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 int 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         return 0;
 579 }
 580
 581 /**
 582  * irdma_setup_kmode_qp - setup initialization for kernel mode qp
 583  * @iwdev: iwarp device
 584  * @iwqp: qp ptr (user or kernel)
 585  * @info: initialize info to return
 586  * @init_attr: Initial QP create attributes
 587  */
 588 static int irdma_setup_kmode_qp(struct irdma_device *iwdev,
 589                                 struct irdma_qp *iwqp,
 590                                 struct irdma_qp_init_info *info,
 591                                 struct ib_qp_init_attr *init_attr)
 592 {
 593         struct irdma_dma_mem *mem = &iwqp->kqp.dma_mem;
 594         u32 sqdepth, rqdepth;
 595         u8 sqshift, rqshift;
 596         u32 size;
 597         enum irdma_status_code status;
 598         struct irdma_qp_uk_init_info *ukinfo = &info->qp_uk_init_info;
 599         struct irdma_uk_attrs *uk_attrs = &iwdev->rf->sc_dev.hw_attrs.uk_attrs;
 600
 601         irdma_get_wqe_shift(uk_attrs,
 602                 uk_attrs->hw_rev >= IRDMA_GEN_2 ? ukinfo->max_sq_frag_cnt + 1 :
 603                                                   ukinfo->max_sq_frag_cnt,
 604                 ukinfo->max_inline_data, &sqshift);
 605         status = irdma_get_sqdepth(uk_attrs, ukinfo->sq_size, sqshift,
 606                                    &sqdepth);
 607         if (status)
 608                 return -ENOMEM;
 609
 610         if (uk_attrs->hw_rev == IRDMA_GEN_1)
 611                 rqshift = IRDMA_MAX_RQ_WQE_SHIFT_GEN1;
 612         else
 613                 irdma_get_wqe_shift(uk_attrs, ukinfo->max_rq_frag_cnt, 0,
 614                                     &rqshift);
 615
 616         status = irdma_get_rqdepth(uk_attrs, ukinfo->rq_size, rqshift,
 617                                    &rqdepth);
 618         if (status)
 619                 return -ENOMEM;
 620
 621         iwqp->kqp.sq_wrid_mem =
 622                 kcalloc(sqdepth, sizeof(*iwqp->kqp.sq_wrid_mem), GFP_KERNEL);
 623         if (!iwqp->kqp.sq_wrid_mem)
 624                 return -ENOMEM;
 625
 626         iwqp->kqp.rq_wrid_mem =
 627                 kcalloc(rqdepth, sizeof(*iwqp->kqp.rq_wrid_mem), GFP_KERNEL);
 628         if (!iwqp->kqp.rq_wrid_mem) {
 629                 kfree(iwqp->kqp.sq_wrid_mem);
 630                 iwqp->kqp.sq_wrid_mem = NULL;
 631                 return -ENOMEM;
 632         }
 633
 634         ukinfo->sq_wrtrk_array = iwqp->kqp.sq_wrid_mem;
 635         ukinfo->rq_wrid_array = iwqp->kqp.rq_wrid_mem;
 636
 637         size = (sqdepth + rqdepth) * IRDMA_QP_WQE_MIN_SIZE;
 638         size += (IRDMA_SHADOW_AREA_SIZE << 3);
 639
 640         mem->size = ALIGN(size, 256);
 641         mem->va = dma_alloc_coherent(iwdev->rf->hw.device, mem->size,
 642                                      &mem->pa, GFP_KERNEL);
 643         if (!mem->va) {
 644                 kfree(iwqp->kqp.sq_wrid_mem);
 645                 iwqp->kqp.sq_wrid_mem = NULL;
 646                 kfree(iwqp->kqp.rq_wrid_mem);
 647                 iwqp->kqp.rq_wrid_mem = NULL;
 648                 return -ENOMEM;
 649         }
 650
 651         ukinfo->sq = mem->va;
 652         info->sq_pa = mem->pa;
 653         ukinfo->rq = &ukinfo->sq[sqdepth];
 654         info->rq_pa = info->sq_pa + (sqdepth * IRDMA_QP_WQE_MIN_SIZE);
 655         ukinfo->shadow_area = ukinfo->rq[rqdepth].elem;
 656         info->shadow_area_pa = info->rq_pa + (rqdepth * IRDMA_QP_WQE_MIN_SIZE);
 657         ukinfo->sq_size = sqdepth >> sqshift;
 658         ukinfo->rq_size = rqdepth >> rqshift;
 659         ukinfo->qp_id = iwqp->ibqp.qp_num;
 660
 661         init_attr->cap.max_send_wr = (sqdepth - IRDMA_SQ_RSVD) >> sqshift;
 662         init_attr->cap.max_recv_wr = (rqdepth - IRDMA_RQ_RSVD) >> rqshift;
 663
 664         return 0;
 665 }
 666
 667 static int irdma_cqp_create_qp_cmd(struct irdma_qp *iwqp)
 668 {
 669         struct irdma_pci_f *rf = iwqp->iwdev->rf;
 670         struct irdma_cqp_request *cqp_request;
 671         struct cqp_cmds_info *cqp_info;
 672         struct irdma_create_qp_info *qp_info;
 673         enum irdma_status_code status;
 674
 675         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
 676         if (!cqp_request)
 677                 return -ENOMEM;
 678
 679         cqp_info = &cqp_request->info;
 680         qp_info = &cqp_request->info.in.u.qp_create.info;
 681         memset(qp_info, 0, sizeof(*qp_info));
 682         qp_info->mac_valid = true;
 683         qp_info->cq_num_valid = true;
 684         qp_info->next_iwarp_state = IRDMA_QP_STATE_IDLE;
 685
 686         cqp_info->cqp_cmd = IRDMA_OP_QP_CREATE;
 687         cqp_info->post_sq = 1;
 688         cqp_info->in.u.qp_create.qp = &iwqp->sc_qp;
 689         cqp_info->in.u.qp_create.scratch = (uintptr_t)cqp_request;
 690         status = irdma_handle_cqp_op(rf, cqp_request);
 691         irdma_put_cqp_request(&rf->cqp, cqp_request);
 692
 693         return status ? -ENOMEM : 0;
 694 }
 695
 696 static void irdma_roce_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
 697                                                struct irdma_qp_host_ctx_info *ctx_info)
 698 {
 699         struct irdma_device *iwdev = iwqp->iwdev;
 700         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 701         struct irdma_roce_offload_info *roce_info;
 702         struct irdma_udp_offload_info *udp_info;
 703
 704         udp_info = &iwqp->udp_info;
 705         udp_info->snd_mss = ib_mtu_enum_to_int(ib_mtu_int_to_enum(iwdev->vsi.mtu));
 706         udp_info->cwnd = iwdev->roce_cwnd;
 707         udp_info->rexmit_thresh = 2;
 708         udp_info->rnr_nak_thresh = 2;
 709         udp_info->src_port = 0xc000;
 710         udp_info->dst_port = ROCE_V2_UDP_DPORT;
 711         roce_info = &iwqp->roce_info;
 712         ether_addr_copy(roce_info->mac_addr, iwdev->netdev->dev_addr);
 713
 714         roce_info->rd_en = true;
 715         roce_info->wr_rdresp_en = true;
 716         roce_info->bind_en = true;
 717         roce_info->dcqcn_en = false;
 718         roce_info->rtomin = 5;
 719
 720         roce_info->ack_credits = iwdev->roce_ackcreds;
 721         roce_info->ird_size = dev->hw_attrs.max_hw_ird;
 722         roce_info->ord_size = dev->hw_attrs.max_hw_ord;
 723
 724         if (!iwqp->user_mode) {
 725                 roce_info->priv_mode_en = true;
 726                 roce_info->fast_reg_en = true;
 727                 roce_info->udprivcq_en = true;
 728         }
 729         roce_info->roce_tver = 0;
 730
 731         ctx_info->roce_info = &iwqp->roce_info;
 732         ctx_info->udp_info = &iwqp->udp_info;
 733         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
 734 }
 735
 736 static void irdma_iw_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
 737                                              struct irdma_qp_host_ctx_info *ctx_info)
 738 {
 739         struct irdma_device *iwdev = iwqp->iwdev;
 740         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 741         struct irdma_iwarp_offload_info *iwarp_info;
 742
 743         iwarp_info = &iwqp->iwarp_info;
 744         ether_addr_copy(iwarp_info->mac_addr, iwdev->netdev->dev_addr);
 745         iwarp_info->rd_en = true;
 746         iwarp_info->wr_rdresp_en = true;
 747         iwarp_info->bind_en = true;
 748         iwarp_info->ecn_en = true;
 749         iwarp_info->rtomin = 5;
 750
 751         if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
 752                 iwarp_info->ib_rd_en = true;
 753         if (!iwqp->user_mode) {
 754                 iwarp_info->priv_mode_en = true;
 755                 iwarp_info->fast_reg_en = true;
 756         }
 757         iwarp_info->ddp_ver = 1;
 758         iwarp_info->rdmap_ver = 1;
 759
 760         ctx_info->iwarp_info = &iwqp->iwarp_info;
 761         ctx_info->iwarp_info_valid = true;
 762         irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
 763         ctx_info->iwarp_info_valid = false;
 764 }
 765
 766 static int irdma_validate_qp_attrs(struct ib_qp_init_attr *init_attr,
 767                                    struct irdma_device *iwdev)
 768 {
 769         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 770         struct irdma_uk_attrs *uk_attrs = &dev->hw_attrs.uk_attrs;
 771
 772         if (init_attr->create_flags)
 773                 return -EOPNOTSUPP;
 774
 775         if (init_attr->cap.max_inline_data > uk_attrs->max_hw_inline ||
 776             init_attr->cap.max_send_sge > uk_attrs->max_hw_wq_frags ||
 777             init_attr->cap.max_recv_sge > uk_attrs->max_hw_wq_frags)
 778                 return -EINVAL;
 779
 780         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 781                 if (init_attr->qp_type != IB_QPT_RC &&
 782                     init_attr->qp_type != IB_QPT_UD &&
 783                     init_attr->qp_type != IB_QPT_GSI)
 784                         return -EOPNOTSUPP;
 785         } else {
 786                 if (init_attr->qp_type != IB_QPT_RC)
 787                         return -EOPNOTSUPP;
 788         }
 789
 790         return 0;
 791 }
 792
 793 /**
 794  * irdma_create_qp - create qp
 795  * @ibpd: ptr of pd
 796  * @init_attr: attributes for qp
 797  * @udata: user data for create qp
 798  */
 799 static struct ib_qp *irdma_create_qp(struct ib_pd *ibpd,
 800                                      struct ib_qp_init_attr *init_attr,
 801                                      struct ib_udata *udata)
 802 {
 803         struct irdma_pd *iwpd = to_iwpd(ibpd);
 804         struct irdma_device *iwdev = to_iwdev(ibpd->device);
 805         struct irdma_pci_f *rf = iwdev->rf;
 806         struct irdma_qp *iwqp;
 807         struct irdma_create_qp_req req;
 808         struct irdma_create_qp_resp uresp = {};
 809         u32 qp_num = 0;
 810         enum irdma_status_code ret;
 811         int err_code;
 812         int sq_size;
 813         int rq_size;
 814         struct irdma_sc_qp *qp;
 815         struct irdma_sc_dev *dev = &rf->sc_dev;
 816         struct irdma_uk_attrs *uk_attrs = &dev->hw_attrs.uk_attrs;
 817         struct irdma_qp_init_info init_info = {};
 818         struct irdma_qp_host_ctx_info *ctx_info;
 819         unsigned long flags;
 820
 821         err_code = irdma_validate_qp_attrs(init_attr, iwdev);
 822         if (err_code)
 823                 return ERR_PTR(err_code);
 824
 825         sq_size = init_attr->cap.max_send_wr;
 826         rq_size = init_attr->cap.max_recv_wr;
 827
 828         init_info.vsi = &iwdev->vsi;
 829         init_info.qp_uk_init_info.uk_attrs = uk_attrs;
 830         init_info.qp_uk_init_info.sq_size = sq_size;
 831         init_info.qp_uk_init_info.rq_size = rq_size;
 832         init_info.qp_uk_init_info.max_sq_frag_cnt = init_attr->cap.max_send_sge;
 833         init_info.qp_uk_init_info.max_rq_frag_cnt = init_attr->cap.max_recv_sge;
 834         init_info.qp_uk_init_info.max_inline_data = init_attr->cap.max_inline_data;
 835
 836         iwqp = kzalloc(sizeof(*iwqp), GFP_KERNEL);
 837         if (!iwqp)
 838                 return ERR_PTR(-ENOMEM);
 839
 840         qp = &iwqp->sc_qp;
 841         qp->qp_uk.back_qp = iwqp;
 842         qp->qp_uk.lock = &iwqp->lock;
 843         qp->push_idx = IRDMA_INVALID_PUSH_PAGE_INDEX;
 844
 845         iwqp->iwdev = iwdev;
 846         iwqp->q2_ctx_mem.size = ALIGN(IRDMA_Q2_BUF_SIZE + IRDMA_QP_CTX_SIZE,
 847                                       256);
 848         iwqp->q2_ctx_mem.va = dma_alloc_coherent(dev->hw->device,
 849                                                  iwqp->q2_ctx_mem.size,
 850                                                  &iwqp->q2_ctx_mem.pa,
 851                                                  GFP_KERNEL);
 852         if (!iwqp->q2_ctx_mem.va) {
 853                 err_code = -ENOMEM;
 854                 goto error;
 855         }
 856
 857         init_info.q2 = iwqp->q2_ctx_mem.va;
 858         init_info.q2_pa = iwqp->q2_ctx_mem.pa;
 859         init_info.host_ctx = (__le64 *)(init_info.q2 + IRDMA_Q2_BUF_SIZE);
 860         init_info.host_ctx_pa = init_info.q2_pa + IRDMA_Q2_BUF_SIZE;
 861
 862         if (init_attr->qp_type == IB_QPT_GSI)
 863                 qp_num = 1;
 864         else
 865                 err_code = irdma_alloc_rsrc(rf, rf->allocated_qps, rf->max_qp,
 866                                             &qp_num, &rf->next_qp);
 867         if (err_code)
 868                 goto error;
 869
 870         iwqp->iwpd = iwpd;
 871         iwqp->ibqp.qp_num = qp_num;
 872         qp = &iwqp->sc_qp;
 873         iwqp->iwscq = to_iwcq(init_attr->send_cq);
 874         iwqp->iwrcq = to_iwcq(init_attr->recv_cq);
 875         iwqp->host_ctx.va = init_info.host_ctx;
 876         iwqp->host_ctx.pa = init_info.host_ctx_pa;
 877         iwqp->host_ctx.size = IRDMA_QP_CTX_SIZE;
 878
 879         init_info.pd = &iwpd->sc_pd;
 880         init_info.qp_uk_init_info.qp_id = iwqp->ibqp.qp_num;
 881         if (!rdma_protocol_roce(&iwdev->ibdev, 1))
 882                 init_info.qp_uk_init_info.first_sq_wq = 1;
 883         iwqp->ctx_info.qp_compl_ctx = (uintptr_t)qp;
 884         init_waitqueue_head(&iwqp->waitq);
 885         init_waitqueue_head(&iwqp->mod_qp_waitq);
 886
 887         if (udata) {
 888                 err_code = ib_copy_from_udata(&req, udata,
 889                                               min(sizeof(req), udata->inlen));
 890                 if (err_code) {
 891                         ibdev_dbg(&iwdev->ibdev,
 892                                   "VERBS: ib_copy_from_data fail\n");
 893                         goto error;
 894                 }
 895
 896                 iwqp->ctx_info.qp_compl_ctx = req.user_compl_ctx;
 897                 iwqp->user_mode = 1;
 898                 if (req.user_wqe_bufs) {
 899                         struct irdma_ucontext *ucontext =
 900                                 rdma_udata_to_drv_context(udata,
 901                                                           struct irdma_ucontext,
 902                                                           ibucontext);
 903
 904                         init_info.qp_uk_init_info.legacy_mode = ucontext->legacy_mode;
 905                         spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
 906                         iwqp->iwpbl = irdma_get_pbl((unsigned long)req.user_wqe_bufs,
 907                                                     &ucontext->qp_reg_mem_list);
 908                         spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
 909
 910                         if (!iwqp->iwpbl) {
 911                                 err_code = -ENODATA;
 912                                 ibdev_dbg(&iwdev->ibdev, "VERBS: no pbl info\n");
 913                                 goto error;
 914                         }
 915                 }
 916                 init_info.qp_uk_init_info.abi_ver = iwpd->sc_pd.abi_ver;
 917                 err_code = irdma_setup_virt_qp(iwdev, iwqp, &init_info);
 918         } else {
 919                 init_info.qp_uk_init_info.abi_ver = IRDMA_ABI_VER;
 920                 err_code = irdma_setup_kmode_qp(iwdev, iwqp, &init_info, init_attr);
 921         }
 922
 923         if (err_code) {
 924                 ibdev_dbg(&iwdev->ibdev, "VERBS: setup qp failed\n");
 925                 goto error;
 926         }
 927
 928         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 929                 if (init_attr->qp_type == IB_QPT_RC) {
 930                         init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_ROCE_RC;
 931                         init_info.qp_uk_init_info.qp_caps = IRDMA_SEND_WITH_IMM |
 932                                                             IRDMA_WRITE_WITH_IMM |
 933                                                             IRDMA_ROCE;
 934                 } else {
 935                         init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_ROCE_UD;
 936                         init_info.qp_uk_init_info.qp_caps = IRDMA_SEND_WITH_IMM |
 937                                                             IRDMA_ROCE;
 938                 }
 939         } else {
 940                 init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_IWARP;
 941                 init_info.qp_uk_init_info.qp_caps = IRDMA_WRITE_WITH_IMM;
 942         }
 943
 944         if (dev->hw_attrs.uk_attrs.hw_rev > IRDMA_GEN_1)
 945                 init_info.qp_uk_init_info.qp_caps |= IRDMA_PUSH_MODE;
 946
 947         ret = irdma_sc_qp_init(qp, &init_info);
 948         if (ret) {
 949                 err_code = -EPROTO;
 950                 ibdev_dbg(&iwdev->ibdev, "VERBS: qp_init fail\n");
 951                 goto error;
 952         }
 953
 954         ctx_info = &iwqp->ctx_info;
 955         ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
 956         ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
 957
 958         if (rdma_protocol_roce(&iwdev->ibdev, 1))
 959                 irdma_roce_fill_and_set_qpctx_info(iwqp, ctx_info);
 960         else
 961                 irdma_iw_fill_and_set_qpctx_info(iwqp, ctx_info);
 962
 963         err_code = irdma_cqp_create_qp_cmd(iwqp);
 964         if (err_code)
 965                 goto error;
 966
 967         refcount_set(&iwqp->refcnt, 1);
 968         spin_lock_init(&iwqp->lock);
 969         spin_lock_init(&iwqp->sc_qp.pfpdu.lock);
 970         iwqp->sig_all = (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ? 1 : 0;
 971         rf->qp_table[qp_num] = iwqp;
 972         iwqp->max_send_wr = sq_size;
 973         iwqp->max_recv_wr = rq_size;
 974
 975         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 976                 if (dev->ws_add(&iwdev->vsi, 0)) {
 977                         irdma_cqp_qp_destroy_cmd(&rf->sc_dev, &iwqp->sc_qp);
 978                         err_code = -EINVAL;
 979                         goto error;
 980                 }
 981
 982                 irdma_qp_add_qos(&iwqp->sc_qp);
 983         }
 984
 985         if (udata) {
 986                 /* GEN_1 legacy support with libi40iw does not have expanded uresp struct */
 987                 if (udata->outlen < sizeof(uresp)) {
 988                         uresp.lsmm = 1;
 989                         uresp.push_idx = IRDMA_INVALID_PUSH_PAGE_INDEX_GEN_1;
 990                 } else {
 991                         if (rdma_protocol_iwarp(&iwdev->ibdev, 1))
 992                                 uresp.lsmm = 1;
 993                 }
 994                 uresp.actual_sq_size = sq_size;
 995                 uresp.actual_rq_size = rq_size;
 996                 uresp.qp_id = qp_num;
 997                 uresp.qp_caps = qp->qp_uk.qp_caps;
 998
 999                 err_code = ib_copy_to_udata(udata, &uresp,
1000                                             min(sizeof(uresp), udata->outlen));
1001                 if (err_code) {
1002                         ibdev_dbg(&iwdev->ibdev, "VERBS: copy_to_udata failed\n");
1003                         irdma_destroy_qp(&iwqp->ibqp, udata);
1004                         return ERR_PTR(err_code);
1005                 }
1006         }
1007
1008         init_completion(&iwqp->free_qp);
1009         return &iwqp->ibqp;
1010
1011 error:
1012         irdma_free_qp_rsrc(iwqp);
1013
1014         return ERR_PTR(err_code);
1015 }
1016
1017 static int irdma_get_ib_acc_flags(struct irdma_qp *iwqp)
1018 {
1019         int acc_flags = 0;
1020
1021         if (rdma_protocol_roce(iwqp->ibqp.device, 1)) {
1022                 if (iwqp->roce_info.wr_rdresp_en) {
1023                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
1024                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
1025                 }
1026                 if (iwqp->roce_info.rd_en)
1027                         acc_flags |= IB_ACCESS_REMOTE_READ;
1028                 if (iwqp->roce_info.bind_en)
1029                         acc_flags |= IB_ACCESS_MW_BIND;
1030         } else {
1031                 if (iwqp->iwarp_info.wr_rdresp_en) {
1032                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
1033                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
1034                 }
1035                 if (iwqp->iwarp_info.rd_en)
1036                         acc_flags |= IB_ACCESS_REMOTE_READ;
1037                 if (iwqp->iwarp_info.bind_en)
1038                         acc_flags |= IB_ACCESS_MW_BIND;
1039         }
1040         return acc_flags;
1041 }
1042
1043 /**
1044  * irdma_query_qp - query qp attributes
1045  * @ibqp: qp pointer
1046  * @attr: attributes pointer
1047  * @attr_mask: Not used
1048  * @init_attr: qp attributes to return
1049  */
1050 static int irdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1051                           int attr_mask, struct ib_qp_init_attr *init_attr)
1052 {
1053         struct irdma_qp *iwqp = to_iwqp(ibqp);
1054         struct irdma_sc_qp *qp = &iwqp->sc_qp;
1055
1056         memset(attr, 0, sizeof(*attr));
1057         memset(init_attr, 0, sizeof(*init_attr));
1058
1059         attr->qp_state = iwqp->ibqp_state;
1060         attr->cur_qp_state = iwqp->ibqp_state;
1061         attr->cap.max_send_wr = iwqp->max_send_wr;
1062         attr->cap.max_recv_wr = iwqp->max_recv_wr;
1063         attr->cap.max_inline_data = qp->qp_uk.max_inline_data;
1064         attr->cap.max_send_sge = qp->qp_uk.max_sq_frag_cnt;
1065         attr->cap.max_recv_sge = qp->qp_uk.max_rq_frag_cnt;
1066         attr->qp_access_flags = irdma_get_ib_acc_flags(iwqp);
1067         attr->port_num = 1;
1068         if (rdma_protocol_roce(ibqp->device, 1)) {
1069                 attr->path_mtu = ib_mtu_int_to_enum(iwqp->udp_info.snd_mss);
1070                 attr->qkey = iwqp->roce_info.qkey;
1071                 attr->rq_psn = iwqp->udp_info.epsn;
1072                 attr->sq_psn = iwqp->udp_info.psn_nxt;
1073                 attr->dest_qp_num = iwqp->roce_info.dest_qp;
1074                 attr->pkey_index = iwqp->roce_info.p_key;
1075                 attr->retry_cnt = iwqp->udp_info.rexmit_thresh;
1076                 attr->rnr_retry = iwqp->udp_info.rnr_nak_thresh;
1077                 attr->max_rd_atomic = iwqp->roce_info.ord_size;
1078                 attr->max_dest_rd_atomic = iwqp->roce_info.ird_size;
1079         }
1080
1081         init_attr->event_handler = iwqp->ibqp.event_handler;
1082         init_attr->qp_context = iwqp->ibqp.qp_context;
1083         init_attr->send_cq = iwqp->ibqp.send_cq;
1084         init_attr->recv_cq = iwqp->ibqp.recv_cq;
1085         init_attr->cap = attr->cap;
1086
1087         return 0;
1088 }
1089
1090 /**
1091  * irdma_query_pkey - Query partition key
1092  * @ibdev: device pointer from stack
1093  * @port: port number
1094  * @index: index of pkey
1095  * @pkey: pointer to store the pkey
1096  */
1097 static int irdma_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1098                             u16 *pkey)
1099 {
1100         if (index >= IRDMA_PKEY_TBL_SZ)
1101                 return -EINVAL;
1102
1103         *pkey = IRDMA_DEFAULT_PKEY;
1104         return 0;
1105 }
1106
1107 /**
1108  * irdma_modify_qp_roce - modify qp request
1109  * @ibqp: qp's pointer for modify
1110  * @attr: access attributes
1111  * @attr_mask: state mask
1112  * @udata: user data
1113  */
1114 int irdma_modify_qp_roce(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1115                          int attr_mask, struct ib_udata *udata)
1116 {
1117         struct irdma_pd *iwpd = to_iwpd(ibqp->pd);
1118         struct irdma_qp *iwqp = to_iwqp(ibqp);
1119         struct irdma_device *iwdev = iwqp->iwdev;
1120         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
1121         struct irdma_qp_host_ctx_info *ctx_info;
1122         struct irdma_roce_offload_info *roce_info;
1123         struct irdma_udp_offload_info *udp_info;
1124         struct irdma_modify_qp_info info = {};
1125         struct irdma_modify_qp_resp uresp = {};
1126         struct irdma_modify_qp_req ureq = {};
1127         unsigned long flags;
1128         u8 issue_modify_qp = 0;
1129         int ret = 0;
1130
1131         ctx_info = &iwqp->ctx_info;
1132         roce_info = &iwqp->roce_info;
1133         udp_info = &iwqp->udp_info;
1134
1135         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1136                 return -EOPNOTSUPP;
1137
1138         if (attr_mask & IB_QP_DEST_QPN)
1139                 roce_info->dest_qp = attr->dest_qp_num;
1140
1141         if (attr_mask & IB_QP_PKEY_INDEX) {
1142                 ret = irdma_query_pkey(ibqp->device, 0, attr->pkey_index,
1143                                        &roce_info->p_key);
1144                 if (ret)
1145                         return ret;
1146         }
1147
1148         if (attr_mask & IB_QP_QKEY)
1149                 roce_info->qkey = attr->qkey;
1150
1151         if (attr_mask & IB_QP_PATH_MTU)
1152                 udp_info->snd_mss = ib_mtu_enum_to_int(attr->path_mtu);
1153
1154         if (attr_mask & IB_QP_SQ_PSN) {
1155                 udp_info->psn_nxt = attr->sq_psn;
1156                 udp_info->lsn =  0xffff;
1157                 udp_info->psn_una = attr->sq_psn;
1158                 udp_info->psn_max = attr->sq_psn;
1159         }
1160
1161         if (attr_mask & IB_QP_RQ_PSN)
1162                 udp_info->epsn = attr->rq_psn;
1163
1164         if (attr_mask & IB_QP_RNR_RETRY)
1165                 udp_info->rnr_nak_thresh = attr->rnr_retry;
1166
1167         if (attr_mask & IB_QP_RETRY_CNT)
1168                 udp_info->rexmit_thresh = attr->retry_cnt;
1169
1170         ctx_info->roce_info->pd_id = iwpd->sc_pd.pd_id;
1171
1172         if (attr_mask & IB_QP_AV) {
1173                 struct irdma_av *av = &iwqp->roce_ah.av;
1174                 const struct ib_gid_attr *sgid_attr;
1175                 u16 vlan_id = VLAN_N_VID;
1176                 u32 local_ip[4];
1177
1178                 memset(&iwqp->roce_ah, 0, sizeof(iwqp->roce_ah));
1179                 if (attr->ah_attr.ah_flags & IB_AH_GRH) {
1180                         udp_info->ttl = attr->ah_attr.grh.hop_limit;
1181                         udp_info->flow_label = attr->ah_attr.grh.flow_label;
1182                         udp_info->tos = attr->ah_attr.grh.traffic_class;
1183                         irdma_qp_rem_qos(&iwqp->sc_qp);
1184                         dev->ws_remove(iwqp->sc_qp.vsi, ctx_info->user_pri);
1185                         ctx_info->user_pri = rt_tos2priority(udp_info->tos);
1186                         iwqp->sc_qp.user_pri = ctx_info->user_pri;
1187                         if (dev->ws_add(iwqp->sc_qp.vsi, ctx_info->user_pri))
1188                                 return -ENOMEM;
1189                         irdma_qp_add_qos(&iwqp->sc_qp);
1190                 }
1191                 sgid_attr = attr->ah_attr.grh.sgid_attr;
1192                 ret = rdma_read_gid_l2_fields(sgid_attr, &vlan_id,
1193                                               ctx_info->roce_info->mac_addr);
1194                 if (ret)
1195                         return ret;
1196
1197                 if (vlan_id >= VLAN_N_VID && iwdev->dcb)
1198                         vlan_id = 0;
1199                 if (vlan_id < VLAN_N_VID) {
1200                         udp_info->insert_vlan_tag = true;
1201                         udp_info->vlan_tag = vlan_id |
1202                                 ctx_info->user_pri << VLAN_PRIO_SHIFT;
1203                 } else {
1204                         udp_info->insert_vlan_tag = false;
1205                 }
1206
1207                 av->attrs = attr->ah_attr;
1208                 rdma_gid2ip((struct sockaddr *)&av->sgid_addr, &sgid_attr->gid);
1209                 rdma_gid2ip((struct sockaddr *)&av->dgid_addr, &attr->ah_attr.grh.dgid);
1210                 roce_info->local_qp = ibqp->qp_num;
1211                 if (av->sgid_addr.saddr.sa_family == AF_INET6) {
1212                         __be32 *daddr =
1213                                 av->dgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32;
1214                         __be32 *saddr =
1215                                 av->sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32;
1216
1217                         irdma_copy_ip_ntohl(&udp_info->dest_ip_addr[0], daddr);
1218                         irdma_copy_ip_ntohl(&udp_info->local_ipaddr[0], saddr);
1219
1220                         udp_info->ipv4 = false;
1221                         irdma_copy_ip_ntohl(local_ip, daddr);
1222
1223                         udp_info->arp_idx = irdma_arp_table(iwdev->rf,
1224                                                             &local_ip[0],
1225                                                             false, NULL,
1226                                                             IRDMA_ARP_RESOLVE);
1227                 } else {
1228                         __be32 saddr = av->sgid_addr.saddr_in.sin_addr.s_addr;
1229                         __be32 daddr = av->dgid_addr.saddr_in.sin_addr.s_addr;
1230
1231                         local_ip[0] = ntohl(daddr);
1232
1233                         udp_info->ipv4 = true;
1234                         udp_info->dest_ip_addr[0] = 0;
1235                         udp_info->dest_ip_addr[1] = 0;
1236                         udp_info->dest_ip_addr[2] = 0;
1237                         udp_info->dest_ip_addr[3] = local_ip[0];
1238
1239                         udp_info->local_ipaddr[0] = 0;
1240                         udp_info->local_ipaddr[1] = 0;
1241                         udp_info->local_ipaddr[2] = 0;
1242                         udp_info->local_ipaddr[3] = ntohl(saddr);
1243                 }
1244                 udp_info->arp_idx =
1245                         irdma_add_arp(iwdev->rf, local_ip, udp_info->ipv4,
1246                                       attr->ah_attr.roce.dmac);
1247         }
1248
1249         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1250                 if (attr->max_rd_atomic > dev->hw_attrs.max_hw_ord) {
1251                         ibdev_err(&iwdev->ibdev,
1252                                   "rd_atomic = %d, above max_hw_ord=%d\n",
1253                                   attr->max_rd_atomic,
1254                                   dev->hw_attrs.max_hw_ord);
1255                         return -EINVAL;
1256                 }
1257                 if (attr->max_rd_atomic)
1258                         roce_info->ord_size = attr->max_rd_atomic;
1259                 info.ord_valid = true;
1260         }
1261
1262         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1263                 if (attr->max_dest_rd_atomic > dev->hw_attrs.max_hw_ird) {
1264                         ibdev_err(&iwdev->ibdev,
1265                                   "rd_atomic = %d, above max_hw_ird=%d\n",
1266                                    attr->max_rd_atomic,
1267                                    dev->hw_attrs.max_hw_ird);
1268                         return -EINVAL;
1269                 }
1270                 if (attr->max_dest_rd_atomic)
1271                         roce_info->ird_size = attr->max_dest_rd_atomic;
1272         }
1273
1274         if (attr_mask & IB_QP_ACCESS_FLAGS) {
1275                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
1276                         roce_info->wr_rdresp_en = true;
1277                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
1278                         roce_info->wr_rdresp_en = true;
1279                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
1280                         roce_info->rd_en = true;
1281         }
1282
1283         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
1284
1285         ibdev_dbg(&iwdev->ibdev,
1286                   "VERBS: caller: %pS qp_id=%d to_ibqpstate=%d ibqpstate=%d irdma_qpstate=%d attr_mask=0x%x\n",
1287                   __builtin_return_address(0), ibqp->qp_num, attr->qp_state,
1288                   iwqp->ibqp_state, iwqp->iwarp_state, attr_mask);
1289
1290         spin_lock_irqsave(&iwqp->lock, flags);
1291         if (attr_mask & IB_QP_STATE) {
1292                 if (!ib_modify_qp_is_ok(iwqp->ibqp_state, attr->qp_state,
1293                                         iwqp->ibqp.qp_type, attr_mask)) {
1294                         ibdev_warn(&iwdev->ibdev, "modify_qp invalid for qp_id=%d, old_state=0x%x, new_state=0x%x\n",
1295                                    iwqp->ibqp.qp_num, iwqp->ibqp_state,
1296                                    attr->qp_state);
1297                         ret = -EINVAL;
1298                         goto exit;
1299                 }
1300                 info.curr_iwarp_state = iwqp->iwarp_state;
1301
1302                 switch (attr->qp_state) {
1303                 case IB_QPS_INIT:
1304                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1305                                 ret = -EINVAL;
1306                                 goto exit;
1307                         }
1308
1309                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
1310                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
1311                                 issue_modify_qp = 1;
1312                         }
1313                         break;
1314                 case IB_QPS_RTR:
1315                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1316                                 ret = -EINVAL;
1317                                 goto exit;
1318                         }
1319                         info.arp_cache_idx_valid = true;
1320                         info.cq_num_valid = true;
1321                         info.next_iwarp_state = IRDMA_QP_STATE_RTR;
1322                         issue_modify_qp = 1;
1323                         break;
1324                 case IB_QPS_RTS:
1325                         if (iwqp->ibqp_state < IB_QPS_RTR ||
1326                             iwqp->ibqp_state == IB_QPS_ERR) {
1327                                 ret = -EINVAL;
1328                                 goto exit;
1329                         }
1330
1331                         info.arp_cache_idx_valid = true;
1332                         info.cq_num_valid = true;
1333                         info.ord_valid = true;
1334                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
1335                         issue_modify_qp = 1;
1336                         if (iwdev->push_mode && udata &&
1337                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1338                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1339                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1340                                 irdma_alloc_push_page(iwqp);
1341                                 spin_lock_irqsave(&iwqp->lock, flags);
1342                         }
1343                         break;
1344                 case IB_QPS_SQD:
1345                         if (iwqp->iwarp_state == IRDMA_QP_STATE_SQD)
1346                                 goto exit;
1347
1348                         if (iwqp->iwarp_state != IRDMA_QP_STATE_RTS) {
1349                                 ret = -EINVAL;
1350                                 goto exit;
1351                         }
1352
1353                         info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1354                         issue_modify_qp = 1;
1355                         break;
1356                 case IB_QPS_SQE:
1357                 case IB_QPS_ERR:
1358                 case IB_QPS_RESET:
1359                         if (iwqp->iwarp_state == IRDMA_QP_STATE_RTS) {
1360                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1361                                 info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1362                                 irdma_hw_modify_qp(iwdev, iwqp, &info, true);
1363                                 spin_lock_irqsave(&iwqp->lock, flags);
1364                         }
1365
1366                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1367                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1368                                 if (udata) {
1369                                         if (ib_copy_from_udata(&ureq, udata,
1370                                             min(sizeof(ureq), udata->inlen)))
1371                                                 return -EINVAL;
1372
1373                                         irdma_flush_wqes(iwqp,
1374                                             (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1375                                             (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1376                                             IRDMA_REFLUSH);
1377                                 }
1378                                 return 0;
1379                         }
1380
1381                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1382                         issue_modify_qp = 1;
1383                         break;
1384                 default:
1385                         ret = -EINVAL;
1386                         goto exit;
1387                 }
1388
1389                 iwqp->ibqp_state = attr->qp_state;
1390         }
1391
1392         ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1393         ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1394         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1395         spin_unlock_irqrestore(&iwqp->lock, flags);
1396
1397         if (attr_mask & IB_QP_STATE) {
1398                 if (issue_modify_qp) {
1399                         ctx_info->rem_endpoint_idx = udp_info->arp_idx;
1400                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1401                                 return -EINVAL;
1402                         spin_lock_irqsave(&iwqp->lock, flags);
1403                         if (iwqp->iwarp_state == info.curr_iwarp_state) {
1404                                 iwqp->iwarp_state = info.next_iwarp_state;
1405                                 iwqp->ibqp_state = attr->qp_state;
1406                         }
1407                         if (iwqp->ibqp_state > IB_QPS_RTS &&
1408                             !iwqp->flush_issued) {
1409                                 iwqp->flush_issued = 1;
1410                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1411                                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_SQ |
1412                                                        IRDMA_FLUSH_RQ |
1413                                                        IRDMA_FLUSH_WAIT);
1414                         } else {
1415                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1416                         }
1417                 } else {
1418                         iwqp->ibqp_state = attr->qp_state;
1419                 }
1420                 if (udata && dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1421                         struct irdma_ucontext *ucontext;
1422
1423                         ucontext = rdma_udata_to_drv_context(udata,
1424                                         struct irdma_ucontext, ibucontext);
1425                         if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1426                             !iwqp->push_wqe_mmap_entry &&
1427                             !irdma_setup_push_mmap_entries(ucontext, iwqp,
1428                                 &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1429                                 uresp.push_valid = 1;
1430                                 uresp.push_offset = iwqp->sc_qp.push_offset;
1431                         }
1432                         ret = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1433                                                udata->outlen));
1434                         if (ret) {
1435                                 irdma_remove_push_mmap_entries(iwqp);
1436                                 ibdev_dbg(&iwdev->ibdev,
1437                                           "VERBS: copy_to_udata failed\n");
1438                                 return ret;
1439                         }
1440                 }
1441         }
1442
1443         return 0;
1444 exit:
1445         spin_unlock_irqrestore(&iwqp->lock, flags);
1446
1447         return ret;
1448 }
1449
1450 /**
1451  * irdma_modify_qp - modify qp request
1452  * @ibqp: qp's pointer for modify
1453  * @attr: access attributes
1454  * @attr_mask: state mask
1455  * @udata: user data
1456  */
1457 int irdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
1458                     struct ib_udata *udata)
1459 {
1460         struct irdma_qp *iwqp = to_iwqp(ibqp);
1461         struct irdma_device *iwdev = iwqp->iwdev;
1462         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
1463         struct irdma_qp_host_ctx_info *ctx_info;
1464         struct irdma_tcp_offload_info *tcp_info;
1465         struct irdma_iwarp_offload_info *offload_info;
1466         struct irdma_modify_qp_info info = {};
1467         struct irdma_modify_qp_resp uresp = {};
1468         struct irdma_modify_qp_req ureq = {};
1469         u8 issue_modify_qp = 0;
1470         u8 dont_wait = 0;
1471         int err;
1472         unsigned long flags;
1473
1474         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1475                 return -EOPNOTSUPP;
1476
1477         ctx_info = &iwqp->ctx_info;
1478         offload_info = &iwqp->iwarp_info;
1479         tcp_info = &iwqp->tcp_info;
1480         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
1481         ibdev_dbg(&iwdev->ibdev,
1482                   "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",
1483                   __builtin_return_address(0), ibqp->qp_num, attr->qp_state,
1484                   iwqp->ibqp_state, iwqp->iwarp_state, iwqp->last_aeq,
1485                   iwqp->hw_tcp_state, iwqp->hw_iwarp_state, attr_mask);
1486
1487         spin_lock_irqsave(&iwqp->lock, flags);
1488         if (attr_mask & IB_QP_STATE) {
1489                 info.curr_iwarp_state = iwqp->iwarp_state;
1490                 switch (attr->qp_state) {
1491                 case IB_QPS_INIT:
1492                 case IB_QPS_RTR:
1493                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1494                                 err = -EINVAL;
1495                                 goto exit;
1496                         }
1497
1498                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
1499                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
1500                                 issue_modify_qp = 1;
1501                         }
1502                         if (iwdev->push_mode && udata &&
1503                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1504                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1505                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1506                                 irdma_alloc_push_page(iwqp);
1507                                 spin_lock_irqsave(&iwqp->lock, flags);
1508                         }
1509                         break;
1510                 case IB_QPS_RTS:
1511                         if (iwqp->iwarp_state > IRDMA_QP_STATE_RTS ||
1512                             !iwqp->cm_id) {
1513                                 err = -EINVAL;
1514                                 goto exit;
1515                         }
1516
1517                         issue_modify_qp = 1;
1518                         iwqp->hw_tcp_state = IRDMA_TCP_STATE_ESTABLISHED;
1519                         iwqp->hte_added = 1;
1520                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
1521                         info.tcp_ctx_valid = true;
1522                         info.ord_valid = true;
1523                         info.arp_cache_idx_valid = true;
1524                         info.cq_num_valid = true;
1525                         break;
1526                 case IB_QPS_SQD:
1527                         if (iwqp->hw_iwarp_state > IRDMA_QP_STATE_RTS) {
1528                                 err = 0;
1529                                 goto exit;
1530                         }
1531
1532                         if (iwqp->iwarp_state == IRDMA_QP_STATE_CLOSING ||
1533                             iwqp->iwarp_state < IRDMA_QP_STATE_RTS) {
1534                                 err = 0;
1535                                 goto exit;
1536                         }
1537
1538                         if (iwqp->iwarp_state > IRDMA_QP_STATE_CLOSING) {
1539                                 err = -EINVAL;
1540                                 goto exit;
1541                         }
1542
1543                         info.next_iwarp_state = IRDMA_QP_STATE_CLOSING;
1544                         issue_modify_qp = 1;
1545                         break;
1546                 case IB_QPS_SQE:
1547                         if (iwqp->iwarp_state >= IRDMA_QP_STATE_TERMINATE) {
1548                                 err = -EINVAL;
1549                                 goto exit;
1550                         }
1551
1552                         info.next_iwarp_state = IRDMA_QP_STATE_TERMINATE;
1553                         issue_modify_qp = 1;
1554                         break;
1555                 case IB_QPS_ERR:
1556                 case IB_QPS_RESET:
1557                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1558                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1559                                 if (udata) {
1560                                         if (ib_copy_from_udata(&ureq, udata,
1561                                             min(sizeof(ureq), udata->inlen)))
1562                                                 return -EINVAL;
1563
1564                                         irdma_flush_wqes(iwqp,
1565                                             (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1566                                             (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1567                                             IRDMA_REFLUSH);
1568                                 }
1569                                 return 0;
1570                         }
1571
1572                         if (iwqp->sc_qp.term_flags) {
1573                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1574                                 irdma_terminate_del_timer(&iwqp->sc_qp);
1575                                 spin_lock_irqsave(&iwqp->lock, flags);
1576                         }
1577                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1578                         if (iwqp->hw_tcp_state > IRDMA_TCP_STATE_CLOSED &&
1579                             iwdev->iw_status &&
1580                             iwqp->hw_tcp_state != IRDMA_TCP_STATE_TIME_WAIT)
1581                                 info.reset_tcp_conn = true;
1582                         else
1583                                 dont_wait = 1;
1584
1585                         issue_modify_qp = 1;
1586                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1587                         break;
1588                 default:
1589                         err = -EINVAL;
1590                         goto exit;
1591                 }
1592
1593                 iwqp->ibqp_state = attr->qp_state;
1594         }
1595         if (attr_mask & IB_QP_ACCESS_FLAGS) {
1596                 ctx_info->iwarp_info_valid = true;
1597                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
1598                         offload_info->wr_rdresp_en = true;
1599                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
1600                         offload_info->wr_rdresp_en = true;
1601                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
1602                         offload_info->rd_en = true;
1603         }
1604
1605         if (ctx_info->iwarp_info_valid) {
1606                 ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1607                 ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1608                 irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1609         }
1610         spin_unlock_irqrestore(&iwqp->lock, flags);
1611
1612         if (attr_mask & IB_QP_STATE) {
1613                 if (issue_modify_qp) {
1614                         ctx_info->rem_endpoint_idx = tcp_info->arp_idx;
1615                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1616                                 return -EINVAL;
1617                 }
1618
1619                 spin_lock_irqsave(&iwqp->lock, flags);
1620                 if (iwqp->iwarp_state == info.curr_iwarp_state) {
1621                         iwqp->iwarp_state = info.next_iwarp_state;
1622                         iwqp->ibqp_state = attr->qp_state;
1623                 }
1624                 spin_unlock_irqrestore(&iwqp->lock, flags);
1625         }
1626
1627         if (issue_modify_qp && iwqp->ibqp_state > IB_QPS_RTS) {
1628                 if (dont_wait) {
1629                         if (iwqp->cm_id && iwqp->hw_tcp_state) {
1630                                 spin_lock_irqsave(&iwqp->lock, flags);
1631                                 iwqp->hw_tcp_state = IRDMA_TCP_STATE_CLOSED;
1632                                 iwqp->last_aeq = IRDMA_AE_RESET_SENT;
1633                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1634                                 irdma_cm_disconn(iwqp);
1635                         }
1636                 } else {
1637                         int close_timer_started;
1638
1639                         spin_lock_irqsave(&iwdev->cm_core.ht_lock, flags);
1640
1641                         if (iwqp->cm_node) {
1642                                 refcount_inc(&iwqp->cm_node->refcnt);
1643                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1644                                 close_timer_started = atomic_inc_return(&iwqp->close_timer_started);
1645                                 if (iwqp->cm_id && close_timer_started == 1)
1646                                         irdma_schedule_cm_timer(iwqp->cm_node,
1647                                                 (struct irdma_puda_buf *)iwqp,
1648                                                 IRDMA_TIMER_TYPE_CLOSE, 1, 0);
1649
1650                                 irdma_rem_ref_cm_node(iwqp->cm_node);
1651                         } else {
1652                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1653                         }
1654                 }
1655         }
1656         if (attr_mask & IB_QP_STATE && udata &&
1657             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1658                 struct irdma_ucontext *ucontext;
1659
1660                 ucontext = rdma_udata_to_drv_context(udata,
1661                                         struct irdma_ucontext, ibucontext);
1662                 if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1663                     !iwqp->push_wqe_mmap_entry &&
1664                     !irdma_setup_push_mmap_entries(ucontext, iwqp,
1665                         &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1666                         uresp.push_valid = 1;
1667                         uresp.push_offset = iwqp->sc_qp.push_offset;
1668                 }
1669
1670                 err = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1671                                        udata->outlen));
1672                 if (err) {
1673                         irdma_remove_push_mmap_entries(iwqp);
1674                         ibdev_dbg(&iwdev->ibdev,
1675                                   "VERBS: copy_to_udata failed\n");
1676                         return err;
1677                 }
1678         }
1679
1680         return 0;
1681 exit:
1682         spin_unlock_irqrestore(&iwqp->lock, flags);
1683
1684         return err;
1685 }
1686
1687 /**
1688  * irdma_cq_free_rsrc - free up resources for cq
1689  * @rf: RDMA PCI function
1690  * @iwcq: cq ptr
1691  */
1692 static void irdma_cq_free_rsrc(struct irdma_pci_f *rf, struct irdma_cq *iwcq)
1693 {
1694         struct irdma_sc_cq *cq = &iwcq->sc_cq;
1695
1696         if (!iwcq->user_mode) {
1697                 dma_free_coherent(rf->sc_dev.hw->device, iwcq->kmem.size,
1698                                   iwcq->kmem.va, iwcq->kmem.pa);
1699                 iwcq->kmem.va = NULL;
1700                 dma_free_coherent(rf->sc_dev.hw->device,
1701                                   iwcq->kmem_shadow.size,
1702                                   iwcq->kmem_shadow.va, iwcq->kmem_shadow.pa);
1703                 iwcq->kmem_shadow.va = NULL;
1704         }
1705
1706         irdma_free_rsrc(rf, rf->allocated_cqs, cq->cq_uk.cq_id);
1707 }
1708
1709 /**
1710  * irdma_free_cqbuf - worker to free a cq buffer
1711  * @work: provides access to the cq buffer to free
1712  */
1713 static void irdma_free_cqbuf(struct work_struct *work)
1714 {
1715         struct irdma_cq_buf *cq_buf = container_of(work, struct irdma_cq_buf, work);
1716
1717         dma_free_coherent(cq_buf->hw->device, cq_buf->kmem_buf.size,
1718                           cq_buf->kmem_buf.va, cq_buf->kmem_buf.pa);
1719         cq_buf->kmem_buf.va = NULL;
1720         kfree(cq_buf);
1721 }
1722
1723 /**
1724  * irdma_process_resize_list - remove resized cq buffers from the resize_list
1725  * @iwcq: cq which owns the resize_list
1726  * @iwdev: irdma device
1727  * @lcqe_buf: the buffer where the last cqe is received
1728  */
1729 static int irdma_process_resize_list(struct irdma_cq *iwcq,
1730                                      struct irdma_device *iwdev,
1731                                      struct irdma_cq_buf *lcqe_buf)
1732 {
1733         struct list_head *tmp_node, *list_node;
1734         struct irdma_cq_buf *cq_buf;
1735         int cnt = 0;
1736
1737         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
1738                 cq_buf = list_entry(list_node, struct irdma_cq_buf, list);
1739                 if (cq_buf == lcqe_buf)
1740                         return cnt;
1741
1742                 list_del(&cq_buf->list);
1743                 queue_work(iwdev->cleanup_wq, &cq_buf->work);
1744                 cnt++;
1745         }
1746
1747         return cnt;
1748 }
1749
1750 /**
1751  * irdma_destroy_cq - destroy cq
1752  * @ib_cq: cq pointer
1753  * @udata: user data
1754  */
1755 static int irdma_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata)
1756 {
1757         struct irdma_device *iwdev = to_iwdev(ib_cq->device);
1758         struct irdma_cq *iwcq = to_iwcq(ib_cq);
1759         struct irdma_sc_cq *cq = &iwcq->sc_cq;
1760         struct irdma_sc_dev *dev = cq->dev;
1761         struct irdma_sc_ceq *ceq = dev->ceq[cq->ceq_id];
1762         struct irdma_ceq *iwceq = container_of(ceq, struct irdma_ceq, sc_ceq);
1763         unsigned long flags;
1764
1765         spin_lock_irqsave(&iwcq->lock, flags);
1766         if (!list_empty(&iwcq->resize_list))
1767                 irdma_process_resize_list(iwcq, iwdev, NULL);
1768         spin_unlock_irqrestore(&iwcq->lock, flags);
1769
1770         irdma_cq_wq_destroy(iwdev->rf, cq);
1771         irdma_cq_free_rsrc(iwdev->rf, iwcq);
1772
1773         spin_lock_irqsave(&iwceq->ce_lock, flags);
1774         irdma_sc_cleanup_ceqes(cq, ceq);
1775         spin_unlock_irqrestore(&iwceq->ce_lock, flags);
1776
1777         return 0;
1778 }
1779
1780 /**
1781  * irdma_resize_cq - resize cq
1782  * @ibcq: cq to be resized
1783  * @entries: desired cq size
1784  * @udata: user data
1785  */
1786 static int irdma_resize_cq(struct ib_cq *ibcq, int entries,
1787                            struct ib_udata *udata)
1788 {
1789         struct irdma_cq *iwcq = to_iwcq(ibcq);
1790         struct irdma_sc_dev *dev = iwcq->sc_cq.dev;
1791         struct irdma_cqp_request *cqp_request;
1792         struct cqp_cmds_info *cqp_info;
1793         struct irdma_modify_cq_info *m_info;
1794         struct irdma_modify_cq_info info = {};
1795         struct irdma_dma_mem kmem_buf;
1796         struct irdma_cq_mr *cqmr_buf;
1797         struct irdma_pbl *iwpbl_buf;
1798         struct irdma_device *iwdev;
1799         struct irdma_pci_f *rf;
1800         struct irdma_cq_buf *cq_buf = NULL;
1801         enum irdma_status_code status = 0;
1802         unsigned long flags;
1803         int ret;
1804
1805         iwdev = to_iwdev(ibcq->device);
1806         rf = iwdev->rf;
1807
1808         if (!(rf->sc_dev.hw_attrs.uk_attrs.feature_flags &
1809             IRDMA_FEATURE_CQ_RESIZE))
1810                 return -EOPNOTSUPP;
1811
1812         if (entries > rf->max_cqe)
1813                 return -EINVAL;
1814
1815         if (!iwcq->user_mode) {
1816                 entries++;
1817                 if (rf->sc_dev.hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
1818                         entries *= 2;
1819         }
1820
1821         info.cq_size = max(entries, 4);
1822
1823         if (info.cq_size == iwcq->sc_cq.cq_uk.cq_size - 1)
1824                 return 0;
1825
1826         if (udata) {
1827                 struct irdma_resize_cq_req req = {};
1828                 struct irdma_ucontext *ucontext =
1829                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
1830                                                   ibucontext);
1831
1832                 /* CQ resize not supported with legacy GEN_1 libi40iw */
1833                 if (ucontext->legacy_mode)
1834                         return -EOPNOTSUPP;
1835
1836                 if (ib_copy_from_udata(&req, udata,
1837                                        min(sizeof(req), udata->inlen)))
1838                         return -EINVAL;
1839
1840                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
1841                 iwpbl_buf = irdma_get_pbl((unsigned long)req.user_cq_buffer,
1842                                           &ucontext->cq_reg_mem_list);
1843                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
1844
1845                 if (!iwpbl_buf)
1846                         return -ENOMEM;
1847
1848                 cqmr_buf = &iwpbl_buf->cq_mr;
1849                 if (iwpbl_buf->pbl_allocated) {
1850                         info.virtual_map = true;
1851                         info.pbl_chunk_size = 1;
1852                         info.first_pm_pbl_idx = cqmr_buf->cq_pbl.idx;
1853                 } else {
1854                         info.cq_pa = cqmr_buf->cq_pbl.addr;
1855                 }
1856         } else {
1857                 /* Kmode CQ resize */
1858                 int rsize;
1859
1860                 rsize = info.cq_size * sizeof(struct irdma_cqe);
1861                 kmem_buf.size = ALIGN(round_up(rsize, 256), 256);
1862                 kmem_buf.va = dma_alloc_coherent(dev->hw->device,
1863                                                  kmem_buf.size, &kmem_buf.pa,
1864                                                  GFP_KERNEL);
1865                 if (!kmem_buf.va)
1866                         return -ENOMEM;
1867
1868                 info.cq_base = kmem_buf.va;
1869                 info.cq_pa = kmem_buf.pa;
1870                 cq_buf = kzalloc(sizeof(*cq_buf), GFP_KERNEL);
1871                 if (!cq_buf) {
1872                         ret = -ENOMEM;
1873                         goto error;
1874                 }
1875         }
1876
1877         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
1878         if (!cqp_request) {
1879                 ret = -ENOMEM;
1880                 goto error;
1881         }
1882
1883         info.shadow_read_threshold = iwcq->sc_cq.shadow_read_threshold;
1884         info.cq_resize = true;
1885
1886         cqp_info = &cqp_request->info;
1887         m_info = &cqp_info->in.u.cq_modify.info;
1888         memcpy(m_info, &info, sizeof(*m_info));
1889
1890         cqp_info->cqp_cmd = IRDMA_OP_CQ_MODIFY;
1891         cqp_info->in.u.cq_modify.cq = &iwcq->sc_cq;
1892         cqp_info->in.u.cq_modify.scratch = (uintptr_t)cqp_request;
1893         cqp_info->post_sq = 1;
1894         status = irdma_handle_cqp_op(rf, cqp_request);
1895         irdma_put_cqp_request(&rf->cqp, cqp_request);
1896         if (status) {
1897                 ret = -EPROTO;
1898                 goto error;
1899         }
1900
1901         spin_lock_irqsave(&iwcq->lock, flags);
1902         if (cq_buf) {
1903                 cq_buf->kmem_buf = iwcq->kmem;
1904                 cq_buf->hw = dev->hw;
1905                 memcpy(&cq_buf->cq_uk, &iwcq->sc_cq.cq_uk, sizeof(cq_buf->cq_uk));
1906                 INIT_WORK(&cq_buf->work, irdma_free_cqbuf);
1907                 list_add_tail(&cq_buf->list, &iwcq->resize_list);
1908                 iwcq->kmem = kmem_buf;
1909         }
1910
1911         irdma_sc_cq_resize(&iwcq->sc_cq, &info);
1912         ibcq->cqe = info.cq_size - 1;
1913         spin_unlock_irqrestore(&iwcq->lock, flags);
1914
1915         return 0;
1916 error:
1917         if (!udata) {
1918                 dma_free_coherent(dev->hw->device, kmem_buf.size, kmem_buf.va,
1919                                   kmem_buf.pa);
1920                 kmem_buf.va = NULL;
1921         }
1922         kfree(cq_buf);
1923
1924         return ret;
1925 }
1926
1927 static inline int cq_validate_flags(u32 flags, u8 hw_rev)
1928 {
1929         /* GEN1 does not support CQ create flags */
1930         if (hw_rev == IRDMA_GEN_1)
1931                 return flags ? -EOPNOTSUPP : 0;
1932
1933         return flags & ~IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION ? -EOPNOTSUPP : 0;
1934 }
1935
1936 /**
1937  * irdma_create_cq - create cq
1938  * @ibcq: CQ allocated
1939  * @attr: attributes for cq
1940  * @udata: user data
1941  */
1942 static int irdma_create_cq(struct ib_cq *ibcq,
1943                            const struct ib_cq_init_attr *attr,
1944                            struct ib_udata *udata)
1945 {
1946         struct ib_device *ibdev = ibcq->device;
1947         struct irdma_device *iwdev = to_iwdev(ibdev);
1948         struct irdma_pci_f *rf = iwdev->rf;
1949         struct irdma_cq *iwcq = to_iwcq(ibcq);
1950         u32 cq_num = 0;
1951         struct irdma_sc_cq *cq;
1952         struct irdma_sc_dev *dev = &rf->sc_dev;
1953         struct irdma_cq_init_info info = {};
1954         enum irdma_status_code status;
1955         struct irdma_cqp_request *cqp_request;
1956         struct cqp_cmds_info *cqp_info;
1957         struct irdma_cq_uk_init_info *ukinfo = &info.cq_uk_init_info;
1958         unsigned long flags;
1959         int err_code;
1960         int entries = attr->cqe;
1961
1962         err_code = cq_validate_flags(attr->flags, dev->hw_attrs.uk_attrs.hw_rev);
1963         if (err_code)
1964                 return err_code;
1965         err_code = irdma_alloc_rsrc(rf, rf->allocated_cqs, rf->max_cq, &cq_num,
1966                                     &rf->next_cq);
1967         if (err_code)
1968                 return err_code;
1969
1970         cq = &iwcq->sc_cq;
1971         cq->back_cq = iwcq;
1972         spin_lock_init(&iwcq->lock);
1973         INIT_LIST_HEAD(&iwcq->resize_list);
1974         info.dev = dev;
1975         ukinfo->cq_size = max(entries, 4);
1976         ukinfo->cq_id = cq_num;
1977         iwcq->ibcq.cqe = info.cq_uk_init_info.cq_size;
1978         if (attr->comp_vector < rf->ceqs_count)
1979                 info.ceq_id = attr->comp_vector;
1980         info.ceq_id_valid = true;
1981         info.ceqe_mask = 1;
1982         info.type = IRDMA_CQ_TYPE_IWARP;
1983         info.vsi = &iwdev->vsi;
1984
1985         if (udata) {
1986                 struct irdma_ucontext *ucontext;
1987                 struct irdma_create_cq_req req = {};
1988                 struct irdma_cq_mr *cqmr;
1989                 struct irdma_pbl *iwpbl;
1990                 struct irdma_pbl *iwpbl_shadow;
1991                 struct irdma_cq_mr *cqmr_shadow;
1992
1993                 iwcq->user_mode = true;
1994                 ucontext =
1995                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
1996                                                   ibucontext);
1997                 if (ib_copy_from_udata(&req, udata,
1998                                        min(sizeof(req), udata->inlen))) {
1999                         err_code = -EFAULT;
2000                         goto cq_free_rsrc;
2001                 }
2002
2003                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2004                 iwpbl = irdma_get_pbl((unsigned long)req.user_cq_buf,
2005                                       &ucontext->cq_reg_mem_list);
2006                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2007                 if (!iwpbl) {
2008                         err_code = -EPROTO;
2009                         goto cq_free_rsrc;
2010                 }
2011
2012                 iwcq->iwpbl = iwpbl;
2013                 iwcq->cq_mem_size = 0;
2014                 cqmr = &iwpbl->cq_mr;
2015
2016                 if (rf->sc_dev.hw_attrs.uk_attrs.feature_flags &
2017                     IRDMA_FEATURE_CQ_RESIZE && !ucontext->legacy_mode) {
2018                         spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2019                         iwpbl_shadow = irdma_get_pbl(
2020                                         (unsigned long)req.user_shadow_area,
2021                                         &ucontext->cq_reg_mem_list);
2022                         spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2023
2024                         if (!iwpbl_shadow) {
2025                                 err_code = -EPROTO;
2026                                 goto cq_free_rsrc;
2027                         }
2028                         iwcq->iwpbl_shadow = iwpbl_shadow;
2029                         cqmr_shadow = &iwpbl_shadow->cq_mr;
2030                         info.shadow_area_pa = cqmr_shadow->cq_pbl.addr;
2031                         cqmr->split = true;
2032                 } else {
2033                         info.shadow_area_pa = cqmr->shadow;
2034                 }
2035                 if (iwpbl->pbl_allocated) {
2036                         info.virtual_map = true;
2037                         info.pbl_chunk_size = 1;
2038                         info.first_pm_pbl_idx = cqmr->cq_pbl.idx;
2039                 } else {
2040                         info.cq_base_pa = cqmr->cq_pbl.addr;
2041                 }
2042         } else {
2043                 /* Kmode allocations */
2044                 int rsize;
2045
2046                 if (entries > rf->max_cqe) {
2047                         err_code = -EINVAL;
2048                         goto cq_free_rsrc;
2049                 }
2050
2051                 entries++;
2052                 if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
2053                         entries *= 2;
2054                 ukinfo->cq_size = entries;
2055
2056                 rsize = info.cq_uk_init_info.cq_size * sizeof(struct irdma_cqe);
2057                 iwcq->kmem.size = ALIGN(round_up(rsize, 256), 256);
2058                 iwcq->kmem.va = dma_alloc_coherent(dev->hw->device,
2059                                                    iwcq->kmem.size,
2060                                                    &iwcq->kmem.pa, GFP_KERNEL);
2061                 if (!iwcq->kmem.va) {
2062                         err_code = -ENOMEM;
2063                         goto cq_free_rsrc;
2064                 }
2065
2066                 iwcq->kmem_shadow.size = ALIGN(IRDMA_SHADOW_AREA_SIZE << 3,
2067                                                64);
2068                 iwcq->kmem_shadow.va = dma_alloc_coherent(dev->hw->device,
2069                                                           iwcq->kmem_shadow.size,
2070                                                           &iwcq->kmem_shadow.pa,
2071                                                           GFP_KERNEL);
2072                 if (!iwcq->kmem_shadow.va) {
2073                         err_code = -ENOMEM;
2074                         goto cq_free_rsrc;
2075                 }
2076                 info.shadow_area_pa = iwcq->kmem_shadow.pa;
2077                 ukinfo->shadow_area = iwcq->kmem_shadow.va;
2078                 ukinfo->cq_base = iwcq->kmem.va;
2079                 info.cq_base_pa = iwcq->kmem.pa;
2080         }
2081
2082         if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
2083                 info.shadow_read_threshold = min(info.cq_uk_init_info.cq_size / 2,
2084                                                  (u32)IRDMA_MAX_CQ_READ_THRESH);
2085
2086         if (irdma_sc_cq_init(cq, &info)) {
2087                 ibdev_dbg(&iwdev->ibdev, "VERBS: init cq fail\n");
2088                 err_code = -EPROTO;
2089                 goto cq_free_rsrc;
2090         }
2091
2092         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
2093         if (!cqp_request) {
2094                 err_code = -ENOMEM;
2095                 goto cq_free_rsrc;
2096         }
2097
2098         cqp_info = &cqp_request->info;
2099         cqp_info->cqp_cmd = IRDMA_OP_CQ_CREATE;
2100         cqp_info->post_sq = 1;
2101         cqp_info->in.u.cq_create.cq = cq;
2102         cqp_info->in.u.cq_create.check_overflow = true;
2103         cqp_info->in.u.cq_create.scratch = (uintptr_t)cqp_request;
2104         status = irdma_handle_cqp_op(rf, cqp_request);
2105         irdma_put_cqp_request(&rf->cqp, cqp_request);
2106         if (status) {
2107                 err_code = -ENOMEM;
2108                 goto cq_free_rsrc;
2109         }
2110
2111         if (udata) {
2112                 struct irdma_create_cq_resp resp = {};
2113
2114                 resp.cq_id = info.cq_uk_init_info.cq_id;
2115                 resp.cq_size = info.cq_uk_init_info.cq_size;
2116                 if (ib_copy_to_udata(udata, &resp,
2117                                      min(sizeof(resp), udata->outlen))) {
2118                         ibdev_dbg(&iwdev->ibdev,
2119                                   "VERBS: copy to user data\n");
2120                         err_code = -EPROTO;
2121                         goto cq_destroy;
2122                 }
2123         }
2124         return 0;
2125 cq_destroy:
2126         irdma_cq_wq_destroy(rf, cq);
2127 cq_free_rsrc:
2128         irdma_cq_free_rsrc(rf, iwcq);
2129
2130         return err_code;
2131 }
2132
2133 /**
2134  * irdma_get_mr_access - get hw MR access permissions from IB access flags
2135  * @access: IB access flags
2136  */
2137 static inline u16 irdma_get_mr_access(int access)
2138 {
2139         u16 hw_access = 0;
2140
2141         hw_access |= (access & IB_ACCESS_LOCAL_WRITE) ?
2142                      IRDMA_ACCESS_FLAGS_LOCALWRITE : 0;
2143         hw_access |= (access & IB_ACCESS_REMOTE_WRITE) ?
2144                      IRDMA_ACCESS_FLAGS_REMOTEWRITE : 0;
2145         hw_access |= (access & IB_ACCESS_REMOTE_READ) ?
2146                      IRDMA_ACCESS_FLAGS_REMOTEREAD : 0;
2147         hw_access |= (access & IB_ACCESS_MW_BIND) ?
2148                      IRDMA_ACCESS_FLAGS_BIND_WINDOW : 0;
2149         hw_access |= (access & IB_ZERO_BASED) ?
2150                      IRDMA_ACCESS_FLAGS_ZERO_BASED : 0;
2151         hw_access |= IRDMA_ACCESS_FLAGS_LOCALREAD;
2152
2153         return hw_access;
2154 }
2155
2156 /**
2157  * irdma_free_stag - free stag resource
2158  * @iwdev: irdma device
2159  * @stag: stag to free
2160  */
2161 static void irdma_free_stag(struct irdma_device *iwdev, u32 stag)
2162 {
2163         u32 stag_idx;
2164
2165         stag_idx = (stag & iwdev->rf->mr_stagmask) >> IRDMA_CQPSQ_STAG_IDX_S;
2166         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_mrs, stag_idx);
2167 }
2168
2169 /**
2170  * irdma_create_stag - create random stag
2171  * @iwdev: irdma device
2172  */
2173 static u32 irdma_create_stag(struct irdma_device *iwdev)
2174 {
2175         u32 stag = 0;
2176         u32 stag_index = 0;
2177         u32 next_stag_index;
2178         u32 driver_key;
2179         u32 random;
2180         u8 consumer_key;
2181         int ret;
2182
2183         get_random_bytes(&random, sizeof(random));
2184         consumer_key = (u8)random;
2185
2186         driver_key = random & ~iwdev->rf->mr_stagmask;
2187         next_stag_index = (random & iwdev->rf->mr_stagmask) >> 8;
2188         next_stag_index %= iwdev->rf->max_mr;
2189
2190         ret = irdma_alloc_rsrc(iwdev->rf, iwdev->rf->allocated_mrs,
2191                                iwdev->rf->max_mr, &stag_index,
2192                                &next_stag_index);
2193         if (ret)
2194                 return stag;
2195         stag = stag_index << IRDMA_CQPSQ_STAG_IDX_S;
2196         stag |= driver_key;
2197         stag += (u32)consumer_key;
2198
2199         return stag;
2200 }
2201
2202 /**
2203  * irdma_next_pbl_addr - Get next pbl address
2204  * @pbl: pointer to a pble
2205  * @pinfo: info pointer
2206  * @idx: index
2207  */
2208 static inline u64 *irdma_next_pbl_addr(u64 *pbl, struct irdma_pble_info **pinfo,
2209                                        u32 *idx)
2210 {
2211         *idx += 1;
2212         if (!(*pinfo) || *idx != (*pinfo)->cnt)
2213                 return ++pbl;
2214         *idx = 0;
2215         (*pinfo)++;
2216
2217         return (*pinfo)->addr;
2218 }
2219
2220 /**
2221  * irdma_copy_user_pgaddrs - copy user page address to pble's os locally
2222  * @iwmr: iwmr for IB's user page addresses
2223  * @pbl: ple pointer to save 1 level or 0 level pble
2224  * @level: indicated level 0, 1 or 2
2225  */
2226 static void irdma_copy_user_pgaddrs(struct irdma_mr *iwmr, u64 *pbl,
2227                                     enum irdma_pble_level level)
2228 {
2229         struct ib_umem *region = iwmr->region;
2230         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2231         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2232         struct irdma_pble_info *pinfo;
2233         struct ib_block_iter biter;
2234         u32 idx = 0;
2235         u32 pbl_cnt = 0;
2236
2237         pinfo = (level == PBLE_LEVEL_1) ? NULL : palloc->level2.leaf;
2238
2239         if (iwmr->type == IRDMA_MEMREG_TYPE_QP)
2240                 iwpbl->qp_mr.sq_page = sg_page(region->sg_head.sgl);
2241
2242         rdma_umem_for_each_dma_block(region, &biter, iwmr->page_size) {
2243                 *pbl = rdma_block_iter_dma_address(&biter);
2244                 if (++pbl_cnt == palloc->total_cnt)
2245                         break;
2246                 pbl = irdma_next_pbl_addr(pbl, &pinfo, &idx);
2247         }
2248 }
2249
2250 /**
2251  * irdma_check_mem_contiguous - check if pbls stored in arr are contiguous
2252  * @arr: lvl1 pbl array
2253  * @npages: page count
2254  * @pg_size: page size
2255  *
2256  */
2257 static bool irdma_check_mem_contiguous(u64 *arr, u32 npages, u32 pg_size)
2258 {
2259         u32 pg_idx;
2260
2261         for (pg_idx = 0; pg_idx < npages; pg_idx++) {
2262                 if ((*arr + (pg_size * pg_idx)) != arr[pg_idx])
2263                         return false;
2264         }
2265
2266         return true;
2267 }
2268
2269 /**
2270  * irdma_check_mr_contiguous - check if MR is physically contiguous
2271  * @palloc: pbl allocation struct
2272  * @pg_size: page size
2273  */
2274 static bool irdma_check_mr_contiguous(struct irdma_pble_alloc *palloc,
2275                                       u32 pg_size)
2276 {
2277         struct irdma_pble_level2 *lvl2 = &palloc->level2;
2278         struct irdma_pble_info *leaf = lvl2->leaf;
2279         u64 *arr = NULL;
2280         u64 *start_addr = NULL;
2281         int i;
2282         bool ret;
2283
2284         if (palloc->level == PBLE_LEVEL_1) {
2285                 arr = palloc->level1.addr;
2286                 ret = irdma_check_mem_contiguous(arr, palloc->total_cnt,
2287                                                  pg_size);
2288                 return ret;
2289         }
2290
2291         start_addr = leaf->addr;
2292
2293         for (i = 0; i < lvl2->leaf_cnt; i++, leaf++) {
2294                 arr = leaf->addr;
2295                 if ((*start_addr + (i * pg_size * PBLE_PER_PAGE)) != *arr)
2296                         return false;
2297                 ret = irdma_check_mem_contiguous(arr, leaf->cnt, pg_size);
2298                 if (!ret)
2299                         return false;
2300         }
2301
2302         return true;
2303 }
2304
2305 /**
2306  * irdma_setup_pbles - copy user pg address to pble's
2307  * @rf: RDMA PCI function
2308  * @iwmr: mr pointer for this memory registration
2309  * @use_pbles: flag if to use pble's
2310  */
2311 static int irdma_setup_pbles(struct irdma_pci_f *rf, struct irdma_mr *iwmr,
2312                              bool use_pbles)
2313 {
2314         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2315         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2316         struct irdma_pble_info *pinfo;
2317         u64 *pbl;
2318         enum irdma_status_code status;
2319         enum irdma_pble_level level = PBLE_LEVEL_1;
2320
2321         if (use_pbles) {
2322                 status = irdma_get_pble(rf->pble_rsrc, palloc, iwmr->page_cnt,
2323                                         false);
2324                 if (status)
2325                         return -ENOMEM;
2326
2327                 iwpbl->pbl_allocated = true;
2328                 level = palloc->level;
2329                 pinfo = (level == PBLE_LEVEL_1) ? &palloc->level1 :
2330                                                   palloc->level2.leaf;
2331                 pbl = pinfo->addr;
2332         } else {
2333                 pbl = iwmr->pgaddrmem;
2334         }
2335
2336         irdma_copy_user_pgaddrs(iwmr, pbl, level);
2337
2338         if (use_pbles)
2339                 iwmr->pgaddrmem[0] = *pbl;
2340
2341         return 0;
2342 }
2343
2344 /**
2345  * irdma_handle_q_mem - handle memory for qp and cq
2346  * @iwdev: irdma device
2347  * @req: information for q memory management
2348  * @iwpbl: pble struct
2349  * @use_pbles: flag to use pble
2350  */
2351 static int irdma_handle_q_mem(struct irdma_device *iwdev,
2352                               struct irdma_mem_reg_req *req,
2353                               struct irdma_pbl *iwpbl, bool use_pbles)
2354 {
2355         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2356         struct irdma_mr *iwmr = iwpbl->iwmr;
2357         struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;
2358         struct irdma_cq_mr *cqmr = &iwpbl->cq_mr;
2359         struct irdma_hmc_pble *hmc_p;
2360         u64 *arr = iwmr->pgaddrmem;
2361         u32 pg_size, total;
2362         int err = 0;
2363         bool ret = true;
2364
2365         pg_size = iwmr->page_size;
2366         err = irdma_setup_pbles(iwdev->rf, iwmr, use_pbles);
2367         if (err)
2368                 return err;
2369
2370         if (use_pbles && palloc->level != PBLE_LEVEL_1) {
2371                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2372                 iwpbl->pbl_allocated = false;
2373                 return -ENOMEM;
2374         }
2375
2376         if (use_pbles)
2377                 arr = palloc->level1.addr;
2378
2379         switch (iwmr->type) {
2380         case IRDMA_MEMREG_TYPE_QP:
2381                 total = req->sq_pages + req->rq_pages;
2382                 hmc_p = &qpmr->sq_pbl;
2383                 qpmr->shadow = (dma_addr_t)arr[total];
2384
2385                 if (use_pbles) {
2386                         ret = irdma_check_mem_contiguous(arr, req->sq_pages,
2387                                                          pg_size);
2388                         if (ret)
2389                                 ret = irdma_check_mem_contiguous(&arr[req->sq_pages],
2390                                                                  req->rq_pages,
2391                                                                  pg_size);
2392                 }
2393
2394                 if (!ret) {
2395                         hmc_p->idx = palloc->level1.idx;
2396                         hmc_p = &qpmr->rq_pbl;
2397                         hmc_p->idx = palloc->level1.idx + req->sq_pages;
2398                 } else {
2399                         hmc_p->addr = arr[0];
2400                         hmc_p = &qpmr->rq_pbl;
2401                         hmc_p->addr = arr[req->sq_pages];
2402                 }
2403                 break;
2404         case IRDMA_MEMREG_TYPE_CQ:
2405                 hmc_p = &cqmr->cq_pbl;
2406
2407                 if (!cqmr->split)
2408                         cqmr->shadow = (dma_addr_t)arr[req->cq_pages];
2409
2410                 if (use_pbles)
2411                         ret = irdma_check_mem_contiguous(arr, req->cq_pages,
2412                                                          pg_size);
2413
2414                 if (!ret)
2415                         hmc_p->idx = palloc->level1.idx;
2416                 else
2417                         hmc_p->addr = arr[0];
2418         break;
2419         default:
2420                 ibdev_dbg(&iwdev->ibdev, "VERBS: MR type error\n");
2421                 err = -EINVAL;
2422         }
2423
2424         if (use_pbles && ret) {
2425                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2426                 iwpbl->pbl_allocated = false;
2427         }
2428
2429         return err;
2430 }
2431
2432 /**
2433  * irdma_hw_alloc_mw - create the hw memory window
2434  * @iwdev: irdma device
2435  * @iwmr: pointer to memory window info
2436  */
2437 static int irdma_hw_alloc_mw(struct irdma_device *iwdev, struct irdma_mr *iwmr)
2438 {
2439         struct irdma_mw_alloc_info *info;
2440         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2441         struct irdma_cqp_request *cqp_request;
2442         struct cqp_cmds_info *cqp_info;
2443         enum irdma_status_code status;
2444
2445         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2446         if (!cqp_request)
2447                 return -ENOMEM;
2448
2449         cqp_info = &cqp_request->info;
2450         info = &cqp_info->in.u.mw_alloc.info;
2451         memset(info, 0, sizeof(*info));
2452         if (iwmr->ibmw.type == IB_MW_TYPE_1)
2453                 info->mw_wide = true;
2454
2455         info->page_size = PAGE_SIZE;
2456         info->mw_stag_index = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2457         info->pd_id = iwpd->sc_pd.pd_id;
2458         info->remote_access = true;
2459         cqp_info->cqp_cmd = IRDMA_OP_MW_ALLOC;
2460         cqp_info->post_sq = 1;
2461         cqp_info->in.u.mw_alloc.dev = &iwdev->rf->sc_dev;
2462         cqp_info->in.u.mw_alloc.scratch = (uintptr_t)cqp_request;
2463         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2464         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2465
2466         return status ? -ENOMEM : 0;
2467 }
2468
2469 /**
2470  * irdma_alloc_mw - Allocate memory window
2471  * @ibmw: Memory Window
2472  * @udata: user data pointer
2473  */
2474 static int irdma_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata)
2475 {
2476         struct irdma_device *iwdev = to_iwdev(ibmw->device);
2477         struct irdma_mr *iwmr = to_iwmw(ibmw);
2478         int err_code;
2479         u32 stag;
2480
2481         stag = irdma_create_stag(iwdev);
2482         if (!stag)
2483                 return -ENOMEM;
2484
2485         iwmr->stag = stag;
2486         ibmw->rkey = stag;
2487
2488         err_code = irdma_hw_alloc_mw(iwdev, iwmr);
2489         if (err_code) {
2490                 irdma_free_stag(iwdev, stag);
2491                 return err_code;
2492         }
2493
2494         return 0;
2495 }
2496
2497 /**
2498  * irdma_dealloc_mw - Dealloc memory window
2499  * @ibmw: memory window structure.
2500  */
2501 static int irdma_dealloc_mw(struct ib_mw *ibmw)
2502 {
2503         struct ib_pd *ibpd = ibmw->pd;
2504         struct irdma_pd *iwpd = to_iwpd(ibpd);
2505         struct irdma_mr *iwmr = to_iwmr((struct ib_mr *)ibmw);
2506         struct irdma_device *iwdev = to_iwdev(ibmw->device);
2507         struct irdma_cqp_request *cqp_request;
2508         struct cqp_cmds_info *cqp_info;
2509         struct irdma_dealloc_stag_info *info;
2510
2511         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2512         if (!cqp_request)
2513                 return -ENOMEM;
2514
2515         cqp_info = &cqp_request->info;
2516         info = &cqp_info->in.u.dealloc_stag.info;
2517         memset(info, 0, sizeof(*info));
2518         info->pd_id = iwpd->sc_pd.pd_id & 0x00007fff;
2519         info->stag_idx = ibmw->rkey >> IRDMA_CQPSQ_STAG_IDX_S;
2520         info->mr = false;
2521         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
2522         cqp_info->post_sq = 1;
2523         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
2524         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
2525         irdma_handle_cqp_op(iwdev->rf, cqp_request);
2526         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2527         irdma_free_stag(iwdev, iwmr->stag);
2528
2529         return 0;
2530 }
2531
2532 /**
2533  * irdma_hw_alloc_stag - cqp command to allocate stag
2534  * @iwdev: irdma device
2535  * @iwmr: irdma mr pointer
2536  */
2537 static int irdma_hw_alloc_stag(struct irdma_device *iwdev,
2538                                struct irdma_mr *iwmr)
2539 {
2540         struct irdma_allocate_stag_info *info;
2541         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2542         enum irdma_status_code status;
2543         int err = 0;
2544         struct irdma_cqp_request *cqp_request;
2545         struct cqp_cmds_info *cqp_info;
2546
2547         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2548         if (!cqp_request)
2549                 return -ENOMEM;
2550
2551         cqp_info = &cqp_request->info;
2552         info = &cqp_info->in.u.alloc_stag.info;
2553         memset(info, 0, sizeof(*info));
2554         info->page_size = PAGE_SIZE;
2555         info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2556         info->pd_id = iwpd->sc_pd.pd_id;
2557         info->total_len = iwmr->len;
2558         info->remote_access = true;
2559         cqp_info->cqp_cmd = IRDMA_OP_ALLOC_STAG;
2560         cqp_info->post_sq = 1;
2561         cqp_info->in.u.alloc_stag.dev = &iwdev->rf->sc_dev;
2562         cqp_info->in.u.alloc_stag.scratch = (uintptr_t)cqp_request;
2563         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2564         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2565         if (status)
2566                 err = -ENOMEM;
2567
2568         return err;
2569 }
2570
2571 /**
2572  * irdma_alloc_mr - register stag for fast memory registration
2573  * @pd: ibpd pointer
2574  * @mr_type: memory for stag registrion
2575  * @max_num_sg: man number of pages
2576  */
2577 static struct ib_mr *irdma_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
2578                                     u32 max_num_sg)
2579 {
2580         struct irdma_device *iwdev = to_iwdev(pd->device);
2581         struct irdma_pble_alloc *palloc;
2582         struct irdma_pbl *iwpbl;
2583         struct irdma_mr *iwmr;
2584         enum irdma_status_code status;
2585         u32 stag;
2586         int err_code = -ENOMEM;
2587
2588         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2589         if (!iwmr)
2590                 return ERR_PTR(-ENOMEM);
2591
2592         stag = irdma_create_stag(iwdev);
2593         if (!stag) {
2594                 err_code = -ENOMEM;
2595                 goto err;
2596         }
2597
2598         iwmr->stag = stag;
2599         iwmr->ibmr.rkey = stag;
2600         iwmr->ibmr.lkey = stag;
2601         iwmr->ibmr.pd = pd;
2602         iwmr->ibmr.device = pd->device;
2603         iwpbl = &iwmr->iwpbl;
2604         iwpbl->iwmr = iwmr;
2605         iwmr->type = IRDMA_MEMREG_TYPE_MEM;
2606         palloc = &iwpbl->pble_alloc;
2607         iwmr->page_cnt = max_num_sg;
2608         status = irdma_get_pble(iwdev->rf->pble_rsrc, palloc, iwmr->page_cnt,
2609                                 true);
2610         if (status)
2611                 goto err_get_pble;
2612
2613         err_code = irdma_hw_alloc_stag(iwdev, iwmr);
2614         if (err_code)
2615                 goto err_alloc_stag;
2616
2617         iwpbl->pbl_allocated = true;
2618
2619         return &iwmr->ibmr;
2620 err_alloc_stag:
2621         irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2622 err_get_pble:
2623         irdma_free_stag(iwdev, stag);
2624 err:
2625         kfree(iwmr);
2626
2627         return ERR_PTR(err_code);
2628 }
2629
2630 /**
2631  * irdma_set_page - populate pbl list for fmr
2632  * @ibmr: ib mem to access iwarp mr pointer
2633  * @addr: page dma address fro pbl list
2634  */
2635 static int irdma_set_page(struct ib_mr *ibmr, u64 addr)
2636 {
2637         struct irdma_mr *iwmr = to_iwmr(ibmr);
2638         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2639         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2640         u64 *pbl;
2641
2642         if (unlikely(iwmr->npages == iwmr->page_cnt))
2643                 return -ENOMEM;
2644
2645         pbl = palloc->level1.addr;
2646         pbl[iwmr->npages++] = addr;
2647
2648         return 0;
2649 }
2650
2651 /**
2652  * irdma_map_mr_sg - map of sg list for fmr
2653  * @ibmr: ib mem to access iwarp mr pointer
2654  * @sg: scatter gather list
2655  * @sg_nents: number of sg pages
2656  * @sg_offset: scatter gather list for fmr
2657  */
2658 static int irdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
2659                            int sg_nents, unsigned int *sg_offset)
2660 {
2661         struct irdma_mr *iwmr = to_iwmr(ibmr);
2662
2663         iwmr->npages = 0;
2664
2665         return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, irdma_set_page);
2666 }
2667
2668 /**
2669  * irdma_hwreg_mr - send cqp command for memory registration
2670  * @iwdev: irdma device
2671  * @iwmr: irdma mr pointer
2672  * @access: access for MR
2673  */
2674 static int irdma_hwreg_mr(struct irdma_device *iwdev, struct irdma_mr *iwmr,
2675                           u16 access)
2676 {
2677         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2678         struct irdma_reg_ns_stag_info *stag_info;
2679         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2680         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2681         enum irdma_status_code status;
2682         int err = 0;
2683         struct irdma_cqp_request *cqp_request;
2684         struct cqp_cmds_info *cqp_info;
2685
2686         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2687         if (!cqp_request)
2688                 return -ENOMEM;
2689
2690         cqp_info = &cqp_request->info;
2691         stag_info = &cqp_info->in.u.mr_reg_non_shared.info;
2692         memset(stag_info, 0, sizeof(*stag_info));
2693         stag_info->va = iwpbl->user_base;
2694         stag_info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2695         stag_info->stag_key = (u8)iwmr->stag;
2696         stag_info->total_len = iwmr->len;
2697         stag_info->access_rights = irdma_get_mr_access(access);
2698         stag_info->pd_id = iwpd->sc_pd.pd_id;
2699         if (stag_info->access_rights & IRDMA_ACCESS_FLAGS_ZERO_BASED)
2700                 stag_info->addr_type = IRDMA_ADDR_TYPE_ZERO_BASED;
2701         else
2702                 stag_info->addr_type = IRDMA_ADDR_TYPE_VA_BASED;
2703         stag_info->page_size = iwmr->page_size;
2704
2705         if (iwpbl->pbl_allocated) {
2706                 if (palloc->level == PBLE_LEVEL_1) {
2707                         stag_info->first_pm_pbl_index = palloc->level1.idx;
2708                         stag_info->chunk_size = 1;
2709                 } else {
2710                         stag_info->first_pm_pbl_index = palloc->level2.root.idx;
2711                         stag_info->chunk_size = 3;
2712                 }
2713         } else {
2714                 stag_info->reg_addr_pa = iwmr->pgaddrmem[0];
2715         }
2716
2717         cqp_info->cqp_cmd = IRDMA_OP_MR_REG_NON_SHARED;
2718         cqp_info->post_sq = 1;
2719         cqp_info->in.u.mr_reg_non_shared.dev = &iwdev->rf->sc_dev;
2720         cqp_info->in.u.mr_reg_non_shared.scratch = (uintptr_t)cqp_request;
2721         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2722         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2723         if (status)
2724                 err = -ENOMEM;
2725
2726         return err;
2727 }
2728
2729 /**
2730  * irdma_reg_user_mr - Register a user memory region
2731  * @pd: ptr of pd
2732  * @start: virtual start address
2733  * @len: length of mr
2734  * @virt: virtual address
2735  * @access: access of mr
2736  * @udata: user data
2737  */
2738 static struct ib_mr *irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,
2739                                        u64 virt, int access,
2740                                        struct ib_udata *udata)
2741 {
2742         struct irdma_device *iwdev = to_iwdev(pd->device);
2743         struct irdma_ucontext *ucontext;
2744         struct irdma_pble_alloc *palloc;
2745         struct irdma_pbl *iwpbl;
2746         struct irdma_mr *iwmr;
2747         struct ib_umem *region;
2748         struct irdma_mem_reg_req req;
2749         u32 total, stag = 0;
2750         u8 shadow_pgcnt = 1;
2751         bool use_pbles = false;
2752         unsigned long flags;
2753         int err = -EINVAL;
2754         int ret;
2755
2756         if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size)
2757                 return ERR_PTR(-EINVAL);
2758
2759         region = ib_umem_get(pd->device, start, len, access);
2760
2761         if (IS_ERR(region)) {
2762                 ibdev_dbg(&iwdev->ibdev,
2763                           "VERBS: Failed to create ib_umem region\n");
2764                 return (struct ib_mr *)region;
2765         }
2766
2767         if (ib_copy_from_udata(&req, udata, min(sizeof(req), udata->inlen))) {
2768                 ib_umem_release(region);
2769                 return ERR_PTR(-EFAULT);
2770         }
2771
2772         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2773         if (!iwmr) {
2774                 ib_umem_release(region);
2775                 return ERR_PTR(-ENOMEM);
2776         }
2777
2778         iwpbl = &iwmr->iwpbl;
2779         iwpbl->iwmr = iwmr;
2780         iwmr->region = region;
2781         iwmr->ibmr.pd = pd;
2782         iwmr->ibmr.device = pd->device;
2783         iwmr->ibmr.iova = virt;
2784         iwmr->page_size = PAGE_SIZE;
2785
2786         if (req.reg_type == IRDMA_MEMREG_TYPE_MEM) {
2787                 iwmr->page_size = ib_umem_find_best_pgsz(region,
2788                                                          SZ_4K | SZ_2M | SZ_1G,
2789                                                          virt);
2790                 if (unlikely(!iwmr->page_size)) {
2791                         kfree(iwmr);
2792                         ib_umem_release(region);
2793                         return ERR_PTR(-EOPNOTSUPP);
2794                 }
2795         }
2796         iwmr->len = region->length;
2797         iwpbl->user_base = virt;
2798         palloc = &iwpbl->pble_alloc;
2799         iwmr->type = req.reg_type;
2800         iwmr->page_cnt = ib_umem_num_dma_blocks(region, iwmr->page_size);
2801
2802         switch (req.reg_type) {
2803         case IRDMA_MEMREG_TYPE_QP:
2804                 total = req.sq_pages + req.rq_pages + shadow_pgcnt;
2805                 if (total > iwmr->page_cnt) {
2806                         err = -EINVAL;
2807                         goto error;
2808                 }
2809                 total = req.sq_pages + req.rq_pages;
2810                 use_pbles = (total > 2);
2811                 err = irdma_handle_q_mem(iwdev, &req, iwpbl, use_pbles);
2812                 if (err)
2813                         goto error;
2814
2815                 ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext,
2816                                                      ibucontext);
2817                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
2818                 list_add_tail(&iwpbl->list, &ucontext->qp_reg_mem_list);
2819                 iwpbl->on_list = true;
2820                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
2821                 break;
2822         case IRDMA_MEMREG_TYPE_CQ:
2823                 if (iwdev->rf->sc_dev.hw_attrs.uk_attrs.feature_flags & IRDMA_FEATURE_CQ_RESIZE)
2824                         shadow_pgcnt = 0;
2825                 total = req.cq_pages + shadow_pgcnt;
2826                 if (total > iwmr->page_cnt) {
2827                         err = -EINVAL;
2828                         goto error;
2829                 }
2830
2831                 use_pbles = (req.cq_pages > 1);
2832                 err = irdma_handle_q_mem(iwdev, &req, iwpbl, use_pbles);
2833                 if (err)
2834                         goto error;
2835
2836                 ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext,
2837                                                      ibucontext);
2838                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2839                 list_add_tail(&iwpbl->list, &ucontext->cq_reg_mem_list);
2840                 iwpbl->on_list = true;
2841                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2842                 break;
2843         case IRDMA_MEMREG_TYPE_MEM:
2844                 use_pbles = (iwmr->page_cnt != 1);
2845
2846                 err = irdma_setup_pbles(iwdev->rf, iwmr, use_pbles);
2847                 if (err)
2848                         goto error;
2849
2850                 if (use_pbles) {
2851                         ret = irdma_check_mr_contiguous(palloc,
2852                                                         iwmr->page_size);
2853                         if (ret) {
2854                                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2855                                 iwpbl->pbl_allocated = false;
2856                         }
2857                 }
2858
2859                 stag = irdma_create_stag(iwdev);
2860                 if (!stag) {
2861                         err = -ENOMEM;
2862                         goto error;
2863                 }
2864
2865                 iwmr->stag = stag;
2866                 iwmr->ibmr.rkey = stag;
2867                 iwmr->ibmr.lkey = stag;
2868                 err = irdma_hwreg_mr(iwdev, iwmr, access);
2869                 if (err) {
2870                         irdma_free_stag(iwdev, stag);
2871                         goto error;
2872                 }
2873
2874                 break;
2875         default:
2876                 goto error;
2877         }
2878
2879         iwmr->type = req.reg_type;
2880
2881         return &iwmr->ibmr;
2882
2883 error:
2884         if (palloc->level != PBLE_LEVEL_0 && iwpbl->pbl_allocated)
2885                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2886         ib_umem_release(region);
2887         kfree(iwmr);
2888
2889         return ERR_PTR(err);
2890 }
2891
2892 /**
2893  * irdma_reg_phys_mr - register kernel physical memory
2894  * @pd: ibpd pointer
2895  * @addr: physical address of memory to register
2896  * @size: size of memory to register
2897  * @access: Access rights
2898  * @iova_start: start of virtual address for physical buffers
2899  */
2900 struct ib_mr *irdma_reg_phys_mr(struct ib_pd *pd, u64 addr, u64 size, int access,
2901                                 u64 *iova_start)
2902 {
2903         struct irdma_device *iwdev = to_iwdev(pd->device);
2904         struct irdma_pbl *iwpbl;
2905         struct irdma_mr *iwmr;
2906         enum irdma_status_code status;
2907         u32 stag;
2908         int ret;
2909
2910         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2911         if (!iwmr)
2912                 return ERR_PTR(-ENOMEM);
2913
2914         iwmr->ibmr.pd = pd;
2915         iwmr->ibmr.device = pd->device;
2916         iwpbl = &iwmr->iwpbl;
2917         iwpbl->iwmr = iwmr;
2918         iwmr->type = IRDMA_MEMREG_TYPE_MEM;
2919         iwpbl->user_base = *iova_start;
2920         stag = irdma_create_stag(iwdev);
2921         if (!stag) {
2922                 ret = -ENOMEM;
2923                 goto err;
2924         }
2925
2926         iwmr->stag = stag;
2927         iwmr->ibmr.iova = *iova_start;
2928         iwmr->ibmr.rkey = stag;
2929         iwmr->ibmr.lkey = stag;
2930         iwmr->page_cnt = 1;
2931         iwmr->pgaddrmem[0] = addr;
2932         iwmr->len = size;
2933         iwmr->page_size = SZ_4K;
2934         status = irdma_hwreg_mr(iwdev, iwmr, access);
2935         if (status) {
2936                 irdma_free_stag(iwdev, stag);
2937                 ret = -ENOMEM;
2938                 goto err;
2939         }
2940
2941         return &iwmr->ibmr;
2942
2943 err:
2944         kfree(iwmr);
2945
2946         return ERR_PTR(ret);
2947 }
2948
2949 /**
2950  * irdma_get_dma_mr - register physical mem
2951  * @pd: ptr of pd
2952  * @acc: access for memory
2953  */
2954 static struct ib_mr *irdma_get_dma_mr(struct ib_pd *pd, int acc)
2955 {
2956         u64 kva = 0;
2957
2958         return irdma_reg_phys_mr(pd, 0, 0, acc, &kva);
2959 }
2960
2961 /**
2962  * irdma_del_memlist - Deleting pbl list entries for CQ/QP
2963  * @iwmr: iwmr for IB's user page addresses
2964  * @ucontext: ptr to user context
2965  */
2966 static void irdma_del_memlist(struct irdma_mr *iwmr,
2967                               struct irdma_ucontext *ucontext)
2968 {
2969         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2970         unsigned long flags;
2971
2972         switch (iwmr->type) {
2973         case IRDMA_MEMREG_TYPE_CQ:
2974                 spin_lock_irqsave(&ucontext->cq_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->cq_reg_mem_list_lock, flags);
2980                 break;
2981         case IRDMA_MEMREG_TYPE_QP:
2982                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
2983                 if (iwpbl->on_list) {
2984                         iwpbl->on_list = false;
2985                         list_del(&iwpbl->list);
2986                 }
2987                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
2988                 break;
2989         default:
2990                 break;
2991         }
2992 }
2993
2994 /**
2995  * irdma_dereg_mr - deregister mr
2996  * @ib_mr: mr ptr for dereg
2997  * @udata: user data
2998  */
2999 static int irdma_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
3000 {
3001         struct ib_pd *ibpd = ib_mr->pd;
3002         struct irdma_pd *iwpd = to_iwpd(ibpd);
3003         struct irdma_mr *iwmr = to_iwmr(ib_mr);
3004         struct irdma_device *iwdev = to_iwdev(ib_mr->device);
3005         struct irdma_dealloc_stag_info *info;
3006         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
3007         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
3008         struct irdma_cqp_request *cqp_request;
3009         struct cqp_cmds_info *cqp_info;
3010
3011         if (iwmr->type != IRDMA_MEMREG_TYPE_MEM) {
3012                 if (iwmr->region) {
3013                         struct irdma_ucontext *ucontext;
3014
3015                         ucontext = rdma_udata_to_drv_context(udata,
3016                                                 struct irdma_ucontext,
3017                                                 ibucontext);
3018                         irdma_del_memlist(iwmr, ucontext);
3019                 }
3020                 goto done;
3021         }
3022
3023         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
3024         if (!cqp_request)
3025                 return -ENOMEM;
3026
3027         cqp_info = &cqp_request->info;
3028         info = &cqp_info->in.u.dealloc_stag.info;
3029         memset(info, 0, sizeof(*info));
3030         info->pd_id = iwpd->sc_pd.pd_id & 0x00007fff;
3031         info->stag_idx = ib_mr->rkey >> IRDMA_CQPSQ_STAG_IDX_S;
3032         info->mr = true;
3033         if (iwpbl->pbl_allocated)
3034                 info->dealloc_pbl = true;
3035
3036         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
3037         cqp_info->post_sq = 1;
3038         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
3039         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
3040         irdma_handle_cqp_op(iwdev->rf, cqp_request);
3041         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
3042         irdma_free_stag(iwdev, iwmr->stag);
3043 done:
3044         if (iwpbl->pbl_allocated)
3045                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
3046         ib_umem_release(iwmr->region);
3047         kfree(iwmr);
3048
3049         return 0;
3050 }
3051
3052 /**
3053  * irdma_copy_sg_list - copy sg list for qp
3054  * @sg_list: copied into sg_list
3055  * @sgl: copy from sgl
3056  * @num_sges: count of sg entries
3057  */
3058 static void irdma_copy_sg_list(struct irdma_sge *sg_list, struct ib_sge *sgl,
3059                                int num_sges)
3060 {
3061         unsigned int i;
3062
3063         for (i = 0; (i < num_sges) && (i < IRDMA_MAX_WQ_FRAGMENT_COUNT); i++) {
3064                 sg_list[i].tag_off = sgl[i].addr;
3065                 sg_list[i].len = sgl[i].length;
3066                 sg_list[i].stag = sgl[i].lkey;
3067         }
3068 }
3069
3070 /**
3071  * irdma_post_send -  kernel application wr
3072  * @ibqp: qp ptr for wr
3073  * @ib_wr: work request ptr
3074  * @bad_wr: return of bad wr if err
3075  */
3076 static int irdma_post_send(struct ib_qp *ibqp,
3077                            const struct ib_send_wr *ib_wr,
3078                            const struct ib_send_wr **bad_wr)
3079 {
3080         struct irdma_qp *iwqp;
3081         struct irdma_qp_uk *ukqp;
3082         struct irdma_sc_dev *dev;
3083         struct irdma_post_sq_info info;
3084         enum irdma_status_code ret;
3085         int err = 0;
3086         unsigned long flags;
3087         bool inv_stag;
3088         struct irdma_ah *ah;
3089         bool reflush = false;
3090
3091         iwqp = to_iwqp(ibqp);
3092         ukqp = &iwqp->sc_qp.qp_uk;
3093         dev = &iwqp->iwdev->rf->sc_dev;
3094
3095         spin_lock_irqsave(&iwqp->lock, flags);
3096         if (iwqp->flush_issued && ukqp->sq_flush_complete)
3097                 reflush = true;
3098         while (ib_wr) {
3099                 memset(&info, 0, sizeof(info));
3100                 inv_stag = false;
3101                 info.wr_id = (ib_wr->wr_id);
3102                 if ((ib_wr->send_flags & IB_SEND_SIGNALED) || iwqp->sig_all)
3103                         info.signaled = true;
3104                 if (ib_wr->send_flags & IB_SEND_FENCE)
3105                         info.read_fence = true;
3106                 switch (ib_wr->opcode) {
3107                 case IB_WR_SEND_WITH_IMM:
3108                         if (ukqp->qp_caps & IRDMA_SEND_WITH_IMM) {
3109                                 info.imm_data_valid = true;
3110                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
3111                         } else {
3112                                 err = -EINVAL;
3113                                 break;
3114                         }
3115                         fallthrough;
3116                 case IB_WR_SEND:
3117                 case IB_WR_SEND_WITH_INV:
3118                         if (ib_wr->opcode == IB_WR_SEND ||
3119                             ib_wr->opcode == IB_WR_SEND_WITH_IMM) {
3120                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
3121                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL;
3122                                 else
3123                                         info.op_type = IRDMA_OP_TYPE_SEND;
3124                         } else {
3125                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
3126                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL_INV;
3127                                 else
3128                                         info.op_type = IRDMA_OP_TYPE_SEND_INV;
3129                                 info.stag_to_inv = ib_wr->ex.invalidate_rkey;
3130                         }
3131
3132                         if (ib_wr->send_flags & IB_SEND_INLINE) {
3133                                 info.op.inline_send.data = (void *)(unsigned long)
3134                                                            ib_wr->sg_list[0].addr;
3135                                 info.op.inline_send.len = ib_wr->sg_list[0].length;
3136                                 if (iwqp->ibqp.qp_type == IB_QPT_UD ||
3137                                     iwqp->ibqp.qp_type == IB_QPT_GSI) {
3138                                         ah = to_iwah(ud_wr(ib_wr)->ah);
3139                                         info.op.inline_send.ah_id = ah->sc_ah.ah_info.ah_idx;
3140                                         info.op.inline_send.qkey = ud_wr(ib_wr)->remote_qkey;
3141                                         info.op.inline_send.dest_qp = ud_wr(ib_wr)->remote_qpn;
3142                                 }
3143                                 ret = irdma_uk_inline_send(ukqp, &info, false);
3144                         } else {
3145                                 info.op.send.num_sges = ib_wr->num_sge;
3146                                 info.op.send.sg_list = (struct irdma_sge *)
3147                                                        ib_wr->sg_list;
3148                                 if (iwqp->ibqp.qp_type == IB_QPT_UD ||
3149                                     iwqp->ibqp.qp_type == IB_QPT_GSI) {
3150                                         ah = to_iwah(ud_wr(ib_wr)->ah);
3151                                         info.op.send.ah_id = ah->sc_ah.ah_info.ah_idx;
3152                                         info.op.send.qkey = ud_wr(ib_wr)->remote_qkey;
3153                                         info.op.send.dest_qp = ud_wr(ib_wr)->remote_qpn;
3154                                 }
3155                                 ret = irdma_uk_send(ukqp, &info, false);
3156                         }
3157
3158                         if (ret) {
3159                                 if (ret == IRDMA_ERR_QP_TOOMANY_WRS_POSTED)
3160                                         err = -ENOMEM;
3161                                 else
3162                                         err = -EINVAL;
3163                         }
3164                         break;
3165                 case IB_WR_RDMA_WRITE_WITH_IMM:
3166                         if (ukqp->qp_caps & IRDMA_WRITE_WITH_IMM) {
3167                                 info.imm_data_valid = true;
3168                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
3169                         } else {
3170                                 err = -EINVAL;
3171                                 break;
3172                         }
3173                         fallthrough;
3174                 case IB_WR_RDMA_WRITE:
3175                         if (ib_wr->send_flags & IB_SEND_SOLICITED)
3176                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE_SOL;
3177                         else
3178                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE;
3179
3180                         if (ib_wr->send_flags & IB_SEND_INLINE) {
3181                                 info.op.inline_rdma_write.data = (void *)(uintptr_t)ib_wr->sg_list[0].addr;
3182                                 info.op.inline_rdma_write.len = ib_wr->sg_list[0].length;
3183                                 info.op.inline_rdma_write.rem_addr.tag_off = rdma_wr(ib_wr)->remote_addr;
3184                                 info.op.inline_rdma_write.rem_addr.stag = rdma_wr(ib_wr)->rkey;
3185                                 ret = irdma_uk_inline_rdma_write(ukqp, &info, false);
3186                         } else {
3187                                 info.op.rdma_write.lo_sg_list = (void *)ib_wr->sg_list;
3188                                 info.op.rdma_write.num_lo_sges = ib_wr->num_sge;
3189                                 info.op.rdma_write.rem_addr.tag_off = rdma_wr(ib_wr)->remote_addr;
3190                                 info.op.rdma_write.rem_addr.stag = rdma_wr(ib_wr)->rkey;
3191                                 ret = irdma_uk_rdma_write(ukqp, &info, false);
3192                         }
3193
3194                         if (ret) {
3195                                 if (ret == IRDMA_ERR_QP_TOOMANY_WRS_POSTED)
3196                                         err = -ENOMEM;
3197                                 else
3198                                         err = -EINVAL;
3199                         }
3200                         break;
3201                 case IB_WR_RDMA_READ_WITH_INV:
3202                         inv_stag = true;
3203                         fallthrough;
3204                 case IB_WR_RDMA_READ:
3205                         if (ib_wr->num_sge >
3206                             dev->hw_attrs.uk_attrs.max_hw_read_sges) {
3207                                 err = -EINVAL;
3208                                 break;
3209                         }
3210                         info.op_type = IRDMA_OP_TYPE_RDMA_READ;
3211                         info.op.rdma_read.rem_addr.tag_off = rdma_wr(ib_wr)->remote_addr;
3212                         info.op.rdma_read.rem_addr.stag = rdma_wr(ib_wr)->rkey;
3213                         info.op.rdma_read.lo_sg_list = (void *)ib_wr->sg_list;
3214                         info.op.rdma_read.num_lo_sges = ib_wr->num_sge;
3215
3216                         ret = irdma_uk_rdma_read(ukqp, &info, inv_stag, false);
3217                         if (ret) {
3218                                 if (ret == IRDMA_ERR_QP_TOOMANY_WRS_POSTED)
3219                                         err = -ENOMEM;
3220                                 else
3221                                         err = -EINVAL;
3222                         }
3223                         break;
3224                 case IB_WR_LOCAL_INV:
3225                         info.op_type = IRDMA_OP_TYPE_INV_STAG;
3226                         info.op.inv_local_stag.target_stag = ib_wr->ex.invalidate_rkey;
3227                         ret = irdma_uk_stag_local_invalidate(ukqp, &info, true);
3228                         if (ret)
3229                                 err = -ENOMEM;
3230                         break;
3231                 case IB_WR_REG_MR: {
3232                         struct irdma_mr *iwmr = to_iwmr(reg_wr(ib_wr)->mr);
3233                         struct irdma_pble_alloc *palloc = &iwmr->iwpbl.pble_alloc;
3234                         struct irdma_fast_reg_stag_info stag_info = {};
3235
3236                         stag_info.signaled = info.signaled;
3237                         stag_info.read_fence = info.read_fence;
3238                         stag_info.access_rights = irdma_get_mr_access(reg_wr(ib_wr)->access);
3239                         stag_info.stag_key = reg_wr(ib_wr)->key & 0xff;
3240                         stag_info.stag_idx = reg_wr(ib_wr)->key >> 8;
3241                         stag_info.page_size = reg_wr(ib_wr)->mr->page_size;
3242                         stag_info.wr_id = ib_wr->wr_id;
3243                         stag_info.addr_type = IRDMA_ADDR_TYPE_VA_BASED;
3244                         stag_info.va = (void *)(uintptr_t)iwmr->ibmr.iova;
3245                         stag_info.total_len = iwmr->ibmr.length;
3246                         stag_info.reg_addr_pa = *palloc->level1.addr;
3247                         stag_info.first_pm_pbl_index = palloc->level1.idx;
3248                         stag_info.local_fence = ib_wr->send_flags & IB_SEND_FENCE;
3249                         if (iwmr->npages > IRDMA_MIN_PAGES_PER_FMR)
3250                                 stag_info.chunk_size = 1;
3251                         ret = irdma_sc_mr_fast_register(&iwqp->sc_qp, &stag_info,
3252                                                         true);
3253                         if (ret)
3254                                 err = -ENOMEM;
3255                         break;
3256                 }
3257                 default:
3258                         err = -EINVAL;
3259                         ibdev_dbg(&iwqp->iwdev->ibdev,
3260                                   "VERBS: upost_send bad opcode = 0x%x\n",
3261                                   ib_wr->opcode);
3262                         break;
3263                 }
3264
3265                 if (err)
3266                         break;
3267                 ib_wr = ib_wr->next;
3268         }
3269
3270         if (!iwqp->flush_issued && iwqp->hw_iwarp_state <= IRDMA_QP_STATE_RTS) {
3271                 irdma_uk_qp_post_wr(ukqp);
3272                 spin_unlock_irqrestore(&iwqp->lock, flags);
3273         } else if (reflush) {
3274                 ukqp->sq_flush_complete = false;
3275                 spin_unlock_irqrestore(&iwqp->lock, flags);
3276                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_SQ | IRDMA_REFLUSH);
3277         } else {
3278                 spin_unlock_irqrestore(&iwqp->lock, flags);
3279         }
3280         if (err)
3281                 *bad_wr = ib_wr;
3282
3283         return err;
3284 }
3285
3286 /**
3287  * irdma_post_recv - post receive wr for kernel application
3288  * @ibqp: ib qp pointer
3289  * @ib_wr: work request for receive
3290  * @bad_wr: bad wr caused an error
3291  */
3292 static int irdma_post_recv(struct ib_qp *ibqp,
3293                            const struct ib_recv_wr *ib_wr,
3294                            const struct ib_recv_wr **bad_wr)
3295 {
3296         struct irdma_qp *iwqp;
3297         struct irdma_qp_uk *ukqp;
3298         struct irdma_post_rq_info post_recv = {};
3299         struct irdma_sge sg_list[IRDMA_MAX_WQ_FRAGMENT_COUNT];
3300         enum irdma_status_code ret = 0;
3301         unsigned long flags;
3302         int err = 0;
3303         bool reflush = false;
3304
3305         iwqp = to_iwqp(ibqp);
3306         ukqp = &iwqp->sc_qp.qp_uk;
3307
3308         spin_lock_irqsave(&iwqp->lock, flags);
3309         if (iwqp->flush_issued && ukqp->rq_flush_complete)
3310                 reflush = true;
3311         while (ib_wr) {
3312                 post_recv.num_sges = ib_wr->num_sge;
3313                 post_recv.wr_id = ib_wr->wr_id;
3314                 irdma_copy_sg_list(sg_list, ib_wr->sg_list, ib_wr->num_sge);
3315                 post_recv.sg_list = sg_list;
3316                 ret = irdma_uk_post_receive(ukqp, &post_recv);
3317                 if (ret) {
3318                         ibdev_dbg(&iwqp->iwdev->ibdev,
3319                                   "VERBS: post_recv err %d\n", ret);
3320                         if (ret == IRDMA_ERR_QP_TOOMANY_WRS_POSTED)
3321                                 err = -ENOMEM;
3322                         else
3323                                 err = -EINVAL;
3324                         goto out;
3325                 }
3326
3327                 ib_wr = ib_wr->next;
3328         }
3329
3330 out:
3331         if (reflush) {
3332                 ukqp->rq_flush_complete = false;
3333                 spin_unlock_irqrestore(&iwqp->lock, flags);
3334                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_RQ | IRDMA_REFLUSH);
3335         } else {
3336                 spin_unlock_irqrestore(&iwqp->lock, flags);
3337         }
3338
3339         if (err)
3340                 *bad_wr = ib_wr;
3341
3342         return err;
3343 }
3344
3345 /**
3346  * irdma_flush_err_to_ib_wc_status - return change flush error code to IB status
3347  * @opcode: iwarp flush code
3348  */
3349 static enum ib_wc_status irdma_flush_err_to_ib_wc_status(enum irdma_flush_opcode opcode)
3350 {
3351         switch (opcode) {
3352         case FLUSH_PROT_ERR:
3353                 return IB_WC_LOC_PROT_ERR;
3354         case FLUSH_REM_ACCESS_ERR:
3355                 return IB_WC_REM_ACCESS_ERR;
3356         case FLUSH_LOC_QP_OP_ERR:
3357                 return IB_WC_LOC_QP_OP_ERR;
3358         case FLUSH_REM_OP_ERR:
3359                 return IB_WC_REM_OP_ERR;
3360         case FLUSH_LOC_LEN_ERR:
3361                 return IB_WC_LOC_LEN_ERR;
3362         case FLUSH_GENERAL_ERR:
3363                 return IB_WC_WR_FLUSH_ERR;
3364         case FLUSH_FATAL_ERR:
3365         default:
3366                 return IB_WC_FATAL_ERR;
3367         }
3368 }
3369
3370 /**
3371  * irdma_process_cqe - process cqe info
3372  * @entry: processed cqe
3373  * @cq_poll_info: cqe info
3374  */
3375 static void irdma_process_cqe(struct ib_wc *entry,
3376                               struct irdma_cq_poll_info *cq_poll_info)
3377 {
3378         struct irdma_qp *iwqp;
3379         struct irdma_sc_qp *qp;
3380
3381         entry->wc_flags = 0;
3382         entry->pkey_index = 0;
3383         entry->wr_id = cq_poll_info->wr_id;
3384
3385         qp = cq_poll_info->qp_handle;
3386         iwqp = qp->qp_uk.back_qp;
3387         entry->qp = qp->qp_uk.back_qp;
3388
3389         if (cq_poll_info->error) {
3390                 entry->status = (cq_poll_info->comp_status == IRDMA_COMPL_STATUS_FLUSHED) ?
3391                                 irdma_flush_err_to_ib_wc_status(cq_poll_info->minor_err) : IB_WC_GENERAL_ERR;
3392
3393                 entry->vendor_err = cq_poll_info->major_err << 16 |
3394                                     cq_poll_info->minor_err;
3395         } else {
3396                 entry->status = IB_WC_SUCCESS;
3397                 if (cq_poll_info->imm_valid) {
3398                         entry->ex.imm_data = htonl(cq_poll_info->imm_data);
3399                         entry->wc_flags |= IB_WC_WITH_IMM;
3400                 }
3401                 if (cq_poll_info->ud_smac_valid) {
3402                         ether_addr_copy(entry->smac, cq_poll_info->ud_smac);
3403                         entry->wc_flags |= IB_WC_WITH_SMAC;
3404                 }
3405
3406                 if (cq_poll_info->ud_vlan_valid) {
3407                         entry->vlan_id = cq_poll_info->ud_vlan & VLAN_VID_MASK;
3408                         entry->wc_flags |= IB_WC_WITH_VLAN;
3409                         entry->sl = cq_poll_info->ud_vlan >> VLAN_PRIO_SHIFT;
3410                 } else {
3411                         entry->sl = 0;
3412                 }
3413         }
3414
3415         switch (cq_poll_info->op_type) {
3416         case IRDMA_OP_TYPE_RDMA_WRITE:
3417         case IRDMA_OP_TYPE_RDMA_WRITE_SOL:
3418                 entry->opcode = IB_WC_RDMA_WRITE;
3419                 break;
3420         case IRDMA_OP_TYPE_RDMA_READ_INV_STAG:
3421         case IRDMA_OP_TYPE_RDMA_READ:
3422                 entry->opcode = IB_WC_RDMA_READ;
3423                 break;
3424         case IRDMA_OP_TYPE_SEND_INV:
3425         case IRDMA_OP_TYPE_SEND_SOL:
3426         case IRDMA_OP_TYPE_SEND_SOL_INV:
3427         case IRDMA_OP_TYPE_SEND:
3428                 entry->opcode = IB_WC_SEND;
3429                 break;
3430         case IRDMA_OP_TYPE_FAST_REG_NSMR:
3431                 entry->opcode = IB_WC_REG_MR;
3432                 break;
3433         case IRDMA_OP_TYPE_INV_STAG:
3434                 entry->opcode = IB_WC_LOCAL_INV;
3435                 break;
3436         case IRDMA_OP_TYPE_REC_IMM:
3437         case IRDMA_OP_TYPE_REC:
3438                 entry->opcode = cq_poll_info->op_type == IRDMA_OP_TYPE_REC_IMM ?
3439                         IB_WC_RECV_RDMA_WITH_IMM : IB_WC_RECV;
3440                 if (qp->qp_uk.qp_type != IRDMA_QP_TYPE_ROCE_UD &&
3441                     cq_poll_info->stag_invalid_set) {
3442                         entry->ex.invalidate_rkey = cq_poll_info->inv_stag;
3443                         entry->wc_flags |= IB_WC_WITH_INVALIDATE;
3444                 }
3445                 break;
3446         default:
3447                 ibdev_err(&iwqp->iwdev->ibdev,
3448                           "Invalid opcode = %d in CQE\n", cq_poll_info->op_type);
3449                 entry->status = IB_WC_GENERAL_ERR;
3450                 return;
3451         }
3452
3453         if (qp->qp_uk.qp_type == IRDMA_QP_TYPE_ROCE_UD) {
3454                 entry->src_qp = cq_poll_info->ud_src_qpn;
3455                 entry->slid = 0;
3456                 entry->wc_flags |=
3457                         (IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE);
3458                 entry->network_hdr_type = cq_poll_info->ipv4 ?
3459                                                   RDMA_NETWORK_IPV4 :
3460                                                   RDMA_NETWORK_IPV6;
3461         } else {
3462                 entry->src_qp = cq_poll_info->qp_id;
3463         }
3464
3465         entry->byte_len = cq_poll_info->bytes_xfered;
3466 }
3467
3468 /**
3469  * irdma_poll_one - poll one entry of the CQ
3470  * @ukcq: ukcq to poll
3471  * @cur_cqe: current CQE info to be filled in
3472  * @entry: ibv_wc object to be filled for non-extended CQ or NULL for extended CQ
3473  *
3474  * Returns the internal irdma device error code or 0 on success
3475  */
3476 static inline int irdma_poll_one(struct irdma_cq_uk *ukcq,
3477                                  struct irdma_cq_poll_info *cur_cqe,
3478                                  struct ib_wc *entry)
3479 {
3480         int ret = irdma_uk_cq_poll_cmpl(ukcq, cur_cqe);
3481
3482         if (ret)
3483                 return ret;
3484
3485         irdma_process_cqe(entry, cur_cqe);
3486
3487         return 0;
3488 }
3489
3490 /**
3491  * __irdma_poll_cq - poll cq for completion (kernel apps)
3492  * @iwcq: cq to poll
3493  * @num_entries: number of entries to poll
3494  * @entry: wr of a completed entry
3495  */
3496 static int __irdma_poll_cq(struct irdma_cq *iwcq, int num_entries, struct ib_wc *entry)
3497 {
3498         struct list_head *tmp_node, *list_node;
3499         struct irdma_cq_buf *last_buf = NULL;
3500         struct irdma_cq_poll_info *cur_cqe = &iwcq->cur_cqe;
3501         struct irdma_cq_buf *cq_buf;
3502         enum irdma_status_code ret;
3503         struct irdma_device *iwdev;
3504         struct irdma_cq_uk *ukcq;
3505         bool cq_new_cqe = false;
3506         int resized_bufs = 0;
3507         int npolled = 0;
3508
3509         iwdev = to_iwdev(iwcq->ibcq.device);
3510         ukcq = &iwcq->sc_cq.cq_uk;
3511
3512         /* go through the list of previously resized CQ buffers */
3513         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
3514                 cq_buf = container_of(list_node, struct irdma_cq_buf, list);
3515                 while (npolled < num_entries) {
3516                         ret = irdma_poll_one(&cq_buf->cq_uk, cur_cqe, entry + npolled);
3517                         if (!ret) {
3518                                 ++npolled;
3519                                 cq_new_cqe = true;
3520                                 continue;
3521                         }
3522                         if (ret == IRDMA_ERR_Q_EMPTY)
3523                                 break;
3524                          /* QP using the CQ is destroyed. Skip reporting this CQE */
3525                         if (ret == IRDMA_ERR_Q_DESTROYED) {
3526                                 cq_new_cqe = true;
3527                                 continue;
3528                         }
3529                         goto error;
3530                 }
3531
3532                 /* save the resized CQ buffer which received the last cqe */
3533                 if (cq_new_cqe)
3534                         last_buf = cq_buf;
3535                 cq_new_cqe = false;
3536         }
3537
3538         /* check the current CQ for new cqes */
3539         while (npolled < num_entries) {
3540                 ret = irdma_poll_one(ukcq, cur_cqe, entry + npolled);
3541                 if (!ret) {
3542                         ++npolled;
3543                         cq_new_cqe = true;
3544                         continue;
3545                 }
3546
3547                 if (ret == IRDMA_ERR_Q_EMPTY)
3548                         break;
3549                 /* QP using the CQ is destroyed. Skip reporting this CQE */
3550                 if (ret == IRDMA_ERR_Q_DESTROYED) {
3551                         cq_new_cqe = true;
3552                         continue;
3553                 }
3554                 goto error;
3555         }
3556
3557         if (cq_new_cqe)
3558                 /* all previous CQ resizes are complete */
3559                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, NULL);
3560         else if (last_buf)
3561                 /* only CQ resizes up to the last_buf are complete */
3562                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, last_buf);
3563         if (resized_bufs)
3564                 /* report to the HW the number of complete CQ resizes */
3565                 irdma_uk_cq_set_resized_cnt(ukcq, resized_bufs);
3566
3567         return npolled;
3568 error:
3569         ibdev_dbg(&iwdev->ibdev, "%s: Error polling CQ, irdma_err: %d\n",
3570                   __func__, ret);
3571
3572         return -EINVAL;
3573 }
3574
3575 /**
3576  * irdma_poll_cq - poll cq for completion (kernel apps)
3577  * @ibcq: cq to poll
3578  * @num_entries: number of entries to poll
3579  * @entry: wr of a completed entry
3580  */
3581 static int irdma_poll_cq(struct ib_cq *ibcq, int num_entries,
3582                          struct ib_wc *entry)
3583 {
3584         struct irdma_cq *iwcq;
3585         unsigned long flags;
3586         int ret;
3587
3588         iwcq = to_iwcq(ibcq);
3589
3590         spin_lock_irqsave(&iwcq->lock, flags);
3591         ret = __irdma_poll_cq(iwcq, num_entries, entry);
3592         spin_unlock_irqrestore(&iwcq->lock, flags);
3593
3594         return ret;
3595 }
3596
3597 /**
3598  * irdma_req_notify_cq - arm cq kernel application
3599  * @ibcq: cq to arm
3600  * @notify_flags: notofication flags
3601  */
3602 static int irdma_req_notify_cq(struct ib_cq *ibcq,
3603                                enum ib_cq_notify_flags notify_flags)
3604 {
3605         struct irdma_cq *iwcq;
3606         struct irdma_cq_uk *ukcq;
3607         unsigned long flags;
3608         enum irdma_cmpl_notify cq_notify = IRDMA_CQ_COMPL_EVENT;
3609
3610         iwcq = to_iwcq(ibcq);
3611         ukcq = &iwcq->sc_cq.cq_uk;
3612         if (notify_flags == IB_CQ_SOLICITED)
3613                 cq_notify = IRDMA_CQ_COMPL_SOLICITED;
3614
3615         spin_lock_irqsave(&iwcq->lock, flags);
3616         irdma_uk_cq_request_notification(ukcq, cq_notify);
3617         spin_unlock_irqrestore(&iwcq->lock, flags);
3618
3619         return 0;
3620 }
3621
3622 static int irdma_roce_port_immutable(struct ib_device *ibdev, u32 port_num,
3623                                      struct ib_port_immutable *immutable)
3624 {
3625         struct ib_port_attr attr;
3626         int err;
3627
3628         immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
3629         err = ib_query_port(ibdev, port_num, &attr);
3630         if (err)
3631                 return err;
3632
3633         immutable->max_mad_size = IB_MGMT_MAD_SIZE;
3634         immutable->pkey_tbl_len = attr.pkey_tbl_len;
3635         immutable->gid_tbl_len = attr.gid_tbl_len;
3636
3637         return 0;
3638 }
3639
3640 static int irdma_iw_port_immutable(struct ib_device *ibdev, u32 port_num,
3641                                    struct ib_port_immutable *immutable)
3642 {
3643         struct ib_port_attr attr;
3644         int err;
3645
3646         immutable->core_cap_flags = RDMA_CORE_PORT_IWARP;
3647         err = ib_query_port(ibdev, port_num, &attr);
3648         if (err)
3649                 return err;
3650         immutable->gid_tbl_len = attr.gid_tbl_len;
3651
3652         return 0;
3653 }
3654
3655 static const char *const irdma_hw_stat_names[] = {
3656         /* 32bit names */
3657         [IRDMA_HW_STAT_INDEX_RXVLANERR] = "rxVlanErrors",
3658         [IRDMA_HW_STAT_INDEX_IP4RXDISCARD] = "ip4InDiscards",
3659         [IRDMA_HW_STAT_INDEX_IP4RXTRUNC] = "ip4InTruncatedPkts",
3660         [IRDMA_HW_STAT_INDEX_IP4TXNOROUTE] = "ip4OutNoRoutes",
3661         [IRDMA_HW_STAT_INDEX_IP6RXDISCARD] = "ip6InDiscards",
3662         [IRDMA_HW_STAT_INDEX_IP6RXTRUNC] = "ip6InTruncatedPkts",
3663         [IRDMA_HW_STAT_INDEX_IP6TXNOROUTE] = "ip6OutNoRoutes",
3664         [IRDMA_HW_STAT_INDEX_TCPRTXSEG] = "tcpRetransSegs",
3665         [IRDMA_HW_STAT_INDEX_TCPRXOPTERR] = "tcpInOptErrors",
3666         [IRDMA_HW_STAT_INDEX_TCPRXPROTOERR] = "tcpInProtoErrors",
3667         [IRDMA_HW_STAT_INDEX_RXRPCNPHANDLED] = "cnpHandled",
3668         [IRDMA_HW_STAT_INDEX_RXRPCNPIGNORED] = "cnpIgnored",
3669         [IRDMA_HW_STAT_INDEX_TXNPCNPSENT] = "cnpSent",
3670
3671         /* 64bit names */
3672         [IRDMA_HW_STAT_INDEX_IP4RXOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3673                 "ip4InOctets",
3674         [IRDMA_HW_STAT_INDEX_IP4RXPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3675                 "ip4InPkts",
3676         [IRDMA_HW_STAT_INDEX_IP4RXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32] =
3677                 "ip4InReasmRqd",
3678         [IRDMA_HW_STAT_INDEX_IP4RXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3679                 "ip4InMcastOctets",
3680         [IRDMA_HW_STAT_INDEX_IP4RXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3681                 "ip4InMcastPkts",
3682         [IRDMA_HW_STAT_INDEX_IP4TXOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3683                 "ip4OutOctets",
3684         [IRDMA_HW_STAT_INDEX_IP4TXPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3685                 "ip4OutPkts",
3686         [IRDMA_HW_STAT_INDEX_IP4TXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32] =
3687                 "ip4OutSegRqd",
3688         [IRDMA_HW_STAT_INDEX_IP4TXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3689                 "ip4OutMcastOctets",
3690         [IRDMA_HW_STAT_INDEX_IP4TXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3691                 "ip4OutMcastPkts",
3692         [IRDMA_HW_STAT_INDEX_IP6RXOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3693                 "ip6InOctets",
3694         [IRDMA_HW_STAT_INDEX_IP6RXPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3695                 "ip6InPkts",
3696         [IRDMA_HW_STAT_INDEX_IP6RXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32] =
3697                 "ip6InReasmRqd",
3698         [IRDMA_HW_STAT_INDEX_IP6RXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3699                 "ip6InMcastOctets",
3700         [IRDMA_HW_STAT_INDEX_IP6RXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3701                 "ip6InMcastPkts",
3702         [IRDMA_HW_STAT_INDEX_IP6TXOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3703                 "ip6OutOctets",
3704         [IRDMA_HW_STAT_INDEX_IP6TXPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3705                 "ip6OutPkts",
3706         [IRDMA_HW_STAT_INDEX_IP6TXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32] =
3707                 "ip6OutSegRqd",
3708         [IRDMA_HW_STAT_INDEX_IP6TXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3709                 "ip6OutMcastOctets",
3710         [IRDMA_HW_STAT_INDEX_IP6TXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3711                 "ip6OutMcastPkts",
3712         [IRDMA_HW_STAT_INDEX_TCPRXSEGS + IRDMA_HW_STAT_INDEX_MAX_32] =
3713                 "tcpInSegs",
3714         [IRDMA_HW_STAT_INDEX_TCPTXSEG + IRDMA_HW_STAT_INDEX_MAX_32] =
3715                 "tcpOutSegs",
3716         [IRDMA_HW_STAT_INDEX_RDMARXRDS + IRDMA_HW_STAT_INDEX_MAX_32] =
3717                 "iwInRdmaReads",
3718         [IRDMA_HW_STAT_INDEX_RDMARXSNDS + IRDMA_HW_STAT_INDEX_MAX_32] =
3719                 "iwInRdmaSends",
3720         [IRDMA_HW_STAT_INDEX_RDMARXWRS + IRDMA_HW_STAT_INDEX_MAX_32] =
3721                 "iwInRdmaWrites",
3722         [IRDMA_HW_STAT_INDEX_RDMATXRDS + IRDMA_HW_STAT_INDEX_MAX_32] =
3723                 "iwOutRdmaReads",
3724         [IRDMA_HW_STAT_INDEX_RDMATXSNDS + IRDMA_HW_STAT_INDEX_MAX_32] =
3725                 "iwOutRdmaSends",
3726         [IRDMA_HW_STAT_INDEX_RDMATXWRS + IRDMA_HW_STAT_INDEX_MAX_32] =
3727                 "iwOutRdmaWrites",
3728         [IRDMA_HW_STAT_INDEX_RDMAVBND + IRDMA_HW_STAT_INDEX_MAX_32] =
3729                 "iwRdmaBnd",
3730         [IRDMA_HW_STAT_INDEX_RDMAVINV + IRDMA_HW_STAT_INDEX_MAX_32] =
3731                 "iwRdmaInv",
3732         [IRDMA_HW_STAT_INDEX_UDPRXPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3733                 "RxUDP",
3734         [IRDMA_HW_STAT_INDEX_UDPTXPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3735                 "TxUDP",
3736         [IRDMA_HW_STAT_INDEX_RXNPECNMARKEDPKTS + IRDMA_HW_STAT_INDEX_MAX_32] =
3737                 "RxECNMrkd",
3738 };
3739
3740 static void irdma_get_dev_fw_str(struct ib_device *dev, char *str)
3741 {
3742         struct irdma_device *iwdev = to_iwdev(dev);
3743
3744         snprintf(str, IB_FW_VERSION_NAME_MAX, "%u.%u",
3745                  irdma_fw_major_ver(&iwdev->rf->sc_dev),
3746                  irdma_fw_minor_ver(&iwdev->rf->sc_dev));
3747 }
3748
3749 /**
3750  * irdma_alloc_hw_port_stats - Allocate a hw stats structure
3751  * @ibdev: device pointer from stack
3752  * @port_num: port number
3753  */
3754 static struct rdma_hw_stats *irdma_alloc_hw_port_stats(struct ib_device *ibdev,
3755                                                        u32 port_num)
3756 {
3757         int num_counters = IRDMA_HW_STAT_INDEX_MAX_32 +
3758                            IRDMA_HW_STAT_INDEX_MAX_64;
3759         unsigned long lifespan = RDMA_HW_STATS_DEFAULT_LIFESPAN;
3760
3761         BUILD_BUG_ON(ARRAY_SIZE(irdma_hw_stat_names) !=
3762                      (IRDMA_HW_STAT_INDEX_MAX_32 + IRDMA_HW_STAT_INDEX_MAX_64));
3763
3764         return rdma_alloc_hw_stats_struct(irdma_hw_stat_names, num_counters,
3765                                           lifespan);
3766 }
3767
3768 /**
3769  * irdma_get_hw_stats - Populates the rdma_hw_stats structure
3770  * @ibdev: device pointer from stack
3771  * @stats: stats pointer from stack
3772  * @port_num: port number
3773  * @index: which hw counter the stack is requesting we update
3774  */
3775 static int irdma_get_hw_stats(struct ib_device *ibdev,
3776                               struct rdma_hw_stats *stats, u32 port_num,
3777                               int index)
3778 {
3779         struct irdma_device *iwdev = to_iwdev(ibdev);
3780         struct irdma_dev_hw_stats *hw_stats = &iwdev->vsi.pestat->hw_stats;
3781
3782         if (iwdev->rf->rdma_ver >= IRDMA_GEN_2)
3783                 irdma_cqp_gather_stats_cmd(&iwdev->rf->sc_dev, iwdev->vsi.pestat, true);
3784         else
3785                 irdma_cqp_gather_stats_gen1(&iwdev->rf->sc_dev, iwdev->vsi.pestat);
3786
3787         memcpy(&stats->value[0], hw_stats, sizeof(*hw_stats));
3788
3789         return stats->num_counters;
3790 }
3791
3792 /**
3793  * irdma_query_gid - Query port GID
3794  * @ibdev: device pointer from stack
3795  * @port: port number
3796  * @index: Entry index
3797  * @gid: Global ID
3798  */
3799 static int irdma_query_gid(struct ib_device *ibdev, u32 port, int index,
3800                            union ib_gid *gid)
3801 {
3802         struct irdma_device *iwdev = to_iwdev(ibdev);
3803
3804         memset(gid->raw, 0, sizeof(gid->raw));
3805         ether_addr_copy(gid->raw, iwdev->netdev->dev_addr);
3806
3807         return 0;
3808 }
3809
3810 /**
3811  * mcast_list_add -  Add a new mcast item to list
3812  * @rf: RDMA PCI function
3813  * @new_elem: pointer to element to add
3814  */
3815 static void mcast_list_add(struct irdma_pci_f *rf,
3816                            struct mc_table_list *new_elem)
3817 {
3818         list_add(&new_elem->list, &rf->mc_qht_list.list);
3819 }
3820
3821 /**
3822  * mcast_list_del - Remove an mcast item from list
3823  * @mc_qht_elem: pointer to mcast table list element
3824  */
3825 static void mcast_list_del(struct mc_table_list *mc_qht_elem)
3826 {
3827         if (mc_qht_elem)
3828                 list_del(&mc_qht_elem->list);
3829 }
3830
3831 /**
3832  * mcast_list_lookup_ip - Search mcast list for address
3833  * @rf: RDMA PCI function
3834  * @ip_mcast: pointer to mcast IP address
3835  */
3836 static struct mc_table_list *mcast_list_lookup_ip(struct irdma_pci_f *rf,
3837                                                   u32 *ip_mcast)
3838 {
3839         struct mc_table_list *mc_qht_el;
3840         struct list_head *pos, *q;
3841
3842         list_for_each_safe (pos, q, &rf->mc_qht_list.list) {
3843                 mc_qht_el = list_entry(pos, struct mc_table_list, list);
3844                 if (!memcmp(mc_qht_el->mc_info.dest_ip, ip_mcast,
3845                             sizeof(mc_qht_el->mc_info.dest_ip)))
3846                         return mc_qht_el;
3847         }
3848
3849         return NULL;
3850 }
3851
3852 /**
3853  * irdma_mcast_cqp_op - perform a mcast cqp operation
3854  * @iwdev: irdma device
3855  * @mc_grp_ctx: mcast group info
3856  * @op: operation
3857  *
3858  * returns error status
3859  */
3860 static int irdma_mcast_cqp_op(struct irdma_device *iwdev,
3861                               struct irdma_mcast_grp_info *mc_grp_ctx, u8 op)
3862 {
3863         struct cqp_cmds_info *cqp_info;
3864         struct irdma_cqp_request *cqp_request;
3865         enum irdma_status_code status;
3866
3867         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
3868         if (!cqp_request)
3869                 return -ENOMEM;
3870
3871         cqp_request->info.in.u.mc_create.info = *mc_grp_ctx;
3872         cqp_info = &cqp_request->info;
3873         cqp_info->cqp_cmd = op;
3874         cqp_info->post_sq = 1;
3875         cqp_info->in.u.mc_create.scratch = (uintptr_t)cqp_request;
3876         cqp_info->in.u.mc_create.cqp = &iwdev->rf->cqp.sc_cqp;
3877         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
3878         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
3879         if (status)
3880                 return -ENOMEM;
3881
3882         return 0;
3883 }
3884
3885 /**
3886  * irdma_mcast_mac - Get the multicast MAC for an IP address
3887  * @ip_addr: IPv4 or IPv6 address
3888  * @mac: pointer to result MAC address
3889  * @ipv4: flag indicating IPv4 or IPv6
3890  *
3891  */
3892 void irdma_mcast_mac(u32 *ip_addr, u8 *mac, bool ipv4)
3893 {
3894         u8 *ip = (u8 *)ip_addr;
3895
3896         if (ipv4) {
3897                 unsigned char mac4[ETH_ALEN] = {0x01, 0x00, 0x5E, 0x00,
3898                                                 0x00, 0x00};
3899
3900                 mac4[3] = ip[2] & 0x7F;
3901                 mac4[4] = ip[1];
3902                 mac4[5] = ip[0];
3903                 ether_addr_copy(mac, mac4);
3904         } else {
3905                 unsigned char mac6[ETH_ALEN] = {0x33, 0x33, 0x00, 0x00,
3906                                                 0x00, 0x00};
3907
3908                 mac6[2] = ip[3];
3909                 mac6[3] = ip[2];
3910                 mac6[4] = ip[1];
3911                 mac6[5] = ip[0];
3912                 ether_addr_copy(mac, mac6);
3913         }
3914 }
3915
3916 /**
3917  * irdma_attach_mcast - attach a qp to a multicast group
3918  * @ibqp: ptr to qp
3919  * @ibgid: pointer to global ID
3920  * @lid: local ID
3921  *
3922  * returns error status
3923  */
3924 static int irdma_attach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
3925 {
3926         struct irdma_qp *iwqp = to_iwqp(ibqp);
3927         struct irdma_device *iwdev = iwqp->iwdev;
3928         struct irdma_pci_f *rf = iwdev->rf;
3929         struct mc_table_list *mc_qht_elem;
3930         struct irdma_mcast_grp_ctx_entry_info mcg_info = {};
3931         unsigned long flags;
3932         u32 ip_addr[4] = {};
3933         u32 mgn;
3934         u32 no_mgs;
3935         int ret = 0;
3936         bool ipv4;
3937         u16 vlan_id;
3938         union {
3939                 struct sockaddr saddr;
3940                 struct sockaddr_in saddr_in;
3941                 struct sockaddr_in6 saddr_in6;
3942         } sgid_addr;
3943         unsigned char dmac[ETH_ALEN];
3944
3945         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
3946
3947         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid)) {
3948                 irdma_copy_ip_ntohl(ip_addr,
3949                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
3950                 irdma_netdev_vlan_ipv6(ip_addr, &vlan_id, NULL);
3951                 ipv4 = false;
3952                 ibdev_dbg(&iwdev->ibdev,
3953                           "VERBS: qp_id=%d, IP6address=%pI6\n", ibqp->qp_num,
3954                           ip_addr);
3955                 irdma_mcast_mac(ip_addr, dmac, false);
3956         } else {
3957                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
3958                 ipv4 = true;
3959                 vlan_id = irdma_get_vlan_ipv4(ip_addr);
3960                 irdma_mcast_mac(ip_addr, dmac, true);
3961                 ibdev_dbg(&iwdev->ibdev,
3962                           "VERBS: qp_id=%d, IP4address=%pI4, MAC=%pM\n",
3963                           ibqp->qp_num, ip_addr, dmac);
3964         }
3965
3966         spin_lock_irqsave(&rf->qh_list_lock, flags);
3967         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
3968         if (!mc_qht_elem) {
3969                 struct irdma_dma_mem *dma_mem_mc;
3970
3971                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3972                 mc_qht_elem = kzalloc(sizeof(*mc_qht_elem), GFP_KERNEL);
3973                 if (!mc_qht_elem)
3974                         return -ENOMEM;
3975
3976                 mc_qht_elem->mc_info.ipv4_valid = ipv4;
3977                 memcpy(mc_qht_elem->mc_info.dest_ip, ip_addr,
3978                        sizeof(mc_qht_elem->mc_info.dest_ip));
3979                 ret = irdma_alloc_rsrc(rf, rf->allocated_mcgs, rf->max_mcg,
3980                                        &mgn, &rf->next_mcg);
3981                 if (ret) {
3982                         kfree(mc_qht_elem);
3983                         return -ENOMEM;
3984                 }
3985
3986                 mc_qht_elem->mc_info.mgn = mgn;
3987                 dma_mem_mc = &mc_qht_elem->mc_grp_ctx.dma_mem_mc;
3988                 dma_mem_mc->size = ALIGN(sizeof(u64) * IRDMA_MAX_MGS_PER_CTX,
3989                                          IRDMA_HW_PAGE_SIZE);
3990                 dma_mem_mc->va = dma_alloc_coherent(rf->hw.device,
3991                                                     dma_mem_mc->size,
3992                                                     &dma_mem_mc->pa,
3993                                                     GFP_KERNEL);
3994                 if (!dma_mem_mc->va) {
3995                         irdma_free_rsrc(rf, rf->allocated_mcgs, mgn);
3996                         kfree(mc_qht_elem);
3997                         return -ENOMEM;
3998                 }
3999
4000                 mc_qht_elem->mc_grp_ctx.mg_id = (u16)mgn;
4001                 memcpy(mc_qht_elem->mc_grp_ctx.dest_ip_addr, ip_addr,
4002                        sizeof(mc_qht_elem->mc_grp_ctx.dest_ip_addr));
4003                 mc_qht_elem->mc_grp_ctx.ipv4_valid = ipv4;
4004                 mc_qht_elem->mc_grp_ctx.vlan_id = vlan_id;
4005                 if (vlan_id < VLAN_N_VID)
4006                         mc_qht_elem->mc_grp_ctx.vlan_valid = true;
4007                 mc_qht_elem->mc_grp_ctx.hmc_fcn_id = iwdev->vsi.fcn_id;
4008                 mc_qht_elem->mc_grp_ctx.qs_handle =
4009                         iwqp->sc_qp.vsi->qos[iwqp->sc_qp.user_pri].qs_handle;
4010                 ether_addr_copy(mc_qht_elem->mc_grp_ctx.dest_mac_addr, dmac);
4011
4012                 spin_lock_irqsave(&rf->qh_list_lock, flags);
4013                 mcast_list_add(rf, mc_qht_elem);
4014         } else {
4015                 if (mc_qht_elem->mc_grp_ctx.no_of_mgs ==
4016                     IRDMA_MAX_MGS_PER_CTX) {
4017                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4018                         return -ENOMEM;
4019                 }
4020         }
4021
4022         mcg_info.qp_id = iwqp->ibqp.qp_num;
4023         no_mgs = mc_qht_elem->mc_grp_ctx.no_of_mgs;
4024         irdma_sc_add_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
4025         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4026
4027         /* Only if there is a change do we need to modify or create */
4028         if (!no_mgs) {
4029                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4030                                          IRDMA_OP_MC_CREATE);
4031         } else if (no_mgs != mc_qht_elem->mc_grp_ctx.no_of_mgs) {
4032                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4033                                          IRDMA_OP_MC_MODIFY);
4034         } else {
4035                 return 0;
4036         }
4037
4038         if (ret)
4039                 goto error;
4040
4041         return 0;
4042
4043 error:
4044         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
4045         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
4046                 mcast_list_del(mc_qht_elem);
4047                 dma_free_coherent(rf->hw.device,
4048                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.size,
4049                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.va,
4050                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.pa);
4051                 mc_qht_elem->mc_grp_ctx.dma_mem_mc.va = NULL;
4052                 irdma_free_rsrc(rf, rf->allocated_mcgs,
4053                                 mc_qht_elem->mc_grp_ctx.mg_id);
4054                 kfree(mc_qht_elem);
4055         }
4056
4057         return ret;
4058 }
4059
4060 /**
4061  * irdma_detach_mcast - detach a qp from a multicast group
4062  * @ibqp: ptr to qp
4063  * @ibgid: pointer to global ID
4064  * @lid: local ID
4065  *
4066  * returns error status
4067  */
4068 static int irdma_detach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
4069 {
4070         struct irdma_qp *iwqp = to_iwqp(ibqp);
4071         struct irdma_device *iwdev = iwqp->iwdev;
4072         struct irdma_pci_f *rf = iwdev->rf;
4073         u32 ip_addr[4] = {};
4074         struct mc_table_list *mc_qht_elem;
4075         struct irdma_mcast_grp_ctx_entry_info mcg_info = {};
4076         int ret;
4077         unsigned long flags;
4078         union {
4079                 struct sockaddr saddr;
4080                 struct sockaddr_in saddr_in;
4081                 struct sockaddr_in6 saddr_in6;
4082         } sgid_addr;
4083
4084         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
4085         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid))
4086                 irdma_copy_ip_ntohl(ip_addr,
4087                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4088         else
4089                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
4090
4091         spin_lock_irqsave(&rf->qh_list_lock, flags);
4092         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
4093         if (!mc_qht_elem) {
4094                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4095                 ibdev_dbg(&iwdev->ibdev,
4096                           "VERBS: address not found MCG\n");
4097                 return 0;
4098         }
4099
4100         mcg_info.qp_id = iwqp->ibqp.qp_num;
4101         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
4102         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
4103                 mcast_list_del(mc_qht_elem);
4104                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4105                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4106                                          IRDMA_OP_MC_DESTROY);
4107                 if (ret) {
4108                         ibdev_dbg(&iwdev->ibdev,
4109                                   "VERBS: failed MC_DESTROY MCG\n");
4110                         spin_lock_irqsave(&rf->qh_list_lock, flags);
4111                         mcast_list_add(rf, mc_qht_elem);
4112                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4113                         return -EAGAIN;
4114                 }
4115
4116                 dma_free_coherent(rf->hw.device,
4117                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.size,
4118                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.va,
4119                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.pa);
4120                 mc_qht_elem->mc_grp_ctx.dma_mem_mc.va = NULL;
4121                 irdma_free_rsrc(rf, rf->allocated_mcgs,
4122                                 mc_qht_elem->mc_grp_ctx.mg_id);
4123                 kfree(mc_qht_elem);
4124         } else {
4125                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4126                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4127                                          IRDMA_OP_MC_MODIFY);
4128                 if (ret) {
4129                         ibdev_dbg(&iwdev->ibdev,
4130                                   "VERBS: failed Modify MCG\n");
4131                         return ret;
4132                 }
4133         }
4134
4135         return 0;
4136 }
4137
4138 /**
4139  * irdma_create_ah - create address handle
4140  * @ibah: address handle
4141  * @attr: address handle attributes
4142  * @udata: User data
4143  *
4144  * returns 0 on success, error otherwise
4145  */
4146 static int irdma_create_ah(struct ib_ah *ibah,
4147                            struct rdma_ah_init_attr *attr,
4148                            struct ib_udata *udata)
4149 {
4150         struct irdma_pd *pd = to_iwpd(ibah->pd);
4151         struct irdma_ah *ah = container_of(ibah, struct irdma_ah, ibah);
4152         struct rdma_ah_attr *ah_attr = attr->ah_attr;
4153         const struct ib_gid_attr *sgid_attr;
4154         struct irdma_device *iwdev = to_iwdev(ibah->pd->device);
4155         struct irdma_pci_f *rf = iwdev->rf;
4156         struct irdma_sc_ah *sc_ah;
4157         u32 ah_id = 0;
4158         struct irdma_ah_info *ah_info;
4159         struct irdma_create_ah_resp uresp;
4160         union {
4161                 struct sockaddr saddr;
4162                 struct sockaddr_in saddr_in;
4163                 struct sockaddr_in6 saddr_in6;
4164         } sgid_addr, dgid_addr;
4165         int err;
4166         u8 dmac[ETH_ALEN];
4167
4168         err = irdma_alloc_rsrc(rf, rf->allocated_ahs, rf->max_ah, &ah_id,
4169                                &rf->next_ah);
4170         if (err)
4171                 return err;
4172
4173         ah->pd = pd;
4174         sc_ah = &ah->sc_ah;
4175         sc_ah->ah_info.ah_idx = ah_id;
4176         sc_ah->ah_info.vsi = &iwdev->vsi;
4177         irdma_sc_init_ah(&rf->sc_dev, sc_ah);
4178         ah->sgid_index = ah_attr->grh.sgid_index;
4179         sgid_attr = ah_attr->grh.sgid_attr;
4180         memcpy(&ah->dgid, &ah_attr->grh.dgid, sizeof(ah->dgid));
4181         rdma_gid2ip((struct sockaddr *)&sgid_addr, &sgid_attr->gid);
4182         rdma_gid2ip((struct sockaddr *)&dgid_addr, &ah_attr->grh.dgid);
4183         ah->av.attrs = *ah_attr;
4184         ah->av.net_type = rdma_gid_attr_network_type(sgid_attr);
4185         ah->av.sgid_addr.saddr = sgid_addr.saddr;
4186         ah->av.dgid_addr.saddr = dgid_addr.saddr;
4187         ah_info = &sc_ah->ah_info;
4188         ah_info->ah_idx = ah_id;
4189         ah_info->pd_idx = pd->sc_pd.pd_id;
4190         if (ah_attr->ah_flags & IB_AH_GRH) {
4191                 ah_info->flow_label = ah_attr->grh.flow_label;
4192                 ah_info->hop_ttl = ah_attr->grh.hop_limit;
4193                 ah_info->tc_tos = ah_attr->grh.traffic_class;
4194         }
4195
4196         ether_addr_copy(dmac, ah_attr->roce.dmac);
4197         if (rdma_gid_attr_network_type(sgid_attr) == RDMA_NETWORK_IPV4) {
4198                 ah_info->ipv4_valid = true;
4199                 ah_info->dest_ip_addr[0] =
4200                         ntohl(dgid_addr.saddr_in.sin_addr.s_addr);
4201                 ah_info->src_ip_addr[0] =
4202                         ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
4203                 ah_info->do_lpbk = irdma_ipv4_is_lpb(ah_info->src_ip_addr[0],
4204                                                      ah_info->dest_ip_addr[0]);
4205                 if (ipv4_is_multicast(dgid_addr.saddr_in.sin_addr.s_addr)) {
4206                         ah_info->do_lpbk = true;
4207                         irdma_mcast_mac(ah_info->dest_ip_addr, dmac, true);
4208                 }
4209         } else {
4210                 irdma_copy_ip_ntohl(ah_info->dest_ip_addr,
4211                                     dgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4212                 irdma_copy_ip_ntohl(ah_info->src_ip_addr,
4213                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4214                 ah_info->do_lpbk = irdma_ipv6_is_lpb(ah_info->src_ip_addr,
4215                                                      ah_info->dest_ip_addr);
4216                 if (rdma_is_multicast_addr(&dgid_addr.saddr_in6.sin6_addr)) {
4217                         ah_info->do_lpbk = true;
4218                         irdma_mcast_mac(ah_info->dest_ip_addr, dmac, false);
4219                 }
4220         }
4221
4222         err = rdma_read_gid_l2_fields(sgid_attr, &ah_info->vlan_tag,
4223                                       ah_info->mac_addr);
4224         if (err)
4225                 goto error;
4226
4227         ah_info->dst_arpindex = irdma_add_arp(iwdev->rf, ah_info->dest_ip_addr,
4228                                               ah_info->ipv4_valid, dmac);
4229
4230         if (ah_info->dst_arpindex == -1) {
4231                 err = -EINVAL;
4232                 goto error;
4233         }
4234
4235         if (ah_info->vlan_tag >= VLAN_N_VID && iwdev->dcb)
4236                 ah_info->vlan_tag = 0;
4237
4238         if (ah_info->vlan_tag < VLAN_N_VID) {
4239                 ah_info->insert_vlan_tag = true;
4240                 ah_info->vlan_tag |=
4241                         rt_tos2priority(ah_info->tc_tos) << VLAN_PRIO_SHIFT;
4242         }
4243
4244         err = irdma_ah_cqp_op(iwdev->rf, sc_ah, IRDMA_OP_AH_CREATE,
4245                               attr->flags & RDMA_CREATE_AH_SLEEPABLE,
4246                               irdma_gsi_ud_qp_ah_cb, sc_ah);
4247
4248         if (err) {
4249                 ibdev_dbg(&iwdev->ibdev,
4250                           "VERBS: CQP-OP Create AH fail");
4251                 goto error;
4252         }
4253
4254         if (!(attr->flags & RDMA_CREATE_AH_SLEEPABLE)) {
4255                 int cnt = CQP_COMPL_WAIT_TIME_MS * CQP_TIMEOUT_THRESHOLD;
4256
4257                 do {
4258                         irdma_cqp_ce_handler(rf, &rf->ccq.sc_cq);
4259                         mdelay(1);
4260                 } while (!sc_ah->ah_info.ah_valid && --cnt);
4261
4262                 if (!cnt) {
4263                         ibdev_dbg(&iwdev->ibdev,
4264                                   "VERBS: CQP create AH timed out");
4265                         err = -ETIMEDOUT;
4266                         goto error;
4267                 }
4268         }
4269
4270         if (udata) {
4271                 uresp.ah_id = ah->sc_ah.ah_info.ah_idx;
4272                 err = ib_copy_to_udata(udata, &uresp,
4273                                        min(sizeof(uresp), udata->outlen));
4274         }
4275         return 0;
4276
4277 error:
4278         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_ahs, ah_id);
4279
4280         return err;
4281 }
4282
4283 /**
4284  * irdma_destroy_ah - Destroy address handle
4285  * @ibah: pointer to address handle
4286  * @ah_flags: flags for sleepable
4287  */
4288 static int irdma_destroy_ah(struct ib_ah *ibah, u32 ah_flags)
4289 {
4290         struct irdma_device *iwdev = to_iwdev(ibah->device);
4291         struct irdma_ah *ah = to_iwah(ibah);
4292
4293         irdma_ah_cqp_op(iwdev->rf, &ah->sc_ah, IRDMA_OP_AH_DESTROY,
4294                         false, NULL, ah);
4295
4296         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_ahs,
4297                         ah->sc_ah.ah_info.ah_idx);
4298
4299         return 0;
4300 }
4301
4302 /**
4303  * irdma_query_ah - Query address handle
4304  * @ibah: pointer to address handle
4305  * @ah_attr: address handle attributes
4306  */
4307 static int irdma_query_ah(struct ib_ah *ibah, struct rdma_ah_attr *ah_attr)
4308 {
4309         struct irdma_ah *ah = to_iwah(ibah);
4310
4311         memset(ah_attr, 0, sizeof(*ah_attr));
4312         if (ah->av.attrs.ah_flags & IB_AH_GRH) {
4313                 ah_attr->ah_flags = IB_AH_GRH;
4314                 ah_attr->grh.flow_label = ah->sc_ah.ah_info.flow_label;
4315                 ah_attr->grh.traffic_class = ah->sc_ah.ah_info.tc_tos;
4316                 ah_attr->grh.hop_limit = ah->sc_ah.ah_info.hop_ttl;
4317                 ah_attr->grh.sgid_index = ah->sgid_index;
4318                 ah_attr->grh.sgid_index = ah->sgid_index;
4319                 memcpy(&ah_attr->grh.dgid, &ah->dgid,
4320                        sizeof(ah_attr->grh.dgid));
4321         }
4322
4323         return 0;
4324 }
4325
4326 static enum rdma_link_layer irdma_get_link_layer(struct ib_device *ibdev,
4327                                                  u32 port_num)
4328 {
4329         return IB_LINK_LAYER_ETHERNET;
4330 }
4331
4332 static __be64 irdma_mac_to_guid(struct net_device *ndev)
4333 {
4334         unsigned char *mac = ndev->dev_addr;
4335         __be64 guid;
4336         unsigned char *dst = (unsigned char *)&guid;
4337
4338         dst[0] = mac[0] ^ 2;
4339         dst[1] = mac[1];
4340         dst[2] = mac[2];
4341         dst[3] = 0xff;
4342         dst[4] = 0xfe;
4343         dst[5] = mac[3];
4344         dst[6] = mac[4];
4345         dst[7] = mac[5];
4346
4347         return guid;
4348 }
4349
4350 static const struct ib_device_ops irdma_roce_dev_ops = {
4351         .attach_mcast = irdma_attach_mcast,
4352         .create_ah = irdma_create_ah,
4353         .create_user_ah = irdma_create_ah,
4354         .destroy_ah = irdma_destroy_ah,
4355         .detach_mcast = irdma_detach_mcast,
4356         .get_link_layer = irdma_get_link_layer,
4357         .get_port_immutable = irdma_roce_port_immutable,
4358         .modify_qp = irdma_modify_qp_roce,
4359         .query_ah = irdma_query_ah,
4360         .query_pkey = irdma_query_pkey,
4361 };
4362
4363 static const struct ib_device_ops irdma_iw_dev_ops = {
4364         .modify_qp = irdma_modify_qp,
4365         .get_port_immutable = irdma_iw_port_immutable,
4366         .query_gid = irdma_query_gid,
4367 };
4368
4369 static const struct ib_device_ops irdma_dev_ops = {
4370         .owner = THIS_MODULE,
4371         .driver_id = RDMA_DRIVER_IRDMA,
4372         .uverbs_abi_ver = IRDMA_ABI_VER,
4373
4374         .alloc_hw_port_stats = irdma_alloc_hw_port_stats,
4375         .alloc_mr = irdma_alloc_mr,
4376         .alloc_mw = irdma_alloc_mw,
4377         .alloc_pd = irdma_alloc_pd,
4378         .alloc_ucontext = irdma_alloc_ucontext,
4379         .create_cq = irdma_create_cq,
4380         .create_qp = irdma_create_qp,
4381         .dealloc_driver = irdma_ib_dealloc_device,
4382         .dealloc_mw = irdma_dealloc_mw,
4383         .dealloc_pd = irdma_dealloc_pd,
4384         .dealloc_ucontext = irdma_dealloc_ucontext,
4385         .dereg_mr = irdma_dereg_mr,
4386         .destroy_cq = irdma_destroy_cq,
4387         .destroy_qp = irdma_destroy_qp,
4388         .disassociate_ucontext = irdma_disassociate_ucontext,
4389         .get_dev_fw_str = irdma_get_dev_fw_str,
4390         .get_dma_mr = irdma_get_dma_mr,
4391         .get_hw_stats = irdma_get_hw_stats,
4392         .map_mr_sg = irdma_map_mr_sg,
4393         .mmap = irdma_mmap,
4394         .mmap_free = irdma_mmap_free,
4395         .poll_cq = irdma_poll_cq,
4396         .post_recv = irdma_post_recv,
4397         .post_send = irdma_post_send,
4398         .query_device = irdma_query_device,
4399         .query_port = irdma_query_port,
4400         .query_qp = irdma_query_qp,
4401         .reg_user_mr = irdma_reg_user_mr,
4402         .req_notify_cq = irdma_req_notify_cq,
4403         .resize_cq = irdma_resize_cq,
4404         INIT_RDMA_OBJ_SIZE(ib_pd, irdma_pd, ibpd),
4405         INIT_RDMA_OBJ_SIZE(ib_ucontext, irdma_ucontext, ibucontext),
4406         INIT_RDMA_OBJ_SIZE(ib_ah, irdma_ah, ibah),
4407         INIT_RDMA_OBJ_SIZE(ib_cq, irdma_cq, ibcq),
4408         INIT_RDMA_OBJ_SIZE(ib_mw, irdma_mr, ibmw),
4409 };
4410
4411 /**
4412  * irdma_init_roce_device - initialization of roce rdma device
4413  * @iwdev: irdma device
4414  */
4415 static void irdma_init_roce_device(struct irdma_device *iwdev)
4416 {
4417         iwdev->ibdev.node_type = RDMA_NODE_IB_CA;
4418         iwdev->ibdev.node_guid = irdma_mac_to_guid(iwdev->netdev);
4419         ib_set_device_ops(&iwdev->ibdev, &irdma_roce_dev_ops);
4420 }
4421
4422 /**
4423  * irdma_init_iw_device - initialization of iwarp rdma device
4424  * @iwdev: irdma device
4425  */
4426 static int irdma_init_iw_device(struct irdma_device *iwdev)
4427 {
4428         struct net_device *netdev = iwdev->netdev;
4429
4430         iwdev->ibdev.node_type = RDMA_NODE_RNIC;
4431         ether_addr_copy((u8 *)&iwdev->ibdev.node_guid, netdev->dev_addr);
4432         iwdev->ibdev.ops.iw_add_ref = irdma_qp_add_ref;
4433         iwdev->ibdev.ops.iw_rem_ref = irdma_qp_rem_ref;
4434         iwdev->ibdev.ops.iw_get_qp = irdma_get_qp;
4435         iwdev->ibdev.ops.iw_connect = irdma_connect;
4436         iwdev->ibdev.ops.iw_accept = irdma_accept;
4437         iwdev->ibdev.ops.iw_reject = irdma_reject;
4438         iwdev->ibdev.ops.iw_create_listen = irdma_create_listen;
4439         iwdev->ibdev.ops.iw_destroy_listen = irdma_destroy_listen;
4440         memcpy(iwdev->ibdev.iw_ifname, netdev->name,
4441                sizeof(iwdev->ibdev.iw_ifname));
4442         ib_set_device_ops(&iwdev->ibdev, &irdma_iw_dev_ops);
4443
4444         return 0;
4445 }
4446
4447 /**
4448  * irdma_init_rdma_device - initialization of rdma device
4449  * @iwdev: irdma device
4450  */
4451 static int irdma_init_rdma_device(struct irdma_device *iwdev)
4452 {
4453         struct pci_dev *pcidev = iwdev->rf->pcidev;
4454         int ret;
4455
4456         if (iwdev->roce_mode) {
4457                 irdma_init_roce_device(iwdev);
4458         } else {
4459                 ret = irdma_init_iw_device(iwdev);
4460                 if (ret)
4461                         return ret;
4462         }
4463         iwdev->ibdev.phys_port_cnt = 1;
4464         iwdev->ibdev.num_comp_vectors = iwdev->rf->ceqs_count;
4465         iwdev->ibdev.dev.parent = &pcidev->dev;
4466         ib_set_device_ops(&iwdev->ibdev, &irdma_dev_ops);
4467
4468         return 0;
4469 }
4470
4471 /**
4472  * irdma_port_ibevent - indicate port event
4473  * @iwdev: irdma device
4474  */
4475 void irdma_port_ibevent(struct irdma_device *iwdev)
4476 {
4477         struct ib_event event;
4478
4479         event.device = &iwdev->ibdev;
4480         event.element.port_num = 1;
4481         event.event =
4482                 iwdev->iw_status ? IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
4483         ib_dispatch_event(&event);
4484 }
4485
4486 /**
4487  * irdma_ib_unregister_device - unregister rdma device from IB
4488  * core
4489  * @iwdev: irdma device
4490  */
4491 void irdma_ib_unregister_device(struct irdma_device *iwdev)
4492 {
4493         iwdev->iw_status = 0;
4494         irdma_port_ibevent(iwdev);
4495         ib_unregister_device(&iwdev->ibdev);
4496 }
4497
4498 /**
4499  * irdma_ib_register_device - register irdma device to IB core
4500  * @iwdev: irdma device
4501  */
4502 int irdma_ib_register_device(struct irdma_device *iwdev)
4503 {
4504         int ret;
4505
4506         ret = irdma_init_rdma_device(iwdev);
4507         if (ret)
4508                 return ret;
4509
4510         ret = ib_device_set_netdev(&iwdev->ibdev, iwdev->netdev, 1);
4511         if (ret)
4512                 goto error;
4513         dma_set_max_seg_size(iwdev->rf->hw.device, UINT_MAX);
4514         ret = ib_register_device(&iwdev->ibdev, "irdma%d", iwdev->rf->hw.device);
4515         if (ret)
4516                 goto error;
4517
4518         iwdev->iw_status = 1;
4519         irdma_port_ibevent(iwdev);
4520
4521         return 0;
4522
4523 error:
4524         if (ret)
4525                 ibdev_dbg(&iwdev->ibdev, "VERBS: Register RDMA device fail\n");
4526
4527         return ret;
4528 }
4529
4530 /**
4531  * irdma_ib_dealloc_device
4532  * @ibdev: ib device
4533  *
4534  * callback from ibdev dealloc_driver to deallocate resources
4535  * unber irdma device
4536  */
4537 void irdma_ib_dealloc_device(struct ib_device *ibdev)
4538 {
4539         struct irdma_device *iwdev = to_iwdev(ibdev);
4540
4541         irdma_rt_deinit_hw(iwdev);
4542         irdma_ctrl_deinit_hw(iwdev->rf);
4543         kfree(iwdev->rf);
4544 }