drivers/infiniband/hw/irdma/verbs.c

   1 // SPDX-License-Identifier: GPL-2.0 or Linux-OpenIB
   2 /* Copyright (c) 2015 - 2021 Intel Corporation */
   3 #include "main.h"
   4
   5 /**
   6  * irdma_query_device - get device attributes
   7  * @ibdev: device pointer from stack
   8  * @props: returning device attributes
   9  * @udata: user data
  10  */
  11 static int irdma_query_device(struct ib_device *ibdev,
  12                               struct ib_device_attr *props,
  13                               struct ib_udata *udata)
  14 {
  15         struct irdma_device *iwdev = to_iwdev(ibdev);
  16         struct irdma_pci_f *rf = iwdev->rf;
  17         struct pci_dev *pcidev = iwdev->rf->pcidev;
  18         struct irdma_hw_attrs *hw_attrs = &rf->sc_dev.hw_attrs;
  19
  20         if (udata->inlen || udata->outlen)
  21                 return -EINVAL;
  22
  23         memset(props, 0, sizeof(*props));
  24         ether_addr_copy((u8 *)&props->sys_image_guid, iwdev->netdev->dev_addr);
  25         props->fw_ver = (u64)irdma_fw_major_ver(&rf->sc_dev) << 32 |
  26                         irdma_fw_minor_ver(&rf->sc_dev);
  27         props->device_cap_flags = iwdev->device_cap_flags;
  28         props->vendor_id = pcidev->vendor;
  29         props->vendor_part_id = pcidev->device;
  30
  31         props->hw_ver = rf->pcidev->revision;
  32         props->page_size_cap = SZ_4K | SZ_2M | SZ_1G;
  33         props->max_mr_size = hw_attrs->max_mr_size;
  34         props->max_qp = rf->max_qp - rf->used_qps;
  35         props->max_qp_wr = hw_attrs->max_qp_wr;
  36         props->max_send_sge = hw_attrs->uk_attrs.max_hw_wq_frags;
  37         props->max_recv_sge = hw_attrs->uk_attrs.max_hw_wq_frags;
  38         props->max_cq = rf->max_cq - rf->used_cqs;
  39         props->max_cqe = rf->max_cqe;
  40         props->max_mr = rf->max_mr - rf->used_mrs;
  41         props->max_mw = props->max_mr;
  42         props->max_pd = rf->max_pd - rf->used_pds;
  43         props->max_sge_rd = hw_attrs->uk_attrs.max_hw_read_sges;
  44         props->max_qp_rd_atom = hw_attrs->max_hw_ird;
  45         props->max_qp_init_rd_atom = hw_attrs->max_hw_ord;
  46         if (rdma_protocol_roce(ibdev, 1))
  47                 props->max_pkeys = IRDMA_PKEY_TBL_SZ;
  48         props->max_ah = rf->max_ah;
  49         props->max_mcast_grp = rf->max_mcg;
  50         props->max_mcast_qp_attach = IRDMA_MAX_MGS_PER_CTX;
  51         props->max_total_mcast_qp_attach = rf->max_qp * IRDMA_MAX_MGS_PER_CTX;
  52         props->max_fast_reg_page_list_len = IRDMA_MAX_PAGES_PER_FMR;
  53 #define HCA_CLOCK_TIMESTAMP_MASK 0x1ffff
  54         if (hw_attrs->uk_attrs.hw_rev >= IRDMA_GEN_2)
  55                 props->timestamp_mask = HCA_CLOCK_TIMESTAMP_MASK;
  56
  57         return 0;
  58 }
  59
  60 /**
  61  * irdma_get_eth_speed_and_width - Get IB port speed and width from netdev speed
  62  * @link_speed: netdev phy link speed
  63  * @active_speed: IB port speed
  64  * @active_width: IB port width
  65  */
  66 static void irdma_get_eth_speed_and_width(u32 link_speed, u16 *active_speed,
  67                                           u8 *active_width)
  68 {
  69         if (link_speed <= SPEED_1000) {
  70                 *active_width = IB_WIDTH_1X;
  71                 *active_speed = IB_SPEED_SDR;
  72         } else if (link_speed <= SPEED_10000) {
  73                 *active_width = IB_WIDTH_1X;
  74                 *active_speed = IB_SPEED_FDR10;
  75         } else if (link_speed <= SPEED_20000) {
  76                 *active_width = IB_WIDTH_4X;
  77                 *active_speed = IB_SPEED_DDR;
  78         } else if (link_speed <= SPEED_25000) {
  79                 *active_width = IB_WIDTH_1X;
  80                 *active_speed = IB_SPEED_EDR;
  81         } else if (link_speed <= SPEED_40000) {
  82                 *active_width = IB_WIDTH_4X;
  83                 *active_speed = IB_SPEED_FDR10;
  84         } else {
  85                 *active_width = IB_WIDTH_4X;
  86                 *active_speed = IB_SPEED_EDR;
  87         }
  88 }
  89
  90 /**
  91  * irdma_query_port - get port attributes
  92  * @ibdev: device pointer from stack
  93  * @port: port number for query
  94  * @props: returning device attributes
  95  */
  96 static int irdma_query_port(struct ib_device *ibdev, u32 port,
  97                             struct ib_port_attr *props)
  98 {
  99         struct irdma_device *iwdev = to_iwdev(ibdev);
 100         struct net_device *netdev = iwdev->netdev;
 101
 102         /* no need to zero out pros here. done by caller */
 103
 104         props->max_mtu = IB_MTU_4096;
 105         props->active_mtu = ib_mtu_int_to_enum(netdev->mtu);
 106         props->lid = 1;
 107         props->lmc = 0;
 108         props->sm_lid = 0;
 109         props->sm_sl = 0;
 110         if (netif_carrier_ok(netdev) && netif_running(netdev)) {
 111                 props->state = IB_PORT_ACTIVE;
 112                 props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
 113         } else {
 114                 props->state = IB_PORT_DOWN;
 115                 props->phys_state = IB_PORT_PHYS_STATE_DISABLED;
 116         }
 117         irdma_get_eth_speed_and_width(SPEED_100000, &props->active_speed,
 118                                       &props->active_width);
 119
 120         if (rdma_protocol_roce(ibdev, 1)) {
 121                 props->gid_tbl_len = 32;
 122                 props->ip_gids = true;
 123                 props->pkey_tbl_len = IRDMA_PKEY_TBL_SZ;
 124         } else {
 125                 props->gid_tbl_len = 1;
 126         }
 127         props->qkey_viol_cntr = 0;
 128         props->port_cap_flags |= IB_PORT_CM_SUP | IB_PORT_REINIT_SUP;
 129         props->max_msg_sz = iwdev->rf->sc_dev.hw_attrs.max_hw_outbound_msg_size;
 130
 131         return 0;
 132 }
 133
 134 /**
 135  * irdma_disassociate_ucontext - Disassociate user context
 136  * @context: ib user context
 137  */
 138 static void irdma_disassociate_ucontext(struct ib_ucontext *context)
 139 {
 140 }
 141
 142 static int irdma_mmap_legacy(struct irdma_ucontext *ucontext,
 143                              struct vm_area_struct *vma)
 144 {
 145         u64 pfn;
 146
 147         if (vma->vm_pgoff || vma->vm_end - vma->vm_start != PAGE_SIZE)
 148                 return -EINVAL;
 149
 150         vma->vm_private_data = ucontext;
 151         pfn = ((uintptr_t)ucontext->iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET] +
 152                pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;
 153
 154         return rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn, PAGE_SIZE,
 155                                  pgprot_noncached(vma->vm_page_prot), NULL);
 156 }
 157
 158 static void irdma_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
 159 {
 160         struct irdma_user_mmap_entry *entry = to_irdma_mmap_entry(rdma_entry);
 161
 162         kfree(entry);
 163 }
 164
 165 static struct rdma_user_mmap_entry*
 166 irdma_user_mmap_entry_insert(struct irdma_ucontext *ucontext, u64 bar_offset,
 167                              enum irdma_mmap_flag mmap_flag, u64 *mmap_offset)
 168 {
 169         struct irdma_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 170         int ret;
 171
 172         if (!entry)
 173                 return NULL;
 174
 175         entry->bar_offset = bar_offset;
 176         entry->mmap_flag = mmap_flag;
 177
 178         ret = rdma_user_mmap_entry_insert(&ucontext->ibucontext,
 179                                           &entry->rdma_entry, PAGE_SIZE);
 180         if (ret) {
 181                 kfree(entry);
 182                 return NULL;
 183         }
 184         *mmap_offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
 185
 186         return &entry->rdma_entry;
 187 }
 188
 189 /**
 190  * irdma_mmap - user memory map
 191  * @context: context created during alloc
 192  * @vma: kernel info for user memory map
 193  */
 194 static int irdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
 195 {
 196         struct rdma_user_mmap_entry *rdma_entry;
 197         struct irdma_user_mmap_entry *entry;
 198         struct irdma_ucontext *ucontext;
 199         u64 pfn;
 200         int ret;
 201
 202         ucontext = to_ucontext(context);
 203
 204         /* Legacy support for libi40iw with hard-coded mmap key */
 205         if (ucontext->legacy_mode)
 206                 return irdma_mmap_legacy(ucontext, vma);
 207
 208         rdma_entry = rdma_user_mmap_entry_get(&ucontext->ibucontext, vma);
 209         if (!rdma_entry) {
 210                 ibdev_dbg(&ucontext->iwdev->ibdev,
 211                           "VERBS: pgoff[0x%lx] does not have valid entry\n",
 212                           vma->vm_pgoff);
 213                 return -EINVAL;
 214         }
 215
 216         entry = to_irdma_mmap_entry(rdma_entry);
 217         ibdev_dbg(&ucontext->iwdev->ibdev,
 218                   "VERBS: bar_offset [0x%llx] mmap_flag [%d]\n",
 219                   entry->bar_offset, entry->mmap_flag);
 220
 221         pfn = (entry->bar_offset +
 222               pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;
 223
 224         switch (entry->mmap_flag) {
 225         case IRDMA_MMAP_IO_NC:
 226                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
 227                                         pgprot_noncached(vma->vm_page_prot),
 228                                         rdma_entry);
 229                 break;
 230         case IRDMA_MMAP_IO_WC:
 231                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
 232                                         pgprot_writecombine(vma->vm_page_prot),
 233                                         rdma_entry);
 234                 break;
 235         default:
 236                 ret = -EINVAL;
 237         }
 238
 239         if (ret)
 240                 ibdev_dbg(&ucontext->iwdev->ibdev,
 241                           "VERBS: bar_offset [0x%llx] mmap_flag[%d] err[%d]\n",
 242                           entry->bar_offset, entry->mmap_flag, ret);
 243         rdma_user_mmap_entry_put(rdma_entry);
 244
 245         return ret;
 246 }
 247
 248 /**
 249  * irdma_alloc_push_page - allocate a push page for qp
 250  * @iwqp: qp pointer
 251  */
 252 static void irdma_alloc_push_page(struct irdma_qp *iwqp)
 253 {
 254         struct irdma_cqp_request *cqp_request;
 255         struct cqp_cmds_info *cqp_info;
 256         struct irdma_device *iwdev = iwqp->iwdev;
 257         struct irdma_sc_qp *qp = &iwqp->sc_qp;
 258         enum irdma_status_code status;
 259
 260         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
 261         if (!cqp_request)
 262                 return;
 263
 264         cqp_info = &cqp_request->info;
 265         cqp_info->cqp_cmd = IRDMA_OP_MANAGE_PUSH_PAGE;
 266         cqp_info->post_sq = 1;
 267         cqp_info->in.u.manage_push_page.info.push_idx = 0;
 268         cqp_info->in.u.manage_push_page.info.qs_handle =
 269                 qp->vsi->qos[qp->user_pri].qs_handle;
 270         cqp_info->in.u.manage_push_page.info.free_page = 0;
 271         cqp_info->in.u.manage_push_page.info.push_page_type = 0;
 272         cqp_info->in.u.manage_push_page.cqp = &iwdev->rf->cqp.sc_cqp;
 273         cqp_info->in.u.manage_push_page.scratch = (uintptr_t)cqp_request;
 274
 275         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
 276         if (!status && cqp_request->compl_info.op_ret_val <
 277             iwdev->rf->sc_dev.hw_attrs.max_hw_device_pages) {
 278                 qp->push_idx = cqp_request->compl_info.op_ret_val;
 279                 qp->push_offset = 0;
 280         }
 281
 282         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
 283 }
 284
 285 /**
 286  * irdma_alloc_ucontext - Allocate the user context data structure
 287  * @uctx: uverbs context pointer
 288  * @udata: user data
 289  *
 290  * This keeps track of all objects associated with a particular
 291  * user-mode client.
 292  */
 293 static int irdma_alloc_ucontext(struct ib_ucontext *uctx,
 294                                 struct ib_udata *udata)
 295 {
 296         struct ib_device *ibdev = uctx->device;
 297         struct irdma_device *iwdev = to_iwdev(ibdev);
 298         struct irdma_alloc_ucontext_req req;
 299         struct irdma_alloc_ucontext_resp uresp = {};
 300         struct irdma_ucontext *ucontext = to_ucontext(uctx);
 301         struct irdma_uk_attrs *uk_attrs;
 302
 303         if (ib_copy_from_udata(&req, udata, min(sizeof(req), udata->inlen)))
 304                 return -EINVAL;
 305
 306         if (req.userspace_ver < 4 || req.userspace_ver > IRDMA_ABI_VER)
 307                 goto ver_error;
 308
 309         ucontext->iwdev = iwdev;
 310         ucontext->abi_ver = req.userspace_ver;
 311
 312         uk_attrs = &iwdev->rf->sc_dev.hw_attrs.uk_attrs;
 313         /* GEN_1 legacy support with libi40iw */
 314         if (udata->outlen < sizeof(uresp)) {
 315                 if (uk_attrs->hw_rev != IRDMA_GEN_1)
 316                         return -EOPNOTSUPP;
 317
 318                 ucontext->legacy_mode = true;
 319                 uresp.max_qps = iwdev->rf->max_qp;
 320                 uresp.max_pds = iwdev->rf->sc_dev.hw_attrs.max_hw_pds;
 321                 uresp.wq_size = iwdev->rf->sc_dev.hw_attrs.max_qp_wr * 2;
 322                 uresp.kernel_ver = req.userspace_ver;
 323                 if (ib_copy_to_udata(udata, &uresp,
 324                                      min(sizeof(uresp), udata->outlen)))
 325                         return -EFAULT;
 326         } else {
 327                 u64 bar_off = (uintptr_t)iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET];
 328
 329                 ucontext->db_mmap_entry =
 330                         irdma_user_mmap_entry_insert(ucontext, bar_off,
 331                                                      IRDMA_MMAP_IO_NC,
 332                                                      &uresp.db_mmap_key);
 333                 if (!ucontext->db_mmap_entry)
 334                         return -ENOMEM;
 335
 336                 uresp.kernel_ver = IRDMA_ABI_VER;
 337                 uresp.feature_flags = uk_attrs->feature_flags;
 338                 uresp.max_hw_wq_frags = uk_attrs->max_hw_wq_frags;
 339                 uresp.max_hw_read_sges = uk_attrs->max_hw_read_sges;
 340                 uresp.max_hw_inline = uk_attrs->max_hw_inline;
 341                 uresp.max_hw_rq_quanta = uk_attrs->max_hw_rq_quanta;
 342                 uresp.max_hw_wq_quanta = uk_attrs->max_hw_wq_quanta;
 343                 uresp.max_hw_sq_chunk = uk_attrs->max_hw_sq_chunk;
 344                 uresp.max_hw_cq_size = uk_attrs->max_hw_cq_size;
 345                 uresp.min_hw_cq_size = uk_attrs->min_hw_cq_size;
 346                 uresp.hw_rev = uk_attrs->hw_rev;
 347                 if (ib_copy_to_udata(udata, &uresp,
 348                                      min(sizeof(uresp), udata->outlen))) {
 349                         rdma_user_mmap_entry_remove(ucontext->db_mmap_entry);
 350                         return -EFAULT;
 351                 }
 352         }
 353
 354         INIT_LIST_HEAD(&ucontext->cq_reg_mem_list);
 355         spin_lock_init(&ucontext->cq_reg_mem_list_lock);
 356         INIT_LIST_HEAD(&ucontext->qp_reg_mem_list);
 357         spin_lock_init(&ucontext->qp_reg_mem_list_lock);
 358
 359         return 0;
 360
 361 ver_error:
 362         ibdev_err(&iwdev->ibdev,
 363                   "Invalid userspace driver version detected. Detected version %d, should be %d\n",
 364                   req.userspace_ver, IRDMA_ABI_VER);
 365         return -EINVAL;
 366 }
 367
 368 /**
 369  * irdma_dealloc_ucontext - deallocate the user context data structure
 370  * @context: user context created during alloc
 371  */
 372 static void irdma_dealloc_ucontext(struct ib_ucontext *context)
 373 {
 374         struct irdma_ucontext *ucontext = to_ucontext(context);
 375
 376         rdma_user_mmap_entry_remove(ucontext->db_mmap_entry);
 377 }
 378
 379 /**
 380  * irdma_alloc_pd - allocate protection domain
 381  * @pd: PD pointer
 382  * @udata: user data
 383  */
 384 static int irdma_alloc_pd(struct ib_pd *pd, struct ib_udata *udata)
 385 {
 386         struct irdma_pd *iwpd = to_iwpd(pd);
 387         struct irdma_device *iwdev = to_iwdev(pd->device);
 388         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 389         struct irdma_pci_f *rf = iwdev->rf;
 390         struct irdma_alloc_pd_resp uresp = {};
 391         struct irdma_sc_pd *sc_pd;
 392         u32 pd_id = 0;
 393         int err;
 394
 395         err = irdma_alloc_rsrc(rf, rf->allocated_pds, rf->max_pd, &pd_id,
 396                                &rf->next_pd);
 397         if (err)
 398                 return err;
 399
 400         sc_pd = &iwpd->sc_pd;
 401         if (udata) {
 402                 struct irdma_ucontext *ucontext =
 403                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
 404                                                   ibucontext);
 405                 irdma_sc_pd_init(dev, sc_pd, pd_id, ucontext->abi_ver);
 406                 uresp.pd_id = pd_id;
 407                 if (ib_copy_to_udata(udata, &uresp,
 408                                      min(sizeof(uresp), udata->outlen))) {
 409                         err = -EFAULT;
 410                         goto error;
 411                 }
 412         } else {
 413                 irdma_sc_pd_init(dev, sc_pd, pd_id, IRDMA_ABI_VER);
 414         }
 415
 416         return 0;
 417 error:
 418         irdma_free_rsrc(rf, rf->allocated_pds, pd_id);
 419
 420         return err;
 421 }
 422
 423 /**
 424  * irdma_dealloc_pd - deallocate pd
 425  * @ibpd: ptr of pd to be deallocated
 426  * @udata: user data
 427  */
 428 static int irdma_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 429 {
 430         struct irdma_pd *iwpd = to_iwpd(ibpd);
 431         struct irdma_device *iwdev = to_iwdev(ibpd->device);
 432
 433         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_pds, iwpd->sc_pd.pd_id);
 434
 435         return 0;
 436 }
 437
 438 /**
 439  * irdma_get_pbl - Retrieve pbl from a list given a virtual
 440  * address
 441  * @va: user virtual address
 442  * @pbl_list: pbl list to search in (QP's or CQ's)
 443  */
 444 static struct irdma_pbl *irdma_get_pbl(unsigned long va,
 445                                        struct list_head *pbl_list)
 446 {
 447         struct irdma_pbl *iwpbl;
 448
 449         list_for_each_entry (iwpbl, pbl_list, list) {
 450                 if (iwpbl->user_base == va) {
 451                         list_del(&iwpbl->list);
 452                         iwpbl->on_list = false;
 453                         return iwpbl;
 454                 }
 455         }
 456
 457         return NULL;
 458 }
 459
 460 /**
 461  * irdma_clean_cqes - clean cq entries for qp
 462  * @iwqp: qp ptr (user or kernel)
 463  * @iwcq: cq ptr
 464  */
 465 static void irdma_clean_cqes(struct irdma_qp *iwqp, struct irdma_cq *iwcq)
 466 {
 467         struct irdma_cq_uk *ukcq = &iwcq->sc_cq.cq_uk;
 468         unsigned long flags;
 469
 470         spin_lock_irqsave(&iwcq->lock, flags);
 471         irdma_uk_clean_cq(&iwqp->sc_qp.qp_uk, ukcq);
 472         spin_unlock_irqrestore(&iwcq->lock, flags);
 473 }
 474
 475 static void irdma_remove_push_mmap_entries(struct irdma_qp *iwqp)
 476 {
 477         if (iwqp->push_db_mmap_entry) {
 478                 rdma_user_mmap_entry_remove(iwqp->push_db_mmap_entry);
 479                 iwqp->push_db_mmap_entry = NULL;
 480         }
 481         if (iwqp->push_wqe_mmap_entry) {
 482                 rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
 483                 iwqp->push_wqe_mmap_entry = NULL;
 484         }
 485 }
 486
 487 static int irdma_setup_push_mmap_entries(struct irdma_ucontext *ucontext,
 488                                          struct irdma_qp *iwqp,
 489                                          u64 *push_wqe_mmap_key,
 490                                          u64 *push_db_mmap_key)
 491 {
 492         struct irdma_device *iwdev = ucontext->iwdev;
 493         u64 rsvd, bar_off;
 494
 495         rsvd = IRDMA_PF_BAR_RSVD;
 496         bar_off = (uintptr_t)iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET];
 497         /* skip over db page */
 498         bar_off += IRDMA_HW_PAGE_SIZE;
 499         /* push wqe page */
 500         bar_off += rsvd + iwqp->sc_qp.push_idx * IRDMA_HW_PAGE_SIZE;
 501         iwqp->push_wqe_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
 502                                         bar_off, IRDMA_MMAP_IO_WC,
 503                                         push_wqe_mmap_key);
 504         if (!iwqp->push_wqe_mmap_entry)
 505                 return -ENOMEM;
 506
 507         /* push doorbell page */
 508         bar_off += IRDMA_HW_PAGE_SIZE;
 509         iwqp->push_db_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
 510                                         bar_off, IRDMA_MMAP_IO_NC,
 511                                         push_db_mmap_key);
 512         if (!iwqp->push_db_mmap_entry) {
 513                 rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
 514                 return -ENOMEM;
 515         }
 516
 517         return 0;
 518 }
 519
 520 /**
 521  * irdma_destroy_qp - destroy qp
 522  * @ibqp: qp's ib pointer also to get to device's qp address
 523  * @udata: user data
 524  */
 525 static int irdma_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
 526 {
 527         struct irdma_qp *iwqp = to_iwqp(ibqp);
 528         struct irdma_device *iwdev = iwqp->iwdev;
 529
 530         iwqp->sc_qp.qp_uk.destroy_pending = true;
 531
 532         if (iwqp->iwarp_state == IRDMA_QP_STATE_RTS)
 533                 irdma_modify_qp_to_err(&iwqp->sc_qp);
 534
 535         irdma_qp_rem_ref(&iwqp->ibqp);
 536         wait_for_completion(&iwqp->free_qp);
 537         irdma_free_lsmm_rsrc(iwqp);
 538         if (!iwdev->reset)
 539                 irdma_cqp_qp_destroy_cmd(&iwdev->rf->sc_dev, &iwqp->sc_qp);
 540
 541         if (!iwqp->user_mode) {
 542                 if (iwqp->iwscq) {
 543                         irdma_clean_cqes(iwqp, iwqp->iwscq);
 544                         if (iwqp->iwrcq != iwqp->iwscq)
 545                                 irdma_clean_cqes(iwqp, iwqp->iwrcq);
 546                 }
 547         }
 548         irdma_remove_push_mmap_entries(iwqp);
 549         irdma_free_qp_rsrc(iwqp);
 550
 551         return 0;
 552 }
 553
 554 /**
 555  * irdma_setup_virt_qp - setup for allocation of virtual qp
 556  * @iwdev: irdma device
 557  * @iwqp: qp ptr
 558  * @init_info: initialize info to return
 559  */
 560 static void irdma_setup_virt_qp(struct irdma_device *iwdev,
 561                                struct irdma_qp *iwqp,
 562                                struct irdma_qp_init_info *init_info)
 563 {
 564         struct irdma_pbl *iwpbl = iwqp->iwpbl;
 565         struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;
 566
 567         iwqp->page = qpmr->sq_page;
 568         init_info->shadow_area_pa = qpmr->shadow;
 569         if (iwpbl->pbl_allocated) {
 570                 init_info->virtual_map = true;
 571                 init_info->sq_pa = qpmr->sq_pbl.idx;
 572                 init_info->rq_pa = qpmr->rq_pbl.idx;
 573         } else {
 574                 init_info->sq_pa = qpmr->sq_pbl.addr;
 575                 init_info->rq_pa = qpmr->rq_pbl.addr;
 576         }
 577 }
 578
 579 /**
 580  * irdma_setup_kmode_qp - setup initialization for kernel mode qp
 581  * @iwdev: iwarp device
 582  * @iwqp: qp ptr (user or kernel)
 583  * @info: initialize info to return
 584  * @init_attr: Initial QP create attributes
 585  */
 586 static int irdma_setup_kmode_qp(struct irdma_device *iwdev,
 587                                 struct irdma_qp *iwqp,
 588                                 struct irdma_qp_init_info *info,
 589                                 struct ib_qp_init_attr *init_attr)
 590 {
 591         struct irdma_dma_mem *mem = &iwqp->kqp.dma_mem;
 592         u32 sqdepth, rqdepth;
 593         u8 sqshift, rqshift;
 594         u32 size;
 595         enum irdma_status_code status;
 596         struct irdma_qp_uk_init_info *ukinfo = &info->qp_uk_init_info;
 597         struct irdma_uk_attrs *uk_attrs = &iwdev->rf->sc_dev.hw_attrs.uk_attrs;
 598
 599         irdma_get_wqe_shift(uk_attrs,
 600                 uk_attrs->hw_rev >= IRDMA_GEN_2 ? ukinfo->max_sq_frag_cnt + 1 :
 601                                                   ukinfo->max_sq_frag_cnt,
 602                 ukinfo->max_inline_data, &sqshift);
 603         status = irdma_get_sqdepth(uk_attrs, ukinfo->sq_size, sqshift,
 604                                    &sqdepth);
 605         if (status)
 606                 return -ENOMEM;
 607
 608         if (uk_attrs->hw_rev == IRDMA_GEN_1)
 609                 rqshift = IRDMA_MAX_RQ_WQE_SHIFT_GEN1;
 610         else
 611                 irdma_get_wqe_shift(uk_attrs, ukinfo->max_rq_frag_cnt, 0,
 612                                     &rqshift);
 613
 614         status = irdma_get_rqdepth(uk_attrs, ukinfo->rq_size, rqshift,
 615                                    &rqdepth);
 616         if (status)
 617                 return -ENOMEM;
 618
 619         iwqp->kqp.sq_wrid_mem =
 620                 kcalloc(sqdepth, sizeof(*iwqp->kqp.sq_wrid_mem), GFP_KERNEL);
 621         if (!iwqp->kqp.sq_wrid_mem)
 622                 return -ENOMEM;
 623
 624         iwqp->kqp.rq_wrid_mem =
 625                 kcalloc(rqdepth, sizeof(*iwqp->kqp.rq_wrid_mem), GFP_KERNEL);
 626         if (!iwqp->kqp.rq_wrid_mem) {
 627                 kfree(iwqp->kqp.sq_wrid_mem);
 628                 iwqp->kqp.sq_wrid_mem = NULL;
 629                 return -ENOMEM;
 630         }
 631
 632         ukinfo->sq_wrtrk_array = iwqp->kqp.sq_wrid_mem;
 633         ukinfo->rq_wrid_array = iwqp->kqp.rq_wrid_mem;
 634
 635         size = (sqdepth + rqdepth) * IRDMA_QP_WQE_MIN_SIZE;
 636         size += (IRDMA_SHADOW_AREA_SIZE << 3);
 637
 638         mem->size = ALIGN(size, 256);
 639         mem->va = dma_alloc_coherent(iwdev->rf->hw.device, mem->size,
 640                                      &mem->pa, GFP_KERNEL);
 641         if (!mem->va) {
 642                 kfree(iwqp->kqp.sq_wrid_mem);
 643                 iwqp->kqp.sq_wrid_mem = NULL;
 644                 kfree(iwqp->kqp.rq_wrid_mem);
 645                 iwqp->kqp.rq_wrid_mem = NULL;
 646                 return -ENOMEM;
 647         }
 648
 649         ukinfo->sq = mem->va;
 650         info->sq_pa = mem->pa;
 651         ukinfo->rq = &ukinfo->sq[sqdepth];
 652         info->rq_pa = info->sq_pa + (sqdepth * IRDMA_QP_WQE_MIN_SIZE);
 653         ukinfo->shadow_area = ukinfo->rq[rqdepth].elem;
 654         info->shadow_area_pa = info->rq_pa + (rqdepth * IRDMA_QP_WQE_MIN_SIZE);
 655         ukinfo->sq_size = sqdepth >> sqshift;
 656         ukinfo->rq_size = rqdepth >> rqshift;
 657         ukinfo->qp_id = iwqp->ibqp.qp_num;
 658
 659         init_attr->cap.max_send_wr = (sqdepth - IRDMA_SQ_RSVD) >> sqshift;
 660         init_attr->cap.max_recv_wr = (rqdepth - IRDMA_RQ_RSVD) >> rqshift;
 661
 662         return 0;
 663 }
 664
 665 static int irdma_cqp_create_qp_cmd(struct irdma_qp *iwqp)
 666 {
 667         struct irdma_pci_f *rf = iwqp->iwdev->rf;
 668         struct irdma_cqp_request *cqp_request;
 669         struct cqp_cmds_info *cqp_info;
 670         struct irdma_create_qp_info *qp_info;
 671         enum irdma_status_code status;
 672
 673         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
 674         if (!cqp_request)
 675                 return -ENOMEM;
 676
 677         cqp_info = &cqp_request->info;
 678         qp_info = &cqp_request->info.in.u.qp_create.info;
 679         memset(qp_info, 0, sizeof(*qp_info));
 680         qp_info->mac_valid = true;
 681         qp_info->cq_num_valid = true;
 682         qp_info->next_iwarp_state = IRDMA_QP_STATE_IDLE;
 683
 684         cqp_info->cqp_cmd = IRDMA_OP_QP_CREATE;
 685         cqp_info->post_sq = 1;
 686         cqp_info->in.u.qp_create.qp = &iwqp->sc_qp;
 687         cqp_info->in.u.qp_create.scratch = (uintptr_t)cqp_request;
 688         status = irdma_handle_cqp_op(rf, cqp_request);
 689         irdma_put_cqp_request(&rf->cqp, cqp_request);
 690
 691         return status ? -ENOMEM : 0;
 692 }
 693
 694 static void irdma_roce_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
 695                                                struct irdma_qp_host_ctx_info *ctx_info)
 696 {
 697         struct irdma_device *iwdev = iwqp->iwdev;
 698         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 699         struct irdma_roce_offload_info *roce_info;
 700         struct irdma_udp_offload_info *udp_info;
 701
 702         udp_info = &iwqp->udp_info;
 703         udp_info->snd_mss = ib_mtu_enum_to_int(ib_mtu_int_to_enum(iwdev->vsi.mtu));
 704         udp_info->cwnd = iwdev->roce_cwnd;
 705         udp_info->rexmit_thresh = 2;
 706         udp_info->rnr_nak_thresh = 2;
 707         udp_info->src_port = 0xc000;
 708         udp_info->dst_port = ROCE_V2_UDP_DPORT;
 709         roce_info = &iwqp->roce_info;
 710         ether_addr_copy(roce_info->mac_addr, iwdev->netdev->dev_addr);
 711
 712         roce_info->rd_en = true;
 713         roce_info->wr_rdresp_en = true;
 714         roce_info->bind_en = true;
 715         roce_info->dcqcn_en = false;
 716         roce_info->rtomin = 5;
 717
 718         roce_info->ack_credits = iwdev->roce_ackcreds;
 719         roce_info->ird_size = dev->hw_attrs.max_hw_ird;
 720         roce_info->ord_size = dev->hw_attrs.max_hw_ord;
 721
 722         if (!iwqp->user_mode) {
 723                 roce_info->priv_mode_en = true;
 724                 roce_info->fast_reg_en = true;
 725                 roce_info->udprivcq_en = true;
 726         }
 727         roce_info->roce_tver = 0;
 728
 729         ctx_info->roce_info = &iwqp->roce_info;
 730         ctx_info->udp_info = &iwqp->udp_info;
 731         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
 732 }
 733
 734 static void irdma_iw_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
 735                                              struct irdma_qp_host_ctx_info *ctx_info)
 736 {
 737         struct irdma_device *iwdev = iwqp->iwdev;
 738         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 739         struct irdma_iwarp_offload_info *iwarp_info;
 740
 741         iwarp_info = &iwqp->iwarp_info;
 742         ether_addr_copy(iwarp_info->mac_addr, iwdev->netdev->dev_addr);
 743         iwarp_info->rd_en = true;
 744         iwarp_info->wr_rdresp_en = true;
 745         iwarp_info->bind_en = true;
 746         iwarp_info->ecn_en = true;
 747         iwarp_info->rtomin = 5;
 748
 749         if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
 750                 iwarp_info->ib_rd_en = true;
 751         if (!iwqp->user_mode) {
 752                 iwarp_info->priv_mode_en = true;
 753                 iwarp_info->fast_reg_en = true;
 754         }
 755         iwarp_info->ddp_ver = 1;
 756         iwarp_info->rdmap_ver = 1;
 757
 758         ctx_info->iwarp_info = &iwqp->iwarp_info;
 759         ctx_info->iwarp_info_valid = true;
 760         irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
 761         ctx_info->iwarp_info_valid = false;
 762 }
 763
 764 static int irdma_validate_qp_attrs(struct ib_qp_init_attr *init_attr,
 765                                    struct irdma_device *iwdev)
 766 {
 767         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 768         struct irdma_uk_attrs *uk_attrs = &dev->hw_attrs.uk_attrs;
 769
 770         if (init_attr->create_flags)
 771                 return -EOPNOTSUPP;
 772
 773         if (init_attr->cap.max_inline_data > uk_attrs->max_hw_inline ||
 774             init_attr->cap.max_send_sge > uk_attrs->max_hw_wq_frags ||
 775             init_attr->cap.max_recv_sge > uk_attrs->max_hw_wq_frags)
 776                 return -EINVAL;
 777
 778         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 779                 if (init_attr->qp_type != IB_QPT_RC &&
 780                     init_attr->qp_type != IB_QPT_UD &&
 781                     init_attr->qp_type != IB_QPT_GSI)
 782                         return -EOPNOTSUPP;
 783         } else {
 784                 if (init_attr->qp_type != IB_QPT_RC)
 785                         return -EOPNOTSUPP;
 786         }
 787
 788         return 0;
 789 }
 790
 791 /**
 792  * irdma_create_qp - create qp
 793  * @ibqp: ptr of qp
 794  * @init_attr: attributes for qp
 795  * @udata: user data for create qp
 796  */
 797 static int irdma_create_qp(struct ib_qp *ibqp,
 798                            struct ib_qp_init_attr *init_attr,
 799                            struct ib_udata *udata)
 800 {
 801         struct ib_pd *ibpd = ibqp->pd;
 802         struct irdma_pd *iwpd = to_iwpd(ibpd);
 803         struct irdma_device *iwdev = to_iwdev(ibpd->device);
 804         struct irdma_pci_f *rf = iwdev->rf;
 805         struct irdma_qp *iwqp = to_iwqp(ibqp);
 806         struct irdma_create_qp_req req;
 807         struct irdma_create_qp_resp uresp = {};
 808         u32 qp_num = 0;
 809         enum irdma_status_code ret;
 810         int err_code;
 811         int sq_size;
 812         int rq_size;
 813         struct irdma_sc_qp *qp;
 814         struct irdma_sc_dev *dev = &rf->sc_dev;
 815         struct irdma_uk_attrs *uk_attrs = &dev->hw_attrs.uk_attrs;
 816         struct irdma_qp_init_info init_info = {};
 817         struct irdma_qp_host_ctx_info *ctx_info;
 818         unsigned long flags;
 819
 820         err_code = irdma_validate_qp_attrs(init_attr, iwdev);
 821         if (err_code)
 822                 return err_code;
 823
 824         sq_size = init_attr->cap.max_send_wr;
 825         rq_size = init_attr->cap.max_recv_wr;
 826
 827         init_info.vsi = &iwdev->vsi;
 828         init_info.qp_uk_init_info.uk_attrs = uk_attrs;
 829         init_info.qp_uk_init_info.sq_size = sq_size;
 830         init_info.qp_uk_init_info.rq_size = rq_size;
 831         init_info.qp_uk_init_info.max_sq_frag_cnt = init_attr->cap.max_send_sge;
 832         init_info.qp_uk_init_info.max_rq_frag_cnt = init_attr->cap.max_recv_sge;
 833         init_info.qp_uk_init_info.max_inline_data = init_attr->cap.max_inline_data;
 834
 835         qp = &iwqp->sc_qp;
 836         qp->qp_uk.back_qp = iwqp;
 837         qp->qp_uk.lock = &iwqp->lock;
 838         qp->push_idx = IRDMA_INVALID_PUSH_PAGE_INDEX;
 839
 840         iwqp->iwdev = iwdev;
 841         iwqp->q2_ctx_mem.size = ALIGN(IRDMA_Q2_BUF_SIZE + IRDMA_QP_CTX_SIZE,
 842                                       256);
 843         iwqp->q2_ctx_mem.va = dma_alloc_coherent(dev->hw->device,
 844                                                  iwqp->q2_ctx_mem.size,
 845                                                  &iwqp->q2_ctx_mem.pa,
 846                                                  GFP_KERNEL);
 847         if (!iwqp->q2_ctx_mem.va)
 848                 return -ENOMEM;
 849
 850         init_info.q2 = iwqp->q2_ctx_mem.va;
 851         init_info.q2_pa = iwqp->q2_ctx_mem.pa;
 852         init_info.host_ctx = (__le64 *)(init_info.q2 + IRDMA_Q2_BUF_SIZE);
 853         init_info.host_ctx_pa = init_info.q2_pa + IRDMA_Q2_BUF_SIZE;
 854
 855         if (init_attr->qp_type == IB_QPT_GSI)
 856                 qp_num = 1;
 857         else
 858                 err_code = irdma_alloc_rsrc(rf, rf->allocated_qps, rf->max_qp,
 859                                             &qp_num, &rf->next_qp);
 860         if (err_code)
 861                 goto error;
 862
 863         iwqp->iwpd = iwpd;
 864         iwqp->ibqp.qp_num = qp_num;
 865         qp = &iwqp->sc_qp;
 866         iwqp->iwscq = to_iwcq(init_attr->send_cq);
 867         iwqp->iwrcq = to_iwcq(init_attr->recv_cq);
 868         iwqp->host_ctx.va = init_info.host_ctx;
 869         iwqp->host_ctx.pa = init_info.host_ctx_pa;
 870         iwqp->host_ctx.size = IRDMA_QP_CTX_SIZE;
 871
 872         init_info.pd = &iwpd->sc_pd;
 873         init_info.qp_uk_init_info.qp_id = iwqp->ibqp.qp_num;
 874         if (!rdma_protocol_roce(&iwdev->ibdev, 1))
 875                 init_info.qp_uk_init_info.first_sq_wq = 1;
 876         iwqp->ctx_info.qp_compl_ctx = (uintptr_t)qp;
 877         init_waitqueue_head(&iwqp->waitq);
 878         init_waitqueue_head(&iwqp->mod_qp_waitq);
 879
 880         if (udata) {
 881                 err_code = ib_copy_from_udata(&req, udata,
 882                                               min(sizeof(req), udata->inlen));
 883                 if (err_code) {
 884                         ibdev_dbg(&iwdev->ibdev,
 885                                   "VERBS: ib_copy_from_data fail\n");
 886                         goto error;
 887                 }
 888
 889                 iwqp->ctx_info.qp_compl_ctx = req.user_compl_ctx;
 890                 iwqp->user_mode = 1;
 891                 if (req.user_wqe_bufs) {
 892                         struct irdma_ucontext *ucontext =
 893                                 rdma_udata_to_drv_context(udata,
 894                                                           struct irdma_ucontext,
 895                                                           ibucontext);
 896
 897                         init_info.qp_uk_init_info.legacy_mode = ucontext->legacy_mode;
 898                         spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
 899                         iwqp->iwpbl = irdma_get_pbl((unsigned long)req.user_wqe_bufs,
 900                                                     &ucontext->qp_reg_mem_list);
 901                         spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
 902
 903                         if (!iwqp->iwpbl) {
 904                                 err_code = -ENODATA;
 905                                 ibdev_dbg(&iwdev->ibdev, "VERBS: no pbl info\n");
 906                                 goto error;
 907                         }
 908                 }
 909                 init_info.qp_uk_init_info.abi_ver = iwpd->sc_pd.abi_ver;
 910                 irdma_setup_virt_qp(iwdev, iwqp, &init_info);
 911         } else {
 912                 init_info.qp_uk_init_info.abi_ver = IRDMA_ABI_VER;
 913                 err_code = irdma_setup_kmode_qp(iwdev, iwqp, &init_info, init_attr);
 914         }
 915
 916         if (err_code) {
 917                 ibdev_dbg(&iwdev->ibdev, "VERBS: setup qp failed\n");
 918                 goto error;
 919         }
 920
 921         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 922                 if (init_attr->qp_type == IB_QPT_RC) {
 923                         init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_ROCE_RC;
 924                         init_info.qp_uk_init_info.qp_caps = IRDMA_SEND_WITH_IMM |
 925                                                             IRDMA_WRITE_WITH_IMM |
 926                                                             IRDMA_ROCE;
 927                 } else {
 928                         init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_ROCE_UD;
 929                         init_info.qp_uk_init_info.qp_caps = IRDMA_SEND_WITH_IMM |
 930                                                             IRDMA_ROCE;
 931                 }
 932         } else {
 933                 init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_IWARP;
 934                 init_info.qp_uk_init_info.qp_caps = IRDMA_WRITE_WITH_IMM;
 935         }
 936
 937         if (dev->hw_attrs.uk_attrs.hw_rev > IRDMA_GEN_1)
 938                 init_info.qp_uk_init_info.qp_caps |= IRDMA_PUSH_MODE;
 939
 940         ret = irdma_sc_qp_init(qp, &init_info);
 941         if (ret) {
 942                 err_code = -EPROTO;
 943                 ibdev_dbg(&iwdev->ibdev, "VERBS: qp_init fail\n");
 944                 goto error;
 945         }
 946
 947         ctx_info = &iwqp->ctx_info;
 948         ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
 949         ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
 950
 951         if (rdma_protocol_roce(&iwdev->ibdev, 1))
 952                 irdma_roce_fill_and_set_qpctx_info(iwqp, ctx_info);
 953         else
 954                 irdma_iw_fill_and_set_qpctx_info(iwqp, ctx_info);
 955
 956         err_code = irdma_cqp_create_qp_cmd(iwqp);
 957         if (err_code)
 958                 goto error;
 959
 960         refcount_set(&iwqp->refcnt, 1);
 961         spin_lock_init(&iwqp->lock);
 962         spin_lock_init(&iwqp->sc_qp.pfpdu.lock);
 963         iwqp->sig_all = (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ? 1 : 0;
 964         rf->qp_table[qp_num] = iwqp;
 965         iwqp->max_send_wr = sq_size;
 966         iwqp->max_recv_wr = rq_size;
 967
 968         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 969                 if (dev->ws_add(&iwdev->vsi, 0)) {
 970                         irdma_cqp_qp_destroy_cmd(&rf->sc_dev, &iwqp->sc_qp);
 971                         err_code = -EINVAL;
 972                         goto error;
 973                 }
 974
 975                 irdma_qp_add_qos(&iwqp->sc_qp);
 976         }
 977
 978         if (udata) {
 979                 /* GEN_1 legacy support with libi40iw does not have expanded uresp struct */
 980                 if (udata->outlen < sizeof(uresp)) {
 981                         uresp.lsmm = 1;
 982                         uresp.push_idx = IRDMA_INVALID_PUSH_PAGE_INDEX_GEN_1;
 983                 } else {
 984                         if (rdma_protocol_iwarp(&iwdev->ibdev, 1))
 985                                 uresp.lsmm = 1;
 986                 }
 987                 uresp.actual_sq_size = sq_size;
 988                 uresp.actual_rq_size = rq_size;
 989                 uresp.qp_id = qp_num;
 990                 uresp.qp_caps = qp->qp_uk.qp_caps;
 991
 992                 err_code = ib_copy_to_udata(udata, &uresp,
 993                                             min(sizeof(uresp), udata->outlen));
 994                 if (err_code) {
 995                         ibdev_dbg(&iwdev->ibdev, "VERBS: copy_to_udata failed\n");
 996                         irdma_destroy_qp(&iwqp->ibqp, udata);
 997                         return err_code;
 998                 }
 999         }
1000
1001         init_completion(&iwqp->free_qp);
1002         return 0;
1003
1004 error:
1005         irdma_free_qp_rsrc(iwqp);
1006         return err_code;
1007 }
1008
1009 static int irdma_get_ib_acc_flags(struct irdma_qp *iwqp)
1010 {
1011         int acc_flags = 0;
1012
1013         if (rdma_protocol_roce(iwqp->ibqp.device, 1)) {
1014                 if (iwqp->roce_info.wr_rdresp_en) {
1015                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
1016                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
1017                 }
1018                 if (iwqp->roce_info.rd_en)
1019                         acc_flags |= IB_ACCESS_REMOTE_READ;
1020                 if (iwqp->roce_info.bind_en)
1021                         acc_flags |= IB_ACCESS_MW_BIND;
1022         } else {
1023                 if (iwqp->iwarp_info.wr_rdresp_en) {
1024                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
1025                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
1026                 }
1027                 if (iwqp->iwarp_info.rd_en)
1028                         acc_flags |= IB_ACCESS_REMOTE_READ;
1029                 if (iwqp->iwarp_info.bind_en)
1030                         acc_flags |= IB_ACCESS_MW_BIND;
1031         }
1032         return acc_flags;
1033 }
1034
1035 /**
1036  * irdma_query_qp - query qp attributes
1037  * @ibqp: qp pointer
1038  * @attr: attributes pointer
1039  * @attr_mask: Not used
1040  * @init_attr: qp attributes to return
1041  */
1042 static int irdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1043                           int attr_mask, struct ib_qp_init_attr *init_attr)
1044 {
1045         struct irdma_qp *iwqp = to_iwqp(ibqp);
1046         struct irdma_sc_qp *qp = &iwqp->sc_qp;
1047
1048         memset(attr, 0, sizeof(*attr));
1049         memset(init_attr, 0, sizeof(*init_attr));
1050
1051         attr->qp_state = iwqp->ibqp_state;
1052         attr->cur_qp_state = iwqp->ibqp_state;
1053         attr->cap.max_send_wr = iwqp->max_send_wr;
1054         attr->cap.max_recv_wr = iwqp->max_recv_wr;
1055         attr->cap.max_inline_data = qp->qp_uk.max_inline_data;
1056         attr->cap.max_send_sge = qp->qp_uk.max_sq_frag_cnt;
1057         attr->cap.max_recv_sge = qp->qp_uk.max_rq_frag_cnt;
1058         attr->qp_access_flags = irdma_get_ib_acc_flags(iwqp);
1059         attr->port_num = 1;
1060         if (rdma_protocol_roce(ibqp->device, 1)) {
1061                 attr->path_mtu = ib_mtu_int_to_enum(iwqp->udp_info.snd_mss);
1062                 attr->qkey = iwqp->roce_info.qkey;
1063                 attr->rq_psn = iwqp->udp_info.epsn;
1064                 attr->sq_psn = iwqp->udp_info.psn_nxt;
1065                 attr->dest_qp_num = iwqp->roce_info.dest_qp;
1066                 attr->pkey_index = iwqp->roce_info.p_key;
1067                 attr->retry_cnt = iwqp->udp_info.rexmit_thresh;
1068                 attr->rnr_retry = iwqp->udp_info.rnr_nak_thresh;
1069                 attr->max_rd_atomic = iwqp->roce_info.ord_size;
1070                 attr->max_dest_rd_atomic = iwqp->roce_info.ird_size;
1071         }
1072
1073         init_attr->event_handler = iwqp->ibqp.event_handler;
1074         init_attr->qp_context = iwqp->ibqp.qp_context;
1075         init_attr->send_cq = iwqp->ibqp.send_cq;
1076         init_attr->recv_cq = iwqp->ibqp.recv_cq;
1077         init_attr->cap = attr->cap;
1078
1079         return 0;
1080 }
1081
1082 /**
1083  * irdma_query_pkey - Query partition key
1084  * @ibdev: device pointer from stack
1085  * @port: port number
1086  * @index: index of pkey
1087  * @pkey: pointer to store the pkey
1088  */
1089 static int irdma_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1090                             u16 *pkey)
1091 {
1092         if (index >= IRDMA_PKEY_TBL_SZ)
1093                 return -EINVAL;
1094
1095         *pkey = IRDMA_DEFAULT_PKEY;
1096         return 0;
1097 }
1098
1099 /**
1100  * irdma_modify_qp_roce - modify qp request
1101  * @ibqp: qp's pointer for modify
1102  * @attr: access attributes
1103  * @attr_mask: state mask
1104  * @udata: user data
1105  */
1106 int irdma_modify_qp_roce(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1107                          int attr_mask, struct ib_udata *udata)
1108 {
1109         struct irdma_pd *iwpd = to_iwpd(ibqp->pd);
1110         struct irdma_qp *iwqp = to_iwqp(ibqp);
1111         struct irdma_device *iwdev = iwqp->iwdev;
1112         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
1113         struct irdma_qp_host_ctx_info *ctx_info;
1114         struct irdma_roce_offload_info *roce_info;
1115         struct irdma_udp_offload_info *udp_info;
1116         struct irdma_modify_qp_info info = {};
1117         struct irdma_modify_qp_resp uresp = {};
1118         struct irdma_modify_qp_req ureq = {};
1119         unsigned long flags;
1120         u8 issue_modify_qp = 0;
1121         int ret = 0;
1122
1123         ctx_info = &iwqp->ctx_info;
1124         roce_info = &iwqp->roce_info;
1125         udp_info = &iwqp->udp_info;
1126
1127         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1128                 return -EOPNOTSUPP;
1129
1130         if (attr_mask & IB_QP_DEST_QPN)
1131                 roce_info->dest_qp = attr->dest_qp_num;
1132
1133         if (attr_mask & IB_QP_PKEY_INDEX) {
1134                 ret = irdma_query_pkey(ibqp->device, 0, attr->pkey_index,
1135                                        &roce_info->p_key);
1136                 if (ret)
1137                         return ret;
1138         }
1139
1140         if (attr_mask & IB_QP_QKEY)
1141                 roce_info->qkey = attr->qkey;
1142
1143         if (attr_mask & IB_QP_PATH_MTU)
1144                 udp_info->snd_mss = ib_mtu_enum_to_int(attr->path_mtu);
1145
1146         if (attr_mask & IB_QP_SQ_PSN) {
1147                 udp_info->psn_nxt = attr->sq_psn;
1148                 udp_info->lsn =  0xffff;
1149                 udp_info->psn_una = attr->sq_psn;
1150                 udp_info->psn_max = attr->sq_psn;
1151         }
1152
1153         if (attr_mask & IB_QP_RQ_PSN)
1154                 udp_info->epsn = attr->rq_psn;
1155
1156         if (attr_mask & IB_QP_RNR_RETRY)
1157                 udp_info->rnr_nak_thresh = attr->rnr_retry;
1158
1159         if (attr_mask & IB_QP_RETRY_CNT)
1160                 udp_info->rexmit_thresh = attr->retry_cnt;
1161
1162         ctx_info->roce_info->pd_id = iwpd->sc_pd.pd_id;
1163
1164         if (attr_mask & IB_QP_AV) {
1165                 struct irdma_av *av = &iwqp->roce_ah.av;
1166                 const struct ib_gid_attr *sgid_attr;
1167                 u16 vlan_id = VLAN_N_VID;
1168                 u32 local_ip[4];
1169
1170                 memset(&iwqp->roce_ah, 0, sizeof(iwqp->roce_ah));
1171                 if (attr->ah_attr.ah_flags & IB_AH_GRH) {
1172                         udp_info->ttl = attr->ah_attr.grh.hop_limit;
1173                         udp_info->flow_label = attr->ah_attr.grh.flow_label;
1174                         udp_info->tos = attr->ah_attr.grh.traffic_class;
1175                         irdma_qp_rem_qos(&iwqp->sc_qp);
1176                         dev->ws_remove(iwqp->sc_qp.vsi, ctx_info->user_pri);
1177                         ctx_info->user_pri = rt_tos2priority(udp_info->tos);
1178                         iwqp->sc_qp.user_pri = ctx_info->user_pri;
1179                         if (dev->ws_add(iwqp->sc_qp.vsi, ctx_info->user_pri))
1180                                 return -ENOMEM;
1181                         irdma_qp_add_qos(&iwqp->sc_qp);
1182                 }
1183                 sgid_attr = attr->ah_attr.grh.sgid_attr;
1184                 ret = rdma_read_gid_l2_fields(sgid_attr, &vlan_id,
1185                                               ctx_info->roce_info->mac_addr);
1186                 if (ret)
1187                         return ret;
1188
1189                 if (vlan_id >= VLAN_N_VID && iwdev->dcb)
1190                         vlan_id = 0;
1191                 if (vlan_id < VLAN_N_VID) {
1192                         udp_info->insert_vlan_tag = true;
1193                         udp_info->vlan_tag = vlan_id |
1194                                 ctx_info->user_pri << VLAN_PRIO_SHIFT;
1195                 } else {
1196                         udp_info->insert_vlan_tag = false;
1197                 }
1198
1199                 av->attrs = attr->ah_attr;
1200                 rdma_gid2ip((struct sockaddr *)&av->sgid_addr, &sgid_attr->gid);
1201                 rdma_gid2ip((struct sockaddr *)&av->dgid_addr, &attr->ah_attr.grh.dgid);
1202                 roce_info->local_qp = ibqp->qp_num;
1203                 if (av->sgid_addr.saddr.sa_family == AF_INET6) {
1204                         __be32 *daddr =
1205                                 av->dgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32;
1206                         __be32 *saddr =
1207                                 av->sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32;
1208
1209                         irdma_copy_ip_ntohl(&udp_info->dest_ip_addr[0], daddr);
1210                         irdma_copy_ip_ntohl(&udp_info->local_ipaddr[0], saddr);
1211
1212                         udp_info->ipv4 = false;
1213                         irdma_copy_ip_ntohl(local_ip, daddr);
1214
1215                         udp_info->arp_idx = irdma_arp_table(iwdev->rf,
1216                                                             &local_ip[0],
1217                                                             false, NULL,
1218                                                             IRDMA_ARP_RESOLVE);
1219                 } else {
1220                         __be32 saddr = av->sgid_addr.saddr_in.sin_addr.s_addr;
1221                         __be32 daddr = av->dgid_addr.saddr_in.sin_addr.s_addr;
1222
1223                         local_ip[0] = ntohl(daddr);
1224
1225                         udp_info->ipv4 = true;
1226                         udp_info->dest_ip_addr[0] = 0;
1227                         udp_info->dest_ip_addr[1] = 0;
1228                         udp_info->dest_ip_addr[2] = 0;
1229                         udp_info->dest_ip_addr[3] = local_ip[0];
1230
1231                         udp_info->local_ipaddr[0] = 0;
1232                         udp_info->local_ipaddr[1] = 0;
1233                         udp_info->local_ipaddr[2] = 0;
1234                         udp_info->local_ipaddr[3] = ntohl(saddr);
1235                 }
1236                 udp_info->arp_idx =
1237                         irdma_add_arp(iwdev->rf, local_ip, udp_info->ipv4,
1238                                       attr->ah_attr.roce.dmac);
1239         }
1240
1241         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1242                 if (attr->max_rd_atomic > dev->hw_attrs.max_hw_ord) {
1243                         ibdev_err(&iwdev->ibdev,
1244                                   "rd_atomic = %d, above max_hw_ord=%d\n",
1245                                   attr->max_rd_atomic,
1246                                   dev->hw_attrs.max_hw_ord);
1247                         return -EINVAL;
1248                 }
1249                 if (attr->max_rd_atomic)
1250                         roce_info->ord_size = attr->max_rd_atomic;
1251                 info.ord_valid = true;
1252         }
1253
1254         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1255                 if (attr->max_dest_rd_atomic > dev->hw_attrs.max_hw_ird) {
1256                         ibdev_err(&iwdev->ibdev,
1257                                   "rd_atomic = %d, above max_hw_ird=%d\n",
1258                                    attr->max_rd_atomic,
1259                                    dev->hw_attrs.max_hw_ird);
1260                         return -EINVAL;
1261                 }
1262                 if (attr->max_dest_rd_atomic)
1263                         roce_info->ird_size = attr->max_dest_rd_atomic;
1264         }
1265
1266         if (attr_mask & IB_QP_ACCESS_FLAGS) {
1267                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
1268                         roce_info->wr_rdresp_en = true;
1269                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
1270                         roce_info->wr_rdresp_en = true;
1271                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
1272                         roce_info->rd_en = true;
1273         }
1274
1275         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
1276
1277         ibdev_dbg(&iwdev->ibdev,
1278                   "VERBS: caller: %pS qp_id=%d to_ibqpstate=%d ibqpstate=%d irdma_qpstate=%d attr_mask=0x%x\n",
1279                   __builtin_return_address(0), ibqp->qp_num, attr->qp_state,
1280                   iwqp->ibqp_state, iwqp->iwarp_state, attr_mask);
1281
1282         spin_lock_irqsave(&iwqp->lock, flags);
1283         if (attr_mask & IB_QP_STATE) {
1284                 if (!ib_modify_qp_is_ok(iwqp->ibqp_state, attr->qp_state,
1285                                         iwqp->ibqp.qp_type, attr_mask)) {
1286                         ibdev_warn(&iwdev->ibdev, "modify_qp invalid for qp_id=%d, old_state=0x%x, new_state=0x%x\n",
1287                                    iwqp->ibqp.qp_num, iwqp->ibqp_state,
1288                                    attr->qp_state);
1289                         ret = -EINVAL;
1290                         goto exit;
1291                 }
1292                 info.curr_iwarp_state = iwqp->iwarp_state;
1293
1294                 switch (attr->qp_state) {
1295                 case IB_QPS_INIT:
1296                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1297                                 ret = -EINVAL;
1298                                 goto exit;
1299                         }
1300
1301                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
1302                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
1303                                 issue_modify_qp = 1;
1304                         }
1305                         break;
1306                 case IB_QPS_RTR:
1307                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1308                                 ret = -EINVAL;
1309                                 goto exit;
1310                         }
1311                         info.arp_cache_idx_valid = true;
1312                         info.cq_num_valid = true;
1313                         info.next_iwarp_state = IRDMA_QP_STATE_RTR;
1314                         issue_modify_qp = 1;
1315                         break;
1316                 case IB_QPS_RTS:
1317                         if (iwqp->ibqp_state < IB_QPS_RTR ||
1318                             iwqp->ibqp_state == IB_QPS_ERR) {
1319                                 ret = -EINVAL;
1320                                 goto exit;
1321                         }
1322
1323                         info.arp_cache_idx_valid = true;
1324                         info.cq_num_valid = true;
1325                         info.ord_valid = true;
1326                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
1327                         issue_modify_qp = 1;
1328                         if (iwdev->push_mode && udata &&
1329                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1330                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1331                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1332                                 irdma_alloc_push_page(iwqp);
1333                                 spin_lock_irqsave(&iwqp->lock, flags);
1334                         }
1335                         break;
1336                 case IB_QPS_SQD:
1337                         if (iwqp->iwarp_state == IRDMA_QP_STATE_SQD)
1338                                 goto exit;
1339
1340                         if (iwqp->iwarp_state != IRDMA_QP_STATE_RTS) {
1341                                 ret = -EINVAL;
1342                                 goto exit;
1343                         }
1344
1345                         info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1346                         issue_modify_qp = 1;
1347                         break;
1348                 case IB_QPS_SQE:
1349                 case IB_QPS_ERR:
1350                 case IB_QPS_RESET:
1351                         if (iwqp->iwarp_state == IRDMA_QP_STATE_RTS) {
1352                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1353                                 info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1354                                 irdma_hw_modify_qp(iwdev, iwqp, &info, true);
1355                                 spin_lock_irqsave(&iwqp->lock, flags);
1356                         }
1357
1358                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1359                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1360                                 if (udata) {
1361                                         if (ib_copy_from_udata(&ureq, udata,
1362                                             min(sizeof(ureq), udata->inlen)))
1363                                                 return -EINVAL;
1364
1365                                         irdma_flush_wqes(iwqp,
1366                                             (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1367                                             (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1368                                             IRDMA_REFLUSH);
1369                                 }
1370                                 return 0;
1371                         }
1372
1373                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1374                         issue_modify_qp = 1;
1375                         break;
1376                 default:
1377                         ret = -EINVAL;
1378                         goto exit;
1379                 }
1380
1381                 iwqp->ibqp_state = attr->qp_state;
1382         }
1383
1384         ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1385         ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1386         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1387         spin_unlock_irqrestore(&iwqp->lock, flags);
1388
1389         if (attr_mask & IB_QP_STATE) {
1390                 if (issue_modify_qp) {
1391                         ctx_info->rem_endpoint_idx = udp_info->arp_idx;
1392                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1393                                 return -EINVAL;
1394                         spin_lock_irqsave(&iwqp->lock, flags);
1395                         if (iwqp->iwarp_state == info.curr_iwarp_state) {
1396                                 iwqp->iwarp_state = info.next_iwarp_state;
1397                                 iwqp->ibqp_state = attr->qp_state;
1398                         }
1399                         if (iwqp->ibqp_state > IB_QPS_RTS &&
1400                             !iwqp->flush_issued) {
1401                                 iwqp->flush_issued = 1;
1402                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1403                                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_SQ |
1404                                                        IRDMA_FLUSH_RQ |
1405                                                        IRDMA_FLUSH_WAIT);
1406                         } else {
1407                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1408                         }
1409                 } else {
1410                         iwqp->ibqp_state = attr->qp_state;
1411                 }
1412                 if (udata && dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1413                         struct irdma_ucontext *ucontext;
1414
1415                         ucontext = rdma_udata_to_drv_context(udata,
1416                                         struct irdma_ucontext, ibucontext);
1417                         if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1418                             !iwqp->push_wqe_mmap_entry &&
1419                             !irdma_setup_push_mmap_entries(ucontext, iwqp,
1420                                 &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1421                                 uresp.push_valid = 1;
1422                                 uresp.push_offset = iwqp->sc_qp.push_offset;
1423                         }
1424                         ret = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1425                                                udata->outlen));
1426                         if (ret) {
1427                                 irdma_remove_push_mmap_entries(iwqp);
1428                                 ibdev_dbg(&iwdev->ibdev,
1429                                           "VERBS: copy_to_udata failed\n");
1430                                 return ret;
1431                         }
1432                 }
1433         }
1434
1435         return 0;
1436 exit:
1437         spin_unlock_irqrestore(&iwqp->lock, flags);
1438
1439         return ret;
1440 }
1441
1442 /**
1443  * irdma_modify_qp - modify qp request
1444  * @ibqp: qp's pointer for modify
1445  * @attr: access attributes
1446  * @attr_mask: state mask
1447  * @udata: user data
1448  */
1449 int irdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
1450                     struct ib_udata *udata)
1451 {
1452         struct irdma_qp *iwqp = to_iwqp(ibqp);
1453         struct irdma_device *iwdev = iwqp->iwdev;
1454         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
1455         struct irdma_qp_host_ctx_info *ctx_info;
1456         struct irdma_tcp_offload_info *tcp_info;
1457         struct irdma_iwarp_offload_info *offload_info;
1458         struct irdma_modify_qp_info info = {};
1459         struct irdma_modify_qp_resp uresp = {};
1460         struct irdma_modify_qp_req ureq = {};
1461         u8 issue_modify_qp = 0;
1462         u8 dont_wait = 0;
1463         int err;
1464         unsigned long flags;
1465
1466         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1467                 return -EOPNOTSUPP;
1468
1469         ctx_info = &iwqp->ctx_info;
1470         offload_info = &iwqp->iwarp_info;
1471         tcp_info = &iwqp->tcp_info;
1472         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
1473         ibdev_dbg(&iwdev->ibdev,
1474                   "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",
1475                   __builtin_return_address(0), ibqp->qp_num, attr->qp_state,
1476                   iwqp->ibqp_state, iwqp->iwarp_state, iwqp->last_aeq,
1477                   iwqp->hw_tcp_state, iwqp->hw_iwarp_state, attr_mask);
1478
1479         spin_lock_irqsave(&iwqp->lock, flags);
1480         if (attr_mask & IB_QP_STATE) {
1481                 info.curr_iwarp_state = iwqp->iwarp_state;
1482                 switch (attr->qp_state) {
1483                 case IB_QPS_INIT:
1484                 case IB_QPS_RTR:
1485                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1486                                 err = -EINVAL;
1487                                 goto exit;
1488                         }
1489
1490                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
1491                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
1492                                 issue_modify_qp = 1;
1493                         }
1494                         if (iwdev->push_mode && udata &&
1495                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1496                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1497                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1498                                 irdma_alloc_push_page(iwqp);
1499                                 spin_lock_irqsave(&iwqp->lock, flags);
1500                         }
1501                         break;
1502                 case IB_QPS_RTS:
1503                         if (iwqp->iwarp_state > IRDMA_QP_STATE_RTS ||
1504                             !iwqp->cm_id) {
1505                                 err = -EINVAL;
1506                                 goto exit;
1507                         }
1508
1509                         issue_modify_qp = 1;
1510                         iwqp->hw_tcp_state = IRDMA_TCP_STATE_ESTABLISHED;
1511                         iwqp->hte_added = 1;
1512                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
1513                         info.tcp_ctx_valid = true;
1514                         info.ord_valid = true;
1515                         info.arp_cache_idx_valid = true;
1516                         info.cq_num_valid = true;
1517                         break;
1518                 case IB_QPS_SQD:
1519                         if (iwqp->hw_iwarp_state > IRDMA_QP_STATE_RTS) {
1520                                 err = 0;
1521                                 goto exit;
1522                         }
1523
1524                         if (iwqp->iwarp_state == IRDMA_QP_STATE_CLOSING ||
1525                             iwqp->iwarp_state < IRDMA_QP_STATE_RTS) {
1526                                 err = 0;
1527                                 goto exit;
1528                         }
1529
1530                         if (iwqp->iwarp_state > IRDMA_QP_STATE_CLOSING) {
1531                                 err = -EINVAL;
1532                                 goto exit;
1533                         }
1534
1535                         info.next_iwarp_state = IRDMA_QP_STATE_CLOSING;
1536                         issue_modify_qp = 1;
1537                         break;
1538                 case IB_QPS_SQE:
1539                         if (iwqp->iwarp_state >= IRDMA_QP_STATE_TERMINATE) {
1540                                 err = -EINVAL;
1541                                 goto exit;
1542                         }
1543
1544                         info.next_iwarp_state = IRDMA_QP_STATE_TERMINATE;
1545                         issue_modify_qp = 1;
1546                         break;
1547                 case IB_QPS_ERR:
1548                 case IB_QPS_RESET:
1549                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1550                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1551                                 if (udata) {
1552                                         if (ib_copy_from_udata(&ureq, udata,
1553                                             min(sizeof(ureq), udata->inlen)))
1554                                                 return -EINVAL;
1555
1556                                         irdma_flush_wqes(iwqp,
1557                                             (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1558                                             (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1559                                             IRDMA_REFLUSH);
1560                                 }
1561                                 return 0;
1562                         }
1563
1564                         if (iwqp->sc_qp.term_flags) {
1565                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1566                                 irdma_terminate_del_timer(&iwqp->sc_qp);
1567                                 spin_lock_irqsave(&iwqp->lock, flags);
1568                         }
1569                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1570                         if (iwqp->hw_tcp_state > IRDMA_TCP_STATE_CLOSED &&
1571                             iwdev->iw_status &&
1572                             iwqp->hw_tcp_state != IRDMA_TCP_STATE_TIME_WAIT)
1573                                 info.reset_tcp_conn = true;
1574                         else
1575                                 dont_wait = 1;
1576
1577                         issue_modify_qp = 1;
1578                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1579                         break;
1580                 default:
1581                         err = -EINVAL;
1582                         goto exit;
1583                 }
1584
1585                 iwqp->ibqp_state = attr->qp_state;
1586         }
1587         if (attr_mask & IB_QP_ACCESS_FLAGS) {
1588                 ctx_info->iwarp_info_valid = true;
1589                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
1590                         offload_info->wr_rdresp_en = true;
1591                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
1592                         offload_info->wr_rdresp_en = true;
1593                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
1594                         offload_info->rd_en = true;
1595         }
1596
1597         if (ctx_info->iwarp_info_valid) {
1598                 ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1599                 ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1600                 irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1601         }
1602         spin_unlock_irqrestore(&iwqp->lock, flags);
1603
1604         if (attr_mask & IB_QP_STATE) {
1605                 if (issue_modify_qp) {
1606                         ctx_info->rem_endpoint_idx = tcp_info->arp_idx;
1607                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1608                                 return -EINVAL;
1609                 }
1610
1611                 spin_lock_irqsave(&iwqp->lock, flags);
1612                 if (iwqp->iwarp_state == info.curr_iwarp_state) {
1613                         iwqp->iwarp_state = info.next_iwarp_state;
1614                         iwqp->ibqp_state = attr->qp_state;
1615                 }
1616                 spin_unlock_irqrestore(&iwqp->lock, flags);
1617         }
1618
1619         if (issue_modify_qp && iwqp->ibqp_state > IB_QPS_RTS) {
1620                 if (dont_wait) {
1621                         if (iwqp->cm_id && iwqp->hw_tcp_state) {
1622                                 spin_lock_irqsave(&iwqp->lock, flags);
1623                                 iwqp->hw_tcp_state = IRDMA_TCP_STATE_CLOSED;
1624                                 iwqp->last_aeq = IRDMA_AE_RESET_SENT;
1625                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1626                                 irdma_cm_disconn(iwqp);
1627                         }
1628                 } else {
1629                         int close_timer_started;
1630
1631                         spin_lock_irqsave(&iwdev->cm_core.ht_lock, flags);
1632
1633                         if (iwqp->cm_node) {
1634                                 refcount_inc(&iwqp->cm_node->refcnt);
1635                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1636                                 close_timer_started = atomic_inc_return(&iwqp->close_timer_started);
1637                                 if (iwqp->cm_id && close_timer_started == 1)
1638                                         irdma_schedule_cm_timer(iwqp->cm_node,
1639                                                 (struct irdma_puda_buf *)iwqp,
1640                                                 IRDMA_TIMER_TYPE_CLOSE, 1, 0);
1641
1642                                 irdma_rem_ref_cm_node(iwqp->cm_node);
1643                         } else {
1644                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1645                         }
1646                 }
1647         }
1648         if (attr_mask & IB_QP_STATE && udata &&
1649             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1650                 struct irdma_ucontext *ucontext;
1651
1652                 ucontext = rdma_udata_to_drv_context(udata,
1653                                         struct irdma_ucontext, ibucontext);
1654                 if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1655                     !iwqp->push_wqe_mmap_entry &&
1656                     !irdma_setup_push_mmap_entries(ucontext, iwqp,
1657                         &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1658                         uresp.push_valid = 1;
1659                         uresp.push_offset = iwqp->sc_qp.push_offset;
1660                 }
1661
1662                 err = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1663                                        udata->outlen));
1664                 if (err) {
1665                         irdma_remove_push_mmap_entries(iwqp);
1666                         ibdev_dbg(&iwdev->ibdev,
1667                                   "VERBS: copy_to_udata failed\n");
1668                         return err;
1669                 }
1670         }
1671
1672         return 0;
1673 exit:
1674         spin_unlock_irqrestore(&iwqp->lock, flags);
1675
1676         return err;
1677 }
1678
1679 /**
1680  * irdma_cq_free_rsrc - free up resources for cq
1681  * @rf: RDMA PCI function
1682  * @iwcq: cq ptr
1683  */
1684 static void irdma_cq_free_rsrc(struct irdma_pci_f *rf, struct irdma_cq *iwcq)
1685 {
1686         struct irdma_sc_cq *cq = &iwcq->sc_cq;
1687
1688         if (!iwcq->user_mode) {
1689                 dma_free_coherent(rf->sc_dev.hw->device, iwcq->kmem.size,
1690                                   iwcq->kmem.va, iwcq->kmem.pa);
1691                 iwcq->kmem.va = NULL;
1692                 dma_free_coherent(rf->sc_dev.hw->device,
1693                                   iwcq->kmem_shadow.size,
1694                                   iwcq->kmem_shadow.va, iwcq->kmem_shadow.pa);
1695                 iwcq->kmem_shadow.va = NULL;
1696         }
1697
1698         irdma_free_rsrc(rf, rf->allocated_cqs, cq->cq_uk.cq_id);
1699 }
1700
1701 /**
1702  * irdma_free_cqbuf - worker to free a cq buffer
1703  * @work: provides access to the cq buffer to free
1704  */
1705 static void irdma_free_cqbuf(struct work_struct *work)
1706 {
1707         struct irdma_cq_buf *cq_buf = container_of(work, struct irdma_cq_buf, work);
1708
1709         dma_free_coherent(cq_buf->hw->device, cq_buf->kmem_buf.size,
1710                           cq_buf->kmem_buf.va, cq_buf->kmem_buf.pa);
1711         cq_buf->kmem_buf.va = NULL;
1712         kfree(cq_buf);
1713 }
1714
1715 /**
1716  * irdma_process_resize_list - remove resized cq buffers from the resize_list
1717  * @iwcq: cq which owns the resize_list
1718  * @iwdev: irdma device
1719  * @lcqe_buf: the buffer where the last cqe is received
1720  */
1721 static int irdma_process_resize_list(struct irdma_cq *iwcq,
1722                                      struct irdma_device *iwdev,
1723                                      struct irdma_cq_buf *lcqe_buf)
1724 {
1725         struct list_head *tmp_node, *list_node;
1726         struct irdma_cq_buf *cq_buf;
1727         int cnt = 0;
1728
1729         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
1730                 cq_buf = list_entry(list_node, struct irdma_cq_buf, list);
1731                 if (cq_buf == lcqe_buf)
1732                         return cnt;
1733
1734                 list_del(&cq_buf->list);
1735                 queue_work(iwdev->cleanup_wq, &cq_buf->work);
1736                 cnt++;
1737         }
1738
1739         return cnt;
1740 }
1741
1742 /**
1743  * irdma_destroy_cq - destroy cq
1744  * @ib_cq: cq pointer
1745  * @udata: user data
1746  */
1747 static int irdma_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata)
1748 {
1749         struct irdma_device *iwdev = to_iwdev(ib_cq->device);
1750         struct irdma_cq *iwcq = to_iwcq(ib_cq);
1751         struct irdma_sc_cq *cq = &iwcq->sc_cq;
1752         struct irdma_sc_dev *dev = cq->dev;
1753         struct irdma_sc_ceq *ceq = dev->ceq[cq->ceq_id];
1754         struct irdma_ceq *iwceq = container_of(ceq, struct irdma_ceq, sc_ceq);
1755         unsigned long flags;
1756
1757         spin_lock_irqsave(&iwcq->lock, flags);
1758         if (!list_empty(&iwcq->resize_list))
1759                 irdma_process_resize_list(iwcq, iwdev, NULL);
1760         spin_unlock_irqrestore(&iwcq->lock, flags);
1761
1762         irdma_cq_wq_destroy(iwdev->rf, cq);
1763         irdma_cq_free_rsrc(iwdev->rf, iwcq);
1764
1765         spin_lock_irqsave(&iwceq->ce_lock, flags);
1766         irdma_sc_cleanup_ceqes(cq, ceq);
1767         spin_unlock_irqrestore(&iwceq->ce_lock, flags);
1768
1769         return 0;
1770 }
1771
1772 /**
1773  * irdma_resize_cq - resize cq
1774  * @ibcq: cq to be resized
1775  * @entries: desired cq size
1776  * @udata: user data
1777  */
1778 static int irdma_resize_cq(struct ib_cq *ibcq, int entries,
1779                            struct ib_udata *udata)
1780 {
1781         struct irdma_cq *iwcq = to_iwcq(ibcq);
1782         struct irdma_sc_dev *dev = iwcq->sc_cq.dev;
1783         struct irdma_cqp_request *cqp_request;
1784         struct cqp_cmds_info *cqp_info;
1785         struct irdma_modify_cq_info *m_info;
1786         struct irdma_modify_cq_info info = {};
1787         struct irdma_dma_mem kmem_buf;
1788         struct irdma_cq_mr *cqmr_buf;
1789         struct irdma_pbl *iwpbl_buf;
1790         struct irdma_device *iwdev;
1791         struct irdma_pci_f *rf;
1792         struct irdma_cq_buf *cq_buf = NULL;
1793         enum irdma_status_code status = 0;
1794         unsigned long flags;
1795         int ret;
1796
1797         iwdev = to_iwdev(ibcq->device);
1798         rf = iwdev->rf;
1799
1800         if (!(rf->sc_dev.hw_attrs.uk_attrs.feature_flags &
1801             IRDMA_FEATURE_CQ_RESIZE))
1802                 return -EOPNOTSUPP;
1803
1804         if (entries > rf->max_cqe)
1805                 return -EINVAL;
1806
1807         if (!iwcq->user_mode) {
1808                 entries++;
1809                 if (rf->sc_dev.hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
1810                         entries *= 2;
1811         }
1812
1813         info.cq_size = max(entries, 4);
1814
1815         if (info.cq_size == iwcq->sc_cq.cq_uk.cq_size - 1)
1816                 return 0;
1817
1818         if (udata) {
1819                 struct irdma_resize_cq_req req = {};
1820                 struct irdma_ucontext *ucontext =
1821                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
1822                                                   ibucontext);
1823
1824                 /* CQ resize not supported with legacy GEN_1 libi40iw */
1825                 if (ucontext->legacy_mode)
1826                         return -EOPNOTSUPP;
1827
1828                 if (ib_copy_from_udata(&req, udata,
1829                                        min(sizeof(req), udata->inlen)))
1830                         return -EINVAL;
1831
1832                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
1833                 iwpbl_buf = irdma_get_pbl((unsigned long)req.user_cq_buffer,
1834                                           &ucontext->cq_reg_mem_list);
1835                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
1836
1837                 if (!iwpbl_buf)
1838                         return -ENOMEM;
1839
1840                 cqmr_buf = &iwpbl_buf->cq_mr;
1841                 if (iwpbl_buf->pbl_allocated) {
1842                         info.virtual_map = true;
1843                         info.pbl_chunk_size = 1;
1844                         info.first_pm_pbl_idx = cqmr_buf->cq_pbl.idx;
1845                 } else {
1846                         info.cq_pa = cqmr_buf->cq_pbl.addr;
1847                 }
1848         } else {
1849                 /* Kmode CQ resize */
1850                 int rsize;
1851
1852                 rsize = info.cq_size * sizeof(struct irdma_cqe);
1853                 kmem_buf.size = ALIGN(round_up(rsize, 256), 256);
1854                 kmem_buf.va = dma_alloc_coherent(dev->hw->device,
1855                                                  kmem_buf.size, &kmem_buf.pa,
1856                                                  GFP_KERNEL);
1857                 if (!kmem_buf.va)
1858                         return -ENOMEM;
1859
1860                 info.cq_base = kmem_buf.va;
1861                 info.cq_pa = kmem_buf.pa;
1862                 cq_buf = kzalloc(sizeof(*cq_buf), GFP_KERNEL);
1863                 if (!cq_buf) {
1864                         ret = -ENOMEM;
1865                         goto error;
1866                 }
1867         }
1868
1869         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
1870         if (!cqp_request) {
1871                 ret = -ENOMEM;
1872                 goto error;
1873         }
1874
1875         info.shadow_read_threshold = iwcq->sc_cq.shadow_read_threshold;
1876         info.cq_resize = true;
1877
1878         cqp_info = &cqp_request->info;
1879         m_info = &cqp_info->in.u.cq_modify.info;
1880         memcpy(m_info, &info, sizeof(*m_info));
1881
1882         cqp_info->cqp_cmd = IRDMA_OP_CQ_MODIFY;
1883         cqp_info->in.u.cq_modify.cq = &iwcq->sc_cq;
1884         cqp_info->in.u.cq_modify.scratch = (uintptr_t)cqp_request;
1885         cqp_info->post_sq = 1;
1886         status = irdma_handle_cqp_op(rf, cqp_request);
1887         irdma_put_cqp_request(&rf->cqp, cqp_request);
1888         if (status) {
1889                 ret = -EPROTO;
1890                 goto error;
1891         }
1892
1893         spin_lock_irqsave(&iwcq->lock, flags);
1894         if (cq_buf) {
1895                 cq_buf->kmem_buf = iwcq->kmem;
1896                 cq_buf->hw = dev->hw;
1897                 memcpy(&cq_buf->cq_uk, &iwcq->sc_cq.cq_uk, sizeof(cq_buf->cq_uk));
1898                 INIT_WORK(&cq_buf->work, irdma_free_cqbuf);
1899                 list_add_tail(&cq_buf->list, &iwcq->resize_list);
1900                 iwcq->kmem = kmem_buf;
1901         }
1902
1903         irdma_sc_cq_resize(&iwcq->sc_cq, &info);
1904         ibcq->cqe = info.cq_size - 1;
1905         spin_unlock_irqrestore(&iwcq->lock, flags);
1906
1907         return 0;
1908 error:
1909         if (!udata) {
1910                 dma_free_coherent(dev->hw->device, kmem_buf.size, kmem_buf.va,
1911                                   kmem_buf.pa);
1912                 kmem_buf.va = NULL;
1913         }
1914         kfree(cq_buf);
1915
1916         return ret;
1917 }
1918
1919 static inline int cq_validate_flags(u32 flags, u8 hw_rev)
1920 {
1921         /* GEN1 does not support CQ create flags */
1922         if (hw_rev == IRDMA_GEN_1)
1923                 return flags ? -EOPNOTSUPP : 0;
1924
1925         return flags & ~IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION ? -EOPNOTSUPP : 0;
1926 }
1927
1928 /**
1929  * irdma_create_cq - create cq
1930  * @ibcq: CQ allocated
1931  * @attr: attributes for cq
1932  * @udata: user data
1933  */
1934 static int irdma_create_cq(struct ib_cq *ibcq,
1935                            const struct ib_cq_init_attr *attr,
1936                            struct ib_udata *udata)
1937 {
1938         struct ib_device *ibdev = ibcq->device;
1939         struct irdma_device *iwdev = to_iwdev(ibdev);
1940         struct irdma_pci_f *rf = iwdev->rf;
1941         struct irdma_cq *iwcq = to_iwcq(ibcq);
1942         u32 cq_num = 0;
1943         struct irdma_sc_cq *cq;
1944         struct irdma_sc_dev *dev = &rf->sc_dev;
1945         struct irdma_cq_init_info info = {};
1946         enum irdma_status_code status;
1947         struct irdma_cqp_request *cqp_request;
1948         struct cqp_cmds_info *cqp_info;
1949         struct irdma_cq_uk_init_info *ukinfo = &info.cq_uk_init_info;
1950         unsigned long flags;
1951         int err_code;
1952         int entries = attr->cqe;
1953
1954         err_code = cq_validate_flags(attr->flags, dev->hw_attrs.uk_attrs.hw_rev);
1955         if (err_code)
1956                 return err_code;
1957         err_code = irdma_alloc_rsrc(rf, rf->allocated_cqs, rf->max_cq, &cq_num,
1958                                     &rf->next_cq);
1959         if (err_code)
1960                 return err_code;
1961
1962         cq = &iwcq->sc_cq;
1963         cq->back_cq = iwcq;
1964         spin_lock_init(&iwcq->lock);
1965         INIT_LIST_HEAD(&iwcq->resize_list);
1966         info.dev = dev;
1967         ukinfo->cq_size = max(entries, 4);
1968         ukinfo->cq_id = cq_num;
1969         iwcq->ibcq.cqe = info.cq_uk_init_info.cq_size;
1970         if (attr->comp_vector < rf->ceqs_count)
1971                 info.ceq_id = attr->comp_vector;
1972         info.ceq_id_valid = true;
1973         info.ceqe_mask = 1;
1974         info.type = IRDMA_CQ_TYPE_IWARP;
1975         info.vsi = &iwdev->vsi;
1976
1977         if (udata) {
1978                 struct irdma_ucontext *ucontext;
1979                 struct irdma_create_cq_req req = {};
1980                 struct irdma_cq_mr *cqmr;
1981                 struct irdma_pbl *iwpbl;
1982                 struct irdma_pbl *iwpbl_shadow;
1983                 struct irdma_cq_mr *cqmr_shadow;
1984
1985                 iwcq->user_mode = true;
1986                 ucontext =
1987                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
1988                                                   ibucontext);
1989                 if (ib_copy_from_udata(&req, udata,
1990                                        min(sizeof(req), udata->inlen))) {
1991                         err_code = -EFAULT;
1992                         goto cq_free_rsrc;
1993                 }
1994
1995                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
1996                 iwpbl = irdma_get_pbl((unsigned long)req.user_cq_buf,
1997                                       &ucontext->cq_reg_mem_list);
1998                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
1999                 if (!iwpbl) {
2000                         err_code = -EPROTO;
2001                         goto cq_free_rsrc;
2002                 }
2003
2004                 iwcq->iwpbl = iwpbl;
2005                 iwcq->cq_mem_size = 0;
2006                 cqmr = &iwpbl->cq_mr;
2007
2008                 if (rf->sc_dev.hw_attrs.uk_attrs.feature_flags &
2009                     IRDMA_FEATURE_CQ_RESIZE && !ucontext->legacy_mode) {
2010                         spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2011                         iwpbl_shadow = irdma_get_pbl(
2012                                         (unsigned long)req.user_shadow_area,
2013                                         &ucontext->cq_reg_mem_list);
2014                         spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2015
2016                         if (!iwpbl_shadow) {
2017                                 err_code = -EPROTO;
2018                                 goto cq_free_rsrc;
2019                         }
2020                         iwcq->iwpbl_shadow = iwpbl_shadow;
2021                         cqmr_shadow = &iwpbl_shadow->cq_mr;
2022                         info.shadow_area_pa = cqmr_shadow->cq_pbl.addr;
2023                         cqmr->split = true;
2024                 } else {
2025                         info.shadow_area_pa = cqmr->shadow;
2026                 }
2027                 if (iwpbl->pbl_allocated) {
2028                         info.virtual_map = true;
2029                         info.pbl_chunk_size = 1;
2030                         info.first_pm_pbl_idx = cqmr->cq_pbl.idx;
2031                 } else {
2032                         info.cq_base_pa = cqmr->cq_pbl.addr;
2033                 }
2034         } else {
2035                 /* Kmode allocations */
2036                 int rsize;
2037
2038                 if (entries > rf->max_cqe) {
2039                         err_code = -EINVAL;
2040                         goto cq_free_rsrc;
2041                 }
2042
2043                 entries++;
2044                 if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
2045                         entries *= 2;
2046                 ukinfo->cq_size = entries;
2047
2048                 rsize = info.cq_uk_init_info.cq_size * sizeof(struct irdma_cqe);
2049                 iwcq->kmem.size = ALIGN(round_up(rsize, 256), 256);
2050                 iwcq->kmem.va = dma_alloc_coherent(dev->hw->device,
2051                                                    iwcq->kmem.size,
2052                                                    &iwcq->kmem.pa, GFP_KERNEL);
2053                 if (!iwcq->kmem.va) {
2054                         err_code = -ENOMEM;
2055                         goto cq_free_rsrc;
2056                 }
2057
2058                 iwcq->kmem_shadow.size = ALIGN(IRDMA_SHADOW_AREA_SIZE << 3,
2059                                                64);
2060                 iwcq->kmem_shadow.va = dma_alloc_coherent(dev->hw->device,
2061                                                           iwcq->kmem_shadow.size,
2062                                                           &iwcq->kmem_shadow.pa,
2063                                                           GFP_KERNEL);
2064                 if (!iwcq->kmem_shadow.va) {
2065                         err_code = -ENOMEM;
2066                         goto cq_free_rsrc;
2067                 }
2068                 info.shadow_area_pa = iwcq->kmem_shadow.pa;
2069                 ukinfo->shadow_area = iwcq->kmem_shadow.va;
2070                 ukinfo->cq_base = iwcq->kmem.va;
2071                 info.cq_base_pa = iwcq->kmem.pa;
2072         }
2073
2074         if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
2075                 info.shadow_read_threshold = min(info.cq_uk_init_info.cq_size / 2,
2076                                                  (u32)IRDMA_MAX_CQ_READ_THRESH);
2077
2078         if (irdma_sc_cq_init(cq, &info)) {
2079                 ibdev_dbg(&iwdev->ibdev, "VERBS: init cq fail\n");
2080                 err_code = -EPROTO;
2081                 goto cq_free_rsrc;
2082         }
2083
2084         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
2085         if (!cqp_request) {
2086                 err_code = -ENOMEM;
2087                 goto cq_free_rsrc;
2088         }
2089
2090         cqp_info = &cqp_request->info;
2091         cqp_info->cqp_cmd = IRDMA_OP_CQ_CREATE;
2092         cqp_info->post_sq = 1;
2093         cqp_info->in.u.cq_create.cq = cq;
2094         cqp_info->in.u.cq_create.check_overflow = true;
2095         cqp_info->in.u.cq_create.scratch = (uintptr_t)cqp_request;
2096         status = irdma_handle_cqp_op(rf, cqp_request);
2097         irdma_put_cqp_request(&rf->cqp, cqp_request);
2098         if (status) {
2099                 err_code = -ENOMEM;
2100                 goto cq_free_rsrc;
2101         }
2102
2103         if (udata) {
2104                 struct irdma_create_cq_resp resp = {};
2105
2106                 resp.cq_id = info.cq_uk_init_info.cq_id;
2107                 resp.cq_size = info.cq_uk_init_info.cq_size;
2108                 if (ib_copy_to_udata(udata, &resp,
2109                                      min(sizeof(resp), udata->outlen))) {
2110                         ibdev_dbg(&iwdev->ibdev,
2111                                   "VERBS: copy to user data\n");
2112                         err_code = -EPROTO;
2113                         goto cq_destroy;
2114                 }
2115         }
2116         return 0;
2117 cq_destroy:
2118         irdma_cq_wq_destroy(rf, cq);
2119 cq_free_rsrc:
2120         irdma_cq_free_rsrc(rf, iwcq);
2121
2122         return err_code;
2123 }
2124
2125 /**
2126  * irdma_get_mr_access - get hw MR access permissions from IB access flags
2127  * @access: IB access flags
2128  */
2129 static inline u16 irdma_get_mr_access(int access)
2130 {
2131         u16 hw_access = 0;
2132
2133         hw_access |= (access & IB_ACCESS_LOCAL_WRITE) ?
2134                      IRDMA_ACCESS_FLAGS_LOCALWRITE : 0;
2135         hw_access |= (access & IB_ACCESS_REMOTE_WRITE) ?
2136                      IRDMA_ACCESS_FLAGS_REMOTEWRITE : 0;
2137         hw_access |= (access & IB_ACCESS_REMOTE_READ) ?
2138                      IRDMA_ACCESS_FLAGS_REMOTEREAD : 0;
2139         hw_access |= (access & IB_ACCESS_MW_BIND) ?
2140                      IRDMA_ACCESS_FLAGS_BIND_WINDOW : 0;
2141         hw_access |= (access & IB_ZERO_BASED) ?
2142                      IRDMA_ACCESS_FLAGS_ZERO_BASED : 0;
2143         hw_access |= IRDMA_ACCESS_FLAGS_LOCALREAD;
2144
2145         return hw_access;
2146 }
2147
2148 /**
2149  * irdma_free_stag - free stag resource
2150  * @iwdev: irdma device
2151  * @stag: stag to free
2152  */
2153 static void irdma_free_stag(struct irdma_device *iwdev, u32 stag)
2154 {
2155         u32 stag_idx;
2156
2157         stag_idx = (stag & iwdev->rf->mr_stagmask) >> IRDMA_CQPSQ_STAG_IDX_S;
2158         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_mrs, stag_idx);
2159 }
2160
2161 /**
2162  * irdma_create_stag - create random stag
2163  * @iwdev: irdma device
2164  */
2165 static u32 irdma_create_stag(struct irdma_device *iwdev)
2166 {
2167         u32 stag = 0;
2168         u32 stag_index = 0;
2169         u32 next_stag_index;
2170         u32 driver_key;
2171         u32 random;
2172         u8 consumer_key;
2173         int ret;
2174
2175         get_random_bytes(&random, sizeof(random));
2176         consumer_key = (u8)random;
2177
2178         driver_key = random & ~iwdev->rf->mr_stagmask;
2179         next_stag_index = (random & iwdev->rf->mr_stagmask) >> 8;
2180         next_stag_index %= iwdev->rf->max_mr;
2181
2182         ret = irdma_alloc_rsrc(iwdev->rf, iwdev->rf->allocated_mrs,
2183                                iwdev->rf->max_mr, &stag_index,
2184                                &next_stag_index);
2185         if (ret)
2186                 return stag;
2187         stag = stag_index << IRDMA_CQPSQ_STAG_IDX_S;
2188         stag |= driver_key;
2189         stag += (u32)consumer_key;
2190
2191         return stag;
2192 }
2193
2194 /**
2195  * irdma_next_pbl_addr - Get next pbl address
2196  * @pbl: pointer to a pble
2197  * @pinfo: info pointer
2198  * @idx: index
2199  */
2200 static inline u64 *irdma_next_pbl_addr(u64 *pbl, struct irdma_pble_info **pinfo,
2201                                        u32 *idx)
2202 {
2203         *idx += 1;
2204         if (!(*pinfo) || *idx != (*pinfo)->cnt)
2205                 return ++pbl;
2206         *idx = 0;
2207         (*pinfo)++;
2208
2209         return (*pinfo)->addr;
2210 }
2211
2212 /**
2213  * irdma_copy_user_pgaddrs - copy user page address to pble's os locally
2214  * @iwmr: iwmr for IB's user page addresses
2215  * @pbl: ple pointer to save 1 level or 0 level pble
2216  * @level: indicated level 0, 1 or 2
2217  */
2218 static void irdma_copy_user_pgaddrs(struct irdma_mr *iwmr, u64 *pbl,
2219                                     enum irdma_pble_level level)
2220 {
2221         struct ib_umem *region = iwmr->region;
2222         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2223         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2224         struct irdma_pble_info *pinfo;
2225         struct ib_block_iter biter;
2226         u32 idx = 0;
2227         u32 pbl_cnt = 0;
2228
2229         pinfo = (level == PBLE_LEVEL_1) ? NULL : palloc->level2.leaf;
2230
2231         if (iwmr->type == IRDMA_MEMREG_TYPE_QP)
2232                 iwpbl->qp_mr.sq_page = sg_page(region->sgt_append.sgt.sgl);
2233
2234         rdma_umem_for_each_dma_block(region, &biter, iwmr->page_size) {
2235                 *pbl = rdma_block_iter_dma_address(&biter);
2236                 if (++pbl_cnt == palloc->total_cnt)
2237                         break;
2238                 pbl = irdma_next_pbl_addr(pbl, &pinfo, &idx);
2239         }
2240 }
2241
2242 /**
2243  * irdma_check_mem_contiguous - check if pbls stored in arr are contiguous
2244  * @arr: lvl1 pbl array
2245  * @npages: page count
2246  * @pg_size: page size
2247  *
2248  */
2249 static bool irdma_check_mem_contiguous(u64 *arr, u32 npages, u32 pg_size)
2250 {
2251         u32 pg_idx;
2252
2253         for (pg_idx = 0; pg_idx < npages; pg_idx++) {
2254                 if ((*arr + (pg_size * pg_idx)) != arr[pg_idx])
2255                         return false;
2256         }
2257
2258         return true;
2259 }
2260
2261 /**
2262  * irdma_check_mr_contiguous - check if MR is physically contiguous
2263  * @palloc: pbl allocation struct
2264  * @pg_size: page size
2265  */
2266 static bool irdma_check_mr_contiguous(struct irdma_pble_alloc *palloc,
2267                                       u32 pg_size)
2268 {
2269         struct irdma_pble_level2 *lvl2 = &palloc->level2;
2270         struct irdma_pble_info *leaf = lvl2->leaf;
2271         u64 *arr = NULL;
2272         u64 *start_addr = NULL;
2273         int i;
2274         bool ret;
2275
2276         if (palloc->level == PBLE_LEVEL_1) {
2277                 arr = palloc->level1.addr;
2278                 ret = irdma_check_mem_contiguous(arr, palloc->total_cnt,
2279                                                  pg_size);
2280                 return ret;
2281         }
2282
2283         start_addr = leaf->addr;
2284
2285         for (i = 0; i < lvl2->leaf_cnt; i++, leaf++) {
2286                 arr = leaf->addr;
2287                 if ((*start_addr + (i * pg_size * PBLE_PER_PAGE)) != *arr)
2288                         return false;
2289                 ret = irdma_check_mem_contiguous(arr, leaf->cnt, pg_size);
2290                 if (!ret)
2291                         return false;
2292         }
2293
2294         return true;
2295 }
2296
2297 /**
2298  * irdma_setup_pbles - copy user pg address to pble's
2299  * @rf: RDMA PCI function
2300  * @iwmr: mr pointer for this memory registration
2301  * @use_pbles: flag if to use pble's
2302  */
2303 static int irdma_setup_pbles(struct irdma_pci_f *rf, struct irdma_mr *iwmr,
2304                              bool use_pbles)
2305 {
2306         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2307         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2308         struct irdma_pble_info *pinfo;
2309         u64 *pbl;
2310         enum irdma_status_code status;
2311         enum irdma_pble_level level = PBLE_LEVEL_1;
2312
2313         if (use_pbles) {
2314                 status = irdma_get_pble(rf->pble_rsrc, palloc, iwmr->page_cnt,
2315                                         false);
2316                 if (status)
2317                         return -ENOMEM;
2318
2319                 iwpbl->pbl_allocated = true;
2320                 level = palloc->level;
2321                 pinfo = (level == PBLE_LEVEL_1) ? &palloc->level1 :
2322                                                   palloc->level2.leaf;
2323                 pbl = pinfo->addr;
2324         } else {
2325                 pbl = iwmr->pgaddrmem;
2326         }
2327
2328         irdma_copy_user_pgaddrs(iwmr, pbl, level);
2329
2330         if (use_pbles)
2331                 iwmr->pgaddrmem[0] = *pbl;
2332
2333         return 0;
2334 }
2335
2336 /**
2337  * irdma_handle_q_mem - handle memory for qp and cq
2338  * @iwdev: irdma device
2339  * @req: information for q memory management
2340  * @iwpbl: pble struct
2341  * @use_pbles: flag to use pble
2342  */
2343 static int irdma_handle_q_mem(struct irdma_device *iwdev,
2344                               struct irdma_mem_reg_req *req,
2345                               struct irdma_pbl *iwpbl, bool use_pbles)
2346 {
2347         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2348         struct irdma_mr *iwmr = iwpbl->iwmr;
2349         struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;
2350         struct irdma_cq_mr *cqmr = &iwpbl->cq_mr;
2351         struct irdma_hmc_pble *hmc_p;
2352         u64 *arr = iwmr->pgaddrmem;
2353         u32 pg_size, total;
2354         int err = 0;
2355         bool ret = true;
2356
2357         pg_size = iwmr->page_size;
2358         err = irdma_setup_pbles(iwdev->rf, iwmr, use_pbles);
2359         if (err)
2360                 return err;
2361
2362         if (use_pbles && palloc->level != PBLE_LEVEL_1) {
2363                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2364                 iwpbl->pbl_allocated = false;
2365                 return -ENOMEM;
2366         }
2367
2368         if (use_pbles)
2369                 arr = palloc->level1.addr;
2370
2371         switch (iwmr->type) {
2372         case IRDMA_MEMREG_TYPE_QP:
2373                 total = req->sq_pages + req->rq_pages;
2374                 hmc_p = &qpmr->sq_pbl;
2375                 qpmr->shadow = (dma_addr_t)arr[total];
2376
2377                 if (use_pbles) {
2378                         ret = irdma_check_mem_contiguous(arr, req->sq_pages,
2379                                                          pg_size);
2380                         if (ret)
2381                                 ret = irdma_check_mem_contiguous(&arr[req->sq_pages],
2382                                                                  req->rq_pages,
2383                                                                  pg_size);
2384                 }
2385
2386                 if (!ret) {
2387                         hmc_p->idx = palloc->level1.idx;
2388                         hmc_p = &qpmr->rq_pbl;
2389                         hmc_p->idx = palloc->level1.idx + req->sq_pages;
2390                 } else {
2391                         hmc_p->addr = arr[0];
2392                         hmc_p = &qpmr->rq_pbl;
2393                         hmc_p->addr = arr[req->sq_pages];
2394                 }
2395                 break;
2396         case IRDMA_MEMREG_TYPE_CQ:
2397                 hmc_p = &cqmr->cq_pbl;
2398
2399                 if (!cqmr->split)
2400                         cqmr->shadow = (dma_addr_t)arr[req->cq_pages];
2401
2402                 if (use_pbles)
2403                         ret = irdma_check_mem_contiguous(arr, req->cq_pages,
2404                                                          pg_size);
2405
2406                 if (!ret)
2407                         hmc_p->idx = palloc->level1.idx;
2408                 else
2409                         hmc_p->addr = arr[0];
2410         break;
2411         default:
2412                 ibdev_dbg(&iwdev->ibdev, "VERBS: MR type error\n");
2413                 err = -EINVAL;
2414         }
2415
2416         if (use_pbles && ret) {
2417                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2418                 iwpbl->pbl_allocated = false;
2419         }
2420
2421         return err;
2422 }
2423
2424 /**
2425  * irdma_hw_alloc_mw - create the hw memory window
2426  * @iwdev: irdma device
2427  * @iwmr: pointer to memory window info
2428  */
2429 static int irdma_hw_alloc_mw(struct irdma_device *iwdev, struct irdma_mr *iwmr)
2430 {
2431         struct irdma_mw_alloc_info *info;
2432         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2433         struct irdma_cqp_request *cqp_request;
2434         struct cqp_cmds_info *cqp_info;
2435         enum irdma_status_code status;
2436
2437         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2438         if (!cqp_request)
2439                 return -ENOMEM;
2440
2441         cqp_info = &cqp_request->info;
2442         info = &cqp_info->in.u.mw_alloc.info;
2443         memset(info, 0, sizeof(*info));
2444         if (iwmr->ibmw.type == IB_MW_TYPE_1)
2445                 info->mw_wide = true;
2446
2447         info->page_size = PAGE_SIZE;
2448         info->mw_stag_index = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2449         info->pd_id = iwpd->sc_pd.pd_id;
2450         info->remote_access = true;
2451         cqp_info->cqp_cmd = IRDMA_OP_MW_ALLOC;
2452         cqp_info->post_sq = 1;
2453         cqp_info->in.u.mw_alloc.dev = &iwdev->rf->sc_dev;
2454         cqp_info->in.u.mw_alloc.scratch = (uintptr_t)cqp_request;
2455         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2456         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2457
2458         return status ? -ENOMEM : 0;
2459 }
2460
2461 /**
2462  * irdma_alloc_mw - Allocate memory window
2463  * @ibmw: Memory Window
2464  * @udata: user data pointer
2465  */
2466 static int irdma_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata)
2467 {
2468         struct irdma_device *iwdev = to_iwdev(ibmw->device);
2469         struct irdma_mr *iwmr = to_iwmw(ibmw);
2470         int err_code;
2471         u32 stag;
2472
2473         stag = irdma_create_stag(iwdev);
2474         if (!stag)
2475                 return -ENOMEM;
2476
2477         iwmr->stag = stag;
2478         ibmw->rkey = stag;
2479
2480         err_code = irdma_hw_alloc_mw(iwdev, iwmr);
2481         if (err_code) {
2482                 irdma_free_stag(iwdev, stag);
2483                 return err_code;
2484         }
2485
2486         return 0;
2487 }
2488
2489 /**
2490  * irdma_dealloc_mw - Dealloc memory window
2491  * @ibmw: memory window structure.
2492  */
2493 static int irdma_dealloc_mw(struct ib_mw *ibmw)
2494 {
2495         struct ib_pd *ibpd = ibmw->pd;
2496         struct irdma_pd *iwpd = to_iwpd(ibpd);
2497         struct irdma_mr *iwmr = to_iwmr((struct ib_mr *)ibmw);
2498         struct irdma_device *iwdev = to_iwdev(ibmw->device);
2499         struct irdma_cqp_request *cqp_request;
2500         struct cqp_cmds_info *cqp_info;
2501         struct irdma_dealloc_stag_info *info;
2502
2503         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2504         if (!cqp_request)
2505                 return -ENOMEM;
2506
2507         cqp_info = &cqp_request->info;
2508         info = &cqp_info->in.u.dealloc_stag.info;
2509         memset(info, 0, sizeof(*info));
2510         info->pd_id = iwpd->sc_pd.pd_id & 0x00007fff;
2511         info->stag_idx = ibmw->rkey >> IRDMA_CQPSQ_STAG_IDX_S;
2512         info->mr = false;
2513         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
2514         cqp_info->post_sq = 1;
2515         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
2516         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
2517         irdma_handle_cqp_op(iwdev->rf, cqp_request);
2518         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2519         irdma_free_stag(iwdev, iwmr->stag);
2520
2521         return 0;
2522 }
2523
2524 /**
2525  * irdma_hw_alloc_stag - cqp command to allocate stag
2526  * @iwdev: irdma device
2527  * @iwmr: irdma mr pointer
2528  */
2529 static int irdma_hw_alloc_stag(struct irdma_device *iwdev,
2530                                struct irdma_mr *iwmr)
2531 {
2532         struct irdma_allocate_stag_info *info;
2533         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2534         enum irdma_status_code status;
2535         int err = 0;
2536         struct irdma_cqp_request *cqp_request;
2537         struct cqp_cmds_info *cqp_info;
2538
2539         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2540         if (!cqp_request)
2541                 return -ENOMEM;
2542
2543         cqp_info = &cqp_request->info;
2544         info = &cqp_info->in.u.alloc_stag.info;
2545         memset(info, 0, sizeof(*info));
2546         info->page_size = PAGE_SIZE;
2547         info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2548         info->pd_id = iwpd->sc_pd.pd_id;
2549         info->total_len = iwmr->len;
2550         info->remote_access = true;
2551         cqp_info->cqp_cmd = IRDMA_OP_ALLOC_STAG;
2552         cqp_info->post_sq = 1;
2553         cqp_info->in.u.alloc_stag.dev = &iwdev->rf->sc_dev;
2554         cqp_info->in.u.alloc_stag.scratch = (uintptr_t)cqp_request;
2555         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2556         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2557         if (status)
2558                 err = -ENOMEM;
2559
2560         return err;
2561 }
2562
2563 /**
2564  * irdma_alloc_mr - register stag for fast memory registration
2565  * @pd: ibpd pointer
2566  * @mr_type: memory for stag registrion
2567  * @max_num_sg: man number of pages
2568  */
2569 static struct ib_mr *irdma_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
2570                                     u32 max_num_sg)
2571 {
2572         struct irdma_device *iwdev = to_iwdev(pd->device);
2573         struct irdma_pble_alloc *palloc;
2574         struct irdma_pbl *iwpbl;
2575         struct irdma_mr *iwmr;
2576         enum irdma_status_code status;
2577         u32 stag;
2578         int err_code = -ENOMEM;
2579
2580         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2581         if (!iwmr)
2582                 return ERR_PTR(-ENOMEM);
2583
2584         stag = irdma_create_stag(iwdev);
2585         if (!stag) {
2586                 err_code = -ENOMEM;
2587                 goto err;
2588         }
2589
2590         iwmr->stag = stag;
2591         iwmr->ibmr.rkey = stag;
2592         iwmr->ibmr.lkey = stag;
2593         iwmr->ibmr.pd = pd;
2594         iwmr->ibmr.device = pd->device;
2595         iwpbl = &iwmr->iwpbl;
2596         iwpbl->iwmr = iwmr;
2597         iwmr->type = IRDMA_MEMREG_TYPE_MEM;
2598         palloc = &iwpbl->pble_alloc;
2599         iwmr->page_cnt = max_num_sg;
2600         status = irdma_get_pble(iwdev->rf->pble_rsrc, palloc, iwmr->page_cnt,
2601                                 true);
2602         if (status)
2603                 goto err_get_pble;
2604
2605         err_code = irdma_hw_alloc_stag(iwdev, iwmr);
2606         if (err_code)
2607                 goto err_alloc_stag;
2608
2609         iwpbl->pbl_allocated = true;
2610
2611         return &iwmr->ibmr;
2612 err_alloc_stag:
2613         irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2614 err_get_pble:
2615         irdma_free_stag(iwdev, stag);
2616 err:
2617         kfree(iwmr);
2618
2619         return ERR_PTR(err_code);
2620 }
2621
2622 /**
2623  * irdma_set_page - populate pbl list for fmr
2624  * @ibmr: ib mem to access iwarp mr pointer
2625  * @addr: page dma address fro pbl list
2626  */
2627 static int irdma_set_page(struct ib_mr *ibmr, u64 addr)
2628 {
2629         struct irdma_mr *iwmr = to_iwmr(ibmr);
2630         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2631         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2632         u64 *pbl;
2633
2634         if (unlikely(iwmr->npages == iwmr->page_cnt))
2635                 return -ENOMEM;
2636
2637         pbl = palloc->level1.addr;
2638         pbl[iwmr->npages++] = addr;
2639
2640         return 0;
2641 }
2642
2643 /**
2644  * irdma_map_mr_sg - map of sg list for fmr
2645  * @ibmr: ib mem to access iwarp mr pointer
2646  * @sg: scatter gather list
2647  * @sg_nents: number of sg pages
2648  * @sg_offset: scatter gather list for fmr
2649  */
2650 static int irdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
2651                            int sg_nents, unsigned int *sg_offset)
2652 {
2653         struct irdma_mr *iwmr = to_iwmr(ibmr);
2654
2655         iwmr->npages = 0;
2656
2657         return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, irdma_set_page);
2658 }
2659
2660 /**
2661  * irdma_hwreg_mr - send cqp command for memory registration
2662  * @iwdev: irdma device
2663  * @iwmr: irdma mr pointer
2664  * @access: access for MR
2665  */
2666 static int irdma_hwreg_mr(struct irdma_device *iwdev, struct irdma_mr *iwmr,
2667                           u16 access)
2668 {
2669         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2670         struct irdma_reg_ns_stag_info *stag_info;
2671         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2672         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2673         enum irdma_status_code status;
2674         int err = 0;
2675         struct irdma_cqp_request *cqp_request;
2676         struct cqp_cmds_info *cqp_info;
2677
2678         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2679         if (!cqp_request)
2680                 return -ENOMEM;
2681
2682         cqp_info = &cqp_request->info;
2683         stag_info = &cqp_info->in.u.mr_reg_non_shared.info;
2684         memset(stag_info, 0, sizeof(*stag_info));
2685         stag_info->va = iwpbl->user_base;
2686         stag_info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2687         stag_info->stag_key = (u8)iwmr->stag;
2688         stag_info->total_len = iwmr->len;
2689         stag_info->access_rights = irdma_get_mr_access(access);
2690         stag_info->pd_id = iwpd->sc_pd.pd_id;
2691         if (stag_info->access_rights & IRDMA_ACCESS_FLAGS_ZERO_BASED)
2692                 stag_info->addr_type = IRDMA_ADDR_TYPE_ZERO_BASED;
2693         else
2694                 stag_info->addr_type = IRDMA_ADDR_TYPE_VA_BASED;
2695         stag_info->page_size = iwmr->page_size;
2696
2697         if (iwpbl->pbl_allocated) {
2698                 if (palloc->level == PBLE_LEVEL_1) {
2699                         stag_info->first_pm_pbl_index = palloc->level1.idx;
2700                         stag_info->chunk_size = 1;
2701                 } else {
2702                         stag_info->first_pm_pbl_index = palloc->level2.root.idx;
2703                         stag_info->chunk_size = 3;
2704                 }
2705         } else {
2706                 stag_info->reg_addr_pa = iwmr->pgaddrmem[0];
2707         }
2708
2709         cqp_info->cqp_cmd = IRDMA_OP_MR_REG_NON_SHARED;
2710         cqp_info->post_sq = 1;
2711         cqp_info->in.u.mr_reg_non_shared.dev = &iwdev->rf->sc_dev;
2712         cqp_info->in.u.mr_reg_non_shared.scratch = (uintptr_t)cqp_request;
2713         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2714         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2715         if (status)
2716                 err = -ENOMEM;
2717
2718         return err;
2719 }
2720
2721 /**
2722  * irdma_reg_user_mr - Register a user memory region
2723  * @pd: ptr of pd
2724  * @start: virtual start address
2725  * @len: length of mr
2726  * @virt: virtual address
2727  * @access: access of mr
2728  * @udata: user data
2729  */
2730 static struct ib_mr *irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,
2731                                        u64 virt, int access,
2732                                        struct ib_udata *udata)
2733 {
2734         struct irdma_device *iwdev = to_iwdev(pd->device);
2735         struct irdma_ucontext *ucontext;
2736         struct irdma_pble_alloc *palloc;
2737         struct irdma_pbl *iwpbl;
2738         struct irdma_mr *iwmr;
2739         struct ib_umem *region;
2740         struct irdma_mem_reg_req req;
2741         u32 total, stag = 0;
2742         u8 shadow_pgcnt = 1;
2743         bool use_pbles = false;
2744         unsigned long flags;
2745         int err = -EINVAL;
2746         int ret;
2747
2748         if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size)
2749                 return ERR_PTR(-EINVAL);
2750
2751         region = ib_umem_get(pd->device, start, len, access);
2752
2753         if (IS_ERR(region)) {
2754                 ibdev_dbg(&iwdev->ibdev,
2755                           "VERBS: Failed to create ib_umem region\n");
2756                 return (struct ib_mr *)region;
2757         }
2758
2759         if (ib_copy_from_udata(&req, udata, min(sizeof(req), udata->inlen))) {
2760                 ib_umem_release(region);
2761                 return ERR_PTR(-EFAULT);
2762         }
2763
2764         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2765         if (!iwmr) {
2766                 ib_umem_release(region);
2767                 return ERR_PTR(-ENOMEM);
2768         }
2769
2770         iwpbl = &iwmr->iwpbl;
2771         iwpbl->iwmr = iwmr;
2772         iwmr->region = region;
2773         iwmr->ibmr.pd = pd;
2774         iwmr->ibmr.device = pd->device;
2775         iwmr->ibmr.iova = virt;
2776         iwmr->page_size = PAGE_SIZE;
2777
2778         if (req.reg_type == IRDMA_MEMREG_TYPE_MEM) {
2779                 iwmr->page_size = ib_umem_find_best_pgsz(region,
2780                                                          SZ_4K | SZ_2M | SZ_1G,
2781                                                          virt);
2782                 if (unlikely(!iwmr->page_size)) {
2783                         kfree(iwmr);
2784                         ib_umem_release(region);
2785                         return ERR_PTR(-EOPNOTSUPP);
2786                 }
2787         }
2788         iwmr->len = region->length;
2789         iwpbl->user_base = virt;
2790         palloc = &iwpbl->pble_alloc;
2791         iwmr->type = req.reg_type;
2792         iwmr->page_cnt = ib_umem_num_dma_blocks(region, iwmr->page_size);
2793
2794         switch (req.reg_type) {
2795         case IRDMA_MEMREG_TYPE_QP:
2796                 total = req.sq_pages + req.rq_pages + shadow_pgcnt;
2797                 if (total > iwmr->page_cnt) {
2798                         err = -EINVAL;
2799                         goto error;
2800                 }
2801                 total = req.sq_pages + req.rq_pages;
2802                 use_pbles = (total > 2);
2803                 err = irdma_handle_q_mem(iwdev, &req, iwpbl, use_pbles);
2804                 if (err)
2805                         goto error;
2806
2807                 ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext,
2808                                                      ibucontext);
2809                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
2810                 list_add_tail(&iwpbl->list, &ucontext->qp_reg_mem_list);
2811                 iwpbl->on_list = true;
2812                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
2813                 break;
2814         case IRDMA_MEMREG_TYPE_CQ:
2815                 if (iwdev->rf->sc_dev.hw_attrs.uk_attrs.feature_flags & IRDMA_FEATURE_CQ_RESIZE)
2816                         shadow_pgcnt = 0;
2817                 total = req.cq_pages + shadow_pgcnt;
2818                 if (total > iwmr->page_cnt) {
2819                         err = -EINVAL;
2820                         goto error;
2821                 }
2822
2823                 use_pbles = (req.cq_pages > 1);
2824                 err = irdma_handle_q_mem(iwdev, &req, iwpbl, use_pbles);
2825                 if (err)
2826                         goto error;
2827
2828                 ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext,
2829                                                      ibucontext);
2830                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2831                 list_add_tail(&iwpbl->list, &ucontext->cq_reg_mem_list);
2832                 iwpbl->on_list = true;
2833                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2834                 break;
2835         case IRDMA_MEMREG_TYPE_MEM:
2836                 use_pbles = (iwmr->page_cnt != 1);
2837
2838                 err = irdma_setup_pbles(iwdev->rf, iwmr, use_pbles);
2839                 if (err)
2840                         goto error;
2841
2842                 if (use_pbles) {
2843                         ret = irdma_check_mr_contiguous(palloc,
2844                                                         iwmr->page_size);
2845                         if (ret) {
2846                                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2847                                 iwpbl->pbl_allocated = false;
2848                         }
2849                 }
2850
2851                 stag = irdma_create_stag(iwdev);
2852                 if (!stag) {
2853                         err = -ENOMEM;
2854                         goto error;
2855                 }
2856
2857                 iwmr->stag = stag;
2858                 iwmr->ibmr.rkey = stag;
2859                 iwmr->ibmr.lkey = stag;
2860                 err = irdma_hwreg_mr(iwdev, iwmr, access);
2861                 if (err) {
2862                         irdma_free_stag(iwdev, stag);
2863                         goto error;
2864                 }
2865
2866                 break;
2867         default:
2868                 goto error;
2869         }
2870
2871         iwmr->type = req.reg_type;
2872
2873         return &iwmr->ibmr;
2874
2875 error:
2876         if (palloc->level != PBLE_LEVEL_0 && iwpbl->pbl_allocated)
2877                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2878         ib_umem_release(region);
2879         kfree(iwmr);
2880
2881         return ERR_PTR(err);
2882 }
2883
2884 /**
2885  * irdma_reg_phys_mr - register kernel physical memory
2886  * @pd: ibpd pointer
2887  * @addr: physical address of memory to register
2888  * @size: size of memory to register
2889  * @access: Access rights
2890  * @iova_start: start of virtual address for physical buffers
2891  */
2892 struct ib_mr *irdma_reg_phys_mr(struct ib_pd *pd, u64 addr, u64 size, int access,
2893                                 u64 *iova_start)
2894 {
2895         struct irdma_device *iwdev = to_iwdev(pd->device);
2896         struct irdma_pbl *iwpbl;
2897         struct irdma_mr *iwmr;
2898         enum irdma_status_code status;
2899         u32 stag;
2900         int ret;
2901
2902         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2903         if (!iwmr)
2904                 return ERR_PTR(-ENOMEM);
2905
2906         iwmr->ibmr.pd = pd;
2907         iwmr->ibmr.device = pd->device;
2908         iwpbl = &iwmr->iwpbl;
2909         iwpbl->iwmr = iwmr;
2910         iwmr->type = IRDMA_MEMREG_TYPE_MEM;
2911         iwpbl->user_base = *iova_start;
2912         stag = irdma_create_stag(iwdev);
2913         if (!stag) {
2914                 ret = -ENOMEM;
2915                 goto err;
2916         }
2917
2918         iwmr->stag = stag;
2919         iwmr->ibmr.iova = *iova_start;
2920         iwmr->ibmr.rkey = stag;
2921         iwmr->ibmr.lkey = stag;
2922         iwmr->page_cnt = 1;
2923         iwmr->pgaddrmem[0] = addr;
2924         iwmr->len = size;
2925         iwmr->page_size = SZ_4K;
2926         status = irdma_hwreg_mr(iwdev, iwmr, access);
2927         if (status) {
2928                 irdma_free_stag(iwdev, stag);
2929                 ret = -ENOMEM;
2930                 goto err;
2931         }
2932
2933         return &iwmr->ibmr;
2934
2935 err:
2936         kfree(iwmr);
2937
2938         return ERR_PTR(ret);
2939 }
2940
2941 /**
2942  * irdma_get_dma_mr - register physical mem
2943  * @pd: ptr of pd
2944  * @acc: access for memory
2945  */
2946 static struct ib_mr *irdma_get_dma_mr(struct ib_pd *pd, int acc)
2947 {
2948         u64 kva = 0;
2949
2950         return irdma_reg_phys_mr(pd, 0, 0, acc, &kva);
2951 }
2952
2953 /**
2954  * irdma_del_memlist - Deleting pbl list entries for CQ/QP
2955  * @iwmr: iwmr for IB's user page addresses
2956  * @ucontext: ptr to user context
2957  */
2958 static void irdma_del_memlist(struct irdma_mr *iwmr,
2959                               struct irdma_ucontext *ucontext)
2960 {
2961         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2962         unsigned long flags;
2963
2964         switch (iwmr->type) {
2965         case IRDMA_MEMREG_TYPE_CQ:
2966                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2967                 if (iwpbl->on_list) {
2968                         iwpbl->on_list = false;
2969                         list_del(&iwpbl->list);
2970                 }
2971                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2972                 break;
2973         case IRDMA_MEMREG_TYPE_QP:
2974                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
2975                 if (iwpbl->on_list) {
2976                         iwpbl->on_list = false;
2977                         list_del(&iwpbl->list);
2978                 }
2979                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
2980                 break;
2981         default:
2982                 break;
2983         }
2984 }
2985
2986 /**
2987  * irdma_dereg_mr - deregister mr
2988  * @ib_mr: mr ptr for dereg
2989  * @udata: user data
2990  */
2991 static int irdma_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
2992 {
2993         struct ib_pd *ibpd = ib_mr->pd;
2994         struct irdma_pd *iwpd = to_iwpd(ibpd);
2995         struct irdma_mr *iwmr = to_iwmr(ib_mr);
2996         struct irdma_device *iwdev = to_iwdev(ib_mr->device);
2997         struct irdma_dealloc_stag_info *info;
2998         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2999         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
3000         struct irdma_cqp_request *cqp_request;
3001         struct cqp_cmds_info *cqp_info;
3002
3003         if (iwmr->type != IRDMA_MEMREG_TYPE_MEM) {
3004                 if (iwmr->region) {
3005                         struct irdma_ucontext *ucontext;
3006
3007                         ucontext = rdma_udata_to_drv_context(udata,
3008                                                 struct irdma_ucontext,
3009                                                 ibucontext);
3010                         irdma_del_memlist(iwmr, ucontext);
3011                 }
3012                 goto done;
3013         }
3014
3015         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
3016         if (!cqp_request)
3017                 return -ENOMEM;
3018
3019         cqp_info = &cqp_request->info;
3020         info = &cqp_info->in.u.dealloc_stag.info;
3021         memset(info, 0, sizeof(*info));
3022         info->pd_id = iwpd->sc_pd.pd_id & 0x00007fff;
3023         info->stag_idx = ib_mr->rkey >> IRDMA_CQPSQ_STAG_IDX_S;
3024         info->mr = true;
3025         if (iwpbl->pbl_allocated)
3026                 info->dealloc_pbl = true;
3027
3028         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
3029         cqp_info->post_sq = 1;
3030         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
3031         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
3032         irdma_handle_cqp_op(iwdev->rf, cqp_request);
3033         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
3034         irdma_free_stag(iwdev, iwmr->stag);
3035 done:
3036         if (iwpbl->pbl_allocated)
3037                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
3038         ib_umem_release(iwmr->region);
3039         kfree(iwmr);
3040
3041         return 0;
3042 }
3043
3044 /**
3045  * irdma_copy_sg_list - copy sg list for qp
3046  * @sg_list: copied into sg_list
3047  * @sgl: copy from sgl
3048  * @num_sges: count of sg entries
3049  */
3050 static void irdma_copy_sg_list(struct irdma_sge *sg_list, struct ib_sge *sgl,
3051                                int num_sges)
3052 {
3053         unsigned int i;
3054
3055         for (i = 0; (i < num_sges) && (i < IRDMA_MAX_WQ_FRAGMENT_COUNT); i++) {
3056                 sg_list[i].tag_off = sgl[i].addr;
3057                 sg_list[i].len = sgl[i].length;
3058                 sg_list[i].stag = sgl[i].lkey;
3059         }
3060 }
3061
3062 /**
3063  * irdma_post_send -  kernel application wr
3064  * @ibqp: qp ptr for wr
3065  * @ib_wr: work request ptr
3066  * @bad_wr: return of bad wr if err
3067  */
3068 static int irdma_post_send(struct ib_qp *ibqp,
3069                            const struct ib_send_wr *ib_wr,
3070                            const struct ib_send_wr **bad_wr)
3071 {
3072         struct irdma_qp *iwqp;
3073         struct irdma_qp_uk *ukqp;
3074         struct irdma_sc_dev *dev;
3075         struct irdma_post_sq_info info;
3076         enum irdma_status_code ret;
3077         int err = 0;
3078         unsigned long flags;
3079         bool inv_stag;
3080         struct irdma_ah *ah;
3081         bool reflush = false;
3082
3083         iwqp = to_iwqp(ibqp);
3084         ukqp = &iwqp->sc_qp.qp_uk;
3085         dev = &iwqp->iwdev->rf->sc_dev;
3086
3087         spin_lock_irqsave(&iwqp->lock, flags);
3088         if (iwqp->flush_issued && ukqp->sq_flush_complete)
3089                 reflush = true;
3090         while (ib_wr) {
3091                 memset(&info, 0, sizeof(info));
3092                 inv_stag = false;
3093                 info.wr_id = (ib_wr->wr_id);
3094                 if ((ib_wr->send_flags & IB_SEND_SIGNALED) || iwqp->sig_all)
3095                         info.signaled = true;
3096                 if (ib_wr->send_flags & IB_SEND_FENCE)
3097                         info.read_fence = true;
3098                 switch (ib_wr->opcode) {
3099                 case IB_WR_SEND_WITH_IMM:
3100                         if (ukqp->qp_caps & IRDMA_SEND_WITH_IMM) {
3101                                 info.imm_data_valid = true;
3102                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
3103                         } else {
3104                                 err = -EINVAL;
3105                                 break;
3106                         }
3107                         fallthrough;
3108                 case IB_WR_SEND:
3109                 case IB_WR_SEND_WITH_INV:
3110                         if (ib_wr->opcode == IB_WR_SEND ||
3111                             ib_wr->opcode == IB_WR_SEND_WITH_IMM) {
3112                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
3113                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL;
3114                                 else
3115                                         info.op_type = IRDMA_OP_TYPE_SEND;
3116                         } else {
3117                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
3118                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL_INV;
3119                                 else
3120                                         info.op_type = IRDMA_OP_TYPE_SEND_INV;
3121                                 info.stag_to_inv = ib_wr->ex.invalidate_rkey;
3122                         }
3123
3124                         if (ib_wr->send_flags & IB_SEND_INLINE) {
3125                                 info.op.inline_send.data = (void *)(unsigned long)
3126                                                            ib_wr->sg_list[0].addr;
3127                                 info.op.inline_send.len = ib_wr->sg_list[0].length;
3128                                 if (iwqp->ibqp.qp_type == IB_QPT_UD ||
3129                                     iwqp->ibqp.qp_type == IB_QPT_GSI) {
3130                                         ah = to_iwah(ud_wr(ib_wr)->ah);
3131                                         info.op.inline_send.ah_id = ah->sc_ah.ah_info.ah_idx;
3132                                         info.op.inline_send.qkey = ud_wr(ib_wr)->remote_qkey;
3133                                         info.op.inline_send.dest_qp = ud_wr(ib_wr)->remote_qpn;
3134                                 }
3135                                 ret = irdma_uk_inline_send(ukqp, &info, false);
3136                         } else {
3137                                 info.op.send.num_sges = ib_wr->num_sge;
3138                                 info.op.send.sg_list = (struct irdma_sge *)
3139                                                        ib_wr->sg_list;
3140                                 if (iwqp->ibqp.qp_type == IB_QPT_UD ||
3141                                     iwqp->ibqp.qp_type == IB_QPT_GSI) {
3142                                         ah = to_iwah(ud_wr(ib_wr)->ah);
3143                                         info.op.send.ah_id = ah->sc_ah.ah_info.ah_idx;
3144                                         info.op.send.qkey = ud_wr(ib_wr)->remote_qkey;
3145                                         info.op.send.dest_qp = ud_wr(ib_wr)->remote_qpn;
3146                                 }
3147                                 ret = irdma_uk_send(ukqp, &info, false);
3148                         }
3149
3150                         if (ret) {
3151                                 if (ret == IRDMA_ERR_QP_TOOMANY_WRS_POSTED)
3152                                         err = -ENOMEM;
3153                                 else
3154                                         err = -EINVAL;
3155                         }
3156                         break;
3157                 case IB_WR_RDMA_WRITE_WITH_IMM:
3158                         if (ukqp->qp_caps & IRDMA_WRITE_WITH_IMM) {
3159                                 info.imm_data_valid = true;
3160                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
3161                         } else {
3162                                 err = -EINVAL;
3163                                 break;
3164                         }
3165                         fallthrough;
3166                 case IB_WR_RDMA_WRITE:
3167                         if (ib_wr->send_flags & IB_SEND_SOLICITED)
3168                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE_SOL;
3169                         else
3170                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE;
3171
3172                         if (ib_wr->send_flags & IB_SEND_INLINE) {
3173                                 info.op.inline_rdma_write.data = (void *)(uintptr_t)ib_wr->sg_list[0].addr;
3174                                 info.op.inline_rdma_write.len = ib_wr->sg_list[0].length;
3175                                 info.op.inline_rdma_write.rem_addr.tag_off = rdma_wr(ib_wr)->remote_addr;
3176                                 info.op.inline_rdma_write.rem_addr.stag = rdma_wr(ib_wr)->rkey;
3177                                 ret = irdma_uk_inline_rdma_write(ukqp, &info, false);
3178                         } else {
3179                                 info.op.rdma_write.lo_sg_list = (void *)ib_wr->sg_list;
3180                                 info.op.rdma_write.num_lo_sges = ib_wr->num_sge;
3181                                 info.op.rdma_write.rem_addr.tag_off = rdma_wr(ib_wr)->remote_addr;
3182                                 info.op.rdma_write.rem_addr.stag = rdma_wr(ib_wr)->rkey;
3183                                 ret = irdma_uk_rdma_write(ukqp, &info, false);
3184                         }
3185
3186                         if (ret) {
3187                                 if (ret == IRDMA_ERR_QP_TOOMANY_WRS_POSTED)
3188                                         err = -ENOMEM;
3189                                 else
3190                                         err = -EINVAL;
3191                         }
3192                         break;
3193                 case IB_WR_RDMA_READ_WITH_INV:
3194                         inv_stag = true;
3195                         fallthrough;
3196                 case IB_WR_RDMA_READ:
3197                         if (ib_wr->num_sge >
3198                             dev->hw_attrs.uk_attrs.max_hw_read_sges) {
3199                                 err = -EINVAL;
3200                                 break;
3201                         }
3202                         info.op_type = IRDMA_OP_TYPE_RDMA_READ;
3203                         info.op.rdma_read.rem_addr.tag_off = rdma_wr(ib_wr)->remote_addr;
3204                         info.op.rdma_read.rem_addr.stag = rdma_wr(ib_wr)->rkey;
3205                         info.op.rdma_read.lo_sg_list = (void *)ib_wr->sg_list;
3206                         info.op.rdma_read.num_lo_sges = ib_wr->num_sge;
3207
3208                         ret = irdma_uk_rdma_read(ukqp, &info, inv_stag, false);
3209                         if (ret) {
3210                                 if (ret == IRDMA_ERR_QP_TOOMANY_WRS_POSTED)
3211                                         err = -ENOMEM;
3212                                 else
3213                                         err = -EINVAL;
3214                         }
3215                         break;
3216                 case IB_WR_LOCAL_INV:
3217                         info.op_type = IRDMA_OP_TYPE_INV_STAG;
3218                         info.op.inv_local_stag.target_stag = ib_wr->ex.invalidate_rkey;
3219                         ret = irdma_uk_stag_local_invalidate(ukqp, &info, true);
3220                         if (ret)
3221                                 err = -ENOMEM;
3222                         break;
3223                 case IB_WR_REG_MR: {
3224                         struct irdma_mr *iwmr = to_iwmr(reg_wr(ib_wr)->mr);
3225                         struct irdma_pble_alloc *palloc = &iwmr->iwpbl.pble_alloc;
3226                         struct irdma_fast_reg_stag_info stag_info = {};
3227
3228                         stag_info.signaled = info.signaled;
3229                         stag_info.read_fence = info.read_fence;
3230                         stag_info.access_rights = irdma_get_mr_access(reg_wr(ib_wr)->access);
3231                         stag_info.stag_key = reg_wr(ib_wr)->key & 0xff;
3232                         stag_info.stag_idx = reg_wr(ib_wr)->key >> 8;
3233                         stag_info.page_size = reg_wr(ib_wr)->mr->page_size;
3234                         stag_info.wr_id = ib_wr->wr_id;
3235                         stag_info.addr_type = IRDMA_ADDR_TYPE_VA_BASED;
3236                         stag_info.va = (void *)(uintptr_t)iwmr->ibmr.iova;
3237                         stag_info.total_len = iwmr->ibmr.length;
3238                         stag_info.reg_addr_pa = *palloc->level1.addr;
3239                         stag_info.first_pm_pbl_index = palloc->level1.idx;
3240                         stag_info.local_fence = ib_wr->send_flags & IB_SEND_FENCE;
3241                         if (iwmr->npages > IRDMA_MIN_PAGES_PER_FMR)
3242                                 stag_info.chunk_size = 1;
3243                         ret = irdma_sc_mr_fast_register(&iwqp->sc_qp, &stag_info,
3244                                                         true);
3245                         if (ret)
3246                                 err = -ENOMEM;
3247                         break;
3248                 }
3249                 default:
3250                         err = -EINVAL;
3251                         ibdev_dbg(&iwqp->iwdev->ibdev,
3252                                   "VERBS: upost_send bad opcode = 0x%x\n",
3253                                   ib_wr->opcode);
3254                         break;
3255                 }
3256
3257                 if (err)
3258                         break;
3259                 ib_wr = ib_wr->next;
3260         }
3261
3262         if (!iwqp->flush_issued && iwqp->hw_iwarp_state <= IRDMA_QP_STATE_RTS) {
3263                 irdma_uk_qp_post_wr(ukqp);
3264                 spin_unlock_irqrestore(&iwqp->lock, flags);
3265         } else if (reflush) {
3266                 ukqp->sq_flush_complete = false;
3267                 spin_unlock_irqrestore(&iwqp->lock, flags);
3268                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_SQ | IRDMA_REFLUSH);
3269         } else {
3270                 spin_unlock_irqrestore(&iwqp->lock, flags);
3271         }
3272         if (err)
3273                 *bad_wr = ib_wr;
3274
3275         return err;
3276 }
3277
3278 /**
3279  * irdma_post_recv - post receive wr for kernel application
3280  * @ibqp: ib qp pointer
3281  * @ib_wr: work request for receive
3282  * @bad_wr: bad wr caused an error
3283  */
3284 static int irdma_post_recv(struct ib_qp *ibqp,
3285                            const struct ib_recv_wr *ib_wr,
3286                            const struct ib_recv_wr **bad_wr)
3287 {
3288         struct irdma_qp *iwqp;
3289         struct irdma_qp_uk *ukqp;
3290         struct irdma_post_rq_info post_recv = {};
3291         struct irdma_sge sg_list[IRDMA_MAX_WQ_FRAGMENT_COUNT];
3292         enum irdma_status_code ret = 0;
3293         unsigned long flags;
3294         int err = 0;
3295         bool reflush = false;
3296
3297         iwqp = to_iwqp(ibqp);
3298         ukqp = &iwqp->sc_qp.qp_uk;
3299
3300         spin_lock_irqsave(&iwqp->lock, flags);
3301         if (iwqp->flush_issued && ukqp->rq_flush_complete)
3302                 reflush = true;
3303         while (ib_wr) {
3304                 post_recv.num_sges = ib_wr->num_sge;
3305                 post_recv.wr_id = ib_wr->wr_id;
3306                 irdma_copy_sg_list(sg_list, ib_wr->sg_list, ib_wr->num_sge);
3307                 post_recv.sg_list = sg_list;
3308                 ret = irdma_uk_post_receive(ukqp, &post_recv);
3309                 if (ret) {
3310                         ibdev_dbg(&iwqp->iwdev->ibdev,
3311                                   "VERBS: post_recv err %d\n", ret);
3312                         if (ret == IRDMA_ERR_QP_TOOMANY_WRS_POSTED)
3313                                 err = -ENOMEM;
3314                         else
3315                                 err = -EINVAL;
3316                         goto out;
3317                 }
3318
3319                 ib_wr = ib_wr->next;
3320         }
3321
3322 out:
3323         if (reflush) {
3324                 ukqp->rq_flush_complete = false;
3325                 spin_unlock_irqrestore(&iwqp->lock, flags);
3326                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_RQ | IRDMA_REFLUSH);
3327         } else {
3328                 spin_unlock_irqrestore(&iwqp->lock, flags);
3329         }
3330
3331         if (err)
3332                 *bad_wr = ib_wr;
3333
3334         return err;
3335 }
3336
3337 /**
3338  * irdma_flush_err_to_ib_wc_status - return change flush error code to IB status
3339  * @opcode: iwarp flush code
3340  */
3341 static enum ib_wc_status irdma_flush_err_to_ib_wc_status(enum irdma_flush_opcode opcode)
3342 {
3343         switch (opcode) {
3344         case FLUSH_PROT_ERR:
3345                 return IB_WC_LOC_PROT_ERR;
3346         case FLUSH_REM_ACCESS_ERR:
3347                 return IB_WC_REM_ACCESS_ERR;
3348         case FLUSH_LOC_QP_OP_ERR:
3349                 return IB_WC_LOC_QP_OP_ERR;
3350         case FLUSH_REM_OP_ERR:
3351                 return IB_WC_REM_OP_ERR;
3352         case FLUSH_LOC_LEN_ERR:
3353                 return IB_WC_LOC_LEN_ERR;
3354         case FLUSH_GENERAL_ERR:
3355                 return IB_WC_WR_FLUSH_ERR;
3356         case FLUSH_FATAL_ERR:
3357         default:
3358                 return IB_WC_FATAL_ERR;
3359         }
3360 }
3361
3362 /**
3363  * irdma_process_cqe - process cqe info
3364  * @entry: processed cqe
3365  * @cq_poll_info: cqe info
3366  */
3367 static void irdma_process_cqe(struct ib_wc *entry,
3368                               struct irdma_cq_poll_info *cq_poll_info)
3369 {
3370         struct irdma_qp *iwqp;
3371         struct irdma_sc_qp *qp;
3372
3373         entry->wc_flags = 0;
3374         entry->pkey_index = 0;
3375         entry->wr_id = cq_poll_info->wr_id;
3376
3377         qp = cq_poll_info->qp_handle;
3378         iwqp = qp->qp_uk.back_qp;
3379         entry->qp = qp->qp_uk.back_qp;
3380
3381         if (cq_poll_info->error) {
3382                 entry->status = (cq_poll_info->comp_status == IRDMA_COMPL_STATUS_FLUSHED) ?
3383                                 irdma_flush_err_to_ib_wc_status(cq_poll_info->minor_err) : IB_WC_GENERAL_ERR;
3384
3385                 entry->vendor_err = cq_poll_info->major_err << 16 |
3386                                     cq_poll_info->minor_err;
3387         } else {
3388                 entry->status = IB_WC_SUCCESS;
3389                 if (cq_poll_info->imm_valid) {
3390                         entry->ex.imm_data = htonl(cq_poll_info->imm_data);
3391                         entry->wc_flags |= IB_WC_WITH_IMM;
3392                 }
3393                 if (cq_poll_info->ud_smac_valid) {
3394                         ether_addr_copy(entry->smac, cq_poll_info->ud_smac);
3395                         entry->wc_flags |= IB_WC_WITH_SMAC;
3396                 }
3397
3398                 if (cq_poll_info->ud_vlan_valid) {
3399                         entry->vlan_id = cq_poll_info->ud_vlan & VLAN_VID_MASK;
3400                         entry->wc_flags |= IB_WC_WITH_VLAN;
3401                         entry->sl = cq_poll_info->ud_vlan >> VLAN_PRIO_SHIFT;
3402                 } else {
3403                         entry->sl = 0;
3404                 }
3405         }
3406
3407         switch (cq_poll_info->op_type) {
3408         case IRDMA_OP_TYPE_RDMA_WRITE:
3409         case IRDMA_OP_TYPE_RDMA_WRITE_SOL:
3410                 entry->opcode = IB_WC_RDMA_WRITE;
3411                 break;
3412         case IRDMA_OP_TYPE_RDMA_READ_INV_STAG:
3413         case IRDMA_OP_TYPE_RDMA_READ:
3414                 entry->opcode = IB_WC_RDMA_READ;
3415                 break;
3416         case IRDMA_OP_TYPE_SEND_INV:
3417         case IRDMA_OP_TYPE_SEND_SOL:
3418         case IRDMA_OP_TYPE_SEND_SOL_INV:
3419         case IRDMA_OP_TYPE_SEND:
3420                 entry->opcode = IB_WC_SEND;
3421                 break;
3422         case IRDMA_OP_TYPE_FAST_REG_NSMR:
3423                 entry->opcode = IB_WC_REG_MR;
3424                 break;
3425         case IRDMA_OP_TYPE_INV_STAG:
3426                 entry->opcode = IB_WC_LOCAL_INV;
3427                 break;
3428         case IRDMA_OP_TYPE_REC_IMM:
3429         case IRDMA_OP_TYPE_REC:
3430                 entry->opcode = cq_poll_info->op_type == IRDMA_OP_TYPE_REC_IMM ?
3431                         IB_WC_RECV_RDMA_WITH_IMM : IB_WC_RECV;
3432                 if (qp->qp_uk.qp_type != IRDMA_QP_TYPE_ROCE_UD &&
3433                     cq_poll_info->stag_invalid_set) {
3434                         entry->ex.invalidate_rkey = cq_poll_info->inv_stag;
3435                         entry->wc_flags |= IB_WC_WITH_INVALIDATE;
3436                 }
3437                 break;
3438         default:
3439                 ibdev_err(&iwqp->iwdev->ibdev,
3440                           "Invalid opcode = %d in CQE\n", cq_poll_info->op_type);
3441                 entry->status = IB_WC_GENERAL_ERR;
3442                 return;
3443         }
3444
3445         if (qp->qp_uk.qp_type == IRDMA_QP_TYPE_ROCE_UD) {
3446                 entry->src_qp = cq_poll_info->ud_src_qpn;
3447                 entry->slid = 0;
3448                 entry->wc_flags |=
3449                         (IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE);
3450                 entry->network_hdr_type = cq_poll_info->ipv4 ?
3451                                                   RDMA_NETWORK_IPV4 :
3452                                                   RDMA_NETWORK_IPV6;
3453         } else {
3454                 entry->src_qp = cq_poll_info->qp_id;
3455         }
3456
3457         entry->byte_len = cq_poll_info->bytes_xfered;
3458 }
3459
3460 /**
3461  * irdma_poll_one - poll one entry of the CQ
3462  * @ukcq: ukcq to poll
3463  * @cur_cqe: current CQE info to be filled in
3464  * @entry: ibv_wc object to be filled for non-extended CQ or NULL for extended CQ
3465  *
3466  * Returns the internal irdma device error code or 0 on success
3467  */
3468 static inline int irdma_poll_one(struct irdma_cq_uk *ukcq,
3469                                  struct irdma_cq_poll_info *cur_cqe,
3470                                  struct ib_wc *entry)
3471 {
3472         int ret = irdma_uk_cq_poll_cmpl(ukcq, cur_cqe);
3473
3474         if (ret)
3475                 return ret;
3476
3477         irdma_process_cqe(entry, cur_cqe);
3478
3479         return 0;
3480 }
3481
3482 /**
3483  * __irdma_poll_cq - poll cq for completion (kernel apps)
3484  * @iwcq: cq to poll
3485  * @num_entries: number of entries to poll
3486  * @entry: wr of a completed entry
3487  */
3488 static int __irdma_poll_cq(struct irdma_cq *iwcq, int num_entries, struct ib_wc *entry)
3489 {
3490         struct list_head *tmp_node, *list_node;
3491         struct irdma_cq_buf *last_buf = NULL;
3492         struct irdma_cq_poll_info *cur_cqe = &iwcq->cur_cqe;
3493         struct irdma_cq_buf *cq_buf;
3494         enum irdma_status_code ret;
3495         struct irdma_device *iwdev;
3496         struct irdma_cq_uk *ukcq;
3497         bool cq_new_cqe = false;
3498         int resized_bufs = 0;
3499         int npolled = 0;
3500
3501         iwdev = to_iwdev(iwcq->ibcq.device);
3502         ukcq = &iwcq->sc_cq.cq_uk;
3503
3504         /* go through the list of previously resized CQ buffers */
3505         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
3506                 cq_buf = container_of(list_node, struct irdma_cq_buf, list);
3507                 while (npolled < num_entries) {
3508                         ret = irdma_poll_one(&cq_buf->cq_uk, cur_cqe, entry + npolled);
3509                         if (!ret) {
3510                                 ++npolled;
3511                                 cq_new_cqe = true;
3512                                 continue;
3513                         }
3514                         if (ret == IRDMA_ERR_Q_EMPTY)
3515                                 break;
3516                          /* QP using the CQ is destroyed. Skip reporting this CQE */
3517                         if (ret == IRDMA_ERR_Q_DESTROYED) {
3518                                 cq_new_cqe = true;
3519                                 continue;
3520                         }
3521                         goto error;
3522                 }
3523
3524                 /* save the resized CQ buffer which received the last cqe */
3525                 if (cq_new_cqe)
3526                         last_buf = cq_buf;
3527                 cq_new_cqe = false;
3528         }
3529
3530         /* check the current CQ for new cqes */
3531         while (npolled < num_entries) {
3532                 ret = irdma_poll_one(ukcq, cur_cqe, entry + npolled);
3533                 if (!ret) {
3534                         ++npolled;
3535                         cq_new_cqe = true;
3536                         continue;
3537                 }
3538
3539                 if (ret == IRDMA_ERR_Q_EMPTY)
3540                         break;
3541                 /* QP using the CQ is destroyed. Skip reporting this CQE */
3542                 if (ret == IRDMA_ERR_Q_DESTROYED) {
3543                         cq_new_cqe = true;
3544                         continue;
3545                 }
3546                 goto error;
3547         }
3548
3549         if (cq_new_cqe)
3550                 /* all previous CQ resizes are complete */
3551                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, NULL);
3552         else if (last_buf)
3553                 /* only CQ resizes up to the last_buf are complete */
3554                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, last_buf);
3555         if (resized_bufs)
3556                 /* report to the HW the number of complete CQ resizes */
3557                 irdma_uk_cq_set_resized_cnt(ukcq, resized_bufs);
3558
3559         return npolled;
3560 error:
3561         ibdev_dbg(&iwdev->ibdev, "%s: Error polling CQ, irdma_err: %d\n",
3562                   __func__, ret);
3563
3564         return -EINVAL;
3565 }
3566
3567 /**
3568  * irdma_poll_cq - poll cq for completion (kernel apps)
3569  * @ibcq: cq to poll
3570  * @num_entries: number of entries to poll
3571  * @entry: wr of a completed entry
3572  */
3573 static int irdma_poll_cq(struct ib_cq *ibcq, int num_entries,
3574                          struct ib_wc *entry)
3575 {
3576         struct irdma_cq *iwcq;
3577         unsigned long flags;
3578         int ret;
3579
3580         iwcq = to_iwcq(ibcq);
3581
3582         spin_lock_irqsave(&iwcq->lock, flags);
3583         ret = __irdma_poll_cq(iwcq, num_entries, entry);
3584         spin_unlock_irqrestore(&iwcq->lock, flags);
3585
3586         return ret;
3587 }
3588
3589 /**
3590  * irdma_req_notify_cq - arm cq kernel application
3591  * @ibcq: cq to arm
3592  * @notify_flags: notofication flags
3593  */
3594 static int irdma_req_notify_cq(struct ib_cq *ibcq,
3595                                enum ib_cq_notify_flags notify_flags)
3596 {
3597         struct irdma_cq *iwcq;
3598         struct irdma_cq_uk *ukcq;
3599         unsigned long flags;
3600         enum irdma_cmpl_notify cq_notify = IRDMA_CQ_COMPL_EVENT;
3601
3602         iwcq = to_iwcq(ibcq);
3603         ukcq = &iwcq->sc_cq.cq_uk;
3604         if (notify_flags == IB_CQ_SOLICITED)
3605                 cq_notify = IRDMA_CQ_COMPL_SOLICITED;
3606
3607         spin_lock_irqsave(&iwcq->lock, flags);
3608         irdma_uk_cq_request_notification(ukcq, cq_notify);
3609         spin_unlock_irqrestore(&iwcq->lock, flags);
3610
3611         return 0;
3612 }
3613
3614 static int irdma_roce_port_immutable(struct ib_device *ibdev, u32 port_num,
3615                                      struct ib_port_immutable *immutable)
3616 {
3617         struct ib_port_attr attr;
3618         int err;
3619
3620         immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
3621         err = ib_query_port(ibdev, port_num, &attr);
3622         if (err)
3623                 return err;
3624
3625         immutable->max_mad_size = IB_MGMT_MAD_SIZE;
3626         immutable->pkey_tbl_len = attr.pkey_tbl_len;
3627         immutable->gid_tbl_len = attr.gid_tbl_len;
3628
3629         return 0;
3630 }
3631
3632 static int irdma_iw_port_immutable(struct ib_device *ibdev, u32 port_num,
3633                                    struct ib_port_immutable *immutable)
3634 {
3635         struct ib_port_attr attr;
3636         int err;
3637
3638         immutable->core_cap_flags = RDMA_CORE_PORT_IWARP;
3639         err = ib_query_port(ibdev, port_num, &attr);
3640         if (err)
3641                 return err;
3642         immutable->gid_tbl_len = attr.gid_tbl_len;
3643
3644         return 0;
3645 }
3646
3647 static const char *const irdma_hw_stat_names[] = {
3648         /* 32bit names */
3649         [IRDMA_HW_STAT_INDEX_RXVLANERR] = "rxVlanErrors",
3650         [IRDMA_HW_STAT_INDEX_IP4RXDISCARD] = "ip4InDiscards",
3651         [IRDMA_HW_STAT_INDEX_IP4RXTRUNC] = "ip4InTruncatedPkts",
3652         [IRDMA_HW_STAT_INDEX_IP4TXNOROUTE] = "ip4OutNoRoutes",
3653         [IRDMA_HW_STAT_INDEX_IP6RXDISCARD] = "ip6InDiscards",
3654         [IRDMA_HW_STAT_INDEX_IP6RXTRUNC] = "ip6InTruncatedPkts",
3655         [IRDMA_HW_STAT_INDEX_IP6TXNOROUTE] = "ip6OutNoRoutes",
3656         [IRDMA_HW_STAT_INDEX_TCPRTXSEG] = "tcpRetransSegs",
3657         [IRDMA_HW_STAT_INDEX_TCPRXOPTERR] = "tcpInOptErrors",
3658         [IRDMA_HW_STAT_INDEX_TCPRXPROTOERR] = "tcpInProtoErrors",
3659         [IRDMA_HW_STAT_INDEX_RXRPCNPHANDLED] = "cnpHandled",
3660         [IRDMA_HW_STAT_INDEX_RXRPCNPIGNORED] = "cnpIgnored",
3661         [IRDMA_HW_STAT_INDEX_TXNPCNPSENT] = "cnpSent",
3662
3663         /* 64bit names */
3664         [IRDMA_HW_STAT_INDEX_IP4RXOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3665                 "ip4InOctets",
3666         [IRDMA_HW_STAT_INDEX_IP4RXPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3667                 "ip4InPkts",
3668         [IRDMA_HW_STAT_INDEX_IP4RXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32] =
3669                 "ip4InReasmRqd",
3670         [IRDMA_HW_STAT_INDEX_IP4RXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3671                 "ip4InMcastOctets",
3672         [IRDMA_HW_STAT_INDEX_IP4RXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3673                 "ip4InMcastPkts",
3674         [IRDMA_HW_STAT_INDEX_IP4TXOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3675                 "ip4OutOctets",
3676         [IRDMA_HW_STAT_INDEX_IP4TXPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3677                 "ip4OutPkts",
3678         [IRDMA_HW_STAT_INDEX_IP4TXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32] =
3679                 "ip4OutSegRqd",
3680         [IRDMA_HW_STAT_INDEX_IP4TXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3681                 "ip4OutMcastOctets",
3682         [IRDMA_HW_STAT_INDEX_IP4TXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3683                 "ip4OutMcastPkts",
3684         [IRDMA_HW_STAT_INDEX_IP6RXOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3685                 "ip6InOctets",
3686         [IRDMA_HW_STAT_INDEX_IP6RXPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3687                 "ip6InPkts",
3688         [IRDMA_HW_STAT_INDEX_IP6RXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32] =
3689                 "ip6InReasmRqd",
3690         [IRDMA_HW_STAT_INDEX_IP6RXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3691                 "ip6InMcastOctets",
3692         [IRDMA_HW_STAT_INDEX_IP6RXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3693                 "ip6InMcastPkts",
3694         [IRDMA_HW_STAT_INDEX_IP6TXOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3695                 "ip6OutOctets",
3696         [IRDMA_HW_STAT_INDEX_IP6TXPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3697                 "ip6OutPkts",
3698         [IRDMA_HW_STAT_INDEX_IP6TXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32] =
3699                 "ip6OutSegRqd",
3700         [IRDMA_HW_STAT_INDEX_IP6TXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3701                 "ip6OutMcastOctets",
3702         [IRDMA_HW_STAT_INDEX_IP6TXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3703                 "ip6OutMcastPkts",
3704         [IRDMA_HW_STAT_INDEX_TCPRXSEGS + IRDMA_HW_STAT_INDEX_MAX_32] =
3705                 "tcpInSegs",
3706         [IRDMA_HW_STAT_INDEX_TCPTXSEG + IRDMA_HW_STAT_INDEX_MAX_32] =
3707                 "tcpOutSegs",
3708         [IRDMA_HW_STAT_INDEX_RDMARXRDS + IRDMA_HW_STAT_INDEX_MAX_32] =
3709                 "iwInRdmaReads",
3710         [IRDMA_HW_STAT_INDEX_RDMARXSNDS + IRDMA_HW_STAT_INDEX_MAX_32] =
3711                 "iwInRdmaSends",
3712         [IRDMA_HW_STAT_INDEX_RDMARXWRS + IRDMA_HW_STAT_INDEX_MAX_32] =
3713                 "iwInRdmaWrites",
3714         [IRDMA_HW_STAT_INDEX_RDMATXRDS + IRDMA_HW_STAT_INDEX_MAX_32] =
3715                 "iwOutRdmaReads",
3716         [IRDMA_HW_STAT_INDEX_RDMATXSNDS + IRDMA_HW_STAT_INDEX_MAX_32] =
3717                 "iwOutRdmaSends",
3718         [IRDMA_HW_STAT_INDEX_RDMATXWRS + IRDMA_HW_STAT_INDEX_MAX_32] =
3719                 "iwOutRdmaWrites",
3720         [IRDMA_HW_STAT_INDEX_RDMAVBND + IRDMA_HW_STAT_INDEX_MAX_32] =
3721                 "iwRdmaBnd",
3722         [IRDMA_HW_STAT_INDEX_RDMAVINV + IRDMA_HW_STAT_INDEX_MAX_32] =
3723                 "iwRdmaInv",
3724         [IRDMA_HW_STAT_INDEX_UDPRXPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3725                 "RxUDP",
3726         [IRDMA_HW_STAT_INDEX_UDPTXPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3727                 "TxUDP",
3728         [IRDMA_HW_STAT_INDEX_RXNPECNMARKEDPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3729                 "RxECNMrkd",
3730 };
3731
3732 static void irdma_get_dev_fw_str(struct ib_device *dev, char *str)
3733 {
3734         struct irdma_device *iwdev = to_iwdev(dev);
3735
3736         snprintf(str, IB_FW_VERSION_NAME_MAX, "%u.%u",
3737                  irdma_fw_major_ver(&iwdev->rf->sc_dev),
3738                  irdma_fw_minor_ver(&iwdev->rf->sc_dev));
3739 }
3740
3741 /**
3742  * irdma_alloc_hw_port_stats - Allocate a hw stats structure
3743  * @ibdev: device pointer from stack
3744  * @port_num: port number
3745  */
3746 static struct rdma_hw_stats *irdma_alloc_hw_port_stats(struct ib_device *ibdev,
3747                                                        u32 port_num)
3748 {
3749         int num_counters = IRDMA_HW_STAT_INDEX_MAX_32 +
3750                            IRDMA_HW_STAT_INDEX_MAX_64;
3751         unsigned long lifespan = RDMA_HW_STATS_DEFAULT_LIFESPAN;
3752
3753         BUILD_BUG_ON(ARRAY_SIZE(irdma_hw_stat_names) !=
3754                      (IRDMA_HW_STAT_INDEX_MAX_32 + IRDMA_HW_STAT_INDEX_MAX_64));
3755
3756         return rdma_alloc_hw_stats_struct(irdma_hw_stat_names, num_counters,
3757                                           lifespan);
3758 }
3759
3760 /**
3761  * irdma_get_hw_stats - Populates the rdma_hw_stats structure
3762  * @ibdev: device pointer from stack
3763  * @stats: stats pointer from stack
3764  * @port_num: port number
3765  * @index: which hw counter the stack is requesting we update
3766  */
3767 static int irdma_get_hw_stats(struct ib_device *ibdev,
3768                               struct rdma_hw_stats *stats, u32 port_num,
3769                               int index)
3770 {
3771         struct irdma_device *iwdev = to_iwdev(ibdev);
3772         struct irdma_dev_hw_stats *hw_stats = &iwdev->vsi.pestat->hw_stats;
3773
3774         if (iwdev->rf->rdma_ver >= IRDMA_GEN_2)
3775                 irdma_cqp_gather_stats_cmd(&iwdev->rf->sc_dev, iwdev->vsi.pestat, true);
3776         else
3777                 irdma_cqp_gather_stats_gen1(&iwdev->rf->sc_dev, iwdev->vsi.pestat);
3778
3779         memcpy(&stats->value[0], hw_stats, sizeof(*hw_stats));
3780
3781         return stats->num_counters;
3782 }
3783
3784 /**
3785  * irdma_query_gid - Query port GID
3786  * @ibdev: device pointer from stack
3787  * @port: port number
3788  * @index: Entry index
3789  * @gid: Global ID
3790  */
3791 static int irdma_query_gid(struct ib_device *ibdev, u32 port, int index,
3792                            union ib_gid *gid)
3793 {
3794         struct irdma_device *iwdev = to_iwdev(ibdev);
3795
3796         memset(gid->raw, 0, sizeof(gid->raw));
3797         ether_addr_copy(gid->raw, iwdev->netdev->dev_addr);
3798
3799         return 0;
3800 }
3801
3802 /**
3803  * mcast_list_add -  Add a new mcast item to list
3804  * @rf: RDMA PCI function
3805  * @new_elem: pointer to element to add
3806  */
3807 static void mcast_list_add(struct irdma_pci_f *rf,
3808                            struct mc_table_list *new_elem)
3809 {
3810         list_add(&new_elem->list, &rf->mc_qht_list.list);
3811 }
3812
3813 /**
3814  * mcast_list_del - Remove an mcast item from list
3815  * @mc_qht_elem: pointer to mcast table list element
3816  */
3817 static void mcast_list_del(struct mc_table_list *mc_qht_elem)
3818 {
3819         if (mc_qht_elem)
3820                 list_del(&mc_qht_elem->list);
3821 }
3822
3823 /**
3824  * mcast_list_lookup_ip - Search mcast list for address
3825  * @rf: RDMA PCI function
3826  * @ip_mcast: pointer to mcast IP address
3827  */
3828 static struct mc_table_list *mcast_list_lookup_ip(struct irdma_pci_f *rf,
3829                                                   u32 *ip_mcast)
3830 {
3831         struct mc_table_list *mc_qht_el;
3832         struct list_head *pos, *q;
3833
3834         list_for_each_safe (pos, q, &rf->mc_qht_list.list) {
3835                 mc_qht_el = list_entry(pos, struct mc_table_list, list);
3836                 if (!memcmp(mc_qht_el->mc_info.dest_ip, ip_mcast,
3837                             sizeof(mc_qht_el->mc_info.dest_ip)))
3838                         return mc_qht_el;
3839         }
3840
3841         return NULL;
3842 }
3843
3844 /**
3845  * irdma_mcast_cqp_op - perform a mcast cqp operation
3846  * @iwdev: irdma device
3847  * @mc_grp_ctx: mcast group info
3848  * @op: operation
3849  *
3850  * returns error status
3851  */
3852 static int irdma_mcast_cqp_op(struct irdma_device *iwdev,
3853                               struct irdma_mcast_grp_info *mc_grp_ctx, u8 op)
3854 {
3855         struct cqp_cmds_info *cqp_info;
3856         struct irdma_cqp_request *cqp_request;
3857         enum irdma_status_code status;
3858
3859         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
3860         if (!cqp_request)
3861                 return -ENOMEM;
3862
3863         cqp_request->info.in.u.mc_create.info = *mc_grp_ctx;
3864         cqp_info = &cqp_request->info;
3865         cqp_info->cqp_cmd = op;
3866         cqp_info->post_sq = 1;
3867         cqp_info->in.u.mc_create.scratch = (uintptr_t)cqp_request;
3868         cqp_info->in.u.mc_create.cqp = &iwdev->rf->cqp.sc_cqp;
3869         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
3870         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
3871         if (status)
3872                 return -ENOMEM;
3873
3874         return 0;
3875 }
3876
3877 /**
3878  * irdma_mcast_mac - Get the multicast MAC for an IP address
3879  * @ip_addr: IPv4 or IPv6 address
3880  * @mac: pointer to result MAC address
3881  * @ipv4: flag indicating IPv4 or IPv6
3882  *
3883  */
3884 void irdma_mcast_mac(u32 *ip_addr, u8 *mac, bool ipv4)
3885 {
3886         u8 *ip = (u8 *)ip_addr;
3887
3888         if (ipv4) {
3889                 unsigned char mac4[ETH_ALEN] = {0x01, 0x00, 0x5E, 0x00,
3890                                                 0x00, 0x00};
3891
3892                 mac4[3] = ip[2] & 0x7F;
3893                 mac4[4] = ip[1];
3894                 mac4[5] = ip[0];
3895                 ether_addr_copy(mac, mac4);
3896         } else {
3897                 unsigned char mac6[ETH_ALEN] = {0x33, 0x33, 0x00, 0x00,
3898                                                 0x00, 0x00};
3899
3900                 mac6[2] = ip[3];
3901                 mac6[3] = ip[2];
3902                 mac6[4] = ip[1];
3903                 mac6[5] = ip[0];
3904                 ether_addr_copy(mac, mac6);
3905         }
3906 }
3907
3908 /**
3909  * irdma_attach_mcast - attach a qp to a multicast group
3910  * @ibqp: ptr to qp
3911  * @ibgid: pointer to global ID
3912  * @lid: local ID
3913  *
3914  * returns error status
3915  */
3916 static int irdma_attach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
3917 {
3918         struct irdma_qp *iwqp = to_iwqp(ibqp);
3919         struct irdma_device *iwdev = iwqp->iwdev;
3920         struct irdma_pci_f *rf = iwdev->rf;
3921         struct mc_table_list *mc_qht_elem;
3922         struct irdma_mcast_grp_ctx_entry_info mcg_info = {};
3923         unsigned long flags;
3924         u32 ip_addr[4] = {};
3925         u32 mgn;
3926         u32 no_mgs;
3927         int ret = 0;
3928         bool ipv4;
3929         u16 vlan_id;
3930         union {
3931                 struct sockaddr saddr;
3932                 struct sockaddr_in saddr_in;
3933                 struct sockaddr_in6 saddr_in6;
3934         } sgid_addr;
3935         unsigned char dmac[ETH_ALEN];
3936
3937         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
3938
3939         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid)) {
3940                 irdma_copy_ip_ntohl(ip_addr,
3941                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
3942                 irdma_netdev_vlan_ipv6(ip_addr, &vlan_id, NULL);
3943                 ipv4 = false;
3944                 ibdev_dbg(&iwdev->ibdev,
3945                           "VERBS: qp_id=%d, IP6address=%pI6\n", ibqp->qp_num,
3946                           ip_addr);
3947                 irdma_mcast_mac(ip_addr, dmac, false);
3948         } else {
3949                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
3950                 ipv4 = true;
3951                 vlan_id = irdma_get_vlan_ipv4(ip_addr);
3952                 irdma_mcast_mac(ip_addr, dmac, true);
3953                 ibdev_dbg(&iwdev->ibdev,
3954                           "VERBS: qp_id=%d, IP4address=%pI4, MAC=%pM\n",
3955                           ibqp->qp_num, ip_addr, dmac);
3956         }
3957
3958         spin_lock_irqsave(&rf->qh_list_lock, flags);
3959         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
3960         if (!mc_qht_elem) {
3961                 struct irdma_dma_mem *dma_mem_mc;
3962
3963                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3964                 mc_qht_elem = kzalloc(sizeof(*mc_qht_elem), GFP_KERNEL);
3965                 if (!mc_qht_elem)
3966                         return -ENOMEM;
3967
3968                 mc_qht_elem->mc_info.ipv4_valid = ipv4;
3969                 memcpy(mc_qht_elem->mc_info.dest_ip, ip_addr,
3970                        sizeof(mc_qht_elem->mc_info.dest_ip));
3971                 ret = irdma_alloc_rsrc(rf, rf->allocated_mcgs, rf->max_mcg,
3972                                        &mgn, &rf->next_mcg);
3973                 if (ret) {
3974                         kfree(mc_qht_elem);
3975                         return -ENOMEM;
3976                 }
3977
3978                 mc_qht_elem->mc_info.mgn = mgn;
3979                 dma_mem_mc = &mc_qht_elem->mc_grp_ctx.dma_mem_mc;
3980                 dma_mem_mc->size = ALIGN(sizeof(u64) * IRDMA_MAX_MGS_PER_CTX,
3981                                          IRDMA_HW_PAGE_SIZE);
3982                 dma_mem_mc->va = dma_alloc_coherent(rf->hw.device,
3983                                                     dma_mem_mc->size,
3984                                                     &dma_mem_mc->pa,
3985                                                     GFP_KERNEL);
3986                 if (!dma_mem_mc->va) {
3987                         irdma_free_rsrc(rf, rf->allocated_mcgs, mgn);
3988                         kfree(mc_qht_elem);
3989                         return -ENOMEM;
3990                 }
3991
3992                 mc_qht_elem->mc_grp_ctx.mg_id = (u16)mgn;
3993                 memcpy(mc_qht_elem->mc_grp_ctx.dest_ip_addr, ip_addr,
3994                        sizeof(mc_qht_elem->mc_grp_ctx.dest_ip_addr));
3995                 mc_qht_elem->mc_grp_ctx.ipv4_valid = ipv4;
3996                 mc_qht_elem->mc_grp_ctx.vlan_id = vlan_id;
3997                 if (vlan_id < VLAN_N_VID)
3998                         mc_qht_elem->mc_grp_ctx.vlan_valid = true;
3999                 mc_qht_elem->mc_grp_ctx.hmc_fcn_id = iwdev->vsi.fcn_id;
4000                 mc_qht_elem->mc_grp_ctx.qs_handle =
4001                         iwqp->sc_qp.vsi->qos[iwqp->sc_qp.user_pri].qs_handle;
4002                 ether_addr_copy(mc_qht_elem->mc_grp_ctx.dest_mac_addr, dmac);
4003
4004                 spin_lock_irqsave(&rf->qh_list_lock, flags);
4005                 mcast_list_add(rf, mc_qht_elem);
4006         } else {
4007                 if (mc_qht_elem->mc_grp_ctx.no_of_mgs ==
4008                     IRDMA_MAX_MGS_PER_CTX) {
4009                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4010                         return -ENOMEM;
4011                 }
4012         }
4013
4014         mcg_info.qp_id = iwqp->ibqp.qp_num;
4015         no_mgs = mc_qht_elem->mc_grp_ctx.no_of_mgs;
4016         irdma_sc_add_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
4017         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4018
4019         /* Only if there is a change do we need to modify or create */
4020         if (!no_mgs) {
4021                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4022                                          IRDMA_OP_MC_CREATE);
4023         } else if (no_mgs != mc_qht_elem->mc_grp_ctx.no_of_mgs) {
4024                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4025                                          IRDMA_OP_MC_MODIFY);
4026         } else {
4027                 return 0;
4028         }
4029
4030         if (ret)
4031                 goto error;
4032
4033         return 0;
4034
4035 error:
4036         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
4037         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
4038                 mcast_list_del(mc_qht_elem);
4039                 dma_free_coherent(rf->hw.device,
4040                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.size,
4041                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.va,
4042                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.pa);
4043                 mc_qht_elem->mc_grp_ctx.dma_mem_mc.va = NULL;
4044                 irdma_free_rsrc(rf, rf->allocated_mcgs,
4045                                 mc_qht_elem->mc_grp_ctx.mg_id);
4046                 kfree(mc_qht_elem);
4047         }
4048
4049         return ret;
4050 }
4051
4052 /**
4053  * irdma_detach_mcast - detach a qp from a multicast group
4054  * @ibqp: ptr to qp
4055  * @ibgid: pointer to global ID
4056  * @lid: local ID
4057  *
4058  * returns error status
4059  */
4060 static int irdma_detach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
4061 {
4062         struct irdma_qp *iwqp = to_iwqp(ibqp);
4063         struct irdma_device *iwdev = iwqp->iwdev;
4064         struct irdma_pci_f *rf = iwdev->rf;
4065         u32 ip_addr[4] = {};
4066         struct mc_table_list *mc_qht_elem;
4067         struct irdma_mcast_grp_ctx_entry_info mcg_info = {};
4068         int ret;
4069         unsigned long flags;
4070         union {
4071                 struct sockaddr saddr;
4072                 struct sockaddr_in saddr_in;
4073                 struct sockaddr_in6 saddr_in6;
4074         } sgid_addr;
4075
4076         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
4077         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid))
4078                 irdma_copy_ip_ntohl(ip_addr,
4079                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4080         else
4081                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
4082
4083         spin_lock_irqsave(&rf->qh_list_lock, flags);
4084         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
4085         if (!mc_qht_elem) {
4086                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4087                 ibdev_dbg(&iwdev->ibdev,
4088                           "VERBS: address not found MCG\n");
4089                 return 0;
4090         }
4091
4092         mcg_info.qp_id = iwqp->ibqp.qp_num;
4093         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
4094         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
4095                 mcast_list_del(mc_qht_elem);
4096                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4097                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4098                                          IRDMA_OP_MC_DESTROY);
4099                 if (ret) {
4100                         ibdev_dbg(&iwdev->ibdev,
4101                                   "VERBS: failed MC_DESTROY MCG\n");
4102                         spin_lock_irqsave(&rf->qh_list_lock, flags);
4103                         mcast_list_add(rf, mc_qht_elem);
4104                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4105                         return -EAGAIN;
4106                 }
4107
4108                 dma_free_coherent(rf->hw.device,
4109                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.size,
4110                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.va,
4111                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.pa);
4112                 mc_qht_elem->mc_grp_ctx.dma_mem_mc.va = NULL;
4113                 irdma_free_rsrc(rf, rf->allocated_mcgs,
4114                                 mc_qht_elem->mc_grp_ctx.mg_id);
4115                 kfree(mc_qht_elem);
4116         } else {
4117                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4118                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4119                                          IRDMA_OP_MC_MODIFY);
4120                 if (ret) {
4121                         ibdev_dbg(&iwdev->ibdev,
4122                                   "VERBS: failed Modify MCG\n");
4123                         return ret;
4124                 }
4125         }
4126
4127         return 0;
4128 }
4129
4130 /**
4131  * irdma_create_ah - create address handle
4132  * @ibah: address handle
4133  * @attr: address handle attributes
4134  * @udata: User data
4135  *
4136  * returns 0 on success, error otherwise
4137  */
4138 static int irdma_create_ah(struct ib_ah *ibah,
4139                            struct rdma_ah_init_attr *attr,
4140                            struct ib_udata *udata)
4141 {
4142         struct irdma_pd *pd = to_iwpd(ibah->pd);
4143         struct irdma_ah *ah = container_of(ibah, struct irdma_ah, ibah);
4144         struct rdma_ah_attr *ah_attr = attr->ah_attr;
4145         const struct ib_gid_attr *sgid_attr;
4146         struct irdma_device *iwdev = to_iwdev(ibah->pd->device);
4147         struct irdma_pci_f *rf = iwdev->rf;
4148         struct irdma_sc_ah *sc_ah;
4149         u32 ah_id = 0;
4150         struct irdma_ah_info *ah_info;
4151         struct irdma_create_ah_resp uresp;
4152         union {
4153                 struct sockaddr saddr;
4154                 struct sockaddr_in saddr_in;
4155                 struct sockaddr_in6 saddr_in6;
4156         } sgid_addr, dgid_addr;
4157         int err;
4158         u8 dmac[ETH_ALEN];
4159
4160         err = irdma_alloc_rsrc(rf, rf->allocated_ahs, rf->max_ah, &ah_id,
4161                                &rf->next_ah);
4162         if (err)
4163                 return err;
4164
4165         ah->pd = pd;
4166         sc_ah = &ah->sc_ah;
4167         sc_ah->ah_info.ah_idx = ah_id;
4168         sc_ah->ah_info.vsi = &iwdev->vsi;
4169         irdma_sc_init_ah(&rf->sc_dev, sc_ah);
4170         ah->sgid_index = ah_attr->grh.sgid_index;
4171         sgid_attr = ah_attr->grh.sgid_attr;
4172         memcpy(&ah->dgid, &ah_attr->grh.dgid, sizeof(ah->dgid));
4173         rdma_gid2ip((struct sockaddr *)&sgid_addr, &sgid_attr->gid);
4174         rdma_gid2ip((struct sockaddr *)&dgid_addr, &ah_attr->grh.dgid);
4175         ah->av.attrs = *ah_attr;
4176         ah->av.net_type = rdma_gid_attr_network_type(sgid_attr);
4177         ah->av.sgid_addr.saddr = sgid_addr.saddr;
4178         ah->av.dgid_addr.saddr = dgid_addr.saddr;
4179         ah_info = &sc_ah->ah_info;
4180         ah_info->ah_idx = ah_id;
4181         ah_info->pd_idx = pd->sc_pd.pd_id;
4182         if (ah_attr->ah_flags & IB_AH_GRH) {
4183                 ah_info->flow_label = ah_attr->grh.flow_label;
4184                 ah_info->hop_ttl = ah_attr->grh.hop_limit;
4185                 ah_info->tc_tos = ah_attr->grh.traffic_class;
4186         }
4187
4188         ether_addr_copy(dmac, ah_attr->roce.dmac);
4189         if (rdma_gid_attr_network_type(sgid_attr) == RDMA_NETWORK_IPV4) {
4190                 ah_info->ipv4_valid = true;
4191                 ah_info->dest_ip_addr[0] =
4192                         ntohl(dgid_addr.saddr_in.sin_addr.s_addr);
4193                 ah_info->src_ip_addr[0] =
4194                         ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
4195                 ah_info->do_lpbk = irdma_ipv4_is_lpb(ah_info->src_ip_addr[0],
4196                                                      ah_info->dest_ip_addr[0]);
4197                 if (ipv4_is_multicast(dgid_addr.saddr_in.sin_addr.s_addr)) {
4198                         ah_info->do_lpbk = true;
4199                         irdma_mcast_mac(ah_info->dest_ip_addr, dmac, true);
4200                 }
4201         } else {
4202                 irdma_copy_ip_ntohl(ah_info->dest_ip_addr,
4203                                     dgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4204                 irdma_copy_ip_ntohl(ah_info->src_ip_addr,
4205                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4206                 ah_info->do_lpbk = irdma_ipv6_is_lpb(ah_info->src_ip_addr,
4207                                                      ah_info->dest_ip_addr);
4208                 if (rdma_is_multicast_addr(&dgid_addr.saddr_in6.sin6_addr)) {
4209                         ah_info->do_lpbk = true;
4210                         irdma_mcast_mac(ah_info->dest_ip_addr, dmac, false);
4211                 }
4212         }
4213
4214         err = rdma_read_gid_l2_fields(sgid_attr, &ah_info->vlan_tag,
4215                                       ah_info->mac_addr);
4216         if (err)
4217                 goto error;
4218
4219         ah_info->dst_arpindex = irdma_add_arp(iwdev->rf, ah_info->dest_ip_addr,
4220                                               ah_info->ipv4_valid, dmac);
4221
4222         if (ah_info->dst_arpindex == -1) {
4223                 err = -EINVAL;
4224                 goto error;
4225         }
4226
4227         if (ah_info->vlan_tag >= VLAN_N_VID && iwdev->dcb)
4228                 ah_info->vlan_tag = 0;
4229
4230         if (ah_info->vlan_tag < VLAN_N_VID) {
4231                 ah_info->insert_vlan_tag = true;
4232                 ah_info->vlan_tag |=
4233                         rt_tos2priority(ah_info->tc_tos) << VLAN_PRIO_SHIFT;
4234         }
4235
4236         err = irdma_ah_cqp_op(iwdev->rf, sc_ah, IRDMA_OP_AH_CREATE,
4237                               attr->flags & RDMA_CREATE_AH_SLEEPABLE,
4238                               irdma_gsi_ud_qp_ah_cb, sc_ah);
4239
4240         if (err) {
4241                 ibdev_dbg(&iwdev->ibdev,
4242                           "VERBS: CQP-OP Create AH fail");
4243                 goto error;
4244         }
4245
4246         if (!(attr->flags & RDMA_CREATE_AH_SLEEPABLE)) {
4247                 int cnt = CQP_COMPL_WAIT_TIME_MS * CQP_TIMEOUT_THRESHOLD;
4248
4249                 do {
4250                         irdma_cqp_ce_handler(rf, &rf->ccq.sc_cq);
4251                         mdelay(1);
4252                 } while (!sc_ah->ah_info.ah_valid && --cnt);
4253
4254                 if (!cnt) {
4255                         ibdev_dbg(&iwdev->ibdev,
4256                                   "VERBS: CQP create AH timed out");
4257                         err = -ETIMEDOUT;
4258                         goto error;
4259                 }
4260         }
4261
4262         if (udata) {
4263                 uresp.ah_id = ah->sc_ah.ah_info.ah_idx;
4264                 err = ib_copy_to_udata(udata, &uresp,
4265                                        min(sizeof(uresp), udata->outlen));
4266         }
4267         return 0;
4268
4269 error:
4270         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_ahs, ah_id);
4271
4272         return err;
4273 }
4274
4275 /**
4276  * irdma_destroy_ah - Destroy address handle
4277  * @ibah: pointer to address handle
4278  * @ah_flags: flags for sleepable
4279  */
4280 static int irdma_destroy_ah(struct ib_ah *ibah, u32 ah_flags)
4281 {
4282         struct irdma_device *iwdev = to_iwdev(ibah->device);
4283         struct irdma_ah *ah = to_iwah(ibah);
4284
4285         irdma_ah_cqp_op(iwdev->rf, &ah->sc_ah, IRDMA_OP_AH_DESTROY,
4286                         false, NULL, ah);
4287
4288         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_ahs,
4289                         ah->sc_ah.ah_info.ah_idx);
4290
4291         return 0;
4292 }
4293
4294 /**
4295  * irdma_query_ah - Query address handle
4296  * @ibah: pointer to address handle
4297  * @ah_attr: address handle attributes
4298  */
4299 static int irdma_query_ah(struct ib_ah *ibah, struct rdma_ah_attr *ah_attr)
4300 {
4301         struct irdma_ah *ah = to_iwah(ibah);
4302
4303         memset(ah_attr, 0, sizeof(*ah_attr));
4304         if (ah->av.attrs.ah_flags & IB_AH_GRH) {
4305                 ah_attr->ah_flags = IB_AH_GRH;
4306                 ah_attr->grh.flow_label = ah->sc_ah.ah_info.flow_label;
4307                 ah_attr->grh.traffic_class = ah->sc_ah.ah_info.tc_tos;
4308                 ah_attr->grh.hop_limit = ah->sc_ah.ah_info.hop_ttl;
4309                 ah_attr->grh.sgid_index = ah->sgid_index;
4310                 ah_attr->grh.sgid_index = ah->sgid_index;
4311                 memcpy(&ah_attr->grh.dgid, &ah->dgid,
4312                        sizeof(ah_attr->grh.dgid));
4313         }
4314
4315         return 0;
4316 }
4317
4318 static enum rdma_link_layer irdma_get_link_layer(struct ib_device *ibdev,
4319                                                  u32 port_num)
4320 {
4321         return IB_LINK_LAYER_ETHERNET;
4322 }
4323
4324 static __be64 irdma_mac_to_guid(struct net_device *ndev)
4325 {
4326         unsigned char *mac = ndev->dev_addr;
4327         __be64 guid;
4328         unsigned char *dst = (unsigned char *)&guid;
4329
4330         dst[0] = mac[0] ^ 2;
4331         dst[1] = mac[1];
4332         dst[2] = mac[2];
4333         dst[3] = 0xff;
4334         dst[4] = 0xfe;
4335         dst[5] = mac[3];
4336         dst[6] = mac[4];
4337         dst[7] = mac[5];
4338
4339         return guid;
4340 }
4341
4342 static const struct ib_device_ops irdma_roce_dev_ops = {
4343         .attach_mcast = irdma_attach_mcast,
4344         .create_ah = irdma_create_ah,
4345         .create_user_ah = irdma_create_ah,
4346         .destroy_ah = irdma_destroy_ah,
4347         .detach_mcast = irdma_detach_mcast,
4348         .get_link_layer = irdma_get_link_layer,
4349         .get_port_immutable = irdma_roce_port_immutable,
4350         .modify_qp = irdma_modify_qp_roce,
4351         .query_ah = irdma_query_ah,
4352         .query_pkey = irdma_query_pkey,
4353 };
4354
4355 static const struct ib_device_ops irdma_iw_dev_ops = {
4356         .modify_qp = irdma_modify_qp,
4357         .get_port_immutable = irdma_iw_port_immutable,
4358         .query_gid = irdma_query_gid,
4359 };
4360
4361 static const struct ib_device_ops irdma_dev_ops = {
4362         .owner = THIS_MODULE,
4363         .driver_id = RDMA_DRIVER_IRDMA,
4364         .uverbs_abi_ver = IRDMA_ABI_VER,
4365
4366         .alloc_hw_port_stats = irdma_alloc_hw_port_stats,
4367         .alloc_mr = irdma_alloc_mr,
4368         .alloc_mw = irdma_alloc_mw,
4369         .alloc_pd = irdma_alloc_pd,
4370         .alloc_ucontext = irdma_alloc_ucontext,
4371         .create_cq = irdma_create_cq,
4372         .create_qp = irdma_create_qp,
4373         .dealloc_driver = irdma_ib_dealloc_device,
4374         .dealloc_mw = irdma_dealloc_mw,
4375         .dealloc_pd = irdma_dealloc_pd,
4376         .dealloc_ucontext = irdma_dealloc_ucontext,
4377         .dereg_mr = irdma_dereg_mr,
4378         .destroy_cq = irdma_destroy_cq,
4379         .destroy_qp = irdma_destroy_qp,
4380         .disassociate_ucontext = irdma_disassociate_ucontext,
4381         .get_dev_fw_str = irdma_get_dev_fw_str,
4382         .get_dma_mr = irdma_get_dma_mr,
4383         .get_hw_stats = irdma_get_hw_stats,
4384         .map_mr_sg = irdma_map_mr_sg,
4385         .mmap = irdma_mmap,
4386         .mmap_free = irdma_mmap_free,
4387         .poll_cq = irdma_poll_cq,
4388         .post_recv = irdma_post_recv,
4389         .post_send = irdma_post_send,
4390         .query_device = irdma_query_device,
4391         .query_port = irdma_query_port,
4392         .query_qp = irdma_query_qp,
4393         .reg_user_mr = irdma_reg_user_mr,
4394         .req_notify_cq = irdma_req_notify_cq,
4395         .resize_cq = irdma_resize_cq,
4396         INIT_RDMA_OBJ_SIZE(ib_pd, irdma_pd, ibpd),
4397         INIT_RDMA_OBJ_SIZE(ib_ucontext, irdma_ucontext, ibucontext),
4398         INIT_RDMA_OBJ_SIZE(ib_ah, irdma_ah, ibah),
4399         INIT_RDMA_OBJ_SIZE(ib_cq, irdma_cq, ibcq),
4400         INIT_RDMA_OBJ_SIZE(ib_mw, irdma_mr, ibmw),
4401         INIT_RDMA_OBJ_SIZE(ib_qp, irdma_qp, ibqp),
4402 };
4403
4404 /**
4405  * irdma_init_roce_device - initialization of roce rdma device
4406  * @iwdev: irdma device
4407  */
4408 static void irdma_init_roce_device(struct irdma_device *iwdev)
4409 {
4410         iwdev->ibdev.node_type = RDMA_NODE_IB_CA;
4411         iwdev->ibdev.node_guid = irdma_mac_to_guid(iwdev->netdev);
4412         ib_set_device_ops(&iwdev->ibdev, &irdma_roce_dev_ops);
4413 }
4414
4415 /**
4416  * irdma_init_iw_device - initialization of iwarp rdma device
4417  * @iwdev: irdma device
4418  */
4419 static int irdma_init_iw_device(struct irdma_device *iwdev)
4420 {
4421         struct net_device *netdev = iwdev->netdev;
4422
4423         iwdev->ibdev.node_type = RDMA_NODE_RNIC;
4424         ether_addr_copy((u8 *)&iwdev->ibdev.node_guid, netdev->dev_addr);
4425         iwdev->ibdev.ops.iw_add_ref = irdma_qp_add_ref;
4426         iwdev->ibdev.ops.iw_rem_ref = irdma_qp_rem_ref;
4427         iwdev->ibdev.ops.iw_get_qp = irdma_get_qp;
4428         iwdev->ibdev.ops.iw_connect = irdma_connect;
4429         iwdev->ibdev.ops.iw_accept = irdma_accept;
4430         iwdev->ibdev.ops.iw_reject = irdma_reject;
4431         iwdev->ibdev.ops.iw_create_listen = irdma_create_listen;
4432         iwdev->ibdev.ops.iw_destroy_listen = irdma_destroy_listen;
4433         memcpy(iwdev->ibdev.iw_ifname, netdev->name,
4434                sizeof(iwdev->ibdev.iw_ifname));
4435         ib_set_device_ops(&iwdev->ibdev, &irdma_iw_dev_ops);
4436
4437         return 0;
4438 }
4439
4440 /**
4441  * irdma_init_rdma_device - initialization of rdma device
4442  * @iwdev: irdma device
4443  */
4444 static int irdma_init_rdma_device(struct irdma_device *iwdev)
4445 {
4446         struct pci_dev *pcidev = iwdev->rf->pcidev;
4447         int ret;
4448
4449         if (iwdev->roce_mode) {
4450                 irdma_init_roce_device(iwdev);
4451         } else {
4452                 ret = irdma_init_iw_device(iwdev);
4453                 if (ret)
4454                         return ret;
4455         }
4456         iwdev->ibdev.phys_port_cnt = 1;
4457         iwdev->ibdev.num_comp_vectors = iwdev->rf->ceqs_count;
4458         iwdev->ibdev.dev.parent = &pcidev->dev;
4459         ib_set_device_ops(&iwdev->ibdev, &irdma_dev_ops);
4460
4461         return 0;
4462 }
4463
4464 /**
4465  * irdma_port_ibevent - indicate port event
4466  * @iwdev: irdma device
4467  */
4468 void irdma_port_ibevent(struct irdma_device *iwdev)
4469 {
4470         struct ib_event event;
4471
4472         event.device = &iwdev->ibdev;
4473         event.element.port_num = 1;
4474         event.event =
4475                 iwdev->iw_status ? IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
4476         ib_dispatch_event(&event);
4477 }
4478
4479 /**
4480  * irdma_ib_unregister_device - unregister rdma device from IB
4481  * core
4482  * @iwdev: irdma device
4483  */
4484 void irdma_ib_unregister_device(struct irdma_device *iwdev)
4485 {
4486         iwdev->iw_status = 0;
4487         irdma_port_ibevent(iwdev);
4488         ib_unregister_device(&iwdev->ibdev);
4489 }
4490
4491 /**
4492  * irdma_ib_register_device - register irdma device to IB core
4493  * @iwdev: irdma device
4494  */
4495 int irdma_ib_register_device(struct irdma_device *iwdev)
4496 {
4497         int ret;
4498
4499         ret = irdma_init_rdma_device(iwdev);
4500         if (ret)
4501                 return ret;
4502
4503         ret = ib_device_set_netdev(&iwdev->ibdev, iwdev->netdev, 1);
4504         if (ret)
4505                 goto error;
4506         dma_set_max_seg_size(iwdev->rf->hw.device, UINT_MAX);
4507         ret = ib_register_device(&iwdev->ibdev, "irdma%d", iwdev->rf->hw.device);
4508         if (ret)
4509                 goto error;
4510
4511         iwdev->iw_status = 1;
4512         irdma_port_ibevent(iwdev);
4513
4514         return 0;
4515
4516 error:
4517         if (ret)
4518                 ibdev_dbg(&iwdev->ibdev, "VERBS: Register RDMA device fail\n");
4519
4520         return ret;
4521 }
4522
4523 /**
4524  * irdma_ib_dealloc_device
4525  * @ibdev: ib device
4526  *
4527  * callback from ibdev dealloc_driver to deallocate resources
4528  * unber irdma device
4529  */
4530 void irdma_ib_dealloc_device(struct ib_device *ibdev)
4531 {
4532         struct irdma_device *iwdev = to_iwdev(ibdev);
4533
4534         irdma_rt_deinit_hw(iwdev);
4535         irdma_ctrl_deinit_hw(iwdev->rf);
4536         kfree(iwdev->rf);
4537 }