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         ether_addr_copy((u8 *)&props->sys_image_guid, iwdev->netdev->dev_addr);
  25         props->fw_ver = (u64)irdma_fw_major_ver(&rf->sc_dev) << 32 |
  26                         irdma_fw_minor_ver(&rf->sc_dev);
  27         props->device_cap_flags = iwdev->device_cap_flags;
  28         props->vendor_id = pcidev->vendor;
  29         props->vendor_part_id = pcidev->device;
  30
  31         props->hw_ver = rf->pcidev->revision;
  32         props->page_size_cap = SZ_4K | SZ_2M | SZ_1G;
  33         props->max_mr_size = hw_attrs->max_mr_size;
  34         props->max_qp = rf->max_qp - rf->used_qps;
  35         props->max_qp_wr = hw_attrs->max_qp_wr;
  36         props->max_send_sge = hw_attrs->uk_attrs.max_hw_wq_frags;
  37         props->max_recv_sge = hw_attrs->uk_attrs.max_hw_wq_frags;
  38         props->max_cq = rf->max_cq - rf->used_cqs;
  39         props->max_cqe = rf->max_cqe;
  40         props->max_mr = rf->max_mr - rf->used_mrs;
  41         props->max_mw = props->max_mr;
  42         props->max_pd = rf->max_pd - rf->used_pds;
  43         props->max_sge_rd = hw_attrs->uk_attrs.max_hw_read_sges;
  44         props->max_qp_rd_atom = hw_attrs->max_hw_ird;
  45         props->max_qp_init_rd_atom = hw_attrs->max_hw_ord;
  46         if (rdma_protocol_roce(ibdev, 1))
  47                 props->max_pkeys = IRDMA_PKEY_TBL_SZ;
  48         props->max_ah = rf->max_ah;
  49         props->max_mcast_grp = rf->max_mcg;
  50         props->max_mcast_qp_attach = IRDMA_MAX_MGS_PER_CTX;
  51         props->max_total_mcast_qp_attach = rf->max_qp * IRDMA_MAX_MGS_PER_CTX;
  52         props->max_fast_reg_page_list_len = IRDMA_MAX_PAGES_PER_FMR;
  53 #define HCA_CLOCK_TIMESTAMP_MASK 0x1ffff
  54         if (hw_attrs->uk_attrs.hw_rev >= IRDMA_GEN_2)
  55                 props->timestamp_mask = HCA_CLOCK_TIMESTAMP_MASK;
  56
  57         return 0;
  58 }
  59
  60 /**
  61  * irdma_get_eth_speed_and_width - Get IB port speed and width from netdev speed
  62  * @link_speed: netdev phy link speed
  63  * @active_speed: IB port speed
  64  * @active_width: IB port width
  65  */
  66 static void irdma_get_eth_speed_and_width(u32 link_speed, u16 *active_speed,
  67                                           u8 *active_width)
  68 {
  69         if (link_speed <= SPEED_1000) {
  70                 *active_width = IB_WIDTH_1X;
  71                 *active_speed = IB_SPEED_SDR;
  72         } else if (link_speed <= SPEED_10000) {
  73                 *active_width = IB_WIDTH_1X;
  74                 *active_speed = IB_SPEED_FDR10;
  75         } else if (link_speed <= SPEED_20000) {
  76                 *active_width = IB_WIDTH_4X;
  77                 *active_speed = IB_SPEED_DDR;
  78         } else if (link_speed <= SPEED_25000) {
  79                 *active_width = IB_WIDTH_1X;
  80                 *active_speed = IB_SPEED_EDR;
  81         } else if (link_speed <= SPEED_40000) {
  82                 *active_width = IB_WIDTH_4X;
  83                 *active_speed = IB_SPEED_FDR10;
  84         } else {
  85                 *active_width = IB_WIDTH_4X;
  86                 *active_speed = IB_SPEED_EDR;
  87         }
  88 }
  89
  90 /**
  91  * irdma_query_port - get port attributes
  92  * @ibdev: device pointer from stack
  93  * @port: port number for query
  94  * @props: returning device attributes
  95  */
  96 static int irdma_query_port(struct ib_device *ibdev, u32 port,
  97                             struct ib_port_attr *props)
  98 {
  99         struct irdma_device *iwdev = to_iwdev(ibdev);
 100         struct net_device *netdev = iwdev->netdev;
 101
 102         /* no need to zero out pros here. done by caller */
 103
 104         props->max_mtu = IB_MTU_4096;
 105         props->active_mtu = ib_mtu_int_to_enum(netdev->mtu);
 106         props->lid = 1;
 107         props->lmc = 0;
 108         props->sm_lid = 0;
 109         props->sm_sl = 0;
 110         if (netif_carrier_ok(netdev) && netif_running(netdev)) {
 111                 props->state = IB_PORT_ACTIVE;
 112                 props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
 113         } else {
 114                 props->state = IB_PORT_DOWN;
 115                 props->phys_state = IB_PORT_PHYS_STATE_DISABLED;
 116         }
 117         irdma_get_eth_speed_and_width(SPEED_100000, &props->active_speed,
 118                                       &props->active_width);
 119
 120         if (rdma_protocol_roce(ibdev, 1)) {
 121                 props->gid_tbl_len = 32;
 122                 props->ip_gids = true;
 123                 props->pkey_tbl_len = IRDMA_PKEY_TBL_SZ;
 124         } else {
 125                 props->gid_tbl_len = 1;
 126         }
 127         props->qkey_viol_cntr = 0;
 128         props->port_cap_flags |= IB_PORT_CM_SUP | IB_PORT_REINIT_SUP;
 129         props->max_msg_sz = iwdev->rf->sc_dev.hw_attrs.max_hw_outbound_msg_size;
 130
 131         return 0;
 132 }
 133
 134 /**
 135  * irdma_disassociate_ucontext - Disassociate user context
 136  * @context: ib user context
 137  */
 138 static void irdma_disassociate_ucontext(struct ib_ucontext *context)
 139 {
 140 }
 141
 142 static int irdma_mmap_legacy(struct irdma_ucontext *ucontext,
 143                              struct vm_area_struct *vma)
 144 {
 145         u64 pfn;
 146
 147         if (vma->vm_pgoff || vma->vm_end - vma->vm_start != PAGE_SIZE)
 148                 return -EINVAL;
 149
 150         vma->vm_private_data = ucontext;
 151         pfn = ((uintptr_t)ucontext->iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET] +
 152                pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;
 153
 154         return rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn, PAGE_SIZE,
 155                                  pgprot_noncached(vma->vm_page_prot), NULL);
 156 }
 157
 158 static void irdma_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
 159 {
 160         struct irdma_user_mmap_entry *entry = to_irdma_mmap_entry(rdma_entry);
 161
 162         kfree(entry);
 163 }
 164
 165 static struct rdma_user_mmap_entry*
 166 irdma_user_mmap_entry_insert(struct irdma_ucontext *ucontext, u64 bar_offset,
 167                              enum irdma_mmap_flag mmap_flag, u64 *mmap_offset)
 168 {
 169         struct irdma_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 170         int ret;
 171
 172         if (!entry)
 173                 return NULL;
 174
 175         entry->bar_offset = bar_offset;
 176         entry->mmap_flag = mmap_flag;
 177
 178         ret = rdma_user_mmap_entry_insert(&ucontext->ibucontext,
 179                                           &entry->rdma_entry, PAGE_SIZE);
 180         if (ret) {
 181                 kfree(entry);
 182                 return NULL;
 183         }
 184         *mmap_offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
 185
 186         return &entry->rdma_entry;
 187 }
 188
 189 /**
 190  * irdma_mmap - user memory map
 191  * @context: context created during alloc
 192  * @vma: kernel info for user memory map
 193  */
 194 static int irdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
 195 {
 196         struct rdma_user_mmap_entry *rdma_entry;
 197         struct irdma_user_mmap_entry *entry;
 198         struct irdma_ucontext *ucontext;
 199         u64 pfn;
 200         int ret;
 201
 202         ucontext = to_ucontext(context);
 203
 204         /* Legacy support for libi40iw with hard-coded mmap key */
 205         if (ucontext->legacy_mode)
 206                 return irdma_mmap_legacy(ucontext, vma);
 207
 208         rdma_entry = rdma_user_mmap_entry_get(&ucontext->ibucontext, vma);
 209         if (!rdma_entry) {
 210                 ibdev_dbg(&ucontext->iwdev->ibdev,
 211                           "VERBS: pgoff[0x%lx] does not have valid entry\n",
 212                           vma->vm_pgoff);
 213                 return -EINVAL;
 214         }
 215
 216         entry = to_irdma_mmap_entry(rdma_entry);
 217         ibdev_dbg(&ucontext->iwdev->ibdev,
 218                   "VERBS: bar_offset [0x%llx] mmap_flag [%d]\n",
 219                   entry->bar_offset, entry->mmap_flag);
 220
 221         pfn = (entry->bar_offset +
 222               pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;
 223
 224         switch (entry->mmap_flag) {
 225         case IRDMA_MMAP_IO_NC:
 226                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
 227                                         pgprot_noncached(vma->vm_page_prot),
 228                                         rdma_entry);
 229                 break;
 230         case IRDMA_MMAP_IO_WC:
 231                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
 232                                         pgprot_writecombine(vma->vm_page_prot),
 233                                         rdma_entry);
 234                 break;
 235         default:
 236                 ret = -EINVAL;
 237         }
 238
 239         if (ret)
 240                 ibdev_dbg(&ucontext->iwdev->ibdev,
 241                           "VERBS: bar_offset [0x%llx] mmap_flag[%d] err[%d]\n",
 242                           entry->bar_offset, entry->mmap_flag, ret);
 243         rdma_user_mmap_entry_put(rdma_entry);
 244
 245         return ret;
 246 }
 247
 248 /**
 249  * irdma_alloc_push_page - allocate a push page for qp
 250  * @iwqp: qp pointer
 251  */
 252 static void irdma_alloc_push_page(struct irdma_qp *iwqp)
 253 {
 254         struct irdma_cqp_request *cqp_request;
 255         struct cqp_cmds_info *cqp_info;
 256         struct irdma_device *iwdev = iwqp->iwdev;
 257         struct irdma_sc_qp *qp = &iwqp->sc_qp;
 258         enum irdma_status_code status;
 259
 260         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
 261         if (!cqp_request)
 262                 return;
 263
 264         cqp_info = &cqp_request->info;
 265         cqp_info->cqp_cmd = IRDMA_OP_MANAGE_PUSH_PAGE;
 266         cqp_info->post_sq = 1;
 267         cqp_info->in.u.manage_push_page.info.push_idx = 0;
 268         cqp_info->in.u.manage_push_page.info.qs_handle =
 269                 qp->vsi->qos[qp->user_pri].qs_handle;
 270         cqp_info->in.u.manage_push_page.info.free_page = 0;
 271         cqp_info->in.u.manage_push_page.info.push_page_type = 0;
 272         cqp_info->in.u.manage_push_page.cqp = &iwdev->rf->cqp.sc_cqp;
 273         cqp_info->in.u.manage_push_page.scratch = (uintptr_t)cqp_request;
 274
 275         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
 276         if (!status && cqp_request->compl_info.op_ret_val <
 277             iwdev->rf->sc_dev.hw_attrs.max_hw_device_pages) {
 278                 qp->push_idx = cqp_request->compl_info.op_ret_val;
 279                 qp->push_offset = 0;
 280         }
 281
 282         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
 283 }
 284
 285 /**
 286  * irdma_alloc_ucontext - Allocate the user context data structure
 287  * @uctx: uverbs context pointer
 288  * @udata: user data
 289  *
 290  * This keeps track of all objects associated with a particular
 291  * user-mode client.
 292  */
 293 static int irdma_alloc_ucontext(struct ib_ucontext *uctx,
 294                                 struct ib_udata *udata)
 295 {
 296         struct ib_device *ibdev = uctx->device;
 297         struct irdma_device *iwdev = to_iwdev(ibdev);
 298         struct irdma_alloc_ucontext_req req;
 299         struct irdma_alloc_ucontext_resp uresp = {};
 300         struct irdma_ucontext *ucontext = to_ucontext(uctx);
 301         struct irdma_uk_attrs *uk_attrs;
 302
 303         if (ib_copy_from_udata(&req, udata, min(sizeof(req), udata->inlen)))
 304                 return -EINVAL;
 305
 306         if (req.userspace_ver < 4 || req.userspace_ver > IRDMA_ABI_VER)
 307                 goto ver_error;
 308
 309         ucontext->iwdev = iwdev;
 310         ucontext->abi_ver = req.userspace_ver;
 311
 312         uk_attrs = &iwdev->rf->sc_dev.hw_attrs.uk_attrs;
 313         /* GEN_1 legacy support with libi40iw */
 314         if (udata->outlen < sizeof(uresp)) {
 315                 if (uk_attrs->hw_rev != IRDMA_GEN_1)
 316                         return -EOPNOTSUPP;
 317
 318                 ucontext->legacy_mode = true;
 319                 uresp.max_qps = iwdev->rf->max_qp;
 320                 uresp.max_pds = iwdev->rf->sc_dev.hw_attrs.max_hw_pds;
 321                 uresp.wq_size = iwdev->rf->sc_dev.hw_attrs.max_qp_wr * 2;
 322                 uresp.kernel_ver = req.userspace_ver;
 323                 if (ib_copy_to_udata(udata, &uresp,
 324                                      min(sizeof(uresp), udata->outlen)))
 325                         return -EFAULT;
 326         } else {
 327                 u64 bar_off = (uintptr_t)iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET];
 328
 329                 ucontext->db_mmap_entry =
 330                         irdma_user_mmap_entry_insert(ucontext, bar_off,
 331                                                      IRDMA_MMAP_IO_NC,
 332                                                      &uresp.db_mmap_key);
 333                 if (!ucontext->db_mmap_entry)
 334                         return -ENOMEM;
 335
 336                 uresp.kernel_ver = IRDMA_ABI_VER;
 337                 uresp.feature_flags = uk_attrs->feature_flags;
 338                 uresp.max_hw_wq_frags = uk_attrs->max_hw_wq_frags;
 339                 uresp.max_hw_read_sges = uk_attrs->max_hw_read_sges;
 340                 uresp.max_hw_inline = uk_attrs->max_hw_inline;
 341                 uresp.max_hw_rq_quanta = uk_attrs->max_hw_rq_quanta;
 342                 uresp.max_hw_wq_quanta = uk_attrs->max_hw_wq_quanta;
 343                 uresp.max_hw_sq_chunk = uk_attrs->max_hw_sq_chunk;
 344                 uresp.max_hw_cq_size = uk_attrs->max_hw_cq_size;
 345                 uresp.min_hw_cq_size = uk_attrs->min_hw_cq_size;
 346                 uresp.hw_rev = uk_attrs->hw_rev;
 347                 if (ib_copy_to_udata(udata, &uresp,
 348                                      min(sizeof(uresp), udata->outlen))) {
 349                         rdma_user_mmap_entry_remove(ucontext->db_mmap_entry);
 350                         return -EFAULT;
 351                 }
 352         }
 353
 354         INIT_LIST_HEAD(&ucontext->cq_reg_mem_list);
 355         spin_lock_init(&ucontext->cq_reg_mem_list_lock);
 356         INIT_LIST_HEAD(&ucontext->qp_reg_mem_list);
 357         spin_lock_init(&ucontext->qp_reg_mem_list_lock);
 358
 359         return 0;
 360
 361 ver_error:
 362         ibdev_err(&iwdev->ibdev,
 363                   "Invalid userspace driver version detected. Detected version %d, should be %d\n",
 364                   req.userspace_ver, IRDMA_ABI_VER);
 365         return -EINVAL;
 366 }
 367
 368 /**
 369  * irdma_dealloc_ucontext - deallocate the user context data structure
 370  * @context: user context created during alloc
 371  */
 372 static void irdma_dealloc_ucontext(struct ib_ucontext *context)
 373 {
 374         struct irdma_ucontext *ucontext = to_ucontext(context);
 375
 376         rdma_user_mmap_entry_remove(ucontext->db_mmap_entry);
 377 }
 378
 379 /**
 380  * irdma_alloc_pd - allocate protection domain
 381  * @pd: PD pointer
 382  * @udata: user data
 383  */
 384 static int irdma_alloc_pd(struct ib_pd *pd, struct ib_udata *udata)
 385 {
 386         struct irdma_pd *iwpd = to_iwpd(pd);
 387         struct irdma_device *iwdev = to_iwdev(pd->device);
 388         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 389         struct irdma_pci_f *rf = iwdev->rf;
 390         struct irdma_alloc_pd_resp uresp = {};
 391         struct irdma_sc_pd *sc_pd;
 392         u32 pd_id = 0;
 393         int err;
 394
 395         err = irdma_alloc_rsrc(rf, rf->allocated_pds, rf->max_pd, &pd_id,
 396                                &rf->next_pd);
 397         if (err)
 398                 return err;
 399
 400         sc_pd = &iwpd->sc_pd;
 401         if (udata) {
 402                 struct irdma_ucontext *ucontext =
 403                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
 404                                                   ibucontext);
 405                 irdma_sc_pd_init(dev, sc_pd, pd_id, ucontext->abi_ver);
 406                 uresp.pd_id = pd_id;
 407                 if (ib_copy_to_udata(udata, &uresp,
 408                                      min(sizeof(uresp), udata->outlen))) {
 409                         err = -EFAULT;
 410                         goto error;
 411                 }
 412         } else {
 413                 irdma_sc_pd_init(dev, sc_pd, pd_id, IRDMA_ABI_VER);
 414         }
 415
 416         return 0;
 417 error:
 418         irdma_free_rsrc(rf, rf->allocated_pds, pd_id);
 419
 420         return err;
 421 }
 422
 423 /**
 424  * irdma_dealloc_pd - deallocate pd
 425  * @ibpd: ptr of pd to be deallocated
 426  * @udata: user data
 427  */
 428 static int irdma_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 429 {
 430         struct irdma_pd *iwpd = to_iwpd(ibpd);
 431         struct irdma_device *iwdev = to_iwdev(ibpd->device);
 432
 433         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_pds, iwpd->sc_pd.pd_id);
 434
 435         return 0;
 436 }
 437
 438 /**
 439  * irdma_get_pbl - Retrieve pbl from a list given a virtual
 440  * address
 441  * @va: user virtual address
 442  * @pbl_list: pbl list to search in (QP's or CQ's)
 443  */
 444 static struct irdma_pbl *irdma_get_pbl(unsigned long va,
 445                                        struct list_head *pbl_list)
 446 {
 447         struct irdma_pbl *iwpbl;
 448
 449         list_for_each_entry (iwpbl, pbl_list, list) {
 450                 if (iwpbl->user_base == va) {
 451                         list_del(&iwpbl->list);
 452                         iwpbl->on_list = false;
 453                         return iwpbl;
 454                 }
 455         }
 456
 457         return NULL;
 458 }
 459
 460 /**
 461  * irdma_clean_cqes - clean cq entries for qp
 462  * @iwqp: qp ptr (user or kernel)
 463  * @iwcq: cq ptr
 464  */
 465 static void irdma_clean_cqes(struct irdma_qp *iwqp, struct irdma_cq *iwcq)
 466 {
 467         struct irdma_cq_uk *ukcq = &iwcq->sc_cq.cq_uk;
 468         unsigned long flags;
 469
 470         spin_lock_irqsave(&iwcq->lock, flags);
 471         irdma_uk_clean_cq(&iwqp->sc_qp.qp_uk, ukcq);
 472         spin_unlock_irqrestore(&iwcq->lock, flags);
 473 }
 474
 475 static void irdma_remove_push_mmap_entries(struct irdma_qp *iwqp)
 476 {
 477         if (iwqp->push_db_mmap_entry) {
 478                 rdma_user_mmap_entry_remove(iwqp->push_db_mmap_entry);
 479                 iwqp->push_db_mmap_entry = NULL;
 480         }
 481         if (iwqp->push_wqe_mmap_entry) {
 482                 rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
 483                 iwqp->push_wqe_mmap_entry = NULL;
 484         }
 485 }
 486
 487 static int irdma_setup_push_mmap_entries(struct irdma_ucontext *ucontext,
 488                                          struct irdma_qp *iwqp,
 489                                          u64 *push_wqe_mmap_key,
 490                                          u64 *push_db_mmap_key)
 491 {
 492         struct irdma_device *iwdev = ucontext->iwdev;
 493         u64 rsvd, bar_off;
 494
 495         rsvd = IRDMA_PF_BAR_RSVD;
 496         bar_off = (uintptr_t)iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET];
 497         /* skip over db page */
 498         bar_off += IRDMA_HW_PAGE_SIZE;
 499         /* push wqe page */
 500         bar_off += rsvd + iwqp->sc_qp.push_idx * IRDMA_HW_PAGE_SIZE;
 501         iwqp->push_wqe_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
 502                                         bar_off, IRDMA_MMAP_IO_WC,
 503                                         push_wqe_mmap_key);
 504         if (!iwqp->push_wqe_mmap_entry)
 505                 return -ENOMEM;
 506
 507         /* push doorbell page */
 508         bar_off += IRDMA_HW_PAGE_SIZE;
 509         iwqp->push_db_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
 510                                         bar_off, IRDMA_MMAP_IO_NC,
 511                                         push_db_mmap_key);
 512         if (!iwqp->push_db_mmap_entry) {
 513                 rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
 514                 return -ENOMEM;
 515         }
 516
 517         return 0;
 518 }
 519
 520 /**
 521  * irdma_destroy_qp - destroy qp
 522  * @ibqp: qp's ib pointer also to get to device's qp address
 523  * @udata: user data
 524  */
 525 static int irdma_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
 526 {
 527         struct irdma_qp *iwqp = to_iwqp(ibqp);
 528         struct irdma_device *iwdev = iwqp->iwdev;
 529
 530         iwqp->sc_qp.qp_uk.destroy_pending = true;
 531
 532         if (iwqp->iwarp_state == IRDMA_QP_STATE_RTS)
 533                 irdma_modify_qp_to_err(&iwqp->sc_qp);
 534
 535         irdma_qp_rem_ref(&iwqp->ibqp);
 536         wait_for_completion(&iwqp->free_qp);
 537         irdma_free_lsmm_rsrc(iwqp);
 538         if (!iwdev->reset)
 539                 irdma_cqp_qp_destroy_cmd(&iwdev->rf->sc_dev, &iwqp->sc_qp);
 540
 541         if (!iwqp->user_mode) {
 542                 if (iwqp->iwscq) {
 543                         irdma_clean_cqes(iwqp, iwqp->iwscq);
 544                         if (iwqp->iwrcq != iwqp->iwscq)
 545                                 irdma_clean_cqes(iwqp, iwqp->iwrcq);
 546                 }
 547         }
 548         irdma_remove_push_mmap_entries(iwqp);
 549         irdma_free_qp_rsrc(iwqp);
 550
 551         return 0;
 552 }
 553
 554 /**
 555  * irdma_setup_virt_qp - setup for allocation of virtual qp
 556  * @iwdev: irdma device
 557  * @iwqp: qp ptr
 558  * @init_info: initialize info to return
 559  */
 560 static void irdma_setup_virt_qp(struct irdma_device *iwdev,
 561                                struct irdma_qp *iwqp,
 562                                struct irdma_qp_init_info *init_info)
 563 {
 564         struct irdma_pbl *iwpbl = iwqp->iwpbl;
 565         struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;
 566
 567         iwqp->page = qpmr->sq_page;
 568         init_info->shadow_area_pa = qpmr->shadow;
 569         if (iwpbl->pbl_allocated) {
 570                 init_info->virtual_map = true;
 571                 init_info->sq_pa = qpmr->sq_pbl.idx;
 572                 init_info->rq_pa = qpmr->rq_pbl.idx;
 573         } else {
 574                 init_info->sq_pa = qpmr->sq_pbl.addr;
 575                 init_info->rq_pa = qpmr->rq_pbl.addr;
 576         }
 577 }
 578
 579 /**
 580  * irdma_setup_kmode_qp - setup initialization for kernel mode qp
 581  * @iwdev: iwarp device
 582  * @iwqp: qp ptr (user or kernel)
 583  * @info: initialize info to return
 584  * @init_attr: Initial QP create attributes
 585  */
 586 static int irdma_setup_kmode_qp(struct irdma_device *iwdev,
 587                                 struct irdma_qp *iwqp,
 588                                 struct irdma_qp_init_info *info,
 589                                 struct ib_qp_init_attr *init_attr)
 590 {
 591         struct irdma_dma_mem *mem = &iwqp->kqp.dma_mem;
 592         u32 sqdepth, rqdepth;
 593         u8 sqshift, rqshift;
 594         u32 size;
 595         enum irdma_status_code status;
 596         struct irdma_qp_uk_init_info *ukinfo = &info->qp_uk_init_info;
 597         struct irdma_uk_attrs *uk_attrs = &iwdev->rf->sc_dev.hw_attrs.uk_attrs;
 598
 599         irdma_get_wqe_shift(uk_attrs,
 600                 uk_attrs->hw_rev >= IRDMA_GEN_2 ? ukinfo->max_sq_frag_cnt + 1 :
 601                                                   ukinfo->max_sq_frag_cnt,
 602                 ukinfo->max_inline_data, &sqshift);
 603         status = irdma_get_sqdepth(uk_attrs, ukinfo->sq_size, sqshift,
 604                                    &sqdepth);
 605         if (status)
 606                 return -ENOMEM;
 607
 608         if (uk_attrs->hw_rev == IRDMA_GEN_1)
 609                 rqshift = IRDMA_MAX_RQ_WQE_SHIFT_GEN1;
 610         else
 611                 irdma_get_wqe_shift(uk_attrs, ukinfo->max_rq_frag_cnt, 0,
 612                                     &rqshift);
 613
 614         status = irdma_get_rqdepth(uk_attrs, ukinfo->rq_size, rqshift,
 615                                    &rqdepth);
 616         if (status)
 617                 return -ENOMEM;
 618
 619         iwqp->kqp.sq_wrid_mem =
 620                 kcalloc(sqdepth, sizeof(*iwqp->kqp.sq_wrid_mem), GFP_KERNEL);
 621         if (!iwqp->kqp.sq_wrid_mem)
 622                 return -ENOMEM;
 623
 624         iwqp->kqp.rq_wrid_mem =
 625                 kcalloc(rqdepth, sizeof(*iwqp->kqp.rq_wrid_mem), GFP_KERNEL);
 626         if (!iwqp->kqp.rq_wrid_mem) {
 627                 kfree(iwqp->kqp.sq_wrid_mem);
 628                 iwqp->kqp.sq_wrid_mem = NULL;
 629                 return -ENOMEM;
 630         }
 631
 632         ukinfo->sq_wrtrk_array = iwqp->kqp.sq_wrid_mem;
 633         ukinfo->rq_wrid_array = iwqp->kqp.rq_wrid_mem;
 634
 635         size = (sqdepth + rqdepth) * IRDMA_QP_WQE_MIN_SIZE;
 636         size += (IRDMA_SHADOW_AREA_SIZE << 3);
 637
 638         mem->size = ALIGN(size, 256);
 639         mem->va = dma_alloc_coherent(iwdev->rf->hw.device, mem->size,
 640                                      &mem->pa, GFP_KERNEL);
 641         if (!mem->va) {
 642                 kfree(iwqp->kqp.sq_wrid_mem);
 643                 iwqp->kqp.sq_wrid_mem = NULL;
 644                 kfree(iwqp->kqp.rq_wrid_mem);
 645                 iwqp->kqp.rq_wrid_mem = NULL;
 646                 return -ENOMEM;
 647         }
 648
 649         ukinfo->sq = mem->va;
 650         info->sq_pa = mem->pa;
 651         ukinfo->rq = &ukinfo->sq[sqdepth];
 652         info->rq_pa = info->sq_pa + (sqdepth * IRDMA_QP_WQE_MIN_SIZE);
 653         ukinfo->shadow_area = ukinfo->rq[rqdepth].elem;
 654         info->shadow_area_pa = info->rq_pa + (rqdepth * IRDMA_QP_WQE_MIN_SIZE);
 655         ukinfo->sq_size = sqdepth >> sqshift;
 656         ukinfo->rq_size = rqdepth >> rqshift;
 657         ukinfo->qp_id = iwqp->ibqp.qp_num;
 658
 659         init_attr->cap.max_send_wr = (sqdepth - IRDMA_SQ_RSVD) >> sqshift;
 660         init_attr->cap.max_recv_wr = (rqdepth - IRDMA_RQ_RSVD) >> rqshift;
 661
 662         return 0;
 663 }
 664
 665 static int irdma_cqp_create_qp_cmd(struct irdma_qp *iwqp)
 666 {
 667         struct irdma_pci_f *rf = iwqp->iwdev->rf;
 668         struct irdma_cqp_request *cqp_request;
 669         struct cqp_cmds_info *cqp_info;
 670         struct irdma_create_qp_info *qp_info;
 671         enum irdma_status_code status;
 672
 673         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
 674         if (!cqp_request)
 675                 return -ENOMEM;
 676
 677         cqp_info = &cqp_request->info;
 678         qp_info = &cqp_request->info.in.u.qp_create.info;
 679         memset(qp_info, 0, sizeof(*qp_info));
 680         qp_info->mac_valid = true;
 681         qp_info->cq_num_valid = true;
 682         qp_info->next_iwarp_state = IRDMA_QP_STATE_IDLE;
 683
 684         cqp_info->cqp_cmd = IRDMA_OP_QP_CREATE;
 685         cqp_info->post_sq = 1;
 686         cqp_info->in.u.qp_create.qp = &iwqp->sc_qp;
 687         cqp_info->in.u.qp_create.scratch = (uintptr_t)cqp_request;
 688         status = irdma_handle_cqp_op(rf, cqp_request);
 689         irdma_put_cqp_request(&rf->cqp, cqp_request);
 690
 691         return status ? -ENOMEM : 0;
 692 }
 693
 694 static void irdma_roce_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
 695                                                struct irdma_qp_host_ctx_info *ctx_info)
 696 {
 697         struct irdma_device *iwdev = iwqp->iwdev;
 698         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 699         struct irdma_roce_offload_info *roce_info;
 700         struct irdma_udp_offload_info *udp_info;
 701
 702         udp_info = &iwqp->udp_info;
 703         udp_info->snd_mss = ib_mtu_enum_to_int(ib_mtu_int_to_enum(iwdev->vsi.mtu));
 704         udp_info->cwnd = iwdev->roce_cwnd;
 705         udp_info->rexmit_thresh = 2;
 706         udp_info->rnr_nak_thresh = 2;
 707         udp_info->src_port = 0xc000;
 708         udp_info->dst_port = ROCE_V2_UDP_DPORT;
 709         roce_info = &iwqp->roce_info;
 710         ether_addr_copy(roce_info->mac_addr, iwdev->netdev->dev_addr);
 711
 712         roce_info->rd_en = true;
 713         roce_info->wr_rdresp_en = true;
 714         roce_info->bind_en = true;
 715         roce_info->dcqcn_en = false;
 716         roce_info->rtomin = 5;
 717
 718         roce_info->ack_credits = iwdev->roce_ackcreds;
 719         roce_info->ird_size = dev->hw_attrs.max_hw_ird;
 720         roce_info->ord_size = dev->hw_attrs.max_hw_ord;
 721
 722         if (!iwqp->user_mode) {
 723                 roce_info->priv_mode_en = true;
 724                 roce_info->fast_reg_en = true;
 725                 roce_info->udprivcq_en = true;
 726         }
 727         roce_info->roce_tver = 0;
 728
 729         ctx_info->roce_info = &iwqp->roce_info;
 730         ctx_info->udp_info = &iwqp->udp_info;
 731         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
 732 }
 733
 734 static void irdma_iw_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
 735                                              struct irdma_qp_host_ctx_info *ctx_info)
 736 {
 737         struct irdma_device *iwdev = iwqp->iwdev;
 738         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 739         struct irdma_iwarp_offload_info *iwarp_info;
 740
 741         iwarp_info = &iwqp->iwarp_info;
 742         ether_addr_copy(iwarp_info->mac_addr, iwdev->netdev->dev_addr);
 743         iwarp_info->rd_en = true;
 744         iwarp_info->wr_rdresp_en = true;
 745         iwarp_info->bind_en = true;
 746         iwarp_info->ecn_en = true;
 747         iwarp_info->rtomin = 5;
 748
 749         if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
 750                 iwarp_info->ib_rd_en = true;
 751         if (!iwqp->user_mode) {
 752                 iwarp_info->priv_mode_en = true;
 753                 iwarp_info->fast_reg_en = true;
 754         }
 755         iwarp_info->ddp_ver = 1;
 756         iwarp_info->rdmap_ver = 1;
 757
 758         ctx_info->iwarp_info = &iwqp->iwarp_info;
 759         ctx_info->iwarp_info_valid = true;
 760         irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
 761         ctx_info->iwarp_info_valid = false;
 762 }
 763
 764 static int irdma_validate_qp_attrs(struct ib_qp_init_attr *init_attr,
 765                                    struct irdma_device *iwdev)
 766 {
 767         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 768         struct irdma_uk_attrs *uk_attrs = &dev->hw_attrs.uk_attrs;
 769
 770         if (init_attr->create_flags)
 771                 return -EOPNOTSUPP;
 772
 773         if (init_attr->cap.max_inline_data > uk_attrs->max_hw_inline ||
 774             init_attr->cap.max_send_sge > uk_attrs->max_hw_wq_frags ||
 775             init_attr->cap.max_recv_sge > uk_attrs->max_hw_wq_frags)
 776                 return -EINVAL;
 777
 778         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 779                 if (init_attr->qp_type != IB_QPT_RC &&
 780                     init_attr->qp_type != IB_QPT_UD &&
 781                     init_attr->qp_type != IB_QPT_GSI)
 782                         return -EOPNOTSUPP;
 783         } else {
 784                 if (init_attr->qp_type != IB_QPT_RC)
 785                         return -EOPNOTSUPP;
 786         }
 787
 788         return 0;
 789 }
 790
 791 /**
 792  * irdma_create_qp - create qp
 793  * @ibpd: ptr of pd
 794  * @init_attr: attributes for qp
 795  * @udata: user data for create qp
 796  */
 797 static struct ib_qp *irdma_create_qp(struct ib_pd *ibpd,
 798                                      struct ib_qp_init_attr *init_attr,
 799                                      struct ib_udata *udata)
 800 {
 801         struct irdma_pd *iwpd = to_iwpd(ibpd);
 802         struct irdma_device *iwdev = to_iwdev(ibpd->device);
 803         struct irdma_pci_f *rf = iwdev->rf;
 804         struct irdma_qp *iwqp;
 805         struct irdma_create_qp_req req;
 806         struct irdma_create_qp_resp uresp = {};
 807         u32 qp_num = 0;
 808         enum irdma_status_code ret;
 809         int err_code;
 810         int sq_size;
 811         int rq_size;
 812         struct irdma_sc_qp *qp;
 813         struct irdma_sc_dev *dev = &rf->sc_dev;
 814         struct irdma_uk_attrs *uk_attrs = &dev->hw_attrs.uk_attrs;
 815         struct irdma_qp_init_info init_info = {};
 816         struct irdma_qp_host_ctx_info *ctx_info;
 817         unsigned long flags;
 818
 819         err_code = irdma_validate_qp_attrs(init_attr, iwdev);
 820         if (err_code)
 821                 return ERR_PTR(err_code);
 822
 823         sq_size = init_attr->cap.max_send_wr;
 824         rq_size = init_attr->cap.max_recv_wr;
 825
 826         init_info.vsi = &iwdev->vsi;
 827         init_info.qp_uk_init_info.uk_attrs = uk_attrs;
 828         init_info.qp_uk_init_info.sq_size = sq_size;
 829         init_info.qp_uk_init_info.rq_size = rq_size;
 830         init_info.qp_uk_init_info.max_sq_frag_cnt = init_attr->cap.max_send_sge;
 831         init_info.qp_uk_init_info.max_rq_frag_cnt = init_attr->cap.max_recv_sge;
 832         init_info.qp_uk_init_info.max_inline_data = init_attr->cap.max_inline_data;
 833
 834         iwqp = kzalloc(sizeof(*iwqp), GFP_KERNEL);
 835         if (!iwqp)
 836                 return ERR_PTR(-ENOMEM);
 837
 838         qp = &iwqp->sc_qp;
 839         qp->qp_uk.back_qp = iwqp;
 840         qp->qp_uk.lock = &iwqp->lock;
 841         qp->push_idx = IRDMA_INVALID_PUSH_PAGE_INDEX;
 842
 843         iwqp->iwdev = iwdev;
 844         iwqp->q2_ctx_mem.size = ALIGN(IRDMA_Q2_BUF_SIZE + IRDMA_QP_CTX_SIZE,
 845                                       256);
 846         iwqp->q2_ctx_mem.va = dma_alloc_coherent(dev->hw->device,
 847                                                  iwqp->q2_ctx_mem.size,
 848                                                  &iwqp->q2_ctx_mem.pa,
 849                                                  GFP_KERNEL);
 850         if (!iwqp->q2_ctx_mem.va) {
 851                 err_code = -ENOMEM;
 852                 goto error;
 853         }
 854
 855         init_info.q2 = iwqp->q2_ctx_mem.va;
 856         init_info.q2_pa = iwqp->q2_ctx_mem.pa;
 857         init_info.host_ctx = (__le64 *)(init_info.q2 + IRDMA_Q2_BUF_SIZE);
 858         init_info.host_ctx_pa = init_info.q2_pa + IRDMA_Q2_BUF_SIZE;
 859
 860         if (init_attr->qp_type == IB_QPT_GSI)
 861                 qp_num = 1;
 862         else
 863                 err_code = irdma_alloc_rsrc(rf, rf->allocated_qps, rf->max_qp,
 864                                             &qp_num, &rf->next_qp);
 865         if (err_code)
 866                 goto error;
 867
 868         iwqp->iwpd = iwpd;
 869         iwqp->ibqp.qp_num = qp_num;
 870         qp = &iwqp->sc_qp;
 871         iwqp->iwscq = to_iwcq(init_attr->send_cq);
 872         iwqp->iwrcq = to_iwcq(init_attr->recv_cq);
 873         iwqp->host_ctx.va = init_info.host_ctx;
 874         iwqp->host_ctx.pa = init_info.host_ctx_pa;
 875         iwqp->host_ctx.size = IRDMA_QP_CTX_SIZE;
 876
 877         init_info.pd = &iwpd->sc_pd;
 878         init_info.qp_uk_init_info.qp_id = iwqp->ibqp.qp_num;
 879         if (!rdma_protocol_roce(&iwdev->ibdev, 1))
 880                 init_info.qp_uk_init_info.first_sq_wq = 1;
 881         iwqp->ctx_info.qp_compl_ctx = (uintptr_t)qp;
 882         init_waitqueue_head(&iwqp->waitq);
 883         init_waitqueue_head(&iwqp->mod_qp_waitq);
 884
 885         if (udata) {
 886                 err_code = ib_copy_from_udata(&req, udata,
 887                                               min(sizeof(req), udata->inlen));
 888                 if (err_code) {
 889                         ibdev_dbg(&iwdev->ibdev,
 890                                   "VERBS: ib_copy_from_data fail\n");
 891                         goto error;
 892                 }
 893
 894                 iwqp->ctx_info.qp_compl_ctx = req.user_compl_ctx;
 895                 iwqp->user_mode = 1;
 896                 if (req.user_wqe_bufs) {
 897                         struct irdma_ucontext *ucontext =
 898                                 rdma_udata_to_drv_context(udata,
 899                                                           struct irdma_ucontext,
 900                                                           ibucontext);
 901
 902                         init_info.qp_uk_init_info.legacy_mode = ucontext->legacy_mode;
 903                         spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
 904                         iwqp->iwpbl = irdma_get_pbl((unsigned long)req.user_wqe_bufs,
 905                                                     &ucontext->qp_reg_mem_list);
 906                         spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
 907
 908                         if (!iwqp->iwpbl) {
 909                                 err_code = -ENODATA;
 910                                 ibdev_dbg(&iwdev->ibdev, "VERBS: no pbl info\n");
 911                                 goto error;
 912                         }
 913                 }
 914                 init_info.qp_uk_init_info.abi_ver = iwpd->sc_pd.abi_ver;
 915                 irdma_setup_virt_qp(iwdev, iwqp, &init_info);
 916         } else {
 917                 init_info.qp_uk_init_info.abi_ver = IRDMA_ABI_VER;
 918                 err_code = irdma_setup_kmode_qp(iwdev, iwqp, &init_info, init_attr);
 919         }
 920
 921         if (err_code) {
 922                 ibdev_dbg(&iwdev->ibdev, "VERBS: setup qp failed\n");
 923                 goto error;
 924         }
 925
 926         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 927                 if (init_attr->qp_type == IB_QPT_RC) {
 928                         init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_ROCE_RC;
 929                         init_info.qp_uk_init_info.qp_caps = IRDMA_SEND_WITH_IMM |
 930                                                             IRDMA_WRITE_WITH_IMM |
 931                                                             IRDMA_ROCE;
 932                 } else {
 933                         init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_ROCE_UD;
 934                         init_info.qp_uk_init_info.qp_caps = IRDMA_SEND_WITH_IMM |
 935                                                             IRDMA_ROCE;
 936                 }
 937         } else {
 938                 init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_IWARP;
 939                 init_info.qp_uk_init_info.qp_caps = IRDMA_WRITE_WITH_IMM;
 940         }
 941
 942         if (dev->hw_attrs.uk_attrs.hw_rev > IRDMA_GEN_1)
 943                 init_info.qp_uk_init_info.qp_caps |= IRDMA_PUSH_MODE;
 944
 945         ret = irdma_sc_qp_init(qp, &init_info);
 946         if (ret) {
 947                 err_code = -EPROTO;
 948                 ibdev_dbg(&iwdev->ibdev, "VERBS: qp_init fail\n");
 949                 goto error;
 950         }
 951
 952         ctx_info = &iwqp->ctx_info;
 953         ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
 954         ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
 955
 956         if (rdma_protocol_roce(&iwdev->ibdev, 1))
 957                 irdma_roce_fill_and_set_qpctx_info(iwqp, ctx_info);
 958         else
 959                 irdma_iw_fill_and_set_qpctx_info(iwqp, ctx_info);
 960
 961         err_code = irdma_cqp_create_qp_cmd(iwqp);
 962         if (err_code)
 963                 goto error;
 964
 965         refcount_set(&iwqp->refcnt, 1);
 966         spin_lock_init(&iwqp->lock);
 967         spin_lock_init(&iwqp->sc_qp.pfpdu.lock);
 968         iwqp->sig_all = (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ? 1 : 0;
 969         rf->qp_table[qp_num] = iwqp;
 970         iwqp->max_send_wr = sq_size;
 971         iwqp->max_recv_wr = rq_size;
 972
 973         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 974                 if (dev->ws_add(&iwdev->vsi, 0)) {
 975                         irdma_cqp_qp_destroy_cmd(&rf->sc_dev, &iwqp->sc_qp);
 976                         err_code = -EINVAL;
 977                         goto error;
 978                 }
 979
 980                 irdma_qp_add_qos(&iwqp->sc_qp);
 981         }
 982
 983         if (udata) {
 984                 /* GEN_1 legacy support with libi40iw does not have expanded uresp struct */
 985                 if (udata->outlen < sizeof(uresp)) {
 986                         uresp.lsmm = 1;
 987                         uresp.push_idx = IRDMA_INVALID_PUSH_PAGE_INDEX_GEN_1;
 988                 } else {
 989                         if (rdma_protocol_iwarp(&iwdev->ibdev, 1))
 990                                 uresp.lsmm = 1;
 991                 }
 992                 uresp.actual_sq_size = sq_size;
 993                 uresp.actual_rq_size = rq_size;
 994                 uresp.qp_id = qp_num;
 995                 uresp.qp_caps = qp->qp_uk.qp_caps;
 996
 997                 err_code = ib_copy_to_udata(udata, &uresp,
 998                                             min(sizeof(uresp), udata->outlen));
 999                 if (err_code) {
1000                         ibdev_dbg(&iwdev->ibdev, "VERBS: copy_to_udata failed\n");
1001                         irdma_destroy_qp(&iwqp->ibqp, udata);
1002                         return ERR_PTR(err_code);
1003                 }
1004         }
1005
1006         init_completion(&iwqp->free_qp);
1007         return &iwqp->ibqp;
1008
1009 error:
1010         irdma_free_qp_rsrc(iwqp);
1011
1012         return ERR_PTR(err_code);
1013 }
1014
1015 static int irdma_get_ib_acc_flags(struct irdma_qp *iwqp)
1016 {
1017         int acc_flags = 0;
1018
1019         if (rdma_protocol_roce(iwqp->ibqp.device, 1)) {
1020                 if (iwqp->roce_info.wr_rdresp_en) {
1021                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
1022                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
1023                 }
1024                 if (iwqp->roce_info.rd_en)
1025                         acc_flags |= IB_ACCESS_REMOTE_READ;
1026                 if (iwqp->roce_info.bind_en)
1027                         acc_flags |= IB_ACCESS_MW_BIND;
1028         } else {
1029                 if (iwqp->iwarp_info.wr_rdresp_en) {
1030                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
1031                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
1032                 }
1033                 if (iwqp->iwarp_info.rd_en)
1034                         acc_flags |= IB_ACCESS_REMOTE_READ;
1035                 if (iwqp->iwarp_info.bind_en)
1036                         acc_flags |= IB_ACCESS_MW_BIND;
1037         }
1038         return acc_flags;
1039 }
1040
1041 /**
1042  * irdma_query_qp - query qp attributes
1043  * @ibqp: qp pointer
1044  * @attr: attributes pointer
1045  * @attr_mask: Not used
1046  * @init_attr: qp attributes to return
1047  */
1048 static int irdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1049                           int attr_mask, struct ib_qp_init_attr *init_attr)
1050 {
1051         struct irdma_qp *iwqp = to_iwqp(ibqp);
1052         struct irdma_sc_qp *qp = &iwqp->sc_qp;
1053
1054         memset(attr, 0, sizeof(*attr));
1055         memset(init_attr, 0, sizeof(*init_attr));
1056
1057         attr->qp_state = iwqp->ibqp_state;
1058         attr->cur_qp_state = iwqp->ibqp_state;
1059         attr->cap.max_send_wr = iwqp->max_send_wr;
1060         attr->cap.max_recv_wr = iwqp->max_recv_wr;
1061         attr->cap.max_inline_data = qp->qp_uk.max_inline_data;
1062         attr->cap.max_send_sge = qp->qp_uk.max_sq_frag_cnt;
1063         attr->cap.max_recv_sge = qp->qp_uk.max_rq_frag_cnt;
1064         attr->qp_access_flags = irdma_get_ib_acc_flags(iwqp);
1065         attr->port_num = 1;
1066         if (rdma_protocol_roce(ibqp->device, 1)) {
1067                 attr->path_mtu = ib_mtu_int_to_enum(iwqp->udp_info.snd_mss);
1068                 attr->qkey = iwqp->roce_info.qkey;
1069                 attr->rq_psn = iwqp->udp_info.epsn;
1070                 attr->sq_psn = iwqp->udp_info.psn_nxt;
1071                 attr->dest_qp_num = iwqp->roce_info.dest_qp;
1072                 attr->pkey_index = iwqp->roce_info.p_key;
1073                 attr->retry_cnt = iwqp->udp_info.rexmit_thresh;
1074                 attr->rnr_retry = iwqp->udp_info.rnr_nak_thresh;
1075                 attr->max_rd_atomic = iwqp->roce_info.ord_size;
1076                 attr->max_dest_rd_atomic = iwqp->roce_info.ird_size;
1077         }
1078
1079         init_attr->event_handler = iwqp->ibqp.event_handler;
1080         init_attr->qp_context = iwqp->ibqp.qp_context;
1081         init_attr->send_cq = iwqp->ibqp.send_cq;
1082         init_attr->recv_cq = iwqp->ibqp.recv_cq;
1083         init_attr->cap = attr->cap;
1084
1085         return 0;
1086 }
1087
1088 /**
1089  * irdma_query_pkey - Query partition key
1090  * @ibdev: device pointer from stack
1091  * @port: port number
1092  * @index: index of pkey
1093  * @pkey: pointer to store the pkey
1094  */
1095 static int irdma_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1096                             u16 *pkey)
1097 {
1098         if (index >= IRDMA_PKEY_TBL_SZ)
1099                 return -EINVAL;
1100
1101         *pkey = IRDMA_DEFAULT_PKEY;
1102         return 0;
1103 }
1104
1105 /**
1106  * irdma_modify_qp_roce - modify qp request
1107  * @ibqp: qp's pointer for modify
1108  * @attr: access attributes
1109  * @attr_mask: state mask
1110  * @udata: user data
1111  */
1112 int irdma_modify_qp_roce(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1113                          int attr_mask, struct ib_udata *udata)
1114 {
1115         struct irdma_pd *iwpd = to_iwpd(ibqp->pd);
1116         struct irdma_qp *iwqp = to_iwqp(ibqp);
1117         struct irdma_device *iwdev = iwqp->iwdev;
1118         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
1119         struct irdma_qp_host_ctx_info *ctx_info;
1120         struct irdma_roce_offload_info *roce_info;
1121         struct irdma_udp_offload_info *udp_info;
1122         struct irdma_modify_qp_info info = {};
1123         struct irdma_modify_qp_resp uresp = {};
1124         struct irdma_modify_qp_req ureq = {};
1125         unsigned long flags;
1126         u8 issue_modify_qp = 0;
1127         int ret = 0;
1128
1129         ctx_info = &iwqp->ctx_info;
1130         roce_info = &iwqp->roce_info;
1131         udp_info = &iwqp->udp_info;
1132
1133         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1134                 return -EOPNOTSUPP;
1135
1136         if (attr_mask & IB_QP_DEST_QPN)
1137                 roce_info->dest_qp = attr->dest_qp_num;
1138
1139         if (attr_mask & IB_QP_PKEY_INDEX) {
1140                 ret = irdma_query_pkey(ibqp->device, 0, attr->pkey_index,
1141                                        &roce_info->p_key);
1142                 if (ret)
1143                         return ret;
1144         }
1145
1146         if (attr_mask & IB_QP_QKEY)
1147                 roce_info->qkey = attr->qkey;
1148
1149         if (attr_mask & IB_QP_PATH_MTU)
1150                 udp_info->snd_mss = ib_mtu_enum_to_int(attr->path_mtu);
1151
1152         if (attr_mask & IB_QP_SQ_PSN) {
1153                 udp_info->psn_nxt = attr->sq_psn;
1154                 udp_info->lsn =  0xffff;
1155                 udp_info->psn_una = attr->sq_psn;
1156                 udp_info->psn_max = attr->sq_psn;
1157         }
1158
1159         if (attr_mask & IB_QP_RQ_PSN)
1160                 udp_info->epsn = attr->rq_psn;
1161
1162         if (attr_mask & IB_QP_RNR_RETRY)
1163                 udp_info->rnr_nak_thresh = attr->rnr_retry;
1164
1165         if (attr_mask & IB_QP_RETRY_CNT)
1166                 udp_info->rexmit_thresh = attr->retry_cnt;
1167
1168         ctx_info->roce_info->pd_id = iwpd->sc_pd.pd_id;
1169
1170         if (attr_mask & IB_QP_AV) {
1171                 struct irdma_av *av = &iwqp->roce_ah.av;
1172                 const struct ib_gid_attr *sgid_attr;
1173                 u16 vlan_id = VLAN_N_VID;
1174                 u32 local_ip[4];
1175
1176                 memset(&iwqp->roce_ah, 0, sizeof(iwqp->roce_ah));
1177                 if (attr->ah_attr.ah_flags & IB_AH_GRH) {
1178                         udp_info->ttl = attr->ah_attr.grh.hop_limit;
1179                         udp_info->flow_label = attr->ah_attr.grh.flow_label;
1180                         udp_info->tos = attr->ah_attr.grh.traffic_class;
1181                         irdma_qp_rem_qos(&iwqp->sc_qp);
1182                         dev->ws_remove(iwqp->sc_qp.vsi, ctx_info->user_pri);
1183                         ctx_info->user_pri = rt_tos2priority(udp_info->tos);
1184                         iwqp->sc_qp.user_pri = ctx_info->user_pri;
1185                         if (dev->ws_add(iwqp->sc_qp.vsi, ctx_info->user_pri))
1186                                 return -ENOMEM;
1187                         irdma_qp_add_qos(&iwqp->sc_qp);
1188                 }
1189                 sgid_attr = attr->ah_attr.grh.sgid_attr;
1190                 ret = rdma_read_gid_l2_fields(sgid_attr, &vlan_id,
1191                                               ctx_info->roce_info->mac_addr);
1192                 if (ret)
1193                         return ret;
1194
1195                 if (vlan_id >= VLAN_N_VID && iwdev->dcb)
1196                         vlan_id = 0;
1197                 if (vlan_id < VLAN_N_VID) {
1198                         udp_info->insert_vlan_tag = true;
1199                         udp_info->vlan_tag = vlan_id |
1200                                 ctx_info->user_pri << VLAN_PRIO_SHIFT;
1201                 } else {
1202                         udp_info->insert_vlan_tag = false;
1203                 }
1204
1205                 av->attrs = attr->ah_attr;
1206                 rdma_gid2ip((struct sockaddr *)&av->sgid_addr, &sgid_attr->gid);
1207                 rdma_gid2ip((struct sockaddr *)&av->dgid_addr, &attr->ah_attr.grh.dgid);
1208                 roce_info->local_qp = ibqp->qp_num;
1209                 if (av->sgid_addr.saddr.sa_family == AF_INET6) {
1210                         __be32 *daddr =
1211                                 av->dgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32;
1212                         __be32 *saddr =
1213                                 av->sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32;
1214
1215                         irdma_copy_ip_ntohl(&udp_info->dest_ip_addr[0], daddr);
1216                         irdma_copy_ip_ntohl(&udp_info->local_ipaddr[0], saddr);
1217
1218                         udp_info->ipv4 = false;
1219                         irdma_copy_ip_ntohl(local_ip, daddr);
1220
1221                         udp_info->arp_idx = irdma_arp_table(iwdev->rf,
1222                                                             &local_ip[0],
1223                                                             false, NULL,
1224                                                             IRDMA_ARP_RESOLVE);
1225                 } else {
1226                         __be32 saddr = av->sgid_addr.saddr_in.sin_addr.s_addr;
1227                         __be32 daddr = av->dgid_addr.saddr_in.sin_addr.s_addr;
1228
1229                         local_ip[0] = ntohl(daddr);
1230
1231                         udp_info->ipv4 = true;
1232                         udp_info->dest_ip_addr[0] = 0;
1233                         udp_info->dest_ip_addr[1] = 0;
1234                         udp_info->dest_ip_addr[2] = 0;
1235                         udp_info->dest_ip_addr[3] = local_ip[0];
1236
1237                         udp_info->local_ipaddr[0] = 0;
1238                         udp_info->local_ipaddr[1] = 0;
1239                         udp_info->local_ipaddr[2] = 0;
1240                         udp_info->local_ipaddr[3] = ntohl(saddr);
1241                 }
1242                 udp_info->arp_idx =
1243                         irdma_add_arp(iwdev->rf, local_ip, udp_info->ipv4,
1244                                       attr->ah_attr.roce.dmac);
1245         }
1246
1247         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1248                 if (attr->max_rd_atomic > dev->hw_attrs.max_hw_ord) {
1249                         ibdev_err(&iwdev->ibdev,
1250                                   "rd_atomic = %d, above max_hw_ord=%d\n",
1251                                   attr->max_rd_atomic,
1252                                   dev->hw_attrs.max_hw_ord);
1253                         return -EINVAL;
1254                 }
1255                 if (attr->max_rd_atomic)
1256                         roce_info->ord_size = attr->max_rd_atomic;
1257                 info.ord_valid = true;
1258         }
1259
1260         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1261                 if (attr->max_dest_rd_atomic > dev->hw_attrs.max_hw_ird) {
1262                         ibdev_err(&iwdev->ibdev,
1263                                   "rd_atomic = %d, above max_hw_ird=%d\n",
1264                                    attr->max_rd_atomic,
1265                                    dev->hw_attrs.max_hw_ird);
1266                         return -EINVAL;
1267                 }
1268                 if (attr->max_dest_rd_atomic)
1269                         roce_info->ird_size = attr->max_dest_rd_atomic;
1270         }
1271
1272         if (attr_mask & IB_QP_ACCESS_FLAGS) {
1273                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
1274                         roce_info->wr_rdresp_en = true;
1275                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
1276                         roce_info->wr_rdresp_en = true;
1277                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
1278                         roce_info->rd_en = true;
1279         }
1280
1281         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
1282
1283         ibdev_dbg(&iwdev->ibdev,
1284                   "VERBS: caller: %pS qp_id=%d to_ibqpstate=%d ibqpstate=%d irdma_qpstate=%d attr_mask=0x%x\n",
1285                   __builtin_return_address(0), ibqp->qp_num, attr->qp_state,
1286                   iwqp->ibqp_state, iwqp->iwarp_state, attr_mask);
1287
1288         spin_lock_irqsave(&iwqp->lock, flags);
1289         if (attr_mask & IB_QP_STATE) {
1290                 if (!ib_modify_qp_is_ok(iwqp->ibqp_state, attr->qp_state,
1291                                         iwqp->ibqp.qp_type, attr_mask)) {
1292                         ibdev_warn(&iwdev->ibdev, "modify_qp invalid for qp_id=%d, old_state=0x%x, new_state=0x%x\n",
1293                                    iwqp->ibqp.qp_num, iwqp->ibqp_state,
1294                                    attr->qp_state);
1295                         ret = -EINVAL;
1296                         goto exit;
1297                 }
1298                 info.curr_iwarp_state = iwqp->iwarp_state;
1299
1300                 switch (attr->qp_state) {
1301                 case IB_QPS_INIT:
1302                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1303                                 ret = -EINVAL;
1304                                 goto exit;
1305                         }
1306
1307                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
1308                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
1309                                 issue_modify_qp = 1;
1310                         }
1311                         break;
1312                 case IB_QPS_RTR:
1313                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1314                                 ret = -EINVAL;
1315                                 goto exit;
1316                         }
1317                         info.arp_cache_idx_valid = true;
1318                         info.cq_num_valid = true;
1319                         info.next_iwarp_state = IRDMA_QP_STATE_RTR;
1320                         issue_modify_qp = 1;
1321                         break;
1322                 case IB_QPS_RTS:
1323                         if (iwqp->ibqp_state < IB_QPS_RTR ||
1324                             iwqp->ibqp_state == IB_QPS_ERR) {
1325                                 ret = -EINVAL;
1326                                 goto exit;
1327                         }
1328
1329                         info.arp_cache_idx_valid = true;
1330                         info.cq_num_valid = true;
1331                         info.ord_valid = true;
1332                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
1333                         issue_modify_qp = 1;
1334                         if (iwdev->push_mode && udata &&
1335                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1336                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1337                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1338                                 irdma_alloc_push_page(iwqp);
1339                                 spin_lock_irqsave(&iwqp->lock, flags);
1340                         }
1341                         break;
1342                 case IB_QPS_SQD:
1343                         if (iwqp->iwarp_state == IRDMA_QP_STATE_SQD)
1344                                 goto exit;
1345
1346                         if (iwqp->iwarp_state != IRDMA_QP_STATE_RTS) {
1347                                 ret = -EINVAL;
1348                                 goto exit;
1349                         }
1350
1351                         info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1352                         issue_modify_qp = 1;
1353                         break;
1354                 case IB_QPS_SQE:
1355                 case IB_QPS_ERR:
1356                 case IB_QPS_RESET:
1357                         if (iwqp->iwarp_state == IRDMA_QP_STATE_RTS) {
1358                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1359                                 info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1360                                 irdma_hw_modify_qp(iwdev, iwqp, &info, true);
1361                                 spin_lock_irqsave(&iwqp->lock, flags);
1362                         }
1363
1364                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1365                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1366                                 if (udata) {
1367                                         if (ib_copy_from_udata(&ureq, udata,
1368                                             min(sizeof(ureq), udata->inlen)))
1369                                                 return -EINVAL;
1370
1371                                         irdma_flush_wqes(iwqp,
1372                                             (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1373                                             (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1374                                             IRDMA_REFLUSH);
1375                                 }
1376                                 return 0;
1377                         }
1378
1379                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1380                         issue_modify_qp = 1;
1381                         break;
1382                 default:
1383                         ret = -EINVAL;
1384                         goto exit;
1385                 }
1386
1387                 iwqp->ibqp_state = attr->qp_state;
1388         }
1389
1390         ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1391         ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1392         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1393         spin_unlock_irqrestore(&iwqp->lock, flags);
1394
1395         if (attr_mask & IB_QP_STATE) {
1396                 if (issue_modify_qp) {
1397                         ctx_info->rem_endpoint_idx = udp_info->arp_idx;
1398                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1399                                 return -EINVAL;
1400                         spin_lock_irqsave(&iwqp->lock, flags);
1401                         if (iwqp->iwarp_state == info.curr_iwarp_state) {
1402                                 iwqp->iwarp_state = info.next_iwarp_state;
1403                                 iwqp->ibqp_state = attr->qp_state;
1404                         }
1405                         if (iwqp->ibqp_state > IB_QPS_RTS &&
1406                             !iwqp->flush_issued) {
1407                                 iwqp->flush_issued = 1;
1408                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1409                                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_SQ |
1410                                                        IRDMA_FLUSH_RQ |
1411                                                        IRDMA_FLUSH_WAIT);
1412                         } else {
1413                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1414                         }
1415                 } else {
1416                         iwqp->ibqp_state = attr->qp_state;
1417                 }
1418                 if (udata && dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1419                         struct irdma_ucontext *ucontext;
1420
1421                         ucontext = rdma_udata_to_drv_context(udata,
1422                                         struct irdma_ucontext, ibucontext);
1423                         if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1424                             !iwqp->push_wqe_mmap_entry &&
1425                             !irdma_setup_push_mmap_entries(ucontext, iwqp,
1426                                 &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1427                                 uresp.push_valid = 1;
1428                                 uresp.push_offset = iwqp->sc_qp.push_offset;
1429                         }
1430                         ret = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1431                                                udata->outlen));
1432                         if (ret) {
1433                                 irdma_remove_push_mmap_entries(iwqp);
1434                                 ibdev_dbg(&iwdev->ibdev,
1435                                           "VERBS: copy_to_udata failed\n");
1436                                 return ret;
1437                         }
1438                 }
1439         }
1440
1441         return 0;
1442 exit:
1443         spin_unlock_irqrestore(&iwqp->lock, flags);
1444
1445         return ret;
1446 }
1447
1448 /**
1449  * irdma_modify_qp - modify qp request
1450  * @ibqp: qp's pointer for modify
1451  * @attr: access attributes
1452  * @attr_mask: state mask
1453  * @udata: user data
1454  */
1455 int irdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
1456                     struct ib_udata *udata)
1457 {
1458         struct irdma_qp *iwqp = to_iwqp(ibqp);
1459         struct irdma_device *iwdev = iwqp->iwdev;
1460         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
1461         struct irdma_qp_host_ctx_info *ctx_info;
1462         struct irdma_tcp_offload_info *tcp_info;
1463         struct irdma_iwarp_offload_info *offload_info;
1464         struct irdma_modify_qp_info info = {};
1465         struct irdma_modify_qp_resp uresp = {};
1466         struct irdma_modify_qp_req ureq = {};
1467         u8 issue_modify_qp = 0;
1468         u8 dont_wait = 0;
1469         int err;
1470         unsigned long flags;
1471
1472         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1473                 return -EOPNOTSUPP;
1474
1475         ctx_info = &iwqp->ctx_info;
1476         offload_info = &iwqp->iwarp_info;
1477         tcp_info = &iwqp->tcp_info;
1478         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
1479         ibdev_dbg(&iwdev->ibdev,
1480                   "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",
1481                   __builtin_return_address(0), ibqp->qp_num, attr->qp_state,
1482                   iwqp->ibqp_state, iwqp->iwarp_state, iwqp->last_aeq,
1483                   iwqp->hw_tcp_state, iwqp->hw_iwarp_state, attr_mask);
1484
1485         spin_lock_irqsave(&iwqp->lock, flags);
1486         if (attr_mask & IB_QP_STATE) {
1487                 info.curr_iwarp_state = iwqp->iwarp_state;
1488                 switch (attr->qp_state) {
1489                 case IB_QPS_INIT:
1490                 case IB_QPS_RTR:
1491                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1492                                 err = -EINVAL;
1493                                 goto exit;
1494                         }
1495
1496                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
1497                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
1498                                 issue_modify_qp = 1;
1499                         }
1500                         if (iwdev->push_mode && udata &&
1501                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1502                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1503                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1504                                 irdma_alloc_push_page(iwqp);
1505                                 spin_lock_irqsave(&iwqp->lock, flags);
1506                         }
1507                         break;
1508                 case IB_QPS_RTS:
1509                         if (iwqp->iwarp_state > IRDMA_QP_STATE_RTS ||
1510                             !iwqp->cm_id) {
1511                                 err = -EINVAL;
1512                                 goto exit;
1513                         }
1514
1515                         issue_modify_qp = 1;
1516                         iwqp->hw_tcp_state = IRDMA_TCP_STATE_ESTABLISHED;
1517                         iwqp->hte_added = 1;
1518                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
1519                         info.tcp_ctx_valid = true;
1520                         info.ord_valid = true;
1521                         info.arp_cache_idx_valid = true;
1522                         info.cq_num_valid = true;
1523                         break;
1524                 case IB_QPS_SQD:
1525                         if (iwqp->hw_iwarp_state > IRDMA_QP_STATE_RTS) {
1526                                 err = 0;
1527                                 goto exit;
1528                         }
1529
1530                         if (iwqp->iwarp_state == IRDMA_QP_STATE_CLOSING ||
1531                             iwqp->iwarp_state < IRDMA_QP_STATE_RTS) {
1532                                 err = 0;
1533                                 goto exit;
1534                         }
1535
1536                         if (iwqp->iwarp_state > IRDMA_QP_STATE_CLOSING) {
1537                                 err = -EINVAL;
1538                                 goto exit;
1539                         }
1540
1541                         info.next_iwarp_state = IRDMA_QP_STATE_CLOSING;
1542                         issue_modify_qp = 1;
1543                         break;
1544                 case IB_QPS_SQE:
1545                         if (iwqp->iwarp_state >= IRDMA_QP_STATE_TERMINATE) {
1546                                 err = -EINVAL;
1547                                 goto exit;
1548                         }
1549
1550                         info.next_iwarp_state = IRDMA_QP_STATE_TERMINATE;
1551                         issue_modify_qp = 1;
1552                         break;
1553                 case IB_QPS_ERR:
1554                 case IB_QPS_RESET:
1555                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1556                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1557                                 if (udata) {
1558                                         if (ib_copy_from_udata(&ureq, udata,
1559                                             min(sizeof(ureq), udata->inlen)))
1560                                                 return -EINVAL;
1561
1562                                         irdma_flush_wqes(iwqp,
1563                                             (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1564                                             (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1565                                             IRDMA_REFLUSH);
1566                                 }
1567                                 return 0;
1568                         }
1569
1570                         if (iwqp->sc_qp.term_flags) {
1571                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1572                                 irdma_terminate_del_timer(&iwqp->sc_qp);
1573                                 spin_lock_irqsave(&iwqp->lock, flags);
1574                         }
1575                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1576                         if (iwqp->hw_tcp_state > IRDMA_TCP_STATE_CLOSED &&
1577                             iwdev->iw_status &&
1578                             iwqp->hw_tcp_state != IRDMA_TCP_STATE_TIME_WAIT)
1579                                 info.reset_tcp_conn = true;
1580                         else
1581                                 dont_wait = 1;
1582
1583                         issue_modify_qp = 1;
1584                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1585                         break;
1586                 default:
1587                         err = -EINVAL;
1588                         goto exit;
1589                 }
1590
1591                 iwqp->ibqp_state = attr->qp_state;
1592         }
1593         if (attr_mask & IB_QP_ACCESS_FLAGS) {
1594                 ctx_info->iwarp_info_valid = true;
1595                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
1596                         offload_info->wr_rdresp_en = true;
1597                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
1598                         offload_info->wr_rdresp_en = true;
1599                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
1600                         offload_info->rd_en = true;
1601         }
1602
1603         if (ctx_info->iwarp_info_valid) {
1604                 ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1605                 ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1606                 irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1607         }
1608         spin_unlock_irqrestore(&iwqp->lock, flags);
1609
1610         if (attr_mask & IB_QP_STATE) {
1611                 if (issue_modify_qp) {
1612                         ctx_info->rem_endpoint_idx = tcp_info->arp_idx;
1613                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1614                                 return -EINVAL;
1615                 }
1616
1617                 spin_lock_irqsave(&iwqp->lock, flags);
1618                 if (iwqp->iwarp_state == info.curr_iwarp_state) {
1619                         iwqp->iwarp_state = info.next_iwarp_state;
1620                         iwqp->ibqp_state = attr->qp_state;
1621                 }
1622                 spin_unlock_irqrestore(&iwqp->lock, flags);
1623         }
1624
1625         if (issue_modify_qp && iwqp->ibqp_state > IB_QPS_RTS) {
1626                 if (dont_wait) {
1627                         if (iwqp->cm_id && iwqp->hw_tcp_state) {
1628                                 spin_lock_irqsave(&iwqp->lock, flags);
1629                                 iwqp->hw_tcp_state = IRDMA_TCP_STATE_CLOSED;
1630                                 iwqp->last_aeq = IRDMA_AE_RESET_SENT;
1631                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1632                                 irdma_cm_disconn(iwqp);
1633                         }
1634                 } else {
1635                         int close_timer_started;
1636
1637                         spin_lock_irqsave(&iwdev->cm_core.ht_lock, flags);
1638
1639                         if (iwqp->cm_node) {
1640                                 refcount_inc(&iwqp->cm_node->refcnt);
1641                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1642                                 close_timer_started = atomic_inc_return(&iwqp->close_timer_started);
1643                                 if (iwqp->cm_id && close_timer_started == 1)
1644                                         irdma_schedule_cm_timer(iwqp->cm_node,
1645                                                 (struct irdma_puda_buf *)iwqp,
1646                                                 IRDMA_TIMER_TYPE_CLOSE, 1, 0);
1647
1648                                 irdma_rem_ref_cm_node(iwqp->cm_node);
1649                         } else {
1650                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1651                         }
1652                 }
1653         }
1654         if (attr_mask & IB_QP_STATE && udata &&
1655             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1656                 struct irdma_ucontext *ucontext;
1657
1658                 ucontext = rdma_udata_to_drv_context(udata,
1659                                         struct irdma_ucontext, ibucontext);
1660                 if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1661                     !iwqp->push_wqe_mmap_entry &&
1662                     !irdma_setup_push_mmap_entries(ucontext, iwqp,
1663                         &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1664                         uresp.push_valid = 1;
1665                         uresp.push_offset = iwqp->sc_qp.push_offset;
1666                 }
1667
1668                 err = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1669                                        udata->outlen));
1670                 if (err) {
1671                         irdma_remove_push_mmap_entries(iwqp);
1672                         ibdev_dbg(&iwdev->ibdev,
1673                                   "VERBS: copy_to_udata failed\n");
1674                         return err;
1675                 }
1676         }
1677
1678         return 0;
1679 exit:
1680         spin_unlock_irqrestore(&iwqp->lock, flags);
1681
1682         return err;
1683 }
1684
1685 /**
1686  * irdma_cq_free_rsrc - free up resources for cq
1687  * @rf: RDMA PCI function
1688  * @iwcq: cq ptr
1689  */
1690 static void irdma_cq_free_rsrc(struct irdma_pci_f *rf, struct irdma_cq *iwcq)
1691 {
1692         struct irdma_sc_cq *cq = &iwcq->sc_cq;
1693
1694         if (!iwcq->user_mode) {
1695                 dma_free_coherent(rf->sc_dev.hw->device, iwcq->kmem.size,
1696                                   iwcq->kmem.va, iwcq->kmem.pa);
1697                 iwcq->kmem.va = NULL;
1698                 dma_free_coherent(rf->sc_dev.hw->device,
1699                                   iwcq->kmem_shadow.size,
1700                                   iwcq->kmem_shadow.va, iwcq->kmem_shadow.pa);
1701                 iwcq->kmem_shadow.va = NULL;
1702         }
1703
1704         irdma_free_rsrc(rf, rf->allocated_cqs, cq->cq_uk.cq_id);
1705 }
1706
1707 /**
1708  * irdma_free_cqbuf - worker to free a cq buffer
1709  * @work: provides access to the cq buffer to free
1710  */
1711 static void irdma_free_cqbuf(struct work_struct *work)
1712 {
1713         struct irdma_cq_buf *cq_buf = container_of(work, struct irdma_cq_buf, work);
1714
1715         dma_free_coherent(cq_buf->hw->device, cq_buf->kmem_buf.size,
1716                           cq_buf->kmem_buf.va, cq_buf->kmem_buf.pa);
1717         cq_buf->kmem_buf.va = NULL;
1718         kfree(cq_buf);
1719 }
1720
1721 /**
1722  * irdma_process_resize_list - remove resized cq buffers from the resize_list
1723  * @iwcq: cq which owns the resize_list
1724  * @iwdev: irdma device
1725  * @lcqe_buf: the buffer where the last cqe is received
1726  */
1727 static int irdma_process_resize_list(struct irdma_cq *iwcq,
1728                                      struct irdma_device *iwdev,
1729                                      struct irdma_cq_buf *lcqe_buf)
1730 {
1731         struct list_head *tmp_node, *list_node;
1732         struct irdma_cq_buf *cq_buf;
1733         int cnt = 0;
1734
1735         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
1736                 cq_buf = list_entry(list_node, struct irdma_cq_buf, list);
1737                 if (cq_buf == lcqe_buf)
1738                         return cnt;
1739
1740                 list_del(&cq_buf->list);
1741                 queue_work(iwdev->cleanup_wq, &cq_buf->work);
1742                 cnt++;
1743         }
1744
1745         return cnt;
1746 }
1747
1748 /**
1749  * irdma_destroy_cq - destroy cq
1750  * @ib_cq: cq pointer
1751  * @udata: user data
1752  */
1753 static int irdma_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata)
1754 {
1755         struct irdma_device *iwdev = to_iwdev(ib_cq->device);
1756         struct irdma_cq *iwcq = to_iwcq(ib_cq);
1757         struct irdma_sc_cq *cq = &iwcq->sc_cq;
1758         struct irdma_sc_dev *dev = cq->dev;
1759         struct irdma_sc_ceq *ceq = dev->ceq[cq->ceq_id];
1760         struct irdma_ceq *iwceq = container_of(ceq, struct irdma_ceq, sc_ceq);
1761         unsigned long flags;
1762
1763         spin_lock_irqsave(&iwcq->lock, flags);
1764         if (!list_empty(&iwcq->resize_list))
1765                 irdma_process_resize_list(iwcq, iwdev, NULL);
1766         spin_unlock_irqrestore(&iwcq->lock, flags);
1767
1768         irdma_cq_wq_destroy(iwdev->rf, cq);
1769         irdma_cq_free_rsrc(iwdev->rf, iwcq);
1770
1771         spin_lock_irqsave(&iwceq->ce_lock, flags);
1772         irdma_sc_cleanup_ceqes(cq, ceq);
1773         spin_unlock_irqrestore(&iwceq->ce_lock, flags);
1774
1775         return 0;
1776 }
1777
1778 /**
1779  * irdma_resize_cq - resize cq
1780  * @ibcq: cq to be resized
1781  * @entries: desired cq size
1782  * @udata: user data
1783  */
1784 static int irdma_resize_cq(struct ib_cq *ibcq, int entries,
1785                            struct ib_udata *udata)
1786 {
1787         struct irdma_cq *iwcq = to_iwcq(ibcq);
1788         struct irdma_sc_dev *dev = iwcq->sc_cq.dev;
1789         struct irdma_cqp_request *cqp_request;
1790         struct cqp_cmds_info *cqp_info;
1791         struct irdma_modify_cq_info *m_info;
1792         struct irdma_modify_cq_info info = {};
1793         struct irdma_dma_mem kmem_buf;
1794         struct irdma_cq_mr *cqmr_buf;
1795         struct irdma_pbl *iwpbl_buf;
1796         struct irdma_device *iwdev;
1797         struct irdma_pci_f *rf;
1798         struct irdma_cq_buf *cq_buf = NULL;
1799         enum irdma_status_code status = 0;
1800         unsigned long flags;
1801         int ret;
1802
1803         iwdev = to_iwdev(ibcq->device);
1804         rf = iwdev->rf;
1805
1806         if (!(rf->sc_dev.hw_attrs.uk_attrs.feature_flags &
1807             IRDMA_FEATURE_CQ_RESIZE))
1808                 return -EOPNOTSUPP;
1809
1810         if (entries > rf->max_cqe)
1811                 return -EINVAL;
1812
1813         if (!iwcq->user_mode) {
1814                 entries++;
1815                 if (rf->sc_dev.hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
1816                         entries *= 2;
1817         }
1818
1819         info.cq_size = max(entries, 4);
1820
1821         if (info.cq_size == iwcq->sc_cq.cq_uk.cq_size - 1)
1822                 return 0;
1823
1824         if (udata) {
1825                 struct irdma_resize_cq_req req = {};
1826                 struct irdma_ucontext *ucontext =
1827                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
1828                                                   ibucontext);
1829
1830                 /* CQ resize not supported with legacy GEN_1 libi40iw */
1831                 if (ucontext->legacy_mode)
1832                         return -EOPNOTSUPP;
1833
1834                 if (ib_copy_from_udata(&req, udata,
1835                                        min(sizeof(req), udata->inlen)))
1836                         return -EINVAL;
1837
1838                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
1839                 iwpbl_buf = irdma_get_pbl((unsigned long)req.user_cq_buffer,
1840                                           &ucontext->cq_reg_mem_list);
1841                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
1842
1843                 if (!iwpbl_buf)
1844                         return -ENOMEM;
1845
1846                 cqmr_buf = &iwpbl_buf->cq_mr;
1847                 if (iwpbl_buf->pbl_allocated) {
1848                         info.virtual_map = true;
1849                         info.pbl_chunk_size = 1;
1850                         info.first_pm_pbl_idx = cqmr_buf->cq_pbl.idx;
1851                 } else {
1852                         info.cq_pa = cqmr_buf->cq_pbl.addr;
1853                 }
1854         } else {
1855                 /* Kmode CQ resize */
1856                 int rsize;
1857
1858                 rsize = info.cq_size * sizeof(struct irdma_cqe);
1859                 kmem_buf.size = ALIGN(round_up(rsize, 256), 256);
1860                 kmem_buf.va = dma_alloc_coherent(dev->hw->device,
1861                                                  kmem_buf.size, &kmem_buf.pa,
1862                                                  GFP_KERNEL);
1863                 if (!kmem_buf.va)
1864                         return -ENOMEM;
1865
1866                 info.cq_base = kmem_buf.va;
1867                 info.cq_pa = kmem_buf.pa;
1868                 cq_buf = kzalloc(sizeof(*cq_buf), GFP_KERNEL);
1869                 if (!cq_buf) {
1870                         ret = -ENOMEM;
1871                         goto error;
1872                 }
1873         }
1874
1875         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
1876         if (!cqp_request) {
1877                 ret = -ENOMEM;
1878                 goto error;
1879         }
1880
1881         info.shadow_read_threshold = iwcq->sc_cq.shadow_read_threshold;
1882         info.cq_resize = true;
1883
1884         cqp_info = &cqp_request->info;
1885         m_info = &cqp_info->in.u.cq_modify.info;
1886         memcpy(m_info, &info, sizeof(*m_info));
1887
1888         cqp_info->cqp_cmd = IRDMA_OP_CQ_MODIFY;
1889         cqp_info->in.u.cq_modify.cq = &iwcq->sc_cq;
1890         cqp_info->in.u.cq_modify.scratch = (uintptr_t)cqp_request;
1891         cqp_info->post_sq = 1;
1892         status = irdma_handle_cqp_op(rf, cqp_request);
1893         irdma_put_cqp_request(&rf->cqp, cqp_request);
1894         if (status) {
1895                 ret = -EPROTO;
1896                 goto error;
1897         }
1898
1899         spin_lock_irqsave(&iwcq->lock, flags);
1900         if (cq_buf) {
1901                 cq_buf->kmem_buf = iwcq->kmem;
1902                 cq_buf->hw = dev->hw;
1903                 memcpy(&cq_buf->cq_uk, &iwcq->sc_cq.cq_uk, sizeof(cq_buf->cq_uk));
1904                 INIT_WORK(&cq_buf->work, irdma_free_cqbuf);
1905                 list_add_tail(&cq_buf->list, &iwcq->resize_list);
1906                 iwcq->kmem = kmem_buf;
1907         }
1908
1909         irdma_sc_cq_resize(&iwcq->sc_cq, &info);
1910         ibcq->cqe = info.cq_size - 1;
1911         spin_unlock_irqrestore(&iwcq->lock, flags);
1912
1913         return 0;
1914 error:
1915         if (!udata) {
1916                 dma_free_coherent(dev->hw->device, kmem_buf.size, kmem_buf.va,
1917                                   kmem_buf.pa);
1918                 kmem_buf.va = NULL;
1919         }
1920         kfree(cq_buf);
1921
1922         return ret;
1923 }
1924
1925 static inline int cq_validate_flags(u32 flags, u8 hw_rev)
1926 {
1927         /* GEN1 does not support CQ create flags */
1928         if (hw_rev == IRDMA_GEN_1)
1929                 return flags ? -EOPNOTSUPP : 0;
1930
1931         return flags & ~IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION ? -EOPNOTSUPP : 0;
1932 }
1933
1934 /**
1935  * irdma_create_cq - create cq
1936  * @ibcq: CQ allocated
1937  * @attr: attributes for cq
1938  * @udata: user data
1939  */
1940 static int irdma_create_cq(struct ib_cq *ibcq,
1941                            const struct ib_cq_init_attr *attr,
1942                            struct ib_udata *udata)
1943 {
1944         struct ib_device *ibdev = ibcq->device;
1945         struct irdma_device *iwdev = to_iwdev(ibdev);
1946         struct irdma_pci_f *rf = iwdev->rf;
1947         struct irdma_cq *iwcq = to_iwcq(ibcq);
1948         u32 cq_num = 0;
1949         struct irdma_sc_cq *cq;
1950         struct irdma_sc_dev *dev = &rf->sc_dev;
1951         struct irdma_cq_init_info info = {};
1952         enum irdma_status_code status;
1953         struct irdma_cqp_request *cqp_request;
1954         struct cqp_cmds_info *cqp_info;
1955         struct irdma_cq_uk_init_info *ukinfo = &info.cq_uk_init_info;
1956         unsigned long flags;
1957         int err_code;
1958         int entries = attr->cqe;
1959
1960         err_code = cq_validate_flags(attr->flags, dev->hw_attrs.uk_attrs.hw_rev);
1961         if (err_code)
1962                 return err_code;
1963         err_code = irdma_alloc_rsrc(rf, rf->allocated_cqs, rf->max_cq, &cq_num,
1964                                     &rf->next_cq);
1965         if (err_code)
1966                 return err_code;
1967
1968         cq = &iwcq->sc_cq;
1969         cq->back_cq = iwcq;
1970         spin_lock_init(&iwcq->lock);
1971         INIT_LIST_HEAD(&iwcq->resize_list);
1972         info.dev = dev;
1973         ukinfo->cq_size = max(entries, 4);
1974         ukinfo->cq_id = cq_num;
1975         iwcq->ibcq.cqe = info.cq_uk_init_info.cq_size;
1976         if (attr->comp_vector < rf->ceqs_count)
1977                 info.ceq_id = attr->comp_vector;
1978         info.ceq_id_valid = true;
1979         info.ceqe_mask = 1;
1980         info.type = IRDMA_CQ_TYPE_IWARP;
1981         info.vsi = &iwdev->vsi;
1982
1983         if (udata) {
1984                 struct irdma_ucontext *ucontext;
1985                 struct irdma_create_cq_req req = {};
1986                 struct irdma_cq_mr *cqmr;
1987                 struct irdma_pbl *iwpbl;
1988                 struct irdma_pbl *iwpbl_shadow;
1989                 struct irdma_cq_mr *cqmr_shadow;
1990
1991                 iwcq->user_mode = true;
1992                 ucontext =
1993                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
1994                                                   ibucontext);
1995                 if (ib_copy_from_udata(&req, udata,
1996                                        min(sizeof(req), udata->inlen))) {
1997                         err_code = -EFAULT;
1998                         goto cq_free_rsrc;
1999                 }
2000
2001                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2002                 iwpbl = irdma_get_pbl((unsigned long)req.user_cq_buf,
2003                                       &ucontext->cq_reg_mem_list);
2004                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2005                 if (!iwpbl) {
2006                         err_code = -EPROTO;
2007                         goto cq_free_rsrc;
2008                 }
2009
2010                 iwcq->iwpbl = iwpbl;
2011                 iwcq->cq_mem_size = 0;
2012                 cqmr = &iwpbl->cq_mr;
2013
2014                 if (rf->sc_dev.hw_attrs.uk_attrs.feature_flags &
2015                     IRDMA_FEATURE_CQ_RESIZE && !ucontext->legacy_mode) {
2016                         spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2017                         iwpbl_shadow = irdma_get_pbl(
2018                                         (unsigned long)req.user_shadow_area,
2019                                         &ucontext->cq_reg_mem_list);
2020                         spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2021
2022                         if (!iwpbl_shadow) {
2023                                 err_code = -EPROTO;
2024                                 goto cq_free_rsrc;
2025                         }
2026                         iwcq->iwpbl_shadow = iwpbl_shadow;
2027                         cqmr_shadow = &iwpbl_shadow->cq_mr;
2028                         info.shadow_area_pa = cqmr_shadow->cq_pbl.addr;
2029                         cqmr->split = true;
2030                 } else {
2031                         info.shadow_area_pa = cqmr->shadow;
2032                 }
2033                 if (iwpbl->pbl_allocated) {
2034                         info.virtual_map = true;
2035                         info.pbl_chunk_size = 1;
2036                         info.first_pm_pbl_idx = cqmr->cq_pbl.idx;
2037                 } else {
2038                         info.cq_base_pa = cqmr->cq_pbl.addr;
2039                 }
2040         } else {
2041                 /* Kmode allocations */
2042                 int rsize;
2043
2044                 if (entries > rf->max_cqe) {
2045                         err_code = -EINVAL;
2046                         goto cq_free_rsrc;
2047                 }
2048
2049                 entries++;
2050                 if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
2051                         entries *= 2;
2052                 ukinfo->cq_size = entries;
2053
2054                 rsize = info.cq_uk_init_info.cq_size * sizeof(struct irdma_cqe);
2055                 iwcq->kmem.size = ALIGN(round_up(rsize, 256), 256);
2056                 iwcq->kmem.va = dma_alloc_coherent(dev->hw->device,
2057                                                    iwcq->kmem.size,
2058                                                    &iwcq->kmem.pa, GFP_KERNEL);
2059                 if (!iwcq->kmem.va) {
2060                         err_code = -ENOMEM;
2061                         goto cq_free_rsrc;
2062                 }
2063
2064                 iwcq->kmem_shadow.size = ALIGN(IRDMA_SHADOW_AREA_SIZE << 3,
2065                                                64);
2066                 iwcq->kmem_shadow.va = dma_alloc_coherent(dev->hw->device,
2067                                                           iwcq->kmem_shadow.size,
2068                                                           &iwcq->kmem_shadow.pa,
2069                                                           GFP_KERNEL);
2070                 if (!iwcq->kmem_shadow.va) {
2071                         err_code = -ENOMEM;
2072                         goto cq_free_rsrc;
2073                 }
2074                 info.shadow_area_pa = iwcq->kmem_shadow.pa;
2075                 ukinfo->shadow_area = iwcq->kmem_shadow.va;
2076                 ukinfo->cq_base = iwcq->kmem.va;
2077                 info.cq_base_pa = iwcq->kmem.pa;
2078         }
2079
2080         if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
2081                 info.shadow_read_threshold = min(info.cq_uk_init_info.cq_size / 2,
2082                                                  (u32)IRDMA_MAX_CQ_READ_THRESH);
2083
2084         if (irdma_sc_cq_init(cq, &info)) {
2085                 ibdev_dbg(&iwdev->ibdev, "VERBS: init cq fail\n");
2086                 err_code = -EPROTO;
2087                 goto cq_free_rsrc;
2088         }
2089
2090         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
2091         if (!cqp_request) {
2092                 err_code = -ENOMEM;
2093                 goto cq_free_rsrc;
2094         }
2095
2096         cqp_info = &cqp_request->info;
2097         cqp_info->cqp_cmd = IRDMA_OP_CQ_CREATE;
2098         cqp_info->post_sq = 1;
2099         cqp_info->in.u.cq_create.cq = cq;
2100         cqp_info->in.u.cq_create.check_overflow = true;
2101         cqp_info->in.u.cq_create.scratch = (uintptr_t)cqp_request;
2102         status = irdma_handle_cqp_op(rf, cqp_request);
2103         irdma_put_cqp_request(&rf->cqp, cqp_request);
2104         if (status) {
2105                 err_code = -ENOMEM;
2106                 goto cq_free_rsrc;
2107         }
2108
2109         if (udata) {
2110                 struct irdma_create_cq_resp resp = {};
2111
2112                 resp.cq_id = info.cq_uk_init_info.cq_id;
2113                 resp.cq_size = info.cq_uk_init_info.cq_size;
2114                 if (ib_copy_to_udata(udata, &resp,
2115                                      min(sizeof(resp), udata->outlen))) {
2116                         ibdev_dbg(&iwdev->ibdev,
2117                                   "VERBS: copy to user data\n");
2118                         err_code = -EPROTO;
2119                         goto cq_destroy;
2120                 }
2121         }
2122         return 0;
2123 cq_destroy:
2124         irdma_cq_wq_destroy(rf, cq);
2125 cq_free_rsrc:
2126         irdma_cq_free_rsrc(rf, iwcq);
2127
2128         return err_code;
2129 }
2130
2131 /**
2132  * irdma_get_mr_access - get hw MR access permissions from IB access flags
2133  * @access: IB access flags
2134  */
2135 static inline u16 irdma_get_mr_access(int access)
2136 {
2137         u16 hw_access = 0;
2138
2139         hw_access |= (access & IB_ACCESS_LOCAL_WRITE) ?
2140                      IRDMA_ACCESS_FLAGS_LOCALWRITE : 0;
2141         hw_access |= (access & IB_ACCESS_REMOTE_WRITE) ?
2142                      IRDMA_ACCESS_FLAGS_REMOTEWRITE : 0;
2143         hw_access |= (access & IB_ACCESS_REMOTE_READ) ?
2144                      IRDMA_ACCESS_FLAGS_REMOTEREAD : 0;
2145         hw_access |= (access & IB_ACCESS_MW_BIND) ?
2146                      IRDMA_ACCESS_FLAGS_BIND_WINDOW : 0;
2147         hw_access |= (access & IB_ZERO_BASED) ?
2148                      IRDMA_ACCESS_FLAGS_ZERO_BASED : 0;
2149         hw_access |= IRDMA_ACCESS_FLAGS_LOCALREAD;
2150
2151         return hw_access;
2152 }
2153
2154 /**
2155  * irdma_free_stag - free stag resource
2156  * @iwdev: irdma device
2157  * @stag: stag to free
2158  */
2159 static void irdma_free_stag(struct irdma_device *iwdev, u32 stag)
2160 {
2161         u32 stag_idx;
2162
2163         stag_idx = (stag & iwdev->rf->mr_stagmask) >> IRDMA_CQPSQ_STAG_IDX_S;
2164         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_mrs, stag_idx);
2165 }
2166
2167 /**
2168  * irdma_create_stag - create random stag
2169  * @iwdev: irdma device
2170  */
2171 static u32 irdma_create_stag(struct irdma_device *iwdev)
2172 {
2173         u32 stag = 0;
2174         u32 stag_index = 0;
2175         u32 next_stag_index;
2176         u32 driver_key;
2177         u32 random;
2178         u8 consumer_key;
2179         int ret;
2180
2181         get_random_bytes(&random, sizeof(random));
2182         consumer_key = (u8)random;
2183
2184         driver_key = random & ~iwdev->rf->mr_stagmask;
2185         next_stag_index = (random & iwdev->rf->mr_stagmask) >> 8;
2186         next_stag_index %= iwdev->rf->max_mr;
2187
2188         ret = irdma_alloc_rsrc(iwdev->rf, iwdev->rf->allocated_mrs,
2189                                iwdev->rf->max_mr, &stag_index,
2190                                &next_stag_index);
2191         if (ret)
2192                 return stag;
2193         stag = stag_index << IRDMA_CQPSQ_STAG_IDX_S;
2194         stag |= driver_key;
2195         stag += (u32)consumer_key;
2196
2197         return stag;
2198 }
2199
2200 /**
2201  * irdma_next_pbl_addr - Get next pbl address
2202  * @pbl: pointer to a pble
2203  * @pinfo: info pointer
2204  * @idx: index
2205  */
2206 static inline u64 *irdma_next_pbl_addr(u64 *pbl, struct irdma_pble_info **pinfo,
2207                                        u32 *idx)
2208 {
2209         *idx += 1;
2210         if (!(*pinfo) || *idx != (*pinfo)->cnt)
2211                 return ++pbl;
2212         *idx = 0;
2213         (*pinfo)++;
2214
2215         return (*pinfo)->addr;
2216 }
2217
2218 /**
2219  * irdma_copy_user_pgaddrs - copy user page address to pble's os locally
2220  * @iwmr: iwmr for IB's user page addresses
2221  * @pbl: ple pointer to save 1 level or 0 level pble
2222  * @level: indicated level 0, 1 or 2
2223  */
2224 static void irdma_copy_user_pgaddrs(struct irdma_mr *iwmr, u64 *pbl,
2225                                     enum irdma_pble_level level)
2226 {
2227         struct ib_umem *region = iwmr->region;
2228         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2229         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2230         struct irdma_pble_info *pinfo;
2231         struct ib_block_iter biter;
2232         u32 idx = 0;
2233         u32 pbl_cnt = 0;
2234
2235         pinfo = (level == PBLE_LEVEL_1) ? NULL : palloc->level2.leaf;
2236
2237         if (iwmr->type == IRDMA_MEMREG_TYPE_QP)
2238                 iwpbl->qp_mr.sq_page = sg_page(region->sg_head.sgl);
2239
2240         rdma_umem_for_each_dma_block(region, &biter, iwmr->page_size) {
2241                 *pbl = rdma_block_iter_dma_address(&biter);
2242                 if (++pbl_cnt == palloc->total_cnt)
2243                         break;
2244                 pbl = irdma_next_pbl_addr(pbl, &pinfo, &idx);
2245         }
2246 }
2247
2248 /**
2249  * irdma_check_mem_contiguous - check if pbls stored in arr are contiguous
2250  * @arr: lvl1 pbl array
2251  * @npages: page count
2252  * @pg_size: page size
2253  *
2254  */
2255 static bool irdma_check_mem_contiguous(u64 *arr, u32 npages, u32 pg_size)
2256 {
2257         u32 pg_idx;
2258
2259         for (pg_idx = 0; pg_idx < npages; pg_idx++) {
2260                 if ((*arr + (pg_size * pg_idx)) != arr[pg_idx])
2261                         return false;
2262         }
2263
2264         return true;
2265 }
2266
2267 /**
2268  * irdma_check_mr_contiguous - check if MR is physically contiguous
2269  * @palloc: pbl allocation struct
2270  * @pg_size: page size
2271  */
2272 static bool irdma_check_mr_contiguous(struct irdma_pble_alloc *palloc,
2273                                       u32 pg_size)
2274 {
2275         struct irdma_pble_level2 *lvl2 = &palloc->level2;
2276         struct irdma_pble_info *leaf = lvl2->leaf;
2277         u64 *arr = NULL;
2278         u64 *start_addr = NULL;
2279         int i;
2280         bool ret;
2281
2282         if (palloc->level == PBLE_LEVEL_1) {
2283                 arr = palloc->level1.addr;
2284                 ret = irdma_check_mem_contiguous(arr, palloc->total_cnt,
2285                                                  pg_size);
2286                 return ret;
2287         }
2288
2289         start_addr = leaf->addr;
2290
2291         for (i = 0; i < lvl2->leaf_cnt; i++, leaf++) {
2292                 arr = leaf->addr;
2293                 if ((*start_addr + (i * pg_size * PBLE_PER_PAGE)) != *arr)
2294                         return false;
2295                 ret = irdma_check_mem_contiguous(arr, leaf->cnt, pg_size);
2296                 if (!ret)
2297                         return false;
2298         }
2299
2300         return true;
2301 }
2302
2303 /**
2304  * irdma_setup_pbles - copy user pg address to pble's
2305  * @rf: RDMA PCI function
2306  * @iwmr: mr pointer for this memory registration
2307  * @use_pbles: flag if to use pble's
2308  */
2309 static int irdma_setup_pbles(struct irdma_pci_f *rf, struct irdma_mr *iwmr,
2310                              bool use_pbles)
2311 {
2312         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2313         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2314         struct irdma_pble_info *pinfo;
2315         u64 *pbl;
2316         enum irdma_status_code status;
2317         enum irdma_pble_level level = PBLE_LEVEL_1;
2318
2319         if (use_pbles) {
2320                 status = irdma_get_pble(rf->pble_rsrc, palloc, iwmr->page_cnt,
2321                                         false);
2322                 if (status)
2323                         return -ENOMEM;
2324
2325                 iwpbl->pbl_allocated = true;
2326                 level = palloc->level;
2327                 pinfo = (level == PBLE_LEVEL_1) ? &palloc->level1 :
2328                                                   palloc->level2.leaf;
2329                 pbl = pinfo->addr;
2330         } else {
2331                 pbl = iwmr->pgaddrmem;
2332         }
2333
2334         irdma_copy_user_pgaddrs(iwmr, pbl, level);
2335
2336         if (use_pbles)
2337                 iwmr->pgaddrmem[0] = *pbl;
2338
2339         return 0;
2340 }
2341
2342 /**
2343  * irdma_handle_q_mem - handle memory for qp and cq
2344  * @iwdev: irdma device
2345  * @req: information for q memory management
2346  * @iwpbl: pble struct
2347  * @use_pbles: flag to use pble
2348  */
2349 static int irdma_handle_q_mem(struct irdma_device *iwdev,
2350                               struct irdma_mem_reg_req *req,
2351                               struct irdma_pbl *iwpbl, bool use_pbles)
2352 {
2353         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2354         struct irdma_mr *iwmr = iwpbl->iwmr;
2355         struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;
2356         struct irdma_cq_mr *cqmr = &iwpbl->cq_mr;
2357         struct irdma_hmc_pble *hmc_p;
2358         u64 *arr = iwmr->pgaddrmem;
2359         u32 pg_size, total;
2360         int err = 0;
2361         bool ret = true;
2362
2363         pg_size = iwmr->page_size;
2364         err = irdma_setup_pbles(iwdev->rf, iwmr, use_pbles);
2365         if (err)
2366                 return err;
2367
2368         if (use_pbles && palloc->level != PBLE_LEVEL_1) {
2369                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2370                 iwpbl->pbl_allocated = false;
2371                 return -ENOMEM;
2372         }
2373
2374         if (use_pbles)
2375                 arr = palloc->level1.addr;
2376
2377         switch (iwmr->type) {
2378         case IRDMA_MEMREG_TYPE_QP:
2379                 total = req->sq_pages + req->rq_pages;
2380                 hmc_p = &qpmr->sq_pbl;
2381                 qpmr->shadow = (dma_addr_t)arr[total];
2382
2383                 if (use_pbles) {
2384                         ret = irdma_check_mem_contiguous(arr, req->sq_pages,
2385                                                          pg_size);
2386                         if (ret)
2387                                 ret = irdma_check_mem_contiguous(&arr[req->sq_pages],
2388                                                                  req->rq_pages,
2389                                                                  pg_size);
2390                 }
2391
2392                 if (!ret) {
2393                         hmc_p->idx = palloc->level1.idx;
2394                         hmc_p = &qpmr->rq_pbl;
2395                         hmc_p->idx = palloc->level1.idx + req->sq_pages;
2396                 } else {
2397                         hmc_p->addr = arr[0];
2398                         hmc_p = &qpmr->rq_pbl;
2399                         hmc_p->addr = arr[req->sq_pages];
2400                 }
2401                 break;
2402         case IRDMA_MEMREG_TYPE_CQ:
2403                 hmc_p = &cqmr->cq_pbl;
2404
2405                 if (!cqmr->split)
2406                         cqmr->shadow = (dma_addr_t)arr[req->cq_pages];
2407
2408                 if (use_pbles)
2409                         ret = irdma_check_mem_contiguous(arr, req->cq_pages,
2410                                                          pg_size);
2411
2412                 if (!ret)
2413                         hmc_p->idx = palloc->level1.idx;
2414                 else
2415                         hmc_p->addr = arr[0];
2416         break;
2417         default:
2418                 ibdev_dbg(&iwdev->ibdev, "VERBS: MR type error\n");
2419                 err = -EINVAL;
2420         }
2421
2422         if (use_pbles && ret) {
2423                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2424                 iwpbl->pbl_allocated = false;
2425         }
2426
2427         return err;
2428 }
2429
2430 /**
2431  * irdma_hw_alloc_mw - create the hw memory window
2432  * @iwdev: irdma device
2433  * @iwmr: pointer to memory window info
2434  */
2435 static int irdma_hw_alloc_mw(struct irdma_device *iwdev, struct irdma_mr *iwmr)
2436 {
2437         struct irdma_mw_alloc_info *info;
2438         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2439         struct irdma_cqp_request *cqp_request;
2440         struct cqp_cmds_info *cqp_info;
2441         enum irdma_status_code status;
2442
2443         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2444         if (!cqp_request)
2445                 return -ENOMEM;
2446
2447         cqp_info = &cqp_request->info;
2448         info = &cqp_info->in.u.mw_alloc.info;
2449         memset(info, 0, sizeof(*info));
2450         if (iwmr->ibmw.type == IB_MW_TYPE_1)
2451                 info->mw_wide = true;
2452
2453         info->page_size = PAGE_SIZE;
2454         info->mw_stag_index = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2455         info->pd_id = iwpd->sc_pd.pd_id;
2456         info->remote_access = true;
2457         cqp_info->cqp_cmd = IRDMA_OP_MW_ALLOC;
2458         cqp_info->post_sq = 1;
2459         cqp_info->in.u.mw_alloc.dev = &iwdev->rf->sc_dev;
2460         cqp_info->in.u.mw_alloc.scratch = (uintptr_t)cqp_request;
2461         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2462         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2463
2464         return status ? -ENOMEM : 0;
2465 }
2466
2467 /**
2468  * irdma_alloc_mw - Allocate memory window
2469  * @ibmw: Memory Window
2470  * @udata: user data pointer
2471  */
2472 static int irdma_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata)
2473 {
2474         struct irdma_device *iwdev = to_iwdev(ibmw->device);
2475         struct irdma_mr *iwmr = to_iwmw(ibmw);
2476         int err_code;
2477         u32 stag;
2478
2479         stag = irdma_create_stag(iwdev);
2480         if (!stag)
2481                 return -ENOMEM;
2482
2483         iwmr->stag = stag;
2484         ibmw->rkey = stag;
2485
2486         err_code = irdma_hw_alloc_mw(iwdev, iwmr);
2487         if (err_code) {
2488                 irdma_free_stag(iwdev, stag);
2489                 return err_code;
2490         }
2491
2492         return 0;
2493 }
2494
2495 /**
2496  * irdma_dealloc_mw - Dealloc memory window
2497  * @ibmw: memory window structure.
2498  */
2499 static int irdma_dealloc_mw(struct ib_mw *ibmw)
2500 {
2501         struct ib_pd *ibpd = ibmw->pd;
2502         struct irdma_pd *iwpd = to_iwpd(ibpd);
2503         struct irdma_mr *iwmr = to_iwmr((struct ib_mr *)ibmw);
2504         struct irdma_device *iwdev = to_iwdev(ibmw->device);
2505         struct irdma_cqp_request *cqp_request;
2506         struct cqp_cmds_info *cqp_info;
2507         struct irdma_dealloc_stag_info *info;
2508
2509         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2510         if (!cqp_request)
2511                 return -ENOMEM;
2512
2513         cqp_info = &cqp_request->info;
2514         info = &cqp_info->in.u.dealloc_stag.info;
2515         memset(info, 0, sizeof(*info));
2516         info->pd_id = iwpd->sc_pd.pd_id & 0x00007fff;
2517         info->stag_idx = ibmw->rkey >> IRDMA_CQPSQ_STAG_IDX_S;
2518         info->mr = false;
2519         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
2520         cqp_info->post_sq = 1;
2521         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
2522         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
2523         irdma_handle_cqp_op(iwdev->rf, cqp_request);
2524         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2525         irdma_free_stag(iwdev, iwmr->stag);
2526
2527         return 0;
2528 }
2529
2530 /**
2531  * irdma_hw_alloc_stag - cqp command to allocate stag
2532  * @iwdev: irdma device
2533  * @iwmr: irdma mr pointer
2534  */
2535 static int irdma_hw_alloc_stag(struct irdma_device *iwdev,
2536                                struct irdma_mr *iwmr)
2537 {
2538         struct irdma_allocate_stag_info *info;
2539         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2540         enum irdma_status_code status;
2541         int err = 0;
2542         struct irdma_cqp_request *cqp_request;
2543         struct cqp_cmds_info *cqp_info;
2544
2545         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2546         if (!cqp_request)
2547                 return -ENOMEM;
2548
2549         cqp_info = &cqp_request->info;
2550         info = &cqp_info->in.u.alloc_stag.info;
2551         memset(info, 0, sizeof(*info));
2552         info->page_size = PAGE_SIZE;
2553         info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2554         info->pd_id = iwpd->sc_pd.pd_id;
2555         info->total_len = iwmr->len;
2556         info->remote_access = true;
2557         cqp_info->cqp_cmd = IRDMA_OP_ALLOC_STAG;
2558         cqp_info->post_sq = 1;
2559         cqp_info->in.u.alloc_stag.dev = &iwdev->rf->sc_dev;
2560         cqp_info->in.u.alloc_stag.scratch = (uintptr_t)cqp_request;
2561         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2562         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2563         if (status)
2564                 err = -ENOMEM;
2565
2566         return err;
2567 }
2568
2569 /**
2570  * irdma_alloc_mr - register stag for fast memory registration
2571  * @pd: ibpd pointer
2572  * @mr_type: memory for stag registrion
2573  * @max_num_sg: man number of pages
2574  */
2575 static struct ib_mr *irdma_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
2576                                     u32 max_num_sg)
2577 {
2578         struct irdma_device *iwdev = to_iwdev(pd->device);
2579         struct irdma_pble_alloc *palloc;
2580         struct irdma_pbl *iwpbl;
2581         struct irdma_mr *iwmr;
2582         enum irdma_status_code status;
2583         u32 stag;
2584         int err_code = -ENOMEM;
2585
2586         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2587         if (!iwmr)
2588                 return ERR_PTR(-ENOMEM);
2589
2590         stag = irdma_create_stag(iwdev);
2591         if (!stag) {
2592                 err_code = -ENOMEM;
2593                 goto err;
2594         }
2595
2596         iwmr->stag = stag;
2597         iwmr->ibmr.rkey = stag;
2598         iwmr->ibmr.lkey = stag;
2599         iwmr->ibmr.pd = pd;
2600         iwmr->ibmr.device = pd->device;
2601         iwpbl = &iwmr->iwpbl;
2602         iwpbl->iwmr = iwmr;
2603         iwmr->type = IRDMA_MEMREG_TYPE_MEM;
2604         palloc = &iwpbl->pble_alloc;
2605         iwmr->page_cnt = max_num_sg;
2606         status = irdma_get_pble(iwdev->rf->pble_rsrc, palloc, iwmr->page_cnt,
2607                                 true);
2608         if (status)
2609                 goto err_get_pble;
2610
2611         err_code = irdma_hw_alloc_stag(iwdev, iwmr);
2612         if (err_code)
2613                 goto err_alloc_stag;
2614
2615         iwpbl->pbl_allocated = true;
2616
2617         return &iwmr->ibmr;
2618 err_alloc_stag:
2619         irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2620 err_get_pble:
2621         irdma_free_stag(iwdev, stag);
2622 err:
2623         kfree(iwmr);
2624
2625         return ERR_PTR(err_code);
2626 }
2627
2628 /**
2629  * irdma_set_page - populate pbl list for fmr
2630  * @ibmr: ib mem to access iwarp mr pointer
2631  * @addr: page dma address fro pbl list
2632  */
2633 static int irdma_set_page(struct ib_mr *ibmr, u64 addr)
2634 {
2635         struct irdma_mr *iwmr = to_iwmr(ibmr);
2636         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2637         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2638         u64 *pbl;
2639
2640         if (unlikely(iwmr->npages == iwmr->page_cnt))
2641                 return -ENOMEM;
2642
2643         pbl = palloc->level1.addr;
2644         pbl[iwmr->npages++] = addr;
2645
2646         return 0;
2647 }
2648
2649 /**
2650  * irdma_map_mr_sg - map of sg list for fmr
2651  * @ibmr: ib mem to access iwarp mr pointer
2652  * @sg: scatter gather list
2653  * @sg_nents: number of sg pages
2654  * @sg_offset: scatter gather list for fmr
2655  */
2656 static int irdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
2657                            int sg_nents, unsigned int *sg_offset)
2658 {
2659         struct irdma_mr *iwmr = to_iwmr(ibmr);
2660
2661         iwmr->npages = 0;
2662
2663         return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, irdma_set_page);
2664 }
2665
2666 /**
2667  * irdma_hwreg_mr - send cqp command for memory registration
2668  * @iwdev: irdma device
2669  * @iwmr: irdma mr pointer
2670  * @access: access for MR
2671  */
2672 static int irdma_hwreg_mr(struct irdma_device *iwdev, struct irdma_mr *iwmr,
2673                           u16 access)
2674 {
2675         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2676         struct irdma_reg_ns_stag_info *stag_info;
2677         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2678         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2679         enum irdma_status_code status;
2680         int err = 0;
2681         struct irdma_cqp_request *cqp_request;
2682         struct cqp_cmds_info *cqp_info;
2683
2684         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2685         if (!cqp_request)
2686                 return -ENOMEM;
2687
2688         cqp_info = &cqp_request->info;
2689         stag_info = &cqp_info->in.u.mr_reg_non_shared.info;
2690         memset(stag_info, 0, sizeof(*stag_info));
2691         stag_info->va = iwpbl->user_base;
2692         stag_info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2693         stag_info->stag_key = (u8)iwmr->stag;
2694         stag_info->total_len = iwmr->len;
2695         stag_info->access_rights = irdma_get_mr_access(access);
2696         stag_info->pd_id = iwpd->sc_pd.pd_id;
2697         if (stag_info->access_rights & IRDMA_ACCESS_FLAGS_ZERO_BASED)
2698                 stag_info->addr_type = IRDMA_ADDR_TYPE_ZERO_BASED;
2699         else
2700                 stag_info->addr_type = IRDMA_ADDR_TYPE_VA_BASED;
2701         stag_info->page_size = iwmr->page_size;
2702
2703         if (iwpbl->pbl_allocated) {
2704                 if (palloc->level == PBLE_LEVEL_1) {
2705                         stag_info->first_pm_pbl_index = palloc->level1.idx;
2706                         stag_info->chunk_size = 1;
2707                 } else {
2708                         stag_info->first_pm_pbl_index = palloc->level2.root.idx;
2709                         stag_info->chunk_size = 3;
2710                 }
2711         } else {
2712                 stag_info->reg_addr_pa = iwmr->pgaddrmem[0];
2713         }
2714
2715         cqp_info->cqp_cmd = IRDMA_OP_MR_REG_NON_SHARED;
2716         cqp_info->post_sq = 1;
2717         cqp_info->in.u.mr_reg_non_shared.dev = &iwdev->rf->sc_dev;
2718         cqp_info->in.u.mr_reg_non_shared.scratch = (uintptr_t)cqp_request;
2719         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2720         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2721         if (status)
2722                 err = -ENOMEM;
2723
2724         return err;
2725 }
2726
2727 /**
2728  * irdma_reg_user_mr - Register a user memory region
2729  * @pd: ptr of pd
2730  * @start: virtual start address
2731  * @len: length of mr
2732  * @virt: virtual address
2733  * @access: access of mr
2734  * @udata: user data
2735  */
2736 static struct ib_mr *irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,
2737                                        u64 virt, int access,
2738                                        struct ib_udata *udata)
2739 {
2740         struct irdma_device *iwdev = to_iwdev(pd->device);
2741         struct irdma_ucontext *ucontext;
2742         struct irdma_pble_alloc *palloc;
2743         struct irdma_pbl *iwpbl;
2744         struct irdma_mr *iwmr;
2745         struct ib_umem *region;
2746         struct irdma_mem_reg_req req;
2747         u32 total, stag = 0;
2748         u8 shadow_pgcnt = 1;
2749         bool use_pbles = false;
2750         unsigned long flags;
2751         int err = -EINVAL;
2752         int ret;
2753
2754         if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size)
2755                 return ERR_PTR(-EINVAL);
2756
2757         region = ib_umem_get(pd->device, start, len, access);
2758
2759         if (IS_ERR(region)) {
2760                 ibdev_dbg(&iwdev->ibdev,
2761                           "VERBS: Failed to create ib_umem region\n");
2762                 return (struct ib_mr *)region;
2763         }
2764
2765         if (ib_copy_from_udata(&req, udata, min(sizeof(req), udata->inlen))) {
2766                 ib_umem_release(region);
2767                 return ERR_PTR(-EFAULT);
2768         }
2769
2770         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2771         if (!iwmr) {
2772                 ib_umem_release(region);
2773                 return ERR_PTR(-ENOMEM);
2774         }
2775
2776         iwpbl = &iwmr->iwpbl;
2777         iwpbl->iwmr = iwmr;
2778         iwmr->region = region;
2779         iwmr->ibmr.pd = pd;
2780         iwmr->ibmr.device = pd->device;
2781         iwmr->ibmr.iova = virt;
2782         iwmr->page_size = PAGE_SIZE;
2783
2784         if (req.reg_type == IRDMA_MEMREG_TYPE_MEM) {
2785                 iwmr->page_size = ib_umem_find_best_pgsz(region,
2786                                                          SZ_4K | SZ_2M | SZ_1G,
2787                                                          virt);
2788                 if (unlikely(!iwmr->page_size)) {
2789                         kfree(iwmr);
2790                         ib_umem_release(region);
2791                         return ERR_PTR(-EOPNOTSUPP);
2792                 }
2793         }
2794         iwmr->len = region->length;
2795         iwpbl->user_base = virt;
2796         palloc = &iwpbl->pble_alloc;
2797         iwmr->type = req.reg_type;
2798         iwmr->page_cnt = ib_umem_num_dma_blocks(region, iwmr->page_size);
2799
2800         switch (req.reg_type) {
2801         case IRDMA_MEMREG_TYPE_QP:
2802                 total = req.sq_pages + req.rq_pages + shadow_pgcnt;
2803                 if (total > iwmr->page_cnt) {
2804                         err = -EINVAL;
2805                         goto error;
2806                 }
2807                 total = req.sq_pages + req.rq_pages;
2808                 use_pbles = (total > 2);
2809                 err = irdma_handle_q_mem(iwdev, &req, iwpbl, use_pbles);
2810                 if (err)
2811                         goto error;
2812
2813                 ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext,
2814                                                      ibucontext);
2815                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
2816                 list_add_tail(&iwpbl->list, &ucontext->qp_reg_mem_list);
2817                 iwpbl->on_list = true;
2818                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
2819                 break;
2820         case IRDMA_MEMREG_TYPE_CQ:
2821                 if (iwdev->rf->sc_dev.hw_attrs.uk_attrs.feature_flags & IRDMA_FEATURE_CQ_RESIZE)
2822                         shadow_pgcnt = 0;
2823                 total = req.cq_pages + shadow_pgcnt;
2824                 if (total > iwmr->page_cnt) {
2825                         err = -EINVAL;
2826                         goto error;
2827                 }
2828
2829                 use_pbles = (req.cq_pages > 1);
2830                 err = irdma_handle_q_mem(iwdev, &req, iwpbl, use_pbles);
2831                 if (err)
2832                         goto error;
2833
2834                 ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext,
2835                                                      ibucontext);
2836                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2837                 list_add_tail(&iwpbl->list, &ucontext->cq_reg_mem_list);
2838                 iwpbl->on_list = true;
2839                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2840                 break;
2841         case IRDMA_MEMREG_TYPE_MEM:
2842                 use_pbles = (iwmr->page_cnt != 1);
2843
2844                 err = irdma_setup_pbles(iwdev->rf, iwmr, use_pbles);
2845                 if (err)
2846                         goto error;
2847
2848                 if (use_pbles) {
2849                         ret = irdma_check_mr_contiguous(palloc,
2850                                                         iwmr->page_size);
2851                         if (ret) {
2852                                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2853                                 iwpbl->pbl_allocated = false;
2854                         }
2855                 }
2856
2857                 stag = irdma_create_stag(iwdev);
2858                 if (!stag) {
2859                         err = -ENOMEM;
2860                         goto error;
2861                 }
2862
2863                 iwmr->stag = stag;
2864                 iwmr->ibmr.rkey = stag;
2865                 iwmr->ibmr.lkey = stag;
2866                 err = irdma_hwreg_mr(iwdev, iwmr, access);
2867                 if (err) {
2868                         irdma_free_stag(iwdev, stag);
2869                         goto error;
2870                 }
2871
2872                 break;
2873         default:
2874                 goto error;
2875         }
2876
2877         iwmr->type = req.reg_type;
2878
2879         return &iwmr->ibmr;
2880
2881 error:
2882         if (palloc->level != PBLE_LEVEL_0 && iwpbl->pbl_allocated)
2883                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2884         ib_umem_release(region);
2885         kfree(iwmr);
2886
2887         return ERR_PTR(err);
2888 }
2889
2890 /**
2891  * irdma_reg_phys_mr - register kernel physical memory
2892  * @pd: ibpd pointer
2893  * @addr: physical address of memory to register
2894  * @size: size of memory to register
2895  * @access: Access rights
2896  * @iova_start: start of virtual address for physical buffers
2897  */
2898 struct ib_mr *irdma_reg_phys_mr(struct ib_pd *pd, u64 addr, u64 size, int access,
2899                                 u64 *iova_start)
2900 {
2901         struct irdma_device *iwdev = to_iwdev(pd->device);
2902         struct irdma_pbl *iwpbl;
2903         struct irdma_mr *iwmr;
2904         enum irdma_status_code status;
2905         u32 stag;
2906         int ret;
2907
2908         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2909         if (!iwmr)
2910                 return ERR_PTR(-ENOMEM);
2911
2912         iwmr->ibmr.pd = pd;
2913         iwmr->ibmr.device = pd->device;
2914         iwpbl = &iwmr->iwpbl;
2915         iwpbl->iwmr = iwmr;
2916         iwmr->type = IRDMA_MEMREG_TYPE_MEM;
2917         iwpbl->user_base = *iova_start;
2918         stag = irdma_create_stag(iwdev);
2919         if (!stag) {
2920                 ret = -ENOMEM;
2921                 goto err;
2922         }
2923
2924         iwmr->stag = stag;
2925         iwmr->ibmr.iova = *iova_start;
2926         iwmr->ibmr.rkey = stag;
2927         iwmr->ibmr.lkey = stag;
2928         iwmr->page_cnt = 1;
2929         iwmr->pgaddrmem[0] = addr;
2930         iwmr->len = size;
2931         iwmr->page_size = SZ_4K;
2932         status = irdma_hwreg_mr(iwdev, iwmr, access);
2933         if (status) {
2934                 irdma_free_stag(iwdev, stag);
2935                 ret = -ENOMEM;
2936                 goto err;
2937         }
2938
2939         return &iwmr->ibmr;
2940
2941 err:
2942         kfree(iwmr);
2943
2944         return ERR_PTR(ret);
2945 }
2946
2947 /**
2948  * irdma_get_dma_mr - register physical mem
2949  * @pd: ptr of pd
2950  * @acc: access for memory
2951  */
2952 static struct ib_mr *irdma_get_dma_mr(struct ib_pd *pd, int acc)
2953 {
2954         u64 kva = 0;
2955
2956         return irdma_reg_phys_mr(pd, 0, 0, acc, &kva);
2957 }
2958
2959 /**
2960  * irdma_del_memlist - Deleting pbl list entries for CQ/QP
2961  * @iwmr: iwmr for IB's user page addresses
2962  * @ucontext: ptr to user context
2963  */
2964 static void irdma_del_memlist(struct irdma_mr *iwmr,
2965                               struct irdma_ucontext *ucontext)
2966 {
2967         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2968         unsigned long flags;
2969
2970         switch (iwmr->type) {
2971         case IRDMA_MEMREG_TYPE_CQ:
2972                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2973                 if (iwpbl->on_list) {
2974                         iwpbl->on_list = false;
2975                         list_del(&iwpbl->list);
2976                 }
2977                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2978                 break;
2979         case IRDMA_MEMREG_TYPE_QP:
2980                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
2981                 if (iwpbl->on_list) {
2982                         iwpbl->on_list = false;
2983                         list_del(&iwpbl->list);
2984                 }
2985                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
2986                 break;
2987         default:
2988                 break;
2989         }
2990 }
2991
2992 /**
2993  * irdma_dereg_mr - deregister mr
2994  * @ib_mr: mr ptr for dereg
2995  * @udata: user data
2996  */
2997 static int irdma_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
2998 {
2999         struct ib_pd *ibpd = ib_mr->pd;
3000         struct irdma_pd *iwpd = to_iwpd(ibpd);
3001         struct irdma_mr *iwmr = to_iwmr(ib_mr);
3002         struct irdma_device *iwdev = to_iwdev(ib_mr->device);
3003         struct irdma_dealloc_stag_info *info;
3004         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
3005         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
3006         struct irdma_cqp_request *cqp_request;
3007         struct cqp_cmds_info *cqp_info;
3008
3009         if (iwmr->type != IRDMA_MEMREG_TYPE_MEM) {
3010                 if (iwmr->region) {
3011                         struct irdma_ucontext *ucontext;
3012
3013                         ucontext = rdma_udata_to_drv_context(udata,
3014                                                 struct irdma_ucontext,
3015                                                 ibucontext);
3016                         irdma_del_memlist(iwmr, ucontext);
3017                 }
3018                 goto done;
3019         }
3020
3021         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
3022         if (!cqp_request)
3023                 return -ENOMEM;
3024
3025         cqp_info = &cqp_request->info;
3026         info = &cqp_info->in.u.dealloc_stag.info;
3027         memset(info, 0, sizeof(*info));
3028         info->pd_id = iwpd->sc_pd.pd_id & 0x00007fff;
3029         info->stag_idx = ib_mr->rkey >> IRDMA_CQPSQ_STAG_IDX_S;
3030         info->mr = true;
3031         if (iwpbl->pbl_allocated)
3032                 info->dealloc_pbl = true;
3033
3034         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
3035         cqp_info->post_sq = 1;
3036         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
3037         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
3038         irdma_handle_cqp_op(iwdev->rf, cqp_request);
3039         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
3040         irdma_free_stag(iwdev, iwmr->stag);
3041 done:
3042         if (iwpbl->pbl_allocated)
3043                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
3044         ib_umem_release(iwmr->region);
3045         kfree(iwmr);
3046
3047         return 0;
3048 }
3049
3050 /**
3051  * irdma_copy_sg_list - copy sg list for qp
3052  * @sg_list: copied into sg_list
3053  * @sgl: copy from sgl
3054  * @num_sges: count of sg entries
3055  */
3056 static void irdma_copy_sg_list(struct irdma_sge *sg_list, struct ib_sge *sgl,
3057                                int num_sges)
3058 {
3059         unsigned int i;
3060
3061         for (i = 0; (i < num_sges) && (i < IRDMA_MAX_WQ_FRAGMENT_COUNT); i++) {
3062                 sg_list[i].tag_off = sgl[i].addr;
3063                 sg_list[i].len = sgl[i].length;
3064                 sg_list[i].stag = sgl[i].lkey;
3065         }
3066 }
3067
3068 /**
3069  * irdma_post_send -  kernel application wr
3070  * @ibqp: qp ptr for wr
3071  * @ib_wr: work request ptr
3072  * @bad_wr: return of bad wr if err
3073  */
3074 static int irdma_post_send(struct ib_qp *ibqp,
3075                            const struct ib_send_wr *ib_wr,
3076                            const struct ib_send_wr **bad_wr)
3077 {
3078         struct irdma_qp *iwqp;
3079         struct irdma_qp_uk *ukqp;
3080         struct irdma_sc_dev *dev;
3081         struct irdma_post_sq_info info;
3082         enum irdma_status_code ret;
3083         int err = 0;
3084         unsigned long flags;
3085         bool inv_stag;
3086         struct irdma_ah *ah;
3087         bool reflush = false;
3088
3089         iwqp = to_iwqp(ibqp);
3090         ukqp = &iwqp->sc_qp.qp_uk;
3091         dev = &iwqp->iwdev->rf->sc_dev;
3092
3093         spin_lock_irqsave(&iwqp->lock, flags);
3094         if (iwqp->flush_issued && ukqp->sq_flush_complete)
3095                 reflush = true;
3096         while (ib_wr) {
3097                 memset(&info, 0, sizeof(info));
3098                 inv_stag = false;
3099                 info.wr_id = (ib_wr->wr_id);
3100                 if ((ib_wr->send_flags & IB_SEND_SIGNALED) || iwqp->sig_all)
3101                         info.signaled = true;
3102                 if (ib_wr->send_flags & IB_SEND_FENCE)
3103                         info.read_fence = true;
3104                 switch (ib_wr->opcode) {
3105                 case IB_WR_SEND_WITH_IMM:
3106                         if (ukqp->qp_caps & IRDMA_SEND_WITH_IMM) {
3107                                 info.imm_data_valid = true;
3108                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
3109                         } else {
3110                                 err = -EINVAL;
3111                                 break;
3112                         }
3113                         fallthrough;
3114                 case IB_WR_SEND:
3115                 case IB_WR_SEND_WITH_INV:
3116                         if (ib_wr->opcode == IB_WR_SEND ||
3117                             ib_wr->opcode == IB_WR_SEND_WITH_IMM) {
3118                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
3119                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL;
3120                                 else
3121                                         info.op_type = IRDMA_OP_TYPE_SEND;
3122                         } else {
3123                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
3124                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL_INV;
3125                                 else
3126                                         info.op_type = IRDMA_OP_TYPE_SEND_INV;
3127                                 info.stag_to_inv = ib_wr->ex.invalidate_rkey;
3128                         }
3129
3130                         if (ib_wr->send_flags & IB_SEND_INLINE) {
3131                                 info.op.inline_send.data = (void *)(unsigned long)
3132                                                            ib_wr->sg_list[0].addr;
3133                                 info.op.inline_send.len = ib_wr->sg_list[0].length;
3134                                 if (iwqp->ibqp.qp_type == IB_QPT_UD ||
3135                                     iwqp->ibqp.qp_type == IB_QPT_GSI) {
3136                                         ah = to_iwah(ud_wr(ib_wr)->ah);
3137                                         info.op.inline_send.ah_id = ah->sc_ah.ah_info.ah_idx;
3138                                         info.op.inline_send.qkey = ud_wr(ib_wr)->remote_qkey;
3139                                         info.op.inline_send.dest_qp = ud_wr(ib_wr)->remote_qpn;
3140                                 }
3141                                 ret = irdma_uk_inline_send(ukqp, &info, false);
3142                         } else {
3143                                 info.op.send.num_sges = ib_wr->num_sge;
3144                                 info.op.send.sg_list = (struct irdma_sge *)
3145                                                        ib_wr->sg_list;
3146                                 if (iwqp->ibqp.qp_type == IB_QPT_UD ||
3147                                     iwqp->ibqp.qp_type == IB_QPT_GSI) {
3148                                         ah = to_iwah(ud_wr(ib_wr)->ah);
3149                                         info.op.send.ah_id = ah->sc_ah.ah_info.ah_idx;
3150                                         info.op.send.qkey = ud_wr(ib_wr)->remote_qkey;
3151                                         info.op.send.dest_qp = ud_wr(ib_wr)->remote_qpn;
3152                                 }
3153                                 ret = irdma_uk_send(ukqp, &info, false);
3154                         }
3155
3156                         if (ret) {
3157                                 if (ret == IRDMA_ERR_QP_TOOMANY_WRS_POSTED)
3158                                         err = -ENOMEM;
3159                                 else
3160                                         err = -EINVAL;
3161                         }
3162                         break;
3163                 case IB_WR_RDMA_WRITE_WITH_IMM:
3164                         if (ukqp->qp_caps & IRDMA_WRITE_WITH_IMM) {
3165                                 info.imm_data_valid = true;
3166                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
3167                         } else {
3168                                 err = -EINVAL;
3169                                 break;
3170                         }
3171                         fallthrough;
3172                 case IB_WR_RDMA_WRITE:
3173                         if (ib_wr->send_flags & IB_SEND_SOLICITED)
3174                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE_SOL;
3175                         else
3176                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE;
3177
3178                         if (ib_wr->send_flags & IB_SEND_INLINE) {
3179                                 info.op.inline_rdma_write.data = (void *)(uintptr_t)ib_wr->sg_list[0].addr;
3180                                 info.op.inline_rdma_write.len = ib_wr->sg_list[0].length;
3181                                 info.op.inline_rdma_write.rem_addr.tag_off = rdma_wr(ib_wr)->remote_addr;
3182                                 info.op.inline_rdma_write.rem_addr.stag = rdma_wr(ib_wr)->rkey;
3183                                 ret = irdma_uk_inline_rdma_write(ukqp, &info, false);
3184                         } else {
3185                                 info.op.rdma_write.lo_sg_list = (void *)ib_wr->sg_list;
3186                                 info.op.rdma_write.num_lo_sges = ib_wr->num_sge;
3187                                 info.op.rdma_write.rem_addr.tag_off = rdma_wr(ib_wr)->remote_addr;
3188                                 info.op.rdma_write.rem_addr.stag = rdma_wr(ib_wr)->rkey;
3189                                 ret = irdma_uk_rdma_write(ukqp, &info, false);
3190                         }
3191
3192                         if (ret) {
3193                                 if (ret == IRDMA_ERR_QP_TOOMANY_WRS_POSTED)
3194                                         err = -ENOMEM;
3195                                 else
3196                                         err = -EINVAL;
3197                         }
3198                         break;
3199                 case IB_WR_RDMA_READ_WITH_INV:
3200                         inv_stag = true;
3201                         fallthrough;
3202                 case IB_WR_RDMA_READ:
3203                         if (ib_wr->num_sge >
3204                             dev->hw_attrs.uk_attrs.max_hw_read_sges) {
3205                                 err = -EINVAL;
3206                                 break;
3207                         }
3208                         info.op_type = IRDMA_OP_TYPE_RDMA_READ;
3209                         info.op.rdma_read.rem_addr.tag_off = rdma_wr(ib_wr)->remote_addr;
3210                         info.op.rdma_read.rem_addr.stag = rdma_wr(ib_wr)->rkey;
3211                         info.op.rdma_read.lo_sg_list = (void *)ib_wr->sg_list;
3212                         info.op.rdma_read.num_lo_sges = ib_wr->num_sge;
3213
3214                         ret = irdma_uk_rdma_read(ukqp, &info, inv_stag, false);
3215                         if (ret) {
3216                                 if (ret == IRDMA_ERR_QP_TOOMANY_WRS_POSTED)
3217                                         err = -ENOMEM;
3218                                 else
3219                                         err = -EINVAL;
3220                         }
3221                         break;
3222                 case IB_WR_LOCAL_INV:
3223                         info.op_type = IRDMA_OP_TYPE_INV_STAG;
3224                         info.op.inv_local_stag.target_stag = ib_wr->ex.invalidate_rkey;
3225                         ret = irdma_uk_stag_local_invalidate(ukqp, &info, true);
3226                         if (ret)
3227                                 err = -ENOMEM;
3228                         break;
3229                 case IB_WR_REG_MR: {
3230                         struct irdma_mr *iwmr = to_iwmr(reg_wr(ib_wr)->mr);
3231                         struct irdma_pble_alloc *palloc = &iwmr->iwpbl.pble_alloc;
3232                         struct irdma_fast_reg_stag_info stag_info = {};
3233
3234                         stag_info.signaled = info.signaled;
3235                         stag_info.read_fence = info.read_fence;
3236                         stag_info.access_rights = irdma_get_mr_access(reg_wr(ib_wr)->access);
3237                         stag_info.stag_key = reg_wr(ib_wr)->key & 0xff;
3238                         stag_info.stag_idx = reg_wr(ib_wr)->key >> 8;
3239                         stag_info.page_size = reg_wr(ib_wr)->mr->page_size;
3240                         stag_info.wr_id = ib_wr->wr_id;
3241                         stag_info.addr_type = IRDMA_ADDR_TYPE_VA_BASED;
3242                         stag_info.va = (void *)(uintptr_t)iwmr->ibmr.iova;
3243                         stag_info.total_len = iwmr->ibmr.length;
3244                         stag_info.reg_addr_pa = *palloc->level1.addr;
3245                         stag_info.first_pm_pbl_index = palloc->level1.idx;
3246                         stag_info.local_fence = ib_wr->send_flags & IB_SEND_FENCE;
3247                         if (iwmr->npages > IRDMA_MIN_PAGES_PER_FMR)
3248                                 stag_info.chunk_size = 1;
3249                         ret = irdma_sc_mr_fast_register(&iwqp->sc_qp, &stag_info,
3250                                                         true);
3251                         if (ret)
3252                                 err = -ENOMEM;
3253                         break;
3254                 }
3255                 default:
3256                         err = -EINVAL;
3257                         ibdev_dbg(&iwqp->iwdev->ibdev,
3258                                   "VERBS: upost_send bad opcode = 0x%x\n",
3259                                   ib_wr->opcode);
3260                         break;
3261                 }
3262
3263                 if (err)
3264                         break;
3265                 ib_wr = ib_wr->next;
3266         }
3267
3268         if (!iwqp->flush_issued && iwqp->hw_iwarp_state <= IRDMA_QP_STATE_RTS) {
3269                 irdma_uk_qp_post_wr(ukqp);
3270                 spin_unlock_irqrestore(&iwqp->lock, flags);
3271         } else if (reflush) {
3272                 ukqp->sq_flush_complete = false;
3273                 spin_unlock_irqrestore(&iwqp->lock, flags);
3274                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_SQ | IRDMA_REFLUSH);
3275         } else {
3276                 spin_unlock_irqrestore(&iwqp->lock, flags);
3277         }
3278         if (err)
3279                 *bad_wr = ib_wr;
3280
3281         return err;
3282 }
3283
3284 /**
3285  * irdma_post_recv - post receive wr for kernel application
3286  * @ibqp: ib qp pointer
3287  * @ib_wr: work request for receive
3288  * @bad_wr: bad wr caused an error
3289  */
3290 static int irdma_post_recv(struct ib_qp *ibqp,
3291                            const struct ib_recv_wr *ib_wr,
3292                            const struct ib_recv_wr **bad_wr)
3293 {
3294         struct irdma_qp *iwqp;
3295         struct irdma_qp_uk *ukqp;
3296         struct irdma_post_rq_info post_recv = {};
3297         struct irdma_sge sg_list[IRDMA_MAX_WQ_FRAGMENT_COUNT];
3298         enum irdma_status_code ret = 0;
3299         unsigned long flags;
3300         int err = 0;
3301         bool reflush = false;
3302
3303         iwqp = to_iwqp(ibqp);
3304         ukqp = &iwqp->sc_qp.qp_uk;
3305
3306         spin_lock_irqsave(&iwqp->lock, flags);
3307         if (iwqp->flush_issued && ukqp->rq_flush_complete)
3308                 reflush = true;
3309         while (ib_wr) {
3310                 post_recv.num_sges = ib_wr->num_sge;
3311                 post_recv.wr_id = ib_wr->wr_id;
3312                 irdma_copy_sg_list(sg_list, ib_wr->sg_list, ib_wr->num_sge);
3313                 post_recv.sg_list = sg_list;
3314                 ret = irdma_uk_post_receive(ukqp, &post_recv);
3315                 if (ret) {
3316                         ibdev_dbg(&iwqp->iwdev->ibdev,
3317                                   "VERBS: post_recv err %d\n", ret);
3318                         if (ret == IRDMA_ERR_QP_TOOMANY_WRS_POSTED)
3319                                 err = -ENOMEM;
3320                         else
3321                                 err = -EINVAL;
3322                         goto out;
3323                 }
3324
3325                 ib_wr = ib_wr->next;
3326         }
3327
3328 out:
3329         if (reflush) {
3330                 ukqp->rq_flush_complete = false;
3331                 spin_unlock_irqrestore(&iwqp->lock, flags);
3332                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_RQ | IRDMA_REFLUSH);
3333         } else {
3334                 spin_unlock_irqrestore(&iwqp->lock, flags);
3335         }
3336
3337         if (err)
3338                 *bad_wr = ib_wr;
3339
3340         return err;
3341 }
3342
3343 /**
3344  * irdma_flush_err_to_ib_wc_status - return change flush error code to IB status
3345  * @opcode: iwarp flush code
3346  */
3347 static enum ib_wc_status irdma_flush_err_to_ib_wc_status(enum irdma_flush_opcode opcode)
3348 {
3349         switch (opcode) {
3350         case FLUSH_PROT_ERR:
3351                 return IB_WC_LOC_PROT_ERR;
3352         case FLUSH_REM_ACCESS_ERR:
3353                 return IB_WC_REM_ACCESS_ERR;
3354         case FLUSH_LOC_QP_OP_ERR:
3355                 return IB_WC_LOC_QP_OP_ERR;
3356         case FLUSH_REM_OP_ERR:
3357                 return IB_WC_REM_OP_ERR;
3358         case FLUSH_LOC_LEN_ERR:
3359                 return IB_WC_LOC_LEN_ERR;
3360         case FLUSH_GENERAL_ERR:
3361                 return IB_WC_WR_FLUSH_ERR;
3362         case FLUSH_FATAL_ERR:
3363         default:
3364                 return IB_WC_FATAL_ERR;
3365         }
3366 }
3367
3368 /**
3369  * irdma_process_cqe - process cqe info
3370  * @entry: processed cqe
3371  * @cq_poll_info: cqe info
3372  */
3373 static void irdma_process_cqe(struct ib_wc *entry,
3374                               struct irdma_cq_poll_info *cq_poll_info)
3375 {
3376         struct irdma_qp *iwqp;
3377         struct irdma_sc_qp *qp;
3378
3379         entry->wc_flags = 0;
3380         entry->pkey_index = 0;
3381         entry->wr_id = cq_poll_info->wr_id;
3382
3383         qp = cq_poll_info->qp_handle;
3384         iwqp = qp->qp_uk.back_qp;
3385         entry->qp = qp->qp_uk.back_qp;
3386
3387         if (cq_poll_info->error) {
3388                 entry->status = (cq_poll_info->comp_status == IRDMA_COMPL_STATUS_FLUSHED) ?
3389                                 irdma_flush_err_to_ib_wc_status(cq_poll_info->minor_err) : IB_WC_GENERAL_ERR;
3390
3391                 entry->vendor_err = cq_poll_info->major_err << 16 |
3392                                     cq_poll_info->minor_err;
3393         } else {
3394                 entry->status = IB_WC_SUCCESS;
3395                 if (cq_poll_info->imm_valid) {
3396                         entry->ex.imm_data = htonl(cq_poll_info->imm_data);
3397                         entry->wc_flags |= IB_WC_WITH_IMM;
3398                 }
3399                 if (cq_poll_info->ud_smac_valid) {
3400                         ether_addr_copy(entry->smac, cq_poll_info->ud_smac);
3401                         entry->wc_flags |= IB_WC_WITH_SMAC;
3402                 }
3403
3404                 if (cq_poll_info->ud_vlan_valid) {
3405                         entry->vlan_id = cq_poll_info->ud_vlan & VLAN_VID_MASK;
3406                         entry->wc_flags |= IB_WC_WITH_VLAN;
3407                         entry->sl = cq_poll_info->ud_vlan >> VLAN_PRIO_SHIFT;
3408                 } else {
3409                         entry->sl = 0;
3410                 }
3411         }
3412
3413         switch (cq_poll_info->op_type) {
3414         case IRDMA_OP_TYPE_RDMA_WRITE:
3415         case IRDMA_OP_TYPE_RDMA_WRITE_SOL:
3416                 entry->opcode = IB_WC_RDMA_WRITE;
3417                 break;
3418         case IRDMA_OP_TYPE_RDMA_READ_INV_STAG:
3419         case IRDMA_OP_TYPE_RDMA_READ:
3420                 entry->opcode = IB_WC_RDMA_READ;
3421                 break;
3422         case IRDMA_OP_TYPE_SEND_INV:
3423         case IRDMA_OP_TYPE_SEND_SOL:
3424         case IRDMA_OP_TYPE_SEND_SOL_INV:
3425         case IRDMA_OP_TYPE_SEND:
3426                 entry->opcode = IB_WC_SEND;
3427                 break;
3428         case IRDMA_OP_TYPE_FAST_REG_NSMR:
3429                 entry->opcode = IB_WC_REG_MR;
3430                 break;
3431         case IRDMA_OP_TYPE_INV_STAG:
3432                 entry->opcode = IB_WC_LOCAL_INV;
3433                 break;
3434         case IRDMA_OP_TYPE_REC_IMM:
3435         case IRDMA_OP_TYPE_REC:
3436                 entry->opcode = cq_poll_info->op_type == IRDMA_OP_TYPE_REC_IMM ?
3437                         IB_WC_RECV_RDMA_WITH_IMM : IB_WC_RECV;
3438                 if (qp->qp_uk.qp_type != IRDMA_QP_TYPE_ROCE_UD &&
3439                     cq_poll_info->stag_invalid_set) {
3440                         entry->ex.invalidate_rkey = cq_poll_info->inv_stag;
3441                         entry->wc_flags |= IB_WC_WITH_INVALIDATE;
3442                 }
3443                 break;
3444         default:
3445                 ibdev_err(&iwqp->iwdev->ibdev,
3446                           "Invalid opcode = %d in CQE\n", cq_poll_info->op_type);
3447                 entry->status = IB_WC_GENERAL_ERR;
3448                 return;
3449         }
3450
3451         if (qp->qp_uk.qp_type == IRDMA_QP_TYPE_ROCE_UD) {
3452                 entry->src_qp = cq_poll_info->ud_src_qpn;
3453                 entry->slid = 0;
3454                 entry->wc_flags |=
3455                         (IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE);
3456                 entry->network_hdr_type = cq_poll_info->ipv4 ?
3457                                                   RDMA_NETWORK_IPV4 :
3458                                                   RDMA_NETWORK_IPV6;
3459         } else {
3460                 entry->src_qp = cq_poll_info->qp_id;
3461         }
3462
3463         entry->byte_len = cq_poll_info->bytes_xfered;
3464 }
3465
3466 /**
3467  * irdma_poll_one - poll one entry of the CQ
3468  * @ukcq: ukcq to poll
3469  * @cur_cqe: current CQE info to be filled in
3470  * @entry: ibv_wc object to be filled for non-extended CQ or NULL for extended CQ
3471  *
3472  * Returns the internal irdma device error code or 0 on success
3473  */
3474 static inline int irdma_poll_one(struct irdma_cq_uk *ukcq,
3475                                  struct irdma_cq_poll_info *cur_cqe,
3476                                  struct ib_wc *entry)
3477 {
3478         int ret = irdma_uk_cq_poll_cmpl(ukcq, cur_cqe);
3479
3480         if (ret)
3481                 return ret;
3482
3483         irdma_process_cqe(entry, cur_cqe);
3484
3485         return 0;
3486 }
3487
3488 /**
3489  * __irdma_poll_cq - poll cq for completion (kernel apps)
3490  * @iwcq: cq to poll
3491  * @num_entries: number of entries to poll
3492  * @entry: wr of a completed entry
3493  */
3494 static int __irdma_poll_cq(struct irdma_cq *iwcq, int num_entries, struct ib_wc *entry)
3495 {
3496         struct list_head *tmp_node, *list_node;
3497         struct irdma_cq_buf *last_buf = NULL;
3498         struct irdma_cq_poll_info *cur_cqe = &iwcq->cur_cqe;
3499         struct irdma_cq_buf *cq_buf;
3500         enum irdma_status_code ret;
3501         struct irdma_device *iwdev;
3502         struct irdma_cq_uk *ukcq;
3503         bool cq_new_cqe = false;
3504         int resized_bufs = 0;
3505         int npolled = 0;
3506
3507         iwdev = to_iwdev(iwcq->ibcq.device);
3508         ukcq = &iwcq->sc_cq.cq_uk;
3509
3510         /* go through the list of previously resized CQ buffers */
3511         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
3512                 cq_buf = container_of(list_node, struct irdma_cq_buf, list);
3513                 while (npolled < num_entries) {
3514                         ret = irdma_poll_one(&cq_buf->cq_uk, cur_cqe, entry + npolled);
3515                         if (!ret) {
3516                                 ++npolled;
3517                                 cq_new_cqe = true;
3518                                 continue;
3519                         }
3520                         if (ret == IRDMA_ERR_Q_EMPTY)
3521                                 break;
3522                          /* QP using the CQ is destroyed. Skip reporting this CQE */
3523                         if (ret == IRDMA_ERR_Q_DESTROYED) {
3524                                 cq_new_cqe = true;
3525                                 continue;
3526                         }
3527                         goto error;
3528                 }
3529
3530                 /* save the resized CQ buffer which received the last cqe */
3531                 if (cq_new_cqe)
3532                         last_buf = cq_buf;
3533                 cq_new_cqe = false;
3534         }
3535
3536         /* check the current CQ for new cqes */
3537         while (npolled < num_entries) {
3538                 ret = irdma_poll_one(ukcq, cur_cqe, entry + npolled);
3539                 if (!ret) {
3540                         ++npolled;
3541                         cq_new_cqe = true;
3542                         continue;
3543                 }
3544
3545                 if (ret == IRDMA_ERR_Q_EMPTY)
3546                         break;
3547                 /* QP using the CQ is destroyed. Skip reporting this CQE */
3548                 if (ret == IRDMA_ERR_Q_DESTROYED) {
3549                         cq_new_cqe = true;
3550                         continue;
3551                 }
3552                 goto error;
3553         }
3554
3555         if (cq_new_cqe)
3556                 /* all previous CQ resizes are complete */
3557                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, NULL);
3558         else if (last_buf)
3559                 /* only CQ resizes up to the last_buf are complete */
3560                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, last_buf);
3561         if (resized_bufs)
3562                 /* report to the HW the number of complete CQ resizes */
3563                 irdma_uk_cq_set_resized_cnt(ukcq, resized_bufs);
3564
3565         return npolled;
3566 error:
3567         ibdev_dbg(&iwdev->ibdev, "%s: Error polling CQ, irdma_err: %d\n",
3568                   __func__, ret);
3569
3570         return -EINVAL;
3571 }
3572
3573 /**
3574  * irdma_poll_cq - poll cq for completion (kernel apps)
3575  * @ibcq: cq to poll
3576  * @num_entries: number of entries to poll
3577  * @entry: wr of a completed entry
3578  */
3579 static int irdma_poll_cq(struct ib_cq *ibcq, int num_entries,
3580                          struct ib_wc *entry)
3581 {
3582         struct irdma_cq *iwcq;
3583         unsigned long flags;
3584         int ret;
3585
3586         iwcq = to_iwcq(ibcq);
3587
3588         spin_lock_irqsave(&iwcq->lock, flags);
3589         ret = __irdma_poll_cq(iwcq, num_entries, entry);
3590         spin_unlock_irqrestore(&iwcq->lock, flags);
3591
3592         return ret;
3593 }
3594
3595 /**
3596  * irdma_req_notify_cq - arm cq kernel application
3597  * @ibcq: cq to arm
3598  * @notify_flags: notofication flags
3599  */
3600 static int irdma_req_notify_cq(struct ib_cq *ibcq,
3601                                enum ib_cq_notify_flags notify_flags)
3602 {
3603         struct irdma_cq *iwcq;
3604         struct irdma_cq_uk *ukcq;
3605         unsigned long flags;
3606         enum irdma_cmpl_notify cq_notify = IRDMA_CQ_COMPL_EVENT;
3607
3608         iwcq = to_iwcq(ibcq);
3609         ukcq = &iwcq->sc_cq.cq_uk;
3610         if (notify_flags == IB_CQ_SOLICITED)
3611                 cq_notify = IRDMA_CQ_COMPL_SOLICITED;
3612
3613         spin_lock_irqsave(&iwcq->lock, flags);
3614         irdma_uk_cq_request_notification(ukcq, cq_notify);
3615         spin_unlock_irqrestore(&iwcq->lock, flags);
3616
3617         return 0;
3618 }
3619
3620 static int irdma_roce_port_immutable(struct ib_device *ibdev, u32 port_num,
3621                                      struct ib_port_immutable *immutable)
3622 {
3623         struct ib_port_attr attr;
3624         int err;
3625
3626         immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
3627         err = ib_query_port(ibdev, port_num, &attr);
3628         if (err)
3629                 return err;
3630
3631         immutable->max_mad_size = IB_MGMT_MAD_SIZE;
3632         immutable->pkey_tbl_len = attr.pkey_tbl_len;
3633         immutable->gid_tbl_len = attr.gid_tbl_len;
3634
3635         return 0;
3636 }
3637
3638 static int irdma_iw_port_immutable(struct ib_device *ibdev, u32 port_num,
3639                                    struct ib_port_immutable *immutable)
3640 {
3641         struct ib_port_attr attr;
3642         int err;
3643
3644         immutable->core_cap_flags = RDMA_CORE_PORT_IWARP;
3645         err = ib_query_port(ibdev, port_num, &attr);
3646         if (err)
3647                 return err;
3648         immutable->gid_tbl_len = attr.gid_tbl_len;
3649
3650         return 0;
3651 }
3652
3653 static const char *const irdma_hw_stat_names[] = {
3654         /* 32bit names */
3655         [IRDMA_HW_STAT_INDEX_RXVLANERR] = "rxVlanErrors",
3656         [IRDMA_HW_STAT_INDEX_IP4RXDISCARD] = "ip4InDiscards",
3657         [IRDMA_HW_STAT_INDEX_IP4RXTRUNC] = "ip4InTruncatedPkts",
3658         [IRDMA_HW_STAT_INDEX_IP4TXNOROUTE] = "ip4OutNoRoutes",
3659         [IRDMA_HW_STAT_INDEX_IP6RXDISCARD] = "ip6InDiscards",
3660         [IRDMA_HW_STAT_INDEX_IP6RXTRUNC] = "ip6InTruncatedPkts",
3661         [IRDMA_HW_STAT_INDEX_IP6TXNOROUTE] = "ip6OutNoRoutes",
3662         [IRDMA_HW_STAT_INDEX_TCPRTXSEG] = "tcpRetransSegs",
3663         [IRDMA_HW_STAT_INDEX_TCPRXOPTERR] = "tcpInOptErrors",
3664         [IRDMA_HW_STAT_INDEX_TCPRXPROTOERR] = "tcpInProtoErrors",
3665         [IRDMA_HW_STAT_INDEX_RXRPCNPHANDLED] = "cnpHandled",
3666         [IRDMA_HW_STAT_INDEX_RXRPCNPIGNORED] = "cnpIgnored",
3667         [IRDMA_HW_STAT_INDEX_TXNPCNPSENT] = "cnpSent",
3668
3669         /* 64bit names */
3670         [IRDMA_HW_STAT_INDEX_IP4RXOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3671                 "ip4InOctets",
3672         [IRDMA_HW_STAT_INDEX_IP4RXPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3673                 "ip4InPkts",
3674         [IRDMA_HW_STAT_INDEX_IP4RXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32] =
3675                 "ip4InReasmRqd",
3676         [IRDMA_HW_STAT_INDEX_IP4RXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3677                 "ip4InMcastOctets",
3678         [IRDMA_HW_STAT_INDEX_IP4RXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3679                 "ip4InMcastPkts",
3680         [IRDMA_HW_STAT_INDEX_IP4TXOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3681                 "ip4OutOctets",
3682         [IRDMA_HW_STAT_INDEX_IP4TXPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3683                 "ip4OutPkts",
3684         [IRDMA_HW_STAT_INDEX_IP4TXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32] =
3685                 "ip4OutSegRqd",
3686         [IRDMA_HW_STAT_INDEX_IP4TXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3687                 "ip4OutMcastOctets",
3688         [IRDMA_HW_STAT_INDEX_IP4TXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3689                 "ip4OutMcastPkts",
3690         [IRDMA_HW_STAT_INDEX_IP6RXOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3691                 "ip6InOctets",
3692         [IRDMA_HW_STAT_INDEX_IP6RXPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3693                 "ip6InPkts",
3694         [IRDMA_HW_STAT_INDEX_IP6RXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32] =
3695                 "ip6InReasmRqd",
3696         [IRDMA_HW_STAT_INDEX_IP6RXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3697                 "ip6InMcastOctets",
3698         [IRDMA_HW_STAT_INDEX_IP6RXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3699                 "ip6InMcastPkts",
3700         [IRDMA_HW_STAT_INDEX_IP6TXOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3701                 "ip6OutOctets",
3702         [IRDMA_HW_STAT_INDEX_IP6TXPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3703                 "ip6OutPkts",
3704         [IRDMA_HW_STAT_INDEX_IP6TXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32] =
3705                 "ip6OutSegRqd",
3706         [IRDMA_HW_STAT_INDEX_IP6TXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3707                 "ip6OutMcastOctets",
3708         [IRDMA_HW_STAT_INDEX_IP6TXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3709                 "ip6OutMcastPkts",
3710         [IRDMA_HW_STAT_INDEX_TCPRXSEGS + IRDMA_HW_STAT_INDEX_MAX_32] =
3711                 "tcpInSegs",
3712         [IRDMA_HW_STAT_INDEX_TCPTXSEG + IRDMA_HW_STAT_INDEX_MAX_32] =
3713                 "tcpOutSegs",
3714         [IRDMA_HW_STAT_INDEX_RDMARXRDS + IRDMA_HW_STAT_INDEX_MAX_32] =
3715                 "iwInRdmaReads",
3716         [IRDMA_HW_STAT_INDEX_RDMARXSNDS + IRDMA_HW_STAT_INDEX_MAX_32] =
3717                 "iwInRdmaSends",
3718         [IRDMA_HW_STAT_INDEX_RDMARXWRS + IRDMA_HW_STAT_INDEX_MAX_32] =
3719                 "iwInRdmaWrites",
3720         [IRDMA_HW_STAT_INDEX_RDMATXRDS + IRDMA_HW_STAT_INDEX_MAX_32] =
3721                 "iwOutRdmaReads",
3722         [IRDMA_HW_STAT_INDEX_RDMATXSNDS + IRDMA_HW_STAT_INDEX_MAX_32] =
3723                 "iwOutRdmaSends",
3724         [IRDMA_HW_STAT_INDEX_RDMATXWRS + IRDMA_HW_STAT_INDEX_MAX_32] =
3725                 "iwOutRdmaWrites",
3726         [IRDMA_HW_STAT_INDEX_RDMAVBND + IRDMA_HW_STAT_INDEX_MAX_32] =
3727                 "iwRdmaBnd",
3728         [IRDMA_HW_STAT_INDEX_RDMAVINV + IRDMA_HW_STAT_INDEX_MAX_32] =
3729                 "iwRdmaInv",
3730         [IRDMA_HW_STAT_INDEX_UDPRXPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3731                 "RxUDP",
3732         [IRDMA_HW_STAT_INDEX_UDPTXPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3733                 "TxUDP",
3734         [IRDMA_HW_STAT_INDEX_RXNPECNMARKEDPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3735                 "RxECNMrkd",
3736 };
3737
3738 static void irdma_get_dev_fw_str(struct ib_device *dev, char *str)
3739 {
3740         struct irdma_device *iwdev = to_iwdev(dev);
3741
3742         snprintf(str, IB_FW_VERSION_NAME_MAX, "%u.%u",
3743                  irdma_fw_major_ver(&iwdev->rf->sc_dev),
3744                  irdma_fw_minor_ver(&iwdev->rf->sc_dev));
3745 }
3746
3747 /**
3748  * irdma_alloc_hw_port_stats - Allocate a hw stats structure
3749  * @ibdev: device pointer from stack
3750  * @port_num: port number
3751  */
3752 static struct rdma_hw_stats *irdma_alloc_hw_port_stats(struct ib_device *ibdev,
3753                                                        u32 port_num)
3754 {
3755         int num_counters = IRDMA_HW_STAT_INDEX_MAX_32 +
3756                            IRDMA_HW_STAT_INDEX_MAX_64;
3757         unsigned long lifespan = RDMA_HW_STATS_DEFAULT_LIFESPAN;
3758
3759         BUILD_BUG_ON(ARRAY_SIZE(irdma_hw_stat_names) !=
3760                      (IRDMA_HW_STAT_INDEX_MAX_32 + IRDMA_HW_STAT_INDEX_MAX_64));
3761
3762         return rdma_alloc_hw_stats_struct(irdma_hw_stat_names, num_counters,
3763                                           lifespan);
3764 }
3765
3766 /**
3767  * irdma_get_hw_stats - Populates the rdma_hw_stats structure
3768  * @ibdev: device pointer from stack
3769  * @stats: stats pointer from stack
3770  * @port_num: port number
3771  * @index: which hw counter the stack is requesting we update
3772  */
3773 static int irdma_get_hw_stats(struct ib_device *ibdev,
3774                               struct rdma_hw_stats *stats, u32 port_num,
3775                               int index)
3776 {
3777         struct irdma_device *iwdev = to_iwdev(ibdev);
3778         struct irdma_dev_hw_stats *hw_stats = &iwdev->vsi.pestat->hw_stats;
3779
3780         if (iwdev->rf->rdma_ver >= IRDMA_GEN_2)
3781                 irdma_cqp_gather_stats_cmd(&iwdev->rf->sc_dev, iwdev->vsi.pestat, true);
3782         else
3783                 irdma_cqp_gather_stats_gen1(&iwdev->rf->sc_dev, iwdev->vsi.pestat);
3784
3785         memcpy(&stats->value[0], hw_stats, sizeof(*hw_stats));
3786
3787         return stats->num_counters;
3788 }
3789
3790 /**
3791  * irdma_query_gid - Query port GID
3792  * @ibdev: device pointer from stack
3793  * @port: port number
3794  * @index: Entry index
3795  * @gid: Global ID
3796  */
3797 static int irdma_query_gid(struct ib_device *ibdev, u32 port, int index,
3798                            union ib_gid *gid)
3799 {
3800         struct irdma_device *iwdev = to_iwdev(ibdev);
3801
3802         memset(gid->raw, 0, sizeof(gid->raw));
3803         ether_addr_copy(gid->raw, iwdev->netdev->dev_addr);
3804
3805         return 0;
3806 }
3807
3808 /**
3809  * mcast_list_add -  Add a new mcast item to list
3810  * @rf: RDMA PCI function
3811  * @new_elem: pointer to element to add
3812  */
3813 static void mcast_list_add(struct irdma_pci_f *rf,
3814                            struct mc_table_list *new_elem)
3815 {
3816         list_add(&new_elem->list, &rf->mc_qht_list.list);
3817 }
3818
3819 /**
3820  * mcast_list_del - Remove an mcast item from list
3821  * @mc_qht_elem: pointer to mcast table list element
3822  */
3823 static void mcast_list_del(struct mc_table_list *mc_qht_elem)
3824 {
3825         if (mc_qht_elem)
3826                 list_del(&mc_qht_elem->list);
3827 }
3828
3829 /**
3830  * mcast_list_lookup_ip - Search mcast list for address
3831  * @rf: RDMA PCI function
3832  * @ip_mcast: pointer to mcast IP address
3833  */
3834 static struct mc_table_list *mcast_list_lookup_ip(struct irdma_pci_f *rf,
3835                                                   u32 *ip_mcast)
3836 {
3837         struct mc_table_list *mc_qht_el;
3838         struct list_head *pos, *q;
3839
3840         list_for_each_safe (pos, q, &rf->mc_qht_list.list) {
3841                 mc_qht_el = list_entry(pos, struct mc_table_list, list);
3842                 if (!memcmp(mc_qht_el->mc_info.dest_ip, ip_mcast,
3843                             sizeof(mc_qht_el->mc_info.dest_ip)))
3844                         return mc_qht_el;
3845         }
3846
3847         return NULL;
3848 }
3849
3850 /**
3851  * irdma_mcast_cqp_op - perform a mcast cqp operation
3852  * @iwdev: irdma device
3853  * @mc_grp_ctx: mcast group info
3854  * @op: operation
3855  *
3856  * returns error status
3857  */
3858 static int irdma_mcast_cqp_op(struct irdma_device *iwdev,
3859                               struct irdma_mcast_grp_info *mc_grp_ctx, u8 op)
3860 {
3861         struct cqp_cmds_info *cqp_info;
3862         struct irdma_cqp_request *cqp_request;
3863         enum irdma_status_code status;
3864
3865         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
3866         if (!cqp_request)
3867                 return -ENOMEM;
3868
3869         cqp_request->info.in.u.mc_create.info = *mc_grp_ctx;
3870         cqp_info = &cqp_request->info;
3871         cqp_info->cqp_cmd = op;
3872         cqp_info->post_sq = 1;
3873         cqp_info->in.u.mc_create.scratch = (uintptr_t)cqp_request;
3874         cqp_info->in.u.mc_create.cqp = &iwdev->rf->cqp.sc_cqp;
3875         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
3876         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
3877         if (status)
3878                 return -ENOMEM;
3879
3880         return 0;
3881 }
3882
3883 /**
3884  * irdma_mcast_mac - Get the multicast MAC for an IP address
3885  * @ip_addr: IPv4 or IPv6 address
3886  * @mac: pointer to result MAC address
3887  * @ipv4: flag indicating IPv4 or IPv6
3888  *
3889  */
3890 void irdma_mcast_mac(u32 *ip_addr, u8 *mac, bool ipv4)
3891 {
3892         u8 *ip = (u8 *)ip_addr;
3893
3894         if (ipv4) {
3895                 unsigned char mac4[ETH_ALEN] = {0x01, 0x00, 0x5E, 0x00,
3896                                                 0x00, 0x00};
3897
3898                 mac4[3] = ip[2] & 0x7F;
3899                 mac4[4] = ip[1];
3900                 mac4[5] = ip[0];
3901                 ether_addr_copy(mac, mac4);
3902         } else {
3903                 unsigned char mac6[ETH_ALEN] = {0x33, 0x33, 0x00, 0x00,
3904                                                 0x00, 0x00};
3905
3906                 mac6[2] = ip[3];
3907                 mac6[3] = ip[2];
3908                 mac6[4] = ip[1];
3909                 mac6[5] = ip[0];
3910                 ether_addr_copy(mac, mac6);
3911         }
3912 }
3913
3914 /**
3915  * irdma_attach_mcast - attach a qp to a multicast group
3916  * @ibqp: ptr to qp
3917  * @ibgid: pointer to global ID
3918  * @lid: local ID
3919  *
3920  * returns error status
3921  */
3922 static int irdma_attach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
3923 {
3924         struct irdma_qp *iwqp = to_iwqp(ibqp);
3925         struct irdma_device *iwdev = iwqp->iwdev;
3926         struct irdma_pci_f *rf = iwdev->rf;
3927         struct mc_table_list *mc_qht_elem;
3928         struct irdma_mcast_grp_ctx_entry_info mcg_info = {};
3929         unsigned long flags;
3930         u32 ip_addr[4] = {};
3931         u32 mgn;
3932         u32 no_mgs;
3933         int ret = 0;
3934         bool ipv4;
3935         u16 vlan_id;
3936         union {
3937                 struct sockaddr saddr;
3938                 struct sockaddr_in saddr_in;
3939                 struct sockaddr_in6 saddr_in6;
3940         } sgid_addr;
3941         unsigned char dmac[ETH_ALEN];
3942
3943         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
3944
3945         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid)) {
3946                 irdma_copy_ip_ntohl(ip_addr,
3947                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
3948                 irdma_netdev_vlan_ipv6(ip_addr, &vlan_id, NULL);
3949                 ipv4 = false;
3950                 ibdev_dbg(&iwdev->ibdev,
3951                           "VERBS: qp_id=%d, IP6address=%pI6\n", ibqp->qp_num,
3952                           ip_addr);
3953                 irdma_mcast_mac(ip_addr, dmac, false);
3954         } else {
3955                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
3956                 ipv4 = true;
3957                 vlan_id = irdma_get_vlan_ipv4(ip_addr);
3958                 irdma_mcast_mac(ip_addr, dmac, true);
3959                 ibdev_dbg(&iwdev->ibdev,
3960                           "VERBS: qp_id=%d, IP4address=%pI4, MAC=%pM\n",
3961                           ibqp->qp_num, ip_addr, dmac);
3962         }
3963
3964         spin_lock_irqsave(&rf->qh_list_lock, flags);
3965         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
3966         if (!mc_qht_elem) {
3967                 struct irdma_dma_mem *dma_mem_mc;
3968
3969                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3970                 mc_qht_elem = kzalloc(sizeof(*mc_qht_elem), GFP_KERNEL);
3971                 if (!mc_qht_elem)
3972                         return -ENOMEM;
3973
3974                 mc_qht_elem->mc_info.ipv4_valid = ipv4;
3975                 memcpy(mc_qht_elem->mc_info.dest_ip, ip_addr,
3976                        sizeof(mc_qht_elem->mc_info.dest_ip));
3977                 ret = irdma_alloc_rsrc(rf, rf->allocated_mcgs, rf->max_mcg,
3978                                        &mgn, &rf->next_mcg);
3979                 if (ret) {
3980                         kfree(mc_qht_elem);
3981                         return -ENOMEM;
3982                 }
3983
3984                 mc_qht_elem->mc_info.mgn = mgn;
3985                 dma_mem_mc = &mc_qht_elem->mc_grp_ctx.dma_mem_mc;
3986                 dma_mem_mc->size = ALIGN(sizeof(u64) * IRDMA_MAX_MGS_PER_CTX,
3987                                          IRDMA_HW_PAGE_SIZE);
3988                 dma_mem_mc->va = dma_alloc_coherent(rf->hw.device,
3989                                                     dma_mem_mc->size,
3990                                                     &dma_mem_mc->pa,
3991                                                     GFP_KERNEL);
3992                 if (!dma_mem_mc->va) {
3993                         irdma_free_rsrc(rf, rf->allocated_mcgs, mgn);
3994                         kfree(mc_qht_elem);
3995                         return -ENOMEM;
3996                 }
3997
3998                 mc_qht_elem->mc_grp_ctx.mg_id = (u16)mgn;
3999                 memcpy(mc_qht_elem->mc_grp_ctx.dest_ip_addr, ip_addr,
4000                        sizeof(mc_qht_elem->mc_grp_ctx.dest_ip_addr));
4001                 mc_qht_elem->mc_grp_ctx.ipv4_valid = ipv4;
4002                 mc_qht_elem->mc_grp_ctx.vlan_id = vlan_id;
4003                 if (vlan_id < VLAN_N_VID)
4004                         mc_qht_elem->mc_grp_ctx.vlan_valid = true;
4005                 mc_qht_elem->mc_grp_ctx.hmc_fcn_id = iwdev->vsi.fcn_id;
4006                 mc_qht_elem->mc_grp_ctx.qs_handle =
4007                         iwqp->sc_qp.vsi->qos[iwqp->sc_qp.user_pri].qs_handle;
4008                 ether_addr_copy(mc_qht_elem->mc_grp_ctx.dest_mac_addr, dmac);
4009
4010                 spin_lock_irqsave(&rf->qh_list_lock, flags);
4011                 mcast_list_add(rf, mc_qht_elem);
4012         } else {
4013                 if (mc_qht_elem->mc_grp_ctx.no_of_mgs ==
4014                     IRDMA_MAX_MGS_PER_CTX) {
4015                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4016                         return -ENOMEM;
4017                 }
4018         }
4019
4020         mcg_info.qp_id = iwqp->ibqp.qp_num;
4021         no_mgs = mc_qht_elem->mc_grp_ctx.no_of_mgs;
4022         irdma_sc_add_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
4023         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4024
4025         /* Only if there is a change do we need to modify or create */
4026         if (!no_mgs) {
4027                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4028                                          IRDMA_OP_MC_CREATE);
4029         } else if (no_mgs != mc_qht_elem->mc_grp_ctx.no_of_mgs) {
4030                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4031                                          IRDMA_OP_MC_MODIFY);
4032         } else {
4033                 return 0;
4034         }
4035
4036         if (ret)
4037                 goto error;
4038
4039         return 0;
4040
4041 error:
4042         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
4043         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
4044                 mcast_list_del(mc_qht_elem);
4045                 dma_free_coherent(rf->hw.device,
4046                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.size,
4047                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.va,
4048                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.pa);
4049                 mc_qht_elem->mc_grp_ctx.dma_mem_mc.va = NULL;
4050                 irdma_free_rsrc(rf, rf->allocated_mcgs,
4051                                 mc_qht_elem->mc_grp_ctx.mg_id);
4052                 kfree(mc_qht_elem);
4053         }
4054
4055         return ret;
4056 }
4057
4058 /**
4059  * irdma_detach_mcast - detach a qp from a multicast group
4060  * @ibqp: ptr to qp
4061  * @ibgid: pointer to global ID
4062  * @lid: local ID
4063  *
4064  * returns error status
4065  */
4066 static int irdma_detach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
4067 {
4068         struct irdma_qp *iwqp = to_iwqp(ibqp);
4069         struct irdma_device *iwdev = iwqp->iwdev;
4070         struct irdma_pci_f *rf = iwdev->rf;
4071         u32 ip_addr[4] = {};
4072         struct mc_table_list *mc_qht_elem;
4073         struct irdma_mcast_grp_ctx_entry_info mcg_info = {};
4074         int ret;
4075         unsigned long flags;
4076         union {
4077                 struct sockaddr saddr;
4078                 struct sockaddr_in saddr_in;
4079                 struct sockaddr_in6 saddr_in6;
4080         } sgid_addr;
4081
4082         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
4083         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid))
4084                 irdma_copy_ip_ntohl(ip_addr,
4085                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4086         else
4087                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
4088
4089         spin_lock_irqsave(&rf->qh_list_lock, flags);
4090         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
4091         if (!mc_qht_elem) {
4092                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4093                 ibdev_dbg(&iwdev->ibdev,
4094                           "VERBS: address not found MCG\n");
4095                 return 0;
4096         }
4097
4098         mcg_info.qp_id = iwqp->ibqp.qp_num;
4099         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
4100         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
4101                 mcast_list_del(mc_qht_elem);
4102                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4103                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4104                                          IRDMA_OP_MC_DESTROY);
4105                 if (ret) {
4106                         ibdev_dbg(&iwdev->ibdev,
4107                                   "VERBS: failed MC_DESTROY MCG\n");
4108                         spin_lock_irqsave(&rf->qh_list_lock, flags);
4109                         mcast_list_add(rf, mc_qht_elem);
4110                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4111                         return -EAGAIN;
4112                 }
4113
4114                 dma_free_coherent(rf->hw.device,
4115                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.size,
4116                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.va,
4117                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.pa);
4118                 mc_qht_elem->mc_grp_ctx.dma_mem_mc.va = NULL;
4119                 irdma_free_rsrc(rf, rf->allocated_mcgs,
4120                                 mc_qht_elem->mc_grp_ctx.mg_id);
4121                 kfree(mc_qht_elem);
4122         } else {
4123                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4124                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4125                                          IRDMA_OP_MC_MODIFY);
4126                 if (ret) {
4127                         ibdev_dbg(&iwdev->ibdev,
4128                                   "VERBS: failed Modify MCG\n");
4129                         return ret;
4130                 }
4131         }
4132
4133         return 0;
4134 }
4135
4136 /**
4137  * irdma_create_ah - create address handle
4138  * @ibah: address handle
4139  * @attr: address handle attributes
4140  * @udata: User data
4141  *
4142  * returns 0 on success, error otherwise
4143  */
4144 static int irdma_create_ah(struct ib_ah *ibah,
4145                            struct rdma_ah_init_attr *attr,
4146                            struct ib_udata *udata)
4147 {
4148         struct irdma_pd *pd = to_iwpd(ibah->pd);
4149         struct irdma_ah *ah = container_of(ibah, struct irdma_ah, ibah);
4150         struct rdma_ah_attr *ah_attr = attr->ah_attr;
4151         const struct ib_gid_attr *sgid_attr;
4152         struct irdma_device *iwdev = to_iwdev(ibah->pd->device);
4153         struct irdma_pci_f *rf = iwdev->rf;
4154         struct irdma_sc_ah *sc_ah;
4155         u32 ah_id = 0;
4156         struct irdma_ah_info *ah_info;
4157         struct irdma_create_ah_resp uresp;
4158         union {
4159                 struct sockaddr saddr;
4160                 struct sockaddr_in saddr_in;
4161                 struct sockaddr_in6 saddr_in6;
4162         } sgid_addr, dgid_addr;
4163         int err;
4164         u8 dmac[ETH_ALEN];
4165
4166         err = irdma_alloc_rsrc(rf, rf->allocated_ahs, rf->max_ah, &ah_id,
4167                                &rf->next_ah);
4168         if (err)
4169                 return err;
4170
4171         ah->pd = pd;
4172         sc_ah = &ah->sc_ah;
4173         sc_ah->ah_info.ah_idx = ah_id;
4174         sc_ah->ah_info.vsi = &iwdev->vsi;
4175         irdma_sc_init_ah(&rf->sc_dev, sc_ah);
4176         ah->sgid_index = ah_attr->grh.sgid_index;
4177         sgid_attr = ah_attr->grh.sgid_attr;
4178         memcpy(&ah->dgid, &ah_attr->grh.dgid, sizeof(ah->dgid));
4179         rdma_gid2ip((struct sockaddr *)&sgid_addr, &sgid_attr->gid);
4180         rdma_gid2ip((struct sockaddr *)&dgid_addr, &ah_attr->grh.dgid);
4181         ah->av.attrs = *ah_attr;
4182         ah->av.net_type = rdma_gid_attr_network_type(sgid_attr);
4183         ah->av.sgid_addr.saddr = sgid_addr.saddr;
4184         ah->av.dgid_addr.saddr = dgid_addr.saddr;
4185         ah_info = &sc_ah->ah_info;
4186         ah_info->ah_idx = ah_id;
4187         ah_info->pd_idx = pd->sc_pd.pd_id;
4188         if (ah_attr->ah_flags & IB_AH_GRH) {
4189                 ah_info->flow_label = ah_attr->grh.flow_label;
4190                 ah_info->hop_ttl = ah_attr->grh.hop_limit;
4191                 ah_info->tc_tos = ah_attr->grh.traffic_class;
4192         }
4193
4194         ether_addr_copy(dmac, ah_attr->roce.dmac);
4195         if (rdma_gid_attr_network_type(sgid_attr) == RDMA_NETWORK_IPV4) {
4196                 ah_info->ipv4_valid = true;
4197                 ah_info->dest_ip_addr[0] =
4198                         ntohl(dgid_addr.saddr_in.sin_addr.s_addr);
4199                 ah_info->src_ip_addr[0] =
4200                         ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
4201                 ah_info->do_lpbk = irdma_ipv4_is_lpb(ah_info->src_ip_addr[0],
4202                                                      ah_info->dest_ip_addr[0]);
4203                 if (ipv4_is_multicast(dgid_addr.saddr_in.sin_addr.s_addr)) {
4204                         ah_info->do_lpbk = true;
4205                         irdma_mcast_mac(ah_info->dest_ip_addr, dmac, true);
4206                 }
4207         } else {
4208                 irdma_copy_ip_ntohl(ah_info->dest_ip_addr,
4209                                     dgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4210                 irdma_copy_ip_ntohl(ah_info->src_ip_addr,
4211                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4212                 ah_info->do_lpbk = irdma_ipv6_is_lpb(ah_info->src_ip_addr,
4213                                                      ah_info->dest_ip_addr);
4214                 if (rdma_is_multicast_addr(&dgid_addr.saddr_in6.sin6_addr)) {
4215                         ah_info->do_lpbk = true;
4216                         irdma_mcast_mac(ah_info->dest_ip_addr, dmac, false);
4217                 }
4218         }
4219
4220         err = rdma_read_gid_l2_fields(sgid_attr, &ah_info->vlan_tag,
4221                                       ah_info->mac_addr);
4222         if (err)
4223                 goto error;
4224
4225         ah_info->dst_arpindex = irdma_add_arp(iwdev->rf, ah_info->dest_ip_addr,
4226                                               ah_info->ipv4_valid, dmac);
4227
4228         if (ah_info->dst_arpindex == -1) {
4229                 err = -EINVAL;
4230                 goto error;
4231         }
4232
4233         if (ah_info->vlan_tag >= VLAN_N_VID && iwdev->dcb)
4234                 ah_info->vlan_tag = 0;
4235
4236         if (ah_info->vlan_tag < VLAN_N_VID) {
4237                 ah_info->insert_vlan_tag = true;
4238                 ah_info->vlan_tag |=
4239                         rt_tos2priority(ah_info->tc_tos) << VLAN_PRIO_SHIFT;
4240         }
4241
4242         err = irdma_ah_cqp_op(iwdev->rf, sc_ah, IRDMA_OP_AH_CREATE,
4243                               attr->flags & RDMA_CREATE_AH_SLEEPABLE,
4244                               irdma_gsi_ud_qp_ah_cb, sc_ah);
4245
4246         if (err) {
4247                 ibdev_dbg(&iwdev->ibdev,
4248                           "VERBS: CQP-OP Create AH fail");
4249                 goto error;
4250         }
4251
4252         if (!(attr->flags & RDMA_CREATE_AH_SLEEPABLE)) {
4253                 int cnt = CQP_COMPL_WAIT_TIME_MS * CQP_TIMEOUT_THRESHOLD;
4254
4255                 do {
4256                         irdma_cqp_ce_handler(rf, &rf->ccq.sc_cq);
4257                         mdelay(1);
4258                 } while (!sc_ah->ah_info.ah_valid && --cnt);
4259
4260                 if (!cnt) {
4261                         ibdev_dbg(&iwdev->ibdev,
4262                                   "VERBS: CQP create AH timed out");
4263                         err = -ETIMEDOUT;
4264                         goto error;
4265                 }
4266         }
4267
4268         if (udata) {
4269                 uresp.ah_id = ah->sc_ah.ah_info.ah_idx;
4270                 err = ib_copy_to_udata(udata, &uresp,
4271                                        min(sizeof(uresp), udata->outlen));
4272         }
4273         return 0;
4274
4275 error:
4276         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_ahs, ah_id);
4277
4278         return err;
4279 }
4280
4281 /**
4282  * irdma_destroy_ah - Destroy address handle
4283  * @ibah: pointer to address handle
4284  * @ah_flags: flags for sleepable
4285  */
4286 static int irdma_destroy_ah(struct ib_ah *ibah, u32 ah_flags)
4287 {
4288         struct irdma_device *iwdev = to_iwdev(ibah->device);
4289         struct irdma_ah *ah = to_iwah(ibah);
4290
4291         irdma_ah_cqp_op(iwdev->rf, &ah->sc_ah, IRDMA_OP_AH_DESTROY,
4292                         false, NULL, ah);
4293
4294         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_ahs,
4295                         ah->sc_ah.ah_info.ah_idx);
4296
4297         return 0;
4298 }
4299
4300 /**
4301  * irdma_query_ah - Query address handle
4302  * @ibah: pointer to address handle
4303  * @ah_attr: address handle attributes
4304  */
4305 static int irdma_query_ah(struct ib_ah *ibah, struct rdma_ah_attr *ah_attr)
4306 {
4307         struct irdma_ah *ah = to_iwah(ibah);
4308
4309         memset(ah_attr, 0, sizeof(*ah_attr));
4310         if (ah->av.attrs.ah_flags & IB_AH_GRH) {
4311                 ah_attr->ah_flags = IB_AH_GRH;
4312                 ah_attr->grh.flow_label = ah->sc_ah.ah_info.flow_label;
4313                 ah_attr->grh.traffic_class = ah->sc_ah.ah_info.tc_tos;
4314                 ah_attr->grh.hop_limit = ah->sc_ah.ah_info.hop_ttl;
4315                 ah_attr->grh.sgid_index = ah->sgid_index;
4316                 ah_attr->grh.sgid_index = ah->sgid_index;
4317                 memcpy(&ah_attr->grh.dgid, &ah->dgid,
4318                        sizeof(ah_attr->grh.dgid));
4319         }
4320
4321         return 0;
4322 }
4323
4324 static enum rdma_link_layer irdma_get_link_layer(struct ib_device *ibdev,
4325                                                  u32 port_num)
4326 {
4327         return IB_LINK_LAYER_ETHERNET;
4328 }
4329
4330 static __be64 irdma_mac_to_guid(struct net_device *ndev)
4331 {
4332         unsigned char *mac = ndev->dev_addr;
4333         __be64 guid;
4334         unsigned char *dst = (unsigned char *)&guid;
4335
4336         dst[0] = mac[0] ^ 2;
4337         dst[1] = mac[1];
4338         dst[2] = mac[2];
4339         dst[3] = 0xff;
4340         dst[4] = 0xfe;
4341         dst[5] = mac[3];
4342         dst[6] = mac[4];
4343         dst[7] = mac[5];
4344
4345         return guid;
4346 }
4347
4348 static const struct ib_device_ops irdma_roce_dev_ops = {
4349         .attach_mcast = irdma_attach_mcast,
4350         .create_ah = irdma_create_ah,
4351         .create_user_ah = irdma_create_ah,
4352         .destroy_ah = irdma_destroy_ah,
4353         .detach_mcast = irdma_detach_mcast,
4354         .get_link_layer = irdma_get_link_layer,
4355         .get_port_immutable = irdma_roce_port_immutable,
4356         .modify_qp = irdma_modify_qp_roce,
4357         .query_ah = irdma_query_ah,
4358         .query_pkey = irdma_query_pkey,
4359 };
4360
4361 static const struct ib_device_ops irdma_iw_dev_ops = {
4362         .modify_qp = irdma_modify_qp,
4363         .get_port_immutable = irdma_iw_port_immutable,
4364         .query_gid = irdma_query_gid,
4365 };
4366
4367 static const struct ib_device_ops irdma_dev_ops = {
4368         .owner = THIS_MODULE,
4369         .driver_id = RDMA_DRIVER_IRDMA,
4370         .uverbs_abi_ver = IRDMA_ABI_VER,
4371
4372         .alloc_hw_port_stats = irdma_alloc_hw_port_stats,
4373         .alloc_mr = irdma_alloc_mr,
4374         .alloc_mw = irdma_alloc_mw,
4375         .alloc_pd = irdma_alloc_pd,
4376         .alloc_ucontext = irdma_alloc_ucontext,
4377         .create_cq = irdma_create_cq,
4378         .create_qp = irdma_create_qp,
4379         .dealloc_driver = irdma_ib_dealloc_device,
4380         .dealloc_mw = irdma_dealloc_mw,
4381         .dealloc_pd = irdma_dealloc_pd,
4382         .dealloc_ucontext = irdma_dealloc_ucontext,
4383         .dereg_mr = irdma_dereg_mr,
4384         .destroy_cq = irdma_destroy_cq,
4385         .destroy_qp = irdma_destroy_qp,
4386         .disassociate_ucontext = irdma_disassociate_ucontext,
4387         .get_dev_fw_str = irdma_get_dev_fw_str,
4388         .get_dma_mr = irdma_get_dma_mr,
4389         .get_hw_stats = irdma_get_hw_stats,
4390         .map_mr_sg = irdma_map_mr_sg,
4391         .mmap = irdma_mmap,
4392         .mmap_free = irdma_mmap_free,
4393         .poll_cq = irdma_poll_cq,
4394         .post_recv = irdma_post_recv,
4395         .post_send = irdma_post_send,
4396         .query_device = irdma_query_device,
4397         .query_port = irdma_query_port,
4398         .query_qp = irdma_query_qp,
4399         .reg_user_mr = irdma_reg_user_mr,
4400         .req_notify_cq = irdma_req_notify_cq,
4401         .resize_cq = irdma_resize_cq,
4402         INIT_RDMA_OBJ_SIZE(ib_pd, irdma_pd, ibpd),
4403         INIT_RDMA_OBJ_SIZE(ib_ucontext, irdma_ucontext, ibucontext),
4404         INIT_RDMA_OBJ_SIZE(ib_ah, irdma_ah, ibah),
4405         INIT_RDMA_OBJ_SIZE(ib_cq, irdma_cq, ibcq),
4406         INIT_RDMA_OBJ_SIZE(ib_mw, irdma_mr, ibmw),
4407 };
4408
4409 /**
4410  * irdma_init_roce_device - initialization of roce rdma device
4411  * @iwdev: irdma device
4412  */
4413 static void irdma_init_roce_device(struct irdma_device *iwdev)
4414 {
4415         iwdev->ibdev.node_type = RDMA_NODE_IB_CA;
4416         iwdev->ibdev.node_guid = irdma_mac_to_guid(iwdev->netdev);
4417         ib_set_device_ops(&iwdev->ibdev, &irdma_roce_dev_ops);
4418 }
4419
4420 /**
4421  * irdma_init_iw_device - initialization of iwarp rdma device
4422  * @iwdev: irdma device
4423  */
4424 static int irdma_init_iw_device(struct irdma_device *iwdev)
4425 {
4426         struct net_device *netdev = iwdev->netdev;
4427
4428         iwdev->ibdev.node_type = RDMA_NODE_RNIC;
4429         ether_addr_copy((u8 *)&iwdev->ibdev.node_guid, netdev->dev_addr);
4430         iwdev->ibdev.ops.iw_add_ref = irdma_qp_add_ref;
4431         iwdev->ibdev.ops.iw_rem_ref = irdma_qp_rem_ref;
4432         iwdev->ibdev.ops.iw_get_qp = irdma_get_qp;
4433         iwdev->ibdev.ops.iw_connect = irdma_connect;
4434         iwdev->ibdev.ops.iw_accept = irdma_accept;
4435         iwdev->ibdev.ops.iw_reject = irdma_reject;
4436         iwdev->ibdev.ops.iw_create_listen = irdma_create_listen;
4437         iwdev->ibdev.ops.iw_destroy_listen = irdma_destroy_listen;
4438         memcpy(iwdev->ibdev.iw_ifname, netdev->name,
4439                sizeof(iwdev->ibdev.iw_ifname));
4440         ib_set_device_ops(&iwdev->ibdev, &irdma_iw_dev_ops);
4441
4442         return 0;
4443 }
4444
4445 /**
4446  * irdma_init_rdma_device - initialization of rdma device
4447  * @iwdev: irdma device
4448  */
4449 static int irdma_init_rdma_device(struct irdma_device *iwdev)
4450 {
4451         struct pci_dev *pcidev = iwdev->rf->pcidev;
4452         int ret;
4453
4454         if (iwdev->roce_mode) {
4455                 irdma_init_roce_device(iwdev);
4456         } else {
4457                 ret = irdma_init_iw_device(iwdev);
4458                 if (ret)
4459                         return ret;
4460         }
4461         iwdev->ibdev.phys_port_cnt = 1;
4462         iwdev->ibdev.num_comp_vectors = iwdev->rf->ceqs_count;
4463         iwdev->ibdev.dev.parent = &pcidev->dev;
4464         ib_set_device_ops(&iwdev->ibdev, &irdma_dev_ops);
4465
4466         return 0;
4467 }
4468
4469 /**
4470  * irdma_port_ibevent - indicate port event
4471  * @iwdev: irdma device
4472  */
4473 void irdma_port_ibevent(struct irdma_device *iwdev)
4474 {
4475         struct ib_event event;
4476
4477         event.device = &iwdev->ibdev;
4478         event.element.port_num = 1;
4479         event.event =
4480                 iwdev->iw_status ? IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
4481         ib_dispatch_event(&event);
4482 }
4483
4484 /**
4485  * irdma_ib_unregister_device - unregister rdma device from IB
4486  * core
4487  * @iwdev: irdma device
4488  */
4489 void irdma_ib_unregister_device(struct irdma_device *iwdev)
4490 {
4491         iwdev->iw_status = 0;
4492         irdma_port_ibevent(iwdev);
4493         ib_unregister_device(&iwdev->ibdev);
4494 }
4495
4496 /**
4497  * irdma_ib_register_device - register irdma device to IB core
4498  * @iwdev: irdma device
4499  */
4500 int irdma_ib_register_device(struct irdma_device *iwdev)
4501 {
4502         int ret;
4503
4504         ret = irdma_init_rdma_device(iwdev);
4505         if (ret)
4506                 return ret;
4507
4508         ret = ib_device_set_netdev(&iwdev->ibdev, iwdev->netdev, 1);
4509         if (ret)
4510                 goto error;
4511         dma_set_max_seg_size(iwdev->rf->hw.device, UINT_MAX);
4512         ret = ib_register_device(&iwdev->ibdev, "irdma%d", iwdev->rf->hw.device);
4513         if (ret)
4514                 goto error;
4515
4516         iwdev->iw_status = 1;
4517         irdma_port_ibevent(iwdev);
4518
4519         return 0;
4520
4521 error:
4522         if (ret)
4523                 ibdev_dbg(&iwdev->ibdev, "VERBS: Register RDMA device fail\n");
4524
4525         return ret;
4526 }
4527
4528 /**
4529  * irdma_ib_dealloc_device
4530  * @ibdev: ib device
4531  *
4532  * callback from ibdev dealloc_driver to deallocate resources
4533  * unber irdma device
4534  */
4535 void irdma_ib_dealloc_device(struct ib_device *ibdev)
4536 {
4537         struct irdma_device *iwdev = to_iwdev(ibdev);
4538
4539         irdma_rt_deinit_hw(iwdev);
4540         irdma_ctrl_deinit_hw(iwdev->rf);
4541         kfree(iwdev->rf);
4542 }