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