_ah.ah_info.dest_ip_addr[3];
4315
4316                         ah->parent_ah = parent_ah;
4317                         hash_add(iwdev->ah_hash_tbl, &parent_ah->list, key);
4318                         refcount_set(&parent_ah->refcnt, 1);
4319                 }
4320         }
4321         mutex_unlock(&iwdev->ah_tbl_lock);
4322
4323         uresp.ah_id = ah->sc_ah.ah_info.ah_idx;
4324         err = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp), udata->outlen));
4325         if (err)
4326                 irdma_destroy_ah(ibah, attr->flags);
4327
4328         return err;
4329 }
4330
4331 /**
4332  * irdma_create_ah - create address handle
4333  * @ibah: address handle
4334  * @attr: address handle attributes
4335  * @udata: NULL
4336  *
4337  * returns 0 on success, error otherwise
4338  */
4339 static int irdma_create_ah(struct ib_ah *ibah, struct rdma_ah_init_attr *attr,
4340                            struct ib_udata *udata)
4341 {
4342         struct irdma_ah *ah = container_of(ibah, struct irdma_ah, ibah);
4343         struct irdma_device *iwdev = to_iwdev(ibah->pd->device);
4344         int err;
4345
4346         err = irdma_setup_ah(ibah, attr);
4347         if (err)
4348                 return err;
4349         err = irdma_create_hw_ah(iwdev, ah, attr->flags & RDMA_CREATE_AH_SLEEPABLE);
4350
4351         return err;
4352 }
4353
4354 /**
4355  * irdma_query_ah - Query address handle
4356  * @ibah: pointer to address handle
4357  * @ah_attr: address handle attributes
4358  */
4359 static int irdma_query_ah(struct ib_ah *ibah, struct rdma_ah_attr *ah_attr)
4360 {
4361         struct irdma_ah *ah = to_iwah(ibah);
4362
4363         memset(ah_attr, 0, sizeof(*ah_attr));
4364         if (ah->av.attrs.ah_flags & IB_AH_GRH) {
4365                 ah_attr->ah_flags = IB_AH_GRH;
4366                 ah_attr->grh.flow_label = ah->sc_ah.ah_info.flow_label;
4367                 ah_attr->grh.traffic_class = ah->sc_ah.ah_info.tc_tos;
4368                 ah_attr->grh.hop_limit = ah->sc_ah.ah_info.hop_ttl;
4369                 ah_attr->grh.sgid_index = ah->sgid_index;
4370                 ah_attr->grh.sgid_index = ah->sgid_index;
4371                 memcpy(&ah_attr->grh.dgid, &ah->dgid,
4372                        sizeof(ah_attr->grh.dgid));
4373         }
4374
4375         return 0;
4376 }
4377
4378 static enum rdma_link_layer irdma_get_link_layer(struct ib_device *ibdev,
4379                                                  u32 port_num)
4380 {
4381         return IB_LINK_LAYER_ETHERNET;
4382 }
4383
4384 static const struct ib_device_ops irdma_roce_dev_ops = {
4385         .attach_mcast = irdma_attach_mcast,
4386         .create_ah = irdma_create_ah,
4387         .create_user_ah = irdma_create_user_ah,
4388         .destroy_ah = irdma_destroy_ah,
4389         .detach_mcast = irdma_detach_mcast,
4390         .get_link_layer = irdma_get_link_layer,
4391         .get_port_immutable = irdma_roce_port_immutable,
4392         .modify_qp = irdma_modify_qp_roce,
4393         .query_ah = irdma_query_ah,
4394         .query_pkey = irdma_query_pkey,
4395 };
4396
4397 static const struct ib_device_ops irdma_iw_dev_ops = {
4398         .modify_qp = irdma_modify_qp,
4399         .get_port_immutable = irdma_iw_port_immutable,
4400         .query_gid = irdma_query_gid,
4401 };
4402
4403 static const struct ib_device_ops irdma_dev_ops = {
4404         .owner = THIS_MODULE,
4405         .driver_id = RDMA_DRIVER_IRDMA,
4406         .uverbs_abi_ver = IRDMA_ABI_VER,
4407
4408         .alloc_hw_port_stats = irdma_alloc_hw_port_stats,
4409         .alloc_mr = irdma_alloc_mr,
4410         .alloc_mw = irdma_alloc_mw,
4411         .alloc_pd = irdma_alloc_pd,
4412         .alloc_ucontext = irdma_alloc_ucontext,
4413         .create_cq = irdma_create_cq,
4414         .create_qp = irdma_create_qp,
4415         .dealloc_driver = irdma_ib_dealloc_device,
4416         .dealloc_mw = irdma_dealloc_mw,
4417         .dealloc_pd = irdma_dealloc_pd,
4418         .dealloc_ucontext = irdma_dealloc_ucontext,
4419         .dereg_mr = irdma_dereg_mr,
4420         .destroy_cq = irdma_destroy_cq,
4421         .destroy_qp = irdma_destroy_qp,
4422         .disassociate_ucontext = irdma_disassociate_ucontext,
4423         .get_dev_fw_str = irdma_get_dev_fw_str,
4424         .get_dma_mr = irdma_get_dma_mr,
4425         .get_hw_stats = irdma_get_hw_stats,
4426         .map_mr_sg = irdma_map_mr_sg,
4427         .mmap = irdma_mmap,
4428         .mmap_free = irdma_mmap_free,
4429         .poll_cq = irdma_poll_cq,
4430         .post_recv = irdma_post_recv,
4431         .post_send = irdma_post_send,
4432         .query_device = irdma_query_device,
4433         .query_port = irdma_query_port,
4434         .query_qp = irdma_query_qp,
4435         .reg_user_mr = irdma_reg_user_mr,
4436         .req_notify_cq = irdma_req_notify_cq,
4437         .resize_cq = irdma_resize_cq,
4438         INIT_RDMA_OBJ_SIZE(ib_pd, irdma_pd, ibpd),
4439         INIT_RDMA_OBJ_SIZE(ib_ucontext, irdma_ucontext, ibucontext),
4440         INIT_RDMA_OBJ_SIZE(ib_ah, irdma_ah, ibah),
4441         INIT_RDMA_OBJ_SIZE(ib_cq, irdma_cq, ibcq),
4442         INIT_RDMA_OBJ_SIZE(ib_mw, irdma_mr, ibmw),
4443         INIT_RDMA_OBJ_SIZE(ib_qp, irdma_qp, ibqp),
4444 };
4445
4446 /**
4447  * irdma_init_roce_device - initialization of roce rdma device
4448  * @iwdev: irdma device
4449  */
4450 static void irdma_init_roce_device(struct irdma_device *iwdev)
4451 {
4452         iwdev->ibdev.node_type = RDMA_NODE_IB_CA;
4453         addrconf_addr_eui48((u8 *)&iwdev->ibdev.node_guid,
4454                             iwdev->netdev->dev_addr);
4455         ib_set_device_ops(&iwdev->ibdev, &irdma_roce_dev_ops);
4456 }
4457
4458 /**
4459  * irdma_init_iw_device - initialization of iwarp rdma device
4460  * @iwdev: irdma device
4461  */
4462 static int irdma_init_iw_device(struct irdma_device *iwdev)
4463 {
4464         struct net_device *netdev = iwdev->netdev;
4465
4466         iwdev->ibdev.node_type = RDMA_NODE_RNIC;
4467         addrconf_addr_eui48((u8 *)&iwdev->ibdev.node_guid,
4468                             netdev->dev_addr);
4469         iwdev->ibdev.ops.iw_add_ref = irdma_qp_add_ref;
4470         iwdev->ibdev.ops.iw_rem_ref = irdma_qp_rem_ref;
4471         iwdev->ibdev.ops.iw_get_qp = irdma_get_qp;
4472         iwdev->ibdev.ops.iw_connect = irdma_connect;
4473         iwdev->ibdev.ops.iw_accept = irdma_accept;
4474         iwdev->ibdev.ops.iw_reject = irdma_reject;
4475         iwdev->ibdev.ops.iw_create_listen = irdma_create_listen;
4476         iwdev->ibdev.ops.iw_destroy_listen = irdma_destroy_listen;
4477         memcpy(iwdev->ibdev.iw_ifname, netdev->name,
4478                sizeof(iwdev->ibdev.iw_ifname));
4479         ib_set_device_ops(&iwdev->ibdev, &irdma_iw_dev_ops);
4480
4481         return 0;
4482 }
4483
4484 /**
4485  * irdma_init_rdma_device - initialization of rdma device
4486  * @iwdev: irdma device
4487  */
4488 static int irdma_init_rdma_device(struct irdma_device *iwdev)
4489 {
4490         struct pci_dev *pcidev = iwdev->rf->pcidev;
4491         int ret;
4492
4493         if (iwdev->roce_mode) {
4494                 irdma_init_roce_device(iwdev);
4495         } else {
4496                 ret = irdma_init_iw_device(iwdev);
4497                 if (ret)
4498                         return ret;
4499         }
4500         iwdev->ibdev.phys_port_cnt = 1;
4501         iwdev->ibdev.num_comp_vectors = iwdev->rf->ceqs_count;
4502         iwdev->ibdev.dev.parent = &pcidev->dev;
4503         ib_set_device_ops(&iwdev->ibdev, &irdma_dev_ops);
4504
4505         return 0;
4506 }
4507
4508 /**
4509  * irdma_port_ibevent - indicate port event
4510  * @iwdev: irdma device
4511  */
4512 void irdma_port_ibevent(struct irdma_device *iwdev)
4513 {
4514         struct ib_event event;
4515
4516         event.device = &iwdev->ibdev;
4517         event.element.port_num = 1;
4518         event.event =
4519                 iwdev->iw_status ? IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
4520         ib_dispatch_event(&event);
4521 }
4522
4523 /**
4524  * irdma_ib_unregister_device - unregister rdma device from IB
4525  * core
4526  * @iwdev: irdma device
4527  */
4528 void irdma_ib_unregister_device(struct irdma_device *iwdev)
4529 {
4530         iwdev->iw_status = 0;
4531         irdma_port_ibevent(iwdev);
4532         ib_unregister_device(&iwdev->ibdev);
4533 }
4534
4535 /**
4536  * irdma_ib_register_device - register irdma device to IB core
4537  * @iwdev: irdma device
4538  */
4539 int irdma_ib_register_device(struct irdma_device *iwdev)
4540 {
4541         int ret;
4542
4543         ret = irdma_init_rdma_device(iwdev);
4544         if (ret)
4545                 return ret;
4546
4547         ret = ib_device_set_netdev(&iwdev->ibdev, iwdev->netdev, 1);
4548         if (ret)
4549                 goto error;
4550         dma_set_max_seg_size(iwdev->rf->hw.device, UINT_MAX);
4551         ret = ib_register_device(&iwdev->ibdev, "irdma%d", iwdev->rf->hw.device);
4552         if (ret)
4553                 goto error;
4554
4555         iwdev->iw_status = 1;
4556         irdma_port_ibevent(iwdev);
4557
4558         return 0;
4559
4560 error:
4561         if (ret)
4562                 ibdev_dbg(&iwdev->ibdev, "VERBS: Register RDMA device fail\n");
4563
4564         return ret;
4565 }
4566
4567 /**
4568  * irdma_ib_dealloc_device
4569  * @ibdev: ib device
4570  *
4571  * callback from ibdev dealloc_driver to deallocate resources
4572  * unber irdma device
4573  */
4574 void irdma_ib_dealloc_device(struct ib_device *ibdev)
4575 {
4576         struct irdma_device *iwdev = to_iwdev(ibdev);
4577
4578         irdma_rt_deinit_hw(iwdev);
4579         irdma_ctrl_deinit_hw(iwdev->rf);
4580         kfree(iwdev->rf);
4581 }