drivers/infiniband/hw/irdma/verbs.c

   1 // SPDX-License-Identifier: GPL-2.0 or Linux-OpenIB
   2 /* Copyright (c) 2015 - 2021 Intel Corporation */
   3 #include "main.h"
   4
   5 /**
   6  * irdma_query_device - get device attributes
   7  * @ibdev: device pointer from stack
   8  * @props: returning device attributes
   9  * @udata: user data
  10  */
  11 static int irdma_query_device(struct ib_device *ibdev,
  12                               struct ib_device_attr *props,
  13                               struct ib_udata *udata)
  14 {
  15         struct irdma_device *iwdev = to_iwdev(ibdev);
  16         struct irdma_pci_f *rf = iwdev->rf;
  17         struct pci_dev *pcidev = iwdev->rf->pcidev;
  18         struct irdma_hw_attrs *hw_attrs = &rf->sc_dev.hw_attrs;
  19
  20         if (udata->inlen || udata->outlen)
  21                 return -EINVAL;
  22
  23         memset(props, 0, sizeof(*props));
  24         addrconf_addr_eui48((u8 *)&props->sys_image_guid,
  25                             iwdev->netdev->dev_addr);
  26         props->fw_ver = (u64)irdma_fw_major_ver(&rf->sc_dev) << 32 |
  27                         irdma_fw_minor_ver(&rf->sc_dev);
  28         props->device_cap_flags = iwdev->device_cap_flags;
  29         props->vendor_id = pcidev->vendor;
  30         props->vendor_part_id = pcidev->device;
  31
  32         props->hw_ver = rf->pcidev->revision;
  33         props->page_size_cap = SZ_4K | SZ_2M | SZ_1G;
  34         props->max_mr_size = hw_attrs->max_mr_size;
  35         props->max_qp = rf->max_qp - rf->used_qps;
  36         props->max_qp_wr = hw_attrs->max_qp_wr;
  37         props->max_send_sge = hw_attrs->uk_attrs.max_hw_wq_frags;
  38         props->max_recv_sge = hw_attrs->uk_attrs.max_hw_wq_frags;
  39         props->max_cq = rf->max_cq - rf->used_cqs;
  40         props->max_cqe = rf->max_cqe;
  41         props->max_mr = rf->max_mr - rf->used_mrs;
  42         props->max_mw = props->max_mr;
  43         props->max_pd = rf->max_pd - rf->used_pds;
  44         props->max_sge_rd = hw_attrs->uk_attrs.max_hw_read_sges;
  45         props->max_qp_rd_atom = hw_attrs->max_hw_ird;
  46         props->max_qp_init_rd_atom = hw_attrs->max_hw_ord;
  47         if (rdma_protocol_roce(ibdev, 1))
  48                 props->max_pkeys = IRDMA_PKEY_TBL_SZ;
  49         props->max_ah = rf->max_ah;
  50         props->max_mcast_grp = rf->max_mcg;
  51         props->max_mcast_qp_attach = IRDMA_MAX_MGS_PER_CTX;
  52         props->max_total_mcast_qp_attach = rf->max_qp * IRDMA_MAX_MGS_PER_CTX;
  53         props->max_fast_reg_page_list_len = IRDMA_MAX_PAGES_PER_FMR;
  54 #define HCA_CLOCK_TIMESTAMP_MASK 0x1ffff
  55         if (hw_attrs->uk_attrs.hw_rev >= IRDMA_GEN_2)
  56                 props->timestamp_mask = HCA_CLOCK_TIMESTAMP_MASK;
  57
  58         return 0;
  59 }
  60
  61 /**
  62  * irdma_get_eth_speed_and_width - Get IB port speed and width from netdev speed
  63  * @link_speed: netdev phy link speed
  64  * @active_speed: IB port speed
  65  * @active_width: IB port width
  66  */
  67 static void irdma_get_eth_speed_and_width(u32 link_speed, u16 *active_speed,
  68                                           u8 *active_width)
  69 {
  70         if (link_speed <= SPEED_1000) {
  71                 *active_width = IB_WIDTH_1X;
  72                 *active_speed = IB_SPEED_SDR;
  73         } else if (link_speed <= SPEED_10000) {
  74                 *active_width = IB_WIDTH_1X;
  75                 *active_speed = IB_SPEED_FDR10;
  76         } else if (link_speed <= SPEED_20000) {
  77                 *active_width = IB_WIDTH_4X;
  78                 *active_speed = IB_SPEED_DDR;
  79         } else if (link_speed <= SPEED_25000) {
  80                 *active_width = IB_WIDTH_1X;
  81                 *active_speed = IB_SPEED_EDR;
  82         } else if (link_speed <= SPEED_40000) {
  83                 *active_width = IB_WIDTH_4X;
  84                 *active_speed = IB_SPEED_FDR10;
  85         } else {
  86                 *active_width = IB_WIDTH_4X;
  87                 *active_speed = IB_SPEED_EDR;
  88         }
  89 }
  90
  91 /**
  92  * irdma_query_port - get port attributes
  93  * @ibdev: device pointer from stack
  94  * @port: port number for query
  95  * @props: returning device attributes
  96  */
  97 static int irdma_query_port(struct ib_device *ibdev, u32 port,
  98                             struct ib_port_attr *props)
  99 {
 100         struct irdma_device *iwdev = to_iwdev(ibdev);
 101         struct net_device *netdev = iwdev->netdev;
 102
 103         /* no need to zero out pros here. done by caller */
 104
 105         props->max_mtu = IB_MTU_4096;
 106         props->active_mtu = ib_mtu_int_to_enum(netdev->mtu);
 107         props->lid = 1;
 108         props->lmc = 0;
 109         props->sm_lid = 0;
 110         props->sm_sl = 0;
 111         if (netif_carrier_ok(netdev) && netif_running(netdev)) {
 112                 props->state = IB_PORT_ACTIVE;
 113                 props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
 114         } else {
 115                 props->state = IB_PORT_DOWN;
 116                 props->phys_state = IB_PORT_PHYS_STATE_DISABLED;
 117         }
 118         irdma_get_eth_speed_and_width(SPEED_100000, &props->active_speed,
 119                                       &props->active_width);
 120
 121         if (rdma_protocol_roce(ibdev, 1)) {
 122                 props->gid_tbl_len = 32;
 123                 props->ip_gids = true;
 124                 props->pkey_tbl_len = IRDMA_PKEY_TBL_SZ;
 125         } else {
 126                 props->gid_tbl_len = 1;
 127         }
 128         props->qkey_viol_cntr = 0;
 129         props->port_cap_flags |= IB_PORT_CM_SUP | IB_PORT_REINIT_SUP;
 130         props->max_msg_sz = iwdev->rf->sc_dev.hw_attrs.max_hw_outbound_msg_size;
 131
 132         return 0;
 133 }
 134
 135 /**
 136  * irdma_disassociate_ucontext - Disassociate user context
 137  * @context: ib user context
 138  */
 139 static void irdma_disassociate_ucontext(struct ib_ucontext *context)
 140 {
 141 }
 142
 143 static int irdma_mmap_legacy(struct irdma_ucontext *ucontext,
 144                              struct vm_area_struct *vma)
 145 {
 146         u64 pfn;
 147
 148         if (vma->vm_pgoff || vma->vm_end - vma->vm_start != PAGE_SIZE)
 149                 return -EINVAL;
 150
 151         vma->vm_private_data = ucontext;
 152         pfn = ((uintptr_t)ucontext->iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET] +
 153                pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;
 154
 155         return rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn, PAGE_SIZE,
 156                                  pgprot_noncached(vma->vm_page_prot), NULL);
 157 }
 158
 159 static void irdma_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
 160 {
 161         struct irdma_user_mmap_entry *entry = to_irdma_mmap_entry(rdma_entry);
 162
 163         kfree(entry);
 164 }
 165
 166 static struct rdma_user_mmap_entry*
 167 irdma_user_mmap_entry_insert(struct irdma_ucontext *ucontext, u64 bar_offset,
 168                              enum irdma_mmap_flag mmap_flag, u64 *mmap_offset)
 169 {
 170         struct irdma_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 171         int ret;
 172
 173         if (!entry)
 174                 return NULL;
 175
 176         entry->bar_offset = bar_offset;
 177         entry->mmap_flag = mmap_flag;
 178
 179         ret = rdma_user_mmap_entry_insert(&ucontext->ibucontext,
 180                                           &entry->rdma_entry, PAGE_SIZE);
 181         if (ret) {
 182                 kfree(entry);
 183                 return NULL;
 184         }
 185         *mmap_offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
 186
 187         return &entry->rdma_entry;
 188 }
 189
 190 /**
 191  * irdma_mmap - user memory map
 192  * @context: context created during alloc
 193  * @vma: kernel info for user memory map
 194  */
 195 static int irdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
 196 {
 197         struct rdma_user_mmap_entry *rdma_entry;
 198         struct irdma_user_mmap_entry *entry;
 199         struct irdma_ucontext *ucontext;
 200         u64 pfn;
 201         int ret;
 202
 203         ucontext = to_ucontext(context);
 204
 205         /* Legacy support for libi40iw with hard-coded mmap key */
 206         if (ucontext->legacy_mode)
 207                 return irdma_mmap_legacy(ucontext, vma);
 208
 209         rdma_entry = rdma_user_mmap_entry_get(&ucontext->ibucontext, vma);
 210         if (!rdma_entry) {
 211                 ibdev_dbg(&ucontext->iwdev->ibdev,
 212                           "VERBS: pgoff[0x%lx] does not have valid entry\n",
 213                           vma->vm_pgoff);
 214                 return -EINVAL;
 215         }
 216
 217         entry = to_irdma_mmap_entry(rdma_entry);
 218         ibdev_dbg(&ucontext->iwdev->ibdev,
 219                   "VERBS: bar_offset [0x%llx] mmap_flag [%d]\n",
 220                   entry->bar_offset, entry->mmap_flag);
 221
 222         pfn = (entry->bar_offset +
 223               pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;
 224
 225         switch (entry->mmap_flag) {
 226         case IRDMA_MMAP_IO_NC:
 227                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
 228                                         pgprot_noncached(vma->vm_page_prot),
 229                                         rdma_entry);
 230                 break;
 231         case IRDMA_MMAP_IO_WC:
 232                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
 233                                         pgprot_writecombine(vma->vm_page_prot),
 234                                         rdma_entry);
 235                 break;
 236         default:
 237                 ret = -EINVAL;
 238         }
 239
 240         if (ret)
 241                 ibdev_dbg(&ucontext->iwdev->ibdev,
 242                           "VERBS: bar_offset [0x%llx] mmap_flag[%d] err[%d]\n",
 243                           entry->bar_offset, entry->mmap_flag, ret);
 244         rdma_user_mmap_entry_put(rdma_entry);
 245
 246         return ret;
 247 }
 248
 249 /**
 250  * irdma_alloc_push_page - allocate a push page for qp
 251  * @iwqp: qp pointer
 252  */
 253 static void irdma_alloc_push_page(struct irdma_qp *iwqp)
 254 {
 255         struct irdma_cqp_request *cqp_request;
 256         struct cqp_cmds_info *cqp_info;
 257         struct irdma_device *iwdev = iwqp->iwdev;
 258         struct irdma_sc_qp *qp = &iwqp->sc_qp;
 259         int status;
 260
 261         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
 262         if (!cqp_request)
 263                 return;
 264
 265         cqp_info = &cqp_request->info;
 266         cqp_info->cqp_cmd = IRDMA_OP_MANAGE_PUSH_PAGE;
 267         cqp_info->post_sq = 1;
 268         cqp_info->in.u.manage_push_page.info.push_idx = 0;
 269         cqp_info->in.u.manage_push_page.info.qs_handle =
 270                 qp->vsi->qos[qp->user_pri].qs_handle;
 271         cqp_info->in.u.manage_push_page.info.free_page = 0;
 272         cqp_info->in.u.manage_push_page.info.push_page_type = 0;
 273         cqp_info->in.u.manage_push_page.cqp = &iwdev->rf->cqp.sc_cqp;
 274         cqp_info->in.u.manage_push_page.scratch = (uintptr_t)cqp_request;
 275
 276         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
 277         if (!status && cqp_request->compl_info.op_ret_val <
 278             iwdev->rf->sc_dev.hw_attrs.max_hw_device_pages) {
 279                 qp->push_idx = cqp_request->compl_info.op_ret_val;
 280                 qp->push_offset = 0;
 281         }
 282
 283         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
 284 }
 285
 286 /**
 287  * irdma_alloc_ucontext - Allocate the user context data structure
 288  * @uctx: uverbs context pointer
 289  * @udata: user data
 290  *
 291  * This keeps track of all objects associated with a particular
 292  * user-mode client.
 293  */
 294 static int irdma_alloc_ucontext(struct ib_ucontext *uctx,
 295                                 struct ib_udata *udata)
 296 {
 297         struct ib_device *ibdev = uctx->device;
 298         struct irdma_device *iwdev = to_iwdev(ibdev);
 299         struct irdma_alloc_ucontext_req req;
 300         struct irdma_alloc_ucontext_resp uresp = {};
 301         struct irdma_ucontext *ucontext = to_ucontext(uctx);
 302         struct irdma_uk_attrs *uk_attrs;
 303
 304         if (ib_copy_from_udata(&req, udata, min(sizeof(req), udata->inlen)))
 305                 return -EINVAL;
 306
 307         if (req.userspace_ver < 4 || req.userspace_ver > IRDMA_ABI_VER)
 308                 goto ver_error;
 309
 310         ucontext->iwdev = iwdev;
 311         ucontext->abi_ver = req.userspace_ver;
 312
 313         uk_attrs = &iwdev->rf->sc_dev.hw_attrs.uk_attrs;
 314         /* GEN_1 legacy support with libi40iw */
 315         if (udata->outlen < sizeof(uresp)) {
 316                 if (uk_attrs->hw_rev != IRDMA_GEN_1)
 317                         return -EOPNOTSUPP;
 318
 319                 ucontext->legacy_mode = true;
 320                 uresp.max_qps = iwdev->rf->max_qp;
 321                 uresp.max_pds = iwdev->rf->sc_dev.hw_attrs.max_hw_pds;
 322                 uresp.wq_size = iwdev->rf->sc_dev.hw_attrs.max_qp_wr * 2;
 323                 uresp.kernel_ver = req.userspace_ver;
 324                 if (ib_copy_to_udata(udata, &uresp,
 325                                      min(sizeof(uresp), udata->outlen)))
 326                         return -EFAULT;
 327         } else {
 328                 u64 bar_off = (uintptr_t)iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET];
 329
 330                 ucontext->db_mmap_entry =
 331                         irdma_user_mmap_entry_insert(ucontext, bar_off,
 332                                                      IRDMA_MMAP_IO_NC,
 333                                                      &uresp.db_mmap_key);
 334                 if (!ucontext->db_mmap_entry)
 335                         return -ENOMEM;
 336
 337                 uresp.kernel_ver = IRDMA_ABI_VER;
 338                 uresp.feature_flags = uk_attrs->feature_flags;
 339                 uresp.max_hw_wq_frags = uk_attrs->max_hw_wq_frags;
 340                 uresp.max_hw_read_sges = uk_attrs->max_hw_read_sges;
 341                 uresp.max_hw_inline = uk_attrs->max_hw_inline;
 342                 uresp.max_hw_rq_quanta = uk_attrs->max_hw_rq_quanta;
 343                 uresp.max_hw_wq_quanta = uk_attrs->max_hw_wq_quanta;
 344                 uresp.max_hw_sq_chunk = uk_attrs->max_hw_sq_chunk;
 345                 uresp.max_hw_cq_size = uk_attrs->max_hw_cq_size;
 346                 uresp.min_hw_cq_size = uk_attrs->min_hw_cq_size;
 347                 uresp.hw_rev = uk_attrs->hw_rev;
 348                 if (ib_copy_to_udata(udata, &uresp,
 349                                      min(sizeof(uresp), udata->outlen))) {
 350                         rdma_user_mmap_entry_remove(ucontext->db_mmap_entry);
 351                         return -EFAULT;
 352                 }
 353         }
 354
 355         INIT_LIST_HEAD(&ucontext->cq_reg_mem_list);
 356         spin_lock_init(&ucontext->cq_reg_mem_list_lock);
 357         INIT_LIST_HEAD(&ucontext->qp_reg_mem_list);
 358         spin_lock_init(&ucontext->qp_reg_mem_list_lock);
 359
 360         return 0;
 361
 362 ver_error:
 363         ibdev_err(&iwdev->ibdev,
 364                   "Invalid userspace driver version detected. Detected version %d, should be %d\n",
 365                   req.userspace_ver, IRDMA_ABI_VER);
 366         return -EINVAL;
 367 }
 368
 369 /**
 370  * irdma_dealloc_ucontext - deallocate the user context data structure
 371  * @context: user context created during alloc
 372  */
 373 static void irdma_dealloc_ucontext(struct ib_ucontext *context)
 374 {
 375         struct irdma_ucontext *ucontext = to_ucontext(context);
 376
 377         rdma_user_mmap_entry_remove(ucontext->db_mmap_entry);
 378 }
 379
 380 /**
 381  * irdma_alloc_pd - allocate protection domain
 382  * @pd: PD pointer
 383  * @udata: user data
 384  */
 385 static int irdma_alloc_pd(struct ib_pd *pd, struct ib_udata *udata)
 386 {
 387         struct irdma_pd *iwpd = to_iwpd(pd);
 388         struct irdma_device *iwdev = to_iwdev(pd->device);
 389         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 390         struct irdma_pci_f *rf = iwdev->rf;
 391         struct irdma_alloc_pd_resp uresp = {};
 392         struct irdma_sc_pd *sc_pd;
 393         u32 pd_id = 0;
 394         int err;
 395
 396         err = irdma_alloc_rsrc(rf, rf->allocated_pds, rf->max_pd, &pd_id,
 397                                &rf->next_pd);
 398         if (err)
 399                 return err;
 400
 401         sc_pd = &iwpd->sc_pd;
 402         if (udata) {
 403                 struct irdma_ucontext *ucontext =
 404                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
 405                                                   ibucontext);
 406                 irdma_sc_pd_init(dev, sc_pd, pd_id, ucontext->abi_ver);
 407                 uresp.pd_id = pd_id;
 408                 if (ib_copy_to_udata(udata, &uresp,
 409                                      min(sizeof(uresp), udata->outlen))) {
 410                         err = -EFAULT;
 411                         goto error;
 412                 }
 413         } else {
 414                 irdma_sc_pd_init(dev, sc_pd, pd_id, IRDMA_ABI_VER);
 415         }
 416
 417         return 0;
 418 error:
 419         irdma_free_rsrc(rf, rf->allocated_pds, pd_id);
 420
 421         return err;
 422 }
 423
 424 /**
 425  * irdma_dealloc_pd - deallocate pd
 426  * @ibpd: ptr of pd to be deallocated
 427  * @udata: user data
 428  */
 429 static int irdma_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 430 {
 431         struct irdma_pd *iwpd = to_iwpd(ibpd);
 432         struct irdma_device *iwdev = to_iwdev(ibpd->device);
 433
 434         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_pds, iwpd->sc_pd.pd_id);
 435
 436         return 0;
 437 }
 438
 439 /**
 440  * irdma_get_pbl - Retrieve pbl from a list given a virtual
 441  * address
 442  * @va: user virtual address
 443  * @pbl_list: pbl list to search in (QP's or CQ's)
 444  */
 445 static struct irdma_pbl *irdma_get_pbl(unsigned long va,
 446                                        struct list_head *pbl_list)
 447 {
 448         struct irdma_pbl *iwpbl;
 449
 450         list_for_each_entry (iwpbl, pbl_list, list) {
 451                 if (iwpbl->user_base == va) {
 452                         list_del(&iwpbl->list);
 453                         iwpbl->on_list = false;
 454                         return iwpbl;
 455                 }
 456         }
 457
 458         return NULL;
 459 }
 460
 461 /**
 462  * irdma_clean_cqes - clean cq entries for qp
 463  * @iwqp: qp ptr (user or kernel)
 464  * @iwcq: cq ptr
 465  */
 466 static void irdma_clean_cqes(struct irdma_qp *iwqp, struct irdma_cq *iwcq)
 467 {
 468         struct irdma_cq_uk *ukcq = &iwcq->sc_cq.cq_uk;
 469         unsigned long flags;
 470
 471         spin_lock_irqsave(&iwcq->lock, flags);
 472         irdma_uk_clean_cq(&iwqp->sc_qp.qp_uk, ukcq);
 473         spin_unlock_irqrestore(&iwcq->lock, flags);
 474 }
 475
 476 static void irdma_remove_push_mmap_entries(struct irdma_qp *iwqp)
 477 {
 478         if (iwqp->push_db_mmap_entry) {
 479                 rdma_user_mmap_entry_remove(iwqp->push_db_mmap_entry);
 480                 iwqp->push_db_mmap_entry = NULL;
 481         }
 482         if (iwqp->push_wqe_mmap_entry) {
 483                 rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
 484                 iwqp->push_wqe_mmap_entry = NULL;
 485         }
 486 }
 487
 488 static int irdma_setup_push_mmap_entries(struct irdma_ucontext *ucontext,
 489                                          struct irdma_qp *iwqp,
 490                                          u64 *push_wqe_mmap_key,
 491                                          u64 *push_db_mmap_key)
 492 {
 493         struct irdma_device *iwdev = ucontext->iwdev;
 494         u64 rsvd, bar_off;
 495
 496         rsvd = IRDMA_PF_BAR_RSVD;
 497         bar_off = (uintptr_t)iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET];
 498         /* skip over db page */
 499         bar_off += IRDMA_HW_PAGE_SIZE;
 500         /* push wqe page */
 501         bar_off += rsvd + iwqp->sc_qp.push_idx * IRDMA_HW_PAGE_SIZE;
 502         iwqp->push_wqe_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
 503                                         bar_off, IRDMA_MMAP_IO_WC,
 504                                         push_wqe_mmap_key);
 505         if (!iwqp->push_wqe_mmap_entry)
 506                 return -ENOMEM;
 507
 508         /* push doorbell page */
 509         bar_off += IRDMA_HW_PAGE_SIZE;
 510         iwqp->push_db_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
 511                                         bar_off, IRDMA_MMAP_IO_NC,
 512                                         push_db_mmap_key);
 513         if (!iwqp->push_db_mmap_entry) {
 514                 rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
 515                 return -ENOMEM;
 516         }
 517
 518         return 0;
 519 }
 520
 521 /**
 522  * irdma_destroy_qp - destroy qp
 523  * @ibqp: qp's ib pointer also to get to device's qp address
 524  * @udata: user data
 525  */
 526 static int irdma_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
 527 {
 528         struct irdma_qp *iwqp = to_iwqp(ibqp);
 529         struct irdma_device *iwdev = iwqp->iwdev;
 530
 531         iwqp->sc_qp.qp_uk.destroy_pending = true;
 532
 533         if (iwqp->iwarp_state == IRDMA_QP_STATE_RTS)
 534                 irdma_modify_qp_to_err(&iwqp->sc_qp);
 535
 536         irdma_qp_rem_ref(&iwqp->ibqp);
 537         wait_for_completion(&iwqp->free_qp);
 538         irdma_free_lsmm_rsrc(iwqp);
 539         irdma_cqp_qp_destroy_cmd(&iwdev->rf->sc_dev, &iwqp->sc_qp);
 540
 541         if (!iwqp->user_mode) {
 542                 if (iwqp->iwscq) {
 543                         irdma_clean_cqes(iwqp, iwqp->iwscq);
 544                         if (iwqp->iwrcq != iwqp->iwscq)
 545                                 irdma_clean_cqes(iwqp, iwqp->iwrcq);
 546                 }
 547         }
 548         irdma_remove_push_mmap_entries(iwqp);
 549         irdma_free_qp_rsrc(iwqp);
 550
 551         return 0;
 552 }
 553
 554 /**
 555  * irdma_setup_virt_qp - setup for allocation of virtual qp
 556  * @iwdev: irdma device
 557  * @iwqp: qp ptr
 558  * @init_info: initialize info to return
 559  */
 560 static void irdma_setup_virt_qp(struct irdma_device *iwdev,
 561                                struct irdma_qp *iwqp,
 562                                struct irdma_qp_init_info *init_info)
 563 {
 564         struct irdma_pbl *iwpbl = iwqp->iwpbl;
 565         struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;
 566
 567         iwqp->page = qpmr->sq_page;
 568         init_info->shadow_area_pa = qpmr->shadow;
 569         if (iwpbl->pbl_allocated) {
 570                 init_info->virtual_map = true;
 571                 init_info->sq_pa = qpmr->sq_pbl.idx;
 572                 init_info->rq_pa = qpmr->rq_pbl.idx;
 573         } else {
 574                 init_info->sq_pa = qpmr->sq_pbl.addr;
 575                 init_info->rq_pa = qpmr->rq_pbl.addr;
 576         }
 577 }
 578
 579 /**
 580  * irdma_setup_kmode_qp - setup initialization for kernel mode qp
 581  * @iwdev: iwarp device
 582  * @iwqp: qp ptr (user or kernel)
 583  * @info: initialize info to return
 584  * @init_attr: Initial QP create attributes
 585  */
 586 static int irdma_setup_kmode_qp(struct irdma_device *iwdev,
 587                                 struct irdma_qp *iwqp,
 588                                 struct irdma_qp_init_info *info,
 589                                 struct ib_qp_init_attr *init_attr)
 590 {
 591         struct irdma_dma_mem *mem = &iwqp->kqp.dma_mem;
 592         u32 sqdepth, rqdepth;
 593         u8 sqshift, rqshift;
 594         u32 size;
 595         int status;
 596         struct irdma_qp_uk_init_info *ukinfo = &info->qp_uk_init_info;
 597         struct irdma_uk_attrs *uk_attrs = &iwdev->rf->sc_dev.hw_attrs.uk_attrs;
 598
 599         irdma_get_wqe_shift(uk_attrs,
 600                 uk_attrs->hw_rev >= IRDMA_GEN_2 ? ukinfo->max_sq_frag_cnt + 1 :
 601                                                   ukinfo->max_sq_frag_cnt,
 602                 ukinfo->max_inline_data, &sqshift);
 603         status = irdma_get_sqdepth(uk_attrs, ukinfo->sq_size, sqshift,
 604                                    &sqdepth);
 605         if (status)
 606                 return status;
 607
 608         if (uk_attrs->hw_rev == IRDMA_GEN_1)
 609                 rqshift = IRDMA_MAX_RQ_WQE_SHIFT_GEN1;
 610         else
 611                 irdma_get_wqe_shift(uk_attrs, ukinfo->max_rq_frag_cnt, 0,
 612                                     &rqshift);
 613
 614         status = irdma_get_rqdepth(uk_attrs, ukinfo->rq_size, rqshift,
 615                                    &rqdepth);
 616         if (status)
 617                 return status;
 618
 619         iwqp->kqp.sq_wrid_mem =
 620                 kcalloc(sqdepth, sizeof(*iwqp->kqp.sq_wrid_mem), GFP_KERNEL);
 621         if (!iwqp->kqp.sq_wrid_mem)
 622                 return -ENOMEM;
 623
 624         iwqp->kqp.rq_wrid_mem =
 625                 kcalloc(rqdepth, sizeof(*iwqp->kqp.rq_wrid_mem), GFP_KERNEL);
 626         if (!iwqp->kqp.rq_wrid_mem) {
 627                 kfree(iwqp->kqp.sq_wrid_mem);
 628                 iwqp->kqp.sq_wrid_mem = NULL;
 629                 return -ENOMEM;
 630         }
 631
 632         ukinfo->sq_wrtrk_array = iwqp->kqp.sq_wrid_mem;
 633         ukinfo->rq_wrid_array = iwqp->kqp.rq_wrid_mem;
 634
 635         size = (sqdepth + rqdepth) * IRDMA_QP_WQE_MIN_SIZE;
 636         size += (IRDMA_SHADOW_AREA_SIZE << 3);
 637
 638         mem->size = ALIGN(size, 256);
 639         mem->va = dma_alloc_coherent(iwdev->rf->hw.device, mem->size,
 640                                      &mem->pa, GFP_KERNEL);
 641         if (!mem->va) {
 642                 kfree(iwqp->kqp.sq_wrid_mem);
 643                 iwqp->kqp.sq_wrid_mem = NULL;
 644                 kfree(iwqp->kqp.rq_wrid_mem);
 645                 iwqp->kqp.rq_wrid_mem = NULL;
 646                 return -ENOMEM;
 647         }
 648
 649         ukinfo->sq = mem->va;
 650         info->sq_pa = mem->pa;
 651         ukinfo->rq = &ukinfo->sq[sqdepth];
 652         info->rq_pa = info->sq_pa + (sqdepth * IRDMA_QP_WQE_MIN_SIZE);
 653         ukinfo->shadow_area = ukinfo->rq[rqdepth].elem;
 654         info->shadow_area_pa = info->rq_pa + (rqdepth * IRDMA_QP_WQE_MIN_SIZE);
 655         ukinfo->sq_size = sqdepth >> sqshift;
 656         ukinfo->rq_size = rqdepth >> rqshift;
 657         ukinfo->qp_id = iwqp->ibqp.qp_num;
 658
 659         init_attr->cap.max_send_wr = (sqdepth - IRDMA_SQ_RSVD) >> sqshift;
 660         init_attr->cap.max_recv_wr = (rqdepth - IRDMA_RQ_RSVD) >> rqshift;
 661
 662         return 0;
 663 }
 664
 665 static int irdma_cqp_create_qp_cmd(struct irdma_qp *iwqp)
 666 {
 667         struct irdma_pci_f *rf = iwqp->iwdev->rf;
 668         struct irdma_cqp_request *cqp_request;
 669         struct cqp_cmds_info *cqp_info;
 670         struct irdma_create_qp_info *qp_info;
 671         int status;
 672
 673         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
 674         if (!cqp_request)
 675                 return -ENOMEM;
 676
 677         cqp_info = &cqp_request->info;
 678         qp_info = &cqp_request->info.in.u.qp_create.info;
 679         memset(qp_info, 0, sizeof(*qp_info));
 680         qp_info->mac_valid = true;
 681         qp_info->cq_num_valid = true;
 682         qp_info->next_iwarp_state = IRDMA_QP_STATE_IDLE;
 683
 684         cqp_info->cqp_cmd = IRDMA_OP_QP_CREATE;
 685         cqp_info->post_sq = 1;
 686         cqp_info->in.u.qp_create.qp = &iwqp->sc_qp;
 687         cqp_info->in.u.qp_create.scratch = (uintptr_t)cqp_request;
 688         status = irdma_handle_cqp_op(rf, cqp_request);
 689         irdma_put_cqp_request(&rf->cqp, cqp_request);
 690
 691         return status;
 692 }
 693
 694 static void irdma_roce_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
 695                                                struct irdma_qp_host_ctx_info *ctx_info)
 696 {
 697         struct irdma_device *iwdev = iwqp->iwdev;
 698         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 699         struct irdma_roce_offload_info *roce_info;
 700         struct irdma_udp_offload_info *udp_info;
 701
 702         udp_info = &iwqp->udp_info;
 703         udp_info->snd_mss = ib_mtu_enum_to_int(ib_mtu_int_to_enum(iwdev->vsi.mtu));
 704         udp_info->cwnd = iwdev->roce_cwnd;
 705         udp_info->rexmit_thresh = 2;
 706         udp_info->rnr_nak_thresh = 2;
 707         udp_info->src_port = 0xc000;
 708         udp_info->dst_port = ROCE_V2_UDP_DPORT;
 709         roce_info = &iwqp->roce_info;
 710         ether_addr_copy(roce_info->mac_addr, iwdev->netdev->dev_addr);
 711
 712         roce_info->rd_en = true;
 713         roce_info->wr_rdresp_en = true;
 714         roce_info->bind_en = true;
 715         roce_info->dcqcn_en = false;
 716         roce_info->rtomin = 5;
 717
 718         roce_info->ack_credits = iwdev->roce_ackcreds;
 719         roce_info->ird_size = dev->hw_attrs.max_hw_ird;
 720         roce_info->ord_size = dev->hw_attrs.max_hw_ord;
 721
 722         if (!iwqp->user_mode) {
 723                 roce_info->priv_mode_en = true;
 724                 roce_info->fast_reg_en = true;
 725                 roce_info->udprivcq_en = true;
 726         }
 727         roce_info->roce_tver = 0;
 728
 729         ctx_info->roce_info = &iwqp->roce_info;
 730         ctx_info->udp_info = &iwqp->udp_info;
 731         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
 732 }
 733
 734 static void irdma_iw_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
 735                                              struct irdma_qp_host_ctx_info *ctx_info)
 736 {
 737         struct irdma_device *iwdev = iwqp->iwdev;
 738         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 739         struct irdma_iwarp_offload_info *iwarp_info;
 740
 741         iwarp_info = &iwqp->iwarp_info;
 742         ether_addr_copy(iwarp_info->mac_addr, iwdev->netdev->dev_addr);
 743         iwarp_info->rd_en = true;
 744         iwarp_info->wr_rdresp_en = true;
 745         iwarp_info->bind_en = true;
 746         iwarp_info->ecn_en = true;
 747         iwarp_info->rtomin = 5;
 748
 749         if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
 750                 iwarp_info->ib_rd_en = true;
 751         if (!iwqp->user_mode) {
 752                 iwarp_info->priv_mode_en = true;
 753                 iwarp_info->fast_reg_en = true;
 754         }
 755         iwarp_info->ddp_ver = 1;
 756         iwarp_info->rdmap_ver = 1;
 757
 758         ctx_info->iwarp_info = &iwqp->iwarp_info;
 759         ctx_info->iwarp_info_valid = true;
 760         irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
 761         ctx_info->iwarp_info_valid = false;
 762 }
 763
 764 static int irdma_validate_qp_attrs(struct ib_qp_init_attr *init_attr,
 765                                    struct irdma_device *iwdev)
 766 {
 767         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 768         struct irdma_uk_attrs *uk_attrs = &dev->hw_attrs.uk_attrs;
 769
 770         if (init_attr->create_flags)
 771                 return -EOPNOTSUPP;
 772
 773         if (init_attr->cap.max_inline_data > uk_attrs->max_hw_inline ||
 774             init_attr->cap.max_send_sge > uk_attrs->max_hw_wq_frags ||
 775             init_attr->cap.max_recv_sge > uk_attrs->max_hw_wq_frags)
 776                 return -EINVAL;
 777
 778         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 779                 if (init_attr->qp_type != IB_QPT_RC &&
 780                     init_attr->qp_type != IB_QPT_UD &&
 781                     init_attr->qp_type != IB_QPT_GSI)
 782                         return -EOPNOTSUPP;
 783         } else {
 784                 if (init_attr->qp_type != IB_QPT_RC)
 785                         return -EOPNOTSUPP;
 786         }
 787
 788         return 0;
 789 }
 790
 791 /**
 792  * irdma_create_qp - create qp
 793  * @ibqp: ptr of qp
 794  * @init_attr: attributes for qp
 795  * @udata: user data for create qp
 796  */
 797 static int irdma_create_qp(struct ib_qp *ibqp,
 798                            struct ib_qp_init_attr *init_attr,
 799                            struct ib_udata *udata)
 800 {
 801         struct ib_pd *ibpd = ibqp->pd;
 802         struct irdma_pd *iwpd = to_iwpd(ibpd);
 803         struct irdma_device *iwdev = to_iwdev(ibpd->device);
 804         struct irdma_pci_f *rf = iwdev->rf;
 805         struct irdma_qp *iwqp = to_iwqp(ibqp);
 806         struct irdma_create_qp_req req;
 807         struct irdma_create_qp_resp uresp = {};
 808         u32 qp_num = 0;
 809         int err_code;
 810         int sq_size;
 811         int rq_size;
 812         struct irdma_sc_qp *qp;
 813         struct irdma_sc_dev *dev = &rf->sc_dev;
 814         struct irdma_uk_attrs *uk_attrs = &dev->hw_attrs.uk_attrs;
 815         struct irdma_qp_init_info init_info = {};
 816         struct irdma_qp_host_ctx_info *ctx_info;
 817         unsigned long flags;
 818
 819         err_code = irdma_validate_qp_attrs(init_attr, iwdev);
 820         if (err_code)
 821                 return err_code;
 822
 823         sq_size = init_attr->cap.max_send_wr;
 824         rq_size = init_attr->cap.max_recv_wr;
 825
 826         init_info.vsi = &iwdev->vsi;
 827         init_info.qp_uk_init_info.uk_attrs = uk_attrs;
 828         init_info.qp_uk_init_info.sq_size = sq_size;
 829         init_info.qp_uk_init_info.rq_size = rq_size;
 830         init_info.qp_uk_init_info.max_sq_frag_cnt = init_attr->cap.max_send_sge;
 831         init_info.qp_uk_init_info.max_rq_frag_cnt = init_attr->cap.max_recv_sge;
 832         init_info.qp_uk_init_info.max_inline_data = init_attr->cap.max_inline_data;
 833
 834         qp = &iwqp->sc_qp;
 835         qp->qp_uk.back_qp = iwqp;
 836         qp->push_idx = IRDMA_INVALID_PUSH_PAGE_INDEX;
 837
 838         iwqp->iwdev = iwdev;
 839         iwqp->q2_ctx_mem.size = ALIGN(IRDMA_Q2_BUF_SIZE + IRDMA_QP_CTX_SIZE,
 840                                       256);
 841         iwqp->q2_ctx_mem.va = dma_alloc_coherent(dev->hw->device,
 842                                                  iwqp->q2_ctx_mem.size,
 843                                                  &iwqp->q2_ctx_mem.pa,
 844                                                  GFP_KERNEL);
 845         if (!iwqp->q2_ctx_mem.va)
 846                 return -ENOMEM;
 847
 848         init_info.q2 = iwqp->q2_ctx_mem.va;
 849         init_info.q2_pa = iwqp->q2_ctx_mem.pa;
 850         init_info.host_ctx = (__le64 *)(init_info.q2 + IRDMA_Q2_BUF_SIZE);
 851         init_info.host_ctx_pa = init_info.q2_pa + IRDMA_Q2_BUF_SIZE;
 852
 853         if (init_attr->qp_type == IB_QPT_GSI)
 854                 qp_num = 1;
 855         else
 856                 err_code = irdma_alloc_rsrc(rf, rf->allocated_qps, rf->max_qp,
 857                                             &qp_num, &rf->next_qp);
 858         if (err_code)
 859                 goto error;
 860
 861         iwqp->iwpd = iwpd;
 862         iwqp->ibqp.qp_num = qp_num;
 863         qp = &iwqp->sc_qp;
 864         iwqp->iwscq = to_iwcq(init_attr->send_cq);
 865         iwqp->iwrcq = to_iwcq(init_attr->recv_cq);
 866         iwqp->host_ctx.va = init_info.host_ctx;
 867         iwqp->host_ctx.pa = init_info.host_ctx_pa;
 868         iwqp->host_ctx.size = IRDMA_QP_CTX_SIZE;
 869
 870         init_info.pd = &iwpd->sc_pd;
 871         init_info.qp_uk_init_info.qp_id = iwqp->ibqp.qp_num;
 872         if (!rdma_protocol_roce(&iwdev->ibdev, 1))
 873                 init_info.qp_uk_init_info.first_sq_wq = 1;
 874         iwqp->ctx_info.qp_compl_ctx = (uintptr_t)qp;
 875         init_waitqueue_head(&iwqp->waitq);
 876         init_waitqueue_head(&iwqp->mod_qp_waitq);
 877
 878         if (udata) {
 879                 err_code = ib_copy_from_udata(&req, udata,
 880                                               min(sizeof(req), udata->inlen));
 881                 if (err_code) {
 882                         ibdev_dbg(&iwdev->ibdev,
 883                                   "VERBS: ib_copy_from_data fail\n");
 884                         goto error;
 885                 }
 886
 887                 iwqp->ctx_info.qp_compl_ctx = req.user_compl_ctx;
 888                 iwqp->user_mode = 1;
 889                 if (req.user_wqe_bufs) {
 890                         struct irdma_ucontext *ucontext =
 891                                 rdma_udata_to_drv_context(udata,
 892                                                           struct irdma_ucontext,
 893                                                           ibucontext);
 894
 895                         init_info.qp_uk_init_info.legacy_mode = ucontext->legacy_mode;
 896                         spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
 897                         iwqp->iwpbl = irdma_get_pbl((unsigned long)req.user_wqe_bufs,
 898                                                     &ucontext->qp_reg_mem_list);
 899                         spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
 900
 901                         if (!iwqp->iwpbl) {
 902                                 err_code = -ENODATA;
 903                                 ibdev_dbg(&iwdev->ibdev, "VERBS: no pbl info\n");
 904                                 goto error;
 905                         }
 906                 }
 907                 init_info.qp_uk_init_info.abi_ver = iwpd->sc_pd.abi_ver;
 908                 irdma_setup_virt_qp(iwdev, iwqp, &init_info);
 909         } else {
 910                 init_info.qp_uk_init_info.abi_ver = IRDMA_ABI_VER;
 911                 err_code = irdma_setup_kmode_qp(iwdev, iwqp, &init_info, init_attr);
 912         }
 913
 914         if (err_code) {
 915                 ibdev_dbg(&iwdev->ibdev, "VERBS: setup qp failed\n");
 916                 goto error;
 917         }
 918
 919         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 920                 if (init_attr->qp_type == IB_QPT_RC) {
 921                         init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_ROCE_RC;
 922                         init_info.qp_uk_init_info.qp_caps = IRDMA_SEND_WITH_IMM |
 923                                                             IRDMA_WRITE_WITH_IMM |
 924                                                             IRDMA_ROCE;
 925                 } else {
 926                         init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_ROCE_UD;
 927                         init_info.qp_uk_init_info.qp_caps = IRDMA_SEND_WITH_IMM |
 928                                                             IRDMA_ROCE;
 929                 }
 930         } else {
 931                 init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_IWARP;
 932                 init_info.qp_uk_init_info.qp_caps = IRDMA_WRITE_WITH_IMM;
 933         }
 934
 935         if (dev->hw_attrs.uk_attrs.hw_rev > IRDMA_GEN_1)
 936                 init_info.qp_uk_init_info.qp_caps |= IRDMA_PUSH_MODE;
 937
 938         err_code = irdma_sc_qp_init(qp, &init_info);
 939         if (err_code) {
 940                 ibdev_dbg(&iwdev->ibdev, "VERBS: qp_init fail\n");
 941                 goto error;
 942         }
 943
 944         ctx_info = &iwqp->ctx_info;
 945         ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
 946         ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
 947
 948         if (rdma_protocol_roce(&iwdev->ibdev, 1))
 949                 irdma_roce_fill_and_set_qpctx_info(iwqp, ctx_info);
 950         else
 951                 irdma_iw_fill_and_set_qpctx_info(iwqp, ctx_info);
 952
 953         err_code = irdma_cqp_create_qp_cmd(iwqp);
 954         if (err_code)
 955                 goto error;
 956
 957         refcount_set(&iwqp->refcnt, 1);
 958         spin_lock_init(&iwqp->lock);
 959         spin_lock_init(&iwqp->sc_qp.pfpdu.lock);
 960         iwqp->sig_all = (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ? 1 : 0;
 961         rf->qp_table[qp_num] = iwqp;
 962         iwqp->max_send_wr = sq_size;
 963         iwqp->max_recv_wr = rq_size;
 964
 965         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 966                 if (dev->ws_add(&iwdev->vsi, 0)) {
 967                         irdma_cqp_qp_destroy_cmd(&rf->sc_dev, &iwqp->sc_qp);
 968                         err_code = -EINVAL;
 969                         goto error;
 970                 }
 971
 972                 irdma_qp_add_qos(&iwqp->sc_qp);
 973         }
 974
 975         if (udata) {
 976                 /* GEN_1 legacy support with libi40iw does not have expanded uresp struct */
 977                 if (udata->outlen < sizeof(uresp)) {
 978                         uresp.lsmm = 1;
 979                         uresp.push_idx = IRDMA_INVALID_PUSH_PAGE_INDEX_GEN_1;
 980                 } else {
 981                         if (rdma_protocol_iwarp(&iwdev->ibdev, 1))
 982                                 uresp.lsmm = 1;
 983                 }
 984                 uresp.actual_sq_size = sq_size;
 985                 uresp.actual_rq_size = rq_size;
 986                 uresp.qp_id = qp_num;
 987                 uresp.qp_caps = qp->qp_uk.qp_caps;
 988
 989                 err_code = ib_copy_to_udata(udata, &uresp,
 990                                             min(sizeof(uresp), udata->outlen));
 991                 if (err_code) {
 992                         ibdev_dbg(&iwdev->ibdev, "VERBS: copy_to_udata failed\n");
 993                         irdma_destroy_qp(&iwqp->ibqp, udata);
 994                         return err_code;
 995                 }
 996         }
 997
 998         init_completion(&iwqp->free_qp);
 999         return 0;
1000
1001 error:
1002         irdma_free_qp_rsrc(iwqp);
1003         return err_code;
1004 }
1005
1006 static int irdma_get_ib_acc_flags(struct irdma_qp *iwqp)
1007 {
1008         int acc_flags = 0;
1009
1010         if (rdma_protocol_roce(iwqp->ibqp.device, 1)) {
1011                 if (iwqp->roce_info.wr_rdresp_en) {
1012                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
1013                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
1014                 }
1015                 if (iwqp->roce_info.rd_en)
1016                         acc_flags |= IB_ACCESS_REMOTE_READ;
1017                 if (iwqp->roce_info.bind_en)
1018                         acc_flags |= IB_ACCESS_MW_BIND;
1019         } else {
1020                 if (iwqp->iwarp_info.wr_rdresp_en) {
1021                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
1022                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
1023                 }
1024                 if (iwqp->iwarp_info.rd_en)
1025                         acc_flags |= IB_ACCESS_REMOTE_READ;
1026                 if (iwqp->iwarp_info.bind_en)
1027                         acc_flags |= IB_ACCESS_MW_BIND;
1028         }
1029         return acc_flags;
1030 }
1031
1032 /**
1033  * irdma_query_qp - query qp attributes
1034  * @ibqp: qp pointer
1035  * @attr: attributes pointer
1036  * @attr_mask: Not used
1037  * @init_attr: qp attributes to return
1038  */
1039 static int irdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1040                           int attr_mask, struct ib_qp_init_attr *init_attr)
1041 {
1042         struct irdma_qp *iwqp = to_iwqp(ibqp);
1043         struct irdma_sc_qp *qp = &iwqp->sc_qp;
1044
1045         memset(attr, 0, sizeof(*attr));
1046         memset(init_attr, 0, sizeof(*init_attr));
1047
1048         attr->qp_state = iwqp->ibqp_state;
1049         attr->cur_qp_state = iwqp->ibqp_state;
1050         attr->cap.max_send_wr = iwqp->max_send_wr;
1051         attr->cap.max_recv_wr = iwqp->max_recv_wr;
1052         attr->cap.max_inline_data = qp->qp_uk.max_inline_data;
1053         attr->cap.max_send_sge = qp->qp_uk.max_sq_frag_cnt;
1054         attr->cap.max_recv_sge = qp->qp_uk.max_rq_frag_cnt;
1055         attr->qp_access_flags = irdma_get_ib_acc_flags(iwqp);
1056         attr->port_num = 1;
1057         if (rdma_protocol_roce(ibqp->device, 1)) {
1058                 attr->path_mtu = ib_mtu_int_to_enum(iwqp->udp_info.snd_mss);
1059                 attr->qkey = iwqp->roce_info.qkey;
1060                 attr->rq_psn = iwqp->udp_info.epsn;
1061                 attr->sq_psn = iwqp->udp_info.psn_nxt;
1062                 attr->dest_qp_num = iwqp->roce_info.dest_qp;
1063                 attr->pkey_index = iwqp->roce_info.p_key;
1064                 attr->retry_cnt = iwqp->udp_info.rexmit_thresh;
1065                 attr->rnr_retry = iwqp->udp_info.rnr_nak_thresh;
1066                 attr->max_rd_atomic = iwqp->roce_info.ord_size;
1067                 attr->max_dest_rd_atomic = iwqp->roce_info.ird_size;
1068         }
1069
1070         init_attr->event_handler = iwqp->ibqp.event_handler;
1071         init_attr->qp_context = iwqp->ibqp.qp_context;
1072         init_attr->send_cq = iwqp->ibqp.send_cq;
1073         init_attr->recv_cq = iwqp->ibqp.recv_cq;
1074         init_attr->cap = attr->cap;
1075
1076         return 0;
1077 }
1078
1079 /**
1080  * irdma_query_pkey - Query partition key
1081  * @ibdev: device pointer from stack
1082  * @port: port number
1083  * @index: index of pkey
1084  * @pkey: pointer to store the pkey
1085  */
1086 static int irdma_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1087                             u16 *pkey)
1088 {
1089         if (index >= IRDMA_PKEY_TBL_SZ)
1090                 return -EINVAL;
1091
1092         *pkey = IRDMA_DEFAULT_PKEY;
1093         return 0;
1094 }
1095
1096 /**
1097  * irdma_modify_qp_roce - modify qp request
1098  * @ibqp: qp's pointer for modify
1099  * @attr: access attributes
1100  * @attr_mask: state mask
1101  * @udata: user data
1102  */
1103 int irdma_modify_qp_roce(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1104                          int attr_mask, struct ib_udata *udata)
1105 {
1106         struct irdma_pd *iwpd = to_iwpd(ibqp->pd);
1107         struct irdma_qp *iwqp = to_iwqp(ibqp);
1108         struct irdma_device *iwdev = iwqp->iwdev;
1109         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
1110         struct irdma_qp_host_ctx_info *ctx_info;
1111         struct irdma_roce_offload_info *roce_info;
1112         struct irdma_udp_offload_info *udp_info;
1113         struct irdma_modify_qp_info info = {};
1114         struct irdma_modify_qp_resp uresp = {};
1115         struct irdma_modify_qp_req ureq = {};
1116         unsigned long flags;
1117         u8 issue_modify_qp = 0;
1118         int ret = 0;
1119
1120         ctx_info = &iwqp->ctx_info;
1121         roce_info = &iwqp->roce_info;
1122         udp_info = &iwqp->udp_info;
1123
1124         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1125                 return -EOPNOTSUPP;
1126
1127         if (attr_mask & IB_QP_DEST_QPN)
1128                 roce_info->dest_qp = attr->dest_qp_num;
1129
1130         if (attr_mask & IB_QP_PKEY_INDEX) {
1131                 ret = irdma_query_pkey(ibqp->device, 0, attr->pkey_index,
1132                                        &roce_info->p_key);
1133                 if (ret)
1134                         return ret;
1135         }
1136
1137         if (attr_mask & IB_QP_QKEY)
1138                 roce_info->qkey = attr->qkey;
1139
1140         if (attr_mask & IB_QP_PATH_MTU)
1141                 udp_info->snd_mss = ib_mtu_enum_to_int(attr->path_mtu);
1142
1143         if (attr_mask & IB_QP_SQ_PSN) {
1144                 udp_info->psn_nxt = attr->sq_psn;
1145                 udp_info->lsn =  0xffff;
1146                 udp_info->psn_una = attr->sq_psn;
1147                 udp_info->psn_max = attr->sq_psn;
1148         }
1149
1150         if (attr_mask & IB_QP_RQ_PSN)
1151                 udp_info->epsn = attr->rq_psn;
1152
1153         if (attr_mask & IB_QP_RNR_RETRY)
1154                 udp_info->rnr_nak_thresh = attr->rnr_retry;
1155
1156         if (attr_mask & IB_QP_RETRY_CNT)
1157                 udp_info->rexmit_thresh = attr->retry_cnt;
1158
1159         ctx_info->roce_info->pd_id = iwpd->sc_pd.pd_id;
1160
1161         if (attr_mask & IB_QP_AV) {
1162                 struct irdma_av *av = &iwqp->roce_ah.av;
1163                 const struct ib_gid_attr *sgid_attr;
1164                 u16 vlan_id = VLAN_N_VID;
1165                 u32 local_ip[4];
1166
1167                 memset(&iwqp->roce_ah, 0, sizeof(iwqp->roce_ah));
1168                 if (attr->ah_attr.ah_flags & IB_AH_GRH) {
1169                         udp_info->ttl = attr->ah_attr.grh.hop_limit;
1170                         udp_info->flow_label = attr->ah_attr.grh.flow_label;
1171                         udp_info->tos = attr->ah_attr.grh.traffic_class;
1172                         udp_info->src_port =
1173                                 rdma_get_udp_sport(udp_info->flow_label,
1174                                                    ibqp->qp_num,
1175                                                    roce_info->dest_qp);
1176                         irdma_qp_rem_qos(&iwqp->sc_qp);
1177                         dev->ws_remove(iwqp->sc_qp.vsi, ctx_info->user_pri);
1178                         ctx_info->user_pri = rt_tos2priority(udp_info->tos);
1179                         iwqp->sc_qp.user_pri = ctx_info->user_pri;
1180                         if (dev->ws_add(iwqp->sc_qp.vsi, ctx_info->user_pri))
1181                                 return -ENOMEM;
1182                         irdma_qp_add_qos(&iwqp->sc_qp);
1183                 }
1184                 sgid_attr = attr->ah_attr.grh.sgid_attr;
1185                 ret = rdma_read_gid_l2_fields(sgid_attr, &vlan_id,
1186                                               ctx_info->roce_info->mac_addr);
1187                 if (ret)
1188                         return ret;
1189
1190                 if (vlan_id >= VLAN_N_VID && iwdev->dcb_vlan_mode)
1191                         vlan_id = 0;
1192                 if (vlan_id < VLAN_N_VID) {
1193                         udp_info->insert_vlan_tag = true;
1194                         udp_info->vlan_tag = vlan_id |
1195                                 ctx_info->user_pri << VLAN_PRIO_SHIFT;
1196                 } else {
1197                         udp_info->insert_vlan_tag = false;
1198                 }
1199
1200                 av->attrs = attr->ah_attr;
1201                 rdma_gid2ip((struct sockaddr *)&av->sgid_addr, &sgid_attr->gid);
1202                 rdma_gid2ip((struct sockaddr *)&av->dgid_addr, &attr->ah_attr.grh.dgid);
1203                 if (av->net_type == RDMA_NETWORK_IPV6) {
1204                         __be32 *daddr =
1205                                 av->dgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32;
1206                         __be32 *saddr =
1207                                 av->sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32;
1208
1209                         irdma_copy_ip_ntohl(&udp_info->dest_ip_addr[0], daddr);
1210                         irdma_copy_ip_ntohl(&udp_info->local_ipaddr[0], saddr);
1211
1212                         udp_info->ipv4 = false;
1213                         irdma_copy_ip_ntohl(local_ip, daddr);
1214
1215                         udp_info->arp_idx = irdma_arp_table(iwdev->rf,
1216                                                             &local_ip[0],
1217                                                             false, NULL,
1218                                                             IRDMA_ARP_RESOLVE);
1219                 } else if (av->net_type == RDMA_NETWORK_IPV4) {
1220                         __be32 saddr = av->sgid_addr.saddr_in.sin_addr.s_addr;
1221                         __be32 daddr = av->dgid_addr.saddr_in.sin_addr.s_addr;
1222
1223                         local_ip[0] = ntohl(daddr);
1224
1225                         udp_info->ipv4 = true;
1226                         udp_info->dest_ip_addr[0] = 0;
1227                         udp_info->dest_ip_addr[1] = 0;
1228                         udp_info->dest_ip_addr[2] = 0;
1229                         udp_info->dest_ip_addr[3] = local_ip[0];
1230
1231                         udp_info->local_ipaddr[0] = 0;
1232                         udp_info->local_ipaddr[1] = 0;
1233                         udp_info->local_ipaddr[2] = 0;
1234                         udp_info->local_ipaddr[3] = ntohl(saddr);
1235                 }
1236                 udp_info->arp_idx =
1237                         irdma_add_arp(iwdev->rf, local_ip, udp_info->ipv4,
1238                                       attr->ah_attr.roce.dmac);
1239         }
1240
1241         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1242                 if (attr->max_rd_atomic > dev->hw_attrs.max_hw_ord) {
1243                         ibdev_err(&iwdev->ibdev,
1244                                   "rd_atomic = %d, above max_hw_ord=%d\n",
1245                                   attr->max_rd_atomic,
1246                                   dev->hw_attrs.max_hw_ord);
1247                         return -EINVAL;
1248                 }
1249                 if (attr->max_rd_atomic)
1250                         roce_info->ord_size = attr->max_rd_atomic;
1251                 info.ord_valid = true;
1252         }
1253
1254         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1255                 if (attr->max_dest_rd_atomic > dev->hw_attrs.max_hw_ird) {
1256                         ibdev_err(&iwdev->ibdev,
1257                                   "rd_atomic = %d, above max_hw_ird=%d\n",
1258                                    attr->max_rd_atomic,
1259                                    dev->hw_attrs.max_hw_ird);
1260                         return -EINVAL;
1261                 }
1262                 if (attr->max_dest_rd_atomic)
1263                         roce_info->ird_size = attr->max_dest_rd_atomic;
1264         }
1265
1266         if (attr_mask & IB_QP_ACCESS_FLAGS) {
1267                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
1268                         roce_info->wr_rdresp_en = true;
1269                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
1270                         roce_info->wr_rdresp_en = true;
1271                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
1272                         roce_info->rd_en = true;
1273         }
1274
1275         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
1276
1277         ibdev_dbg(&iwdev->ibdev,
1278                   "VERBS: caller: %pS qp_id=%d to_ibqpstate=%d ibqpstate=%d irdma_qpstate=%d attr_mask=0x%x\n",
1279                   __builtin_return_address(0), ibqp->qp_num, attr->qp_state,
1280                   iwqp->ibqp_state, iwqp->iwarp_state, attr_mask);
1281
1282         spin_lock_irqsave(&iwqp->lock, flags);
1283         if (attr_mask & IB_QP_STATE) {
1284                 if (!ib_modify_qp_is_ok(iwqp->ibqp_state, attr->qp_state,
1285                                         iwqp->ibqp.qp_type, attr_mask)) {
1286                         ibdev_warn(&iwdev->ibdev, "modify_qp invalid for qp_id=%d, old_state=0x%x, new_state=0x%x\n",
1287                                    iwqp->ibqp.qp_num, iwqp->ibqp_state,
1288                                    attr->qp_state);
1289                         ret = -EINVAL;
1290                         goto exit;
1291                 }
1292                 info.curr_iwarp_state = iwqp->iwarp_state;
1293
1294                 switch (attr->qp_state) {
1295                 case IB_QPS_INIT:
1296                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1297                                 ret = -EINVAL;
1298                                 goto exit;
1299                         }
1300
1301                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
1302                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
1303                                 issue_modify_qp = 1;
1304                         }
1305                         break;
1306                 case IB_QPS_RTR:
1307                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1308                                 ret = -EINVAL;
1309                                 goto exit;
1310                         }
1311                         info.arp_cache_idx_valid = true;
1312                         info.cq_num_valid = true;
1313                         info.next_iwarp_state = IRDMA_QP_STATE_RTR;
1314                         issue_modify_qp = 1;
1315                         break;
1316                 case IB_QPS_RTS:
1317                         if (iwqp->ibqp_state < IB_QPS_RTR ||
1318                             iwqp->ibqp_state == IB_QPS_ERR) {
1319                                 ret = -EINVAL;
1320                                 goto exit;
1321                         }
1322
1323                         info.arp_cache_idx_valid = true;
1324                         info.cq_num_valid = true;
1325                         info.ord_valid = true;
1326                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
1327                         issue_modify_qp = 1;
1328                         if (iwdev->push_mode && udata &&
1329                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1330                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1331                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1332                                 irdma_alloc_push_page(iwqp);
1333                                 spin_lock_irqsave(&iwqp->lock, flags);
1334                         }
1335                         break;
1336                 case IB_QPS_SQD:
1337                         if (iwqp->iwarp_state == IRDMA_QP_STATE_SQD)
1338                                 goto exit;
1339
1340                         if (iwqp->iwarp_state != IRDMA_QP_STATE_RTS) {
1341                                 ret = -EINVAL;
1342                                 goto exit;
1343                         }
1344
1345                         info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1346                         issue_modify_qp = 1;
1347                         break;
1348                 case IB_QPS_SQE:
1349                 case IB_QPS_ERR:
1350                 case IB_QPS_RESET:
1351                         if (iwqp->iwarp_state == IRDMA_QP_STATE_RTS) {
1352                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1353                                 info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1354                                 irdma_hw_modify_qp(iwdev, iwqp, &info, true);
1355                                 spin_lock_irqsave(&iwqp->lock, flags);
1356                         }
1357
1358                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1359                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1360                                 if (udata) {
1361                                         if (ib_copy_from_udata(&ureq, udata,
1362                                             min(sizeof(ureq), udata->inlen)))
1363                                                 return -EINVAL;
1364
1365                                         irdma_flush_wqes(iwqp,
1366                                             (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1367                                             (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1368                                             IRDMA_REFLUSH);
1369                                 }
1370                                 return 0;
1371                         }
1372
1373                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1374                         issue_modify_qp = 1;
1375                         break;
1376                 default:
1377                         ret = -EINVAL;
1378                         goto exit;
1379                 }
1380
1381                 iwqp->ibqp_state = attr->qp_state;
1382         }
1383
1384         ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1385         ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1386         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1387         spin_unlock_irqrestore(&iwqp->lock, flags);
1388
1389         if (attr_mask & IB_QP_STATE) {
1390                 if (issue_modify_qp) {
1391                         ctx_info->rem_endpoint_idx = udp_info->arp_idx;
1392                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1393                                 return -EINVAL;
1394                         spin_lock_irqsave(&iwqp->lock, flags);
1395                         if (iwqp->iwarp_state == info.curr_iwarp_state) {
1396                                 iwqp->iwarp_state = info.next_iwarp_state;
1397                                 iwqp->ibqp_state = attr->qp_state;
1398                         }
1399                         if (iwqp->ibqp_state > IB_QPS_RTS &&
1400                             !iwqp->flush_issued) {
1401                                 iwqp->flush_issued = 1;
1402                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1403                                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_SQ |
1404                                                        IRDMA_FLUSH_RQ |
1405                                                        IRDMA_FLUSH_WAIT);
1406                         } else {
1407                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1408                         }
1409                 } else {
1410                         iwqp->ibqp_state = attr->qp_state;
1411                 }
1412                 if (udata && dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1413                         struct irdma_ucontext *ucontext;
1414
1415                         ucontext = rdma_udata_to_drv_context(udata,
1416                                         struct irdma_ucontext, ibucontext);
1417                         if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1418                             !iwqp->push_wqe_mmap_entry &&
1419                             !irdma_setup_push_mmap_entries(ucontext, iwqp,
1420                                 &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1421                                 uresp.push_valid = 1;
1422                                 uresp.push_offset = iwqp->sc_qp.push_offset;
1423                         }
1424                         ret = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1425                                                udata->outlen));
1426                         if (ret) {
1427                                 irdma_remove_push_mmap_entries(iwqp);
1428                                 ibdev_dbg(&iwdev->ibdev,
1429                                           "VERBS: copy_to_udata failed\n");
1430                                 return ret;
1431                         }
1432                 }
1433         }
1434
1435         return 0;
1436 exit:
1437         spin_unlock_irqrestore(&iwqp->lock, flags);
1438
1439         return ret;
1440 }
1441
1442 /**
1443  * irdma_modify_qp - modify qp request
1444  * @ibqp: qp's pointer for modify
1445  * @attr: access attributes
1446  * @attr_mask: state mask
1447  * @udata: user data
1448  */
1449 int irdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
1450                     struct ib_udata *udata)
1451 {
1452         struct irdma_qp *iwqp = to_iwqp(ibqp);
1453         struct irdma_device *iwdev = iwqp->iwdev;
1454         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
1455         struct irdma_qp_host_ctx_info *ctx_info;
1456         struct irdma_tcp_offload_info *tcp_info;
1457         struct irdma_iwarp_offload_info *offload_info;
1458         struct irdma_modify_qp_info info = {};
1459         struct irdma_modify_qp_resp uresp = {};
1460         struct irdma_modify_qp_req ureq = {};
1461         u8 issue_modify_qp = 0;
1462         u8 dont_wait = 0;
1463         int err;
1464         unsigned long flags;
1465
1466         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1467                 return -EOPNOTSUPP;
1468
1469         ctx_info = &iwqp->ctx_info;
1470         offload_info = &iwqp->iwarp_info;
1471         tcp_info = &iwqp->tcp_info;
1472         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
1473         ibdev_dbg(&iwdev->ibdev,
1474                   "VERBS: caller: %pS qp_id=%d to_ibqpstate=%d ibqpstate=%d irdma_qpstate=%d last_aeq=%d hw_tcp_state=%d hw_iwarp_state=%d attr_mask=0x%x\n",
1475                   __builtin_return_address(0), ibqp->qp_num, attr->qp_state,
1476                   iwqp->ibqp_state, iwqp->iwarp_state, iwqp->last_aeq,
1477                   iwqp->hw_tcp_state, iwqp->hw_iwarp_state, attr_mask);
1478
1479         spin_lock_irqsave(&iwqp->lock, flags);
1480         if (attr_mask & IB_QP_STATE) {
1481                 info.curr_iwarp_state = iwqp->iwarp_state;
1482                 switch (attr->qp_state) {
1483                 case IB_QPS_INIT:
1484                 case IB_QPS_RTR:
1485                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1486                                 err = -EINVAL;
1487                                 goto exit;
1488                         }
1489
1490                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
1491                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
1492                                 issue_modify_qp = 1;
1493                         }
1494                         if (iwdev->push_mode && udata &&
1495                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1496                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1497                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1498                                 irdma_alloc_push_page(iwqp);
1499                                 spin_lock_irqsave(&iwqp->lock, flags);
1500                         }
1501                         break;
1502                 case IB_QPS_RTS:
1503                         if (iwqp->iwarp_state > IRDMA_QP_STATE_RTS ||
1504                             !iwqp->cm_id) {
1505                                 err = -EINVAL;
1506                                 goto exit;
1507                         }
1508
1509                         issue_modify_qp = 1;
1510                         iwqp->hw_tcp_state = IRDMA_TCP_STATE_ESTABLISHED;
1511                         iwqp->hte_added = 1;
1512                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
1513                         info.tcp_ctx_valid = true;
1514                         info.ord_valid = true;
1515                         info.arp_cache_idx_valid = true;
1516                         info.cq_num_valid = true;
1517                         break;
1518                 case IB_QPS_SQD:
1519                         if (iwqp->hw_iwarp_state > IRDMA_QP_STATE_RTS) {
1520                                 err = 0;
1521                                 goto exit;
1522                         }
1523
1524                         if (iwqp->iwarp_state == IRDMA_QP_STATE_CLOSING ||
1525                             iwqp->iwarp_state < IRDMA_QP_STATE_RTS) {
1526                                 err = 0;
1527                                 goto exit;
1528                         }
1529
1530                         if (iwqp->iwarp_state > IRDMA_QP_STATE_CLOSING) {
1531                                 err = -EINVAL;
1532                                 goto exit;
1533                         }
1534
1535                         info.next_iwarp_state = IRDMA_QP_STATE_CLOSING;
1536                         issue_modify_qp = 1;
1537                         break;
1538                 case IB_QPS_SQE:
1539                         if (iwqp->iwarp_state >= IRDMA_QP_STATE_TERMINATE) {
1540                                 err = -EINVAL;
1541                                 goto exit;
1542                         }
1543
1544                         info.next_iwarp_state = IRDMA_QP_STATE_TERMINATE;
1545                         issue_modify_qp = 1;
1546                         break;
1547                 case IB_QPS_ERR:
1548                 case IB_QPS_RESET:
1549                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1550                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1551                                 if (udata) {
1552                                         if (ib_copy_from_udata(&ureq, udata,
1553                                             min(sizeof(ureq), udata->inlen)))
1554                                                 return -EINVAL;
1555
1556                                         irdma_flush_wqes(iwqp,
1557                                             (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1558                                             (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1559                                             IRDMA_REFLUSH);
1560                                 }
1561                                 return 0;
1562                         }
1563
1564                         if (iwqp->sc_qp.term_flags) {
1565                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1566                                 irdma_terminate_del_timer(&iwqp->sc_qp);
1567                                 spin_lock_irqsave(&iwqp->lock, flags);
1568                         }
1569                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1570                         if (iwqp->hw_tcp_state > IRDMA_TCP_STATE_CLOSED &&
1571                             iwdev->iw_status &&
1572                             iwqp->hw_tcp_state != IRDMA_TCP_STATE_TIME_WAIT)
1573                                 info.reset_tcp_conn = true;
1574                         else
1575                                 dont_wait = 1;
1576
1577                         issue_modify_qp = 1;
1578                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1579                         break;
1580                 default:
1581                         err = -EINVAL;
1582                         goto exit;
1583                 }
1584
1585                 iwqp->ibqp_state = attr->qp_state;
1586         }
1587         if (attr_mask & IB_QP_ACCESS_FLAGS) {
1588                 ctx_info->iwarp_info_valid = true;
1589                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
1590                         offload_info->wr_rdresp_en = true;
1591                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
1592                         offload_info->wr_rdresp_en = true;
1593                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
1594                         offload_info->rd_en = true;
1595         }
1596
1597         if (ctx_info->iwarp_info_valid) {
1598                 ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1599                 ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1600                 irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1601         }
1602         spin_unlock_irqrestore(&iwqp->lock, flags);
1603
1604         if (attr_mask & IB_QP_STATE) {
1605                 if (issue_modify_qp) {
1606                         ctx_info->rem_endpoint_idx = tcp_info->arp_idx;
1607                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1608                                 return -EINVAL;
1609                 }
1610
1611                 spin_lock_irqsave(&iwqp->lock, flags);
1612                 if (iwqp->iwarp_state == info.curr_iwarp_state) {
1613                         iwqp->iwarp_state = info.next_iwarp_state;
1614                         iwqp->ibqp_state = attr->qp_state;
1615                 }
1616                 spin_unlock_irqrestore(&iwqp->lock, flags);
1617         }
1618
1619         if (issue_modify_qp && iwqp->ibqp_state > IB_QPS_RTS) {
1620                 if (dont_wait) {
1621                         if (iwqp->cm_id && iwqp->hw_tcp_state) {
1622                                 spin_lock_irqsave(&iwqp->lock, flags);
1623                                 iwqp->hw_tcp_state = IRDMA_TCP_STATE_CLOSED;
1624                                 iwqp->last_aeq = IRDMA_AE_RESET_SENT;
1625                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1626                                 irdma_cm_disconn(iwqp);
1627                         }
1628                 } else {
1629                         int close_timer_started;
1630
1631                         spin_lock_irqsave(&iwdev->cm_core.ht_lock, flags);
1632
1633                         if (iwqp->cm_node) {
1634                                 refcount_inc(&iwqp->cm_node->refcnt);
1635                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1636                                 close_timer_started = atomic_inc_return(&iwqp->close_timer_started);
1637                                 if (iwqp->cm_id && close_timer_started == 1)
1638                                         irdma_schedule_cm_timer(iwqp->cm_node,
1639                                                 (struct irdma_puda_buf *)iwqp,
1640                                                 IRDMA_TIMER_TYPE_CLOSE, 1, 0);
1641
1642                                 irdma_rem_ref_cm_node(iwqp->cm_node);
1643                         } else {
1644                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1645                         }
1646                 }
1647         }
1648         if (attr_mask & IB_QP_STATE && udata &&
1649             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1650                 struct irdma_ucontext *ucontext;
1651
1652                 ucontext = rdma_udata_to_drv_context(udata,
1653                                         struct irdma_ucontext, ibucontext);
1654                 if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1655                     !iwqp->push_wqe_mmap_entry &&
1656                     !irdma_setup_push_mmap_entries(ucontext, iwqp,
1657                         &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1658                         uresp.push_valid = 1;
1659                         uresp.push_offset = iwqp->sc_qp.push_offset;
1660                 }
1661
1662                 err = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1663                                        udata->outlen));
1664                 if (err) {
1665                         irdma_remove_push_mmap_entries(iwqp);
1666                         ibdev_dbg(&iwdev->ibdev,
1667                                   "VERBS: copy_to_udata failed\n");
1668                         return err;
1669                 }
1670         }
1671
1672         return 0;
1673 exit:
1674         spin_unlock_irqrestore(&iwqp->lock, flags);
1675
1676         return err;
1677 }
1678
1679 /**
1680  * irdma_cq_free_rsrc - free up resources for cq
1681  * @rf: RDMA PCI function
1682  * @iwcq: cq ptr
1683  */
1684 static void irdma_cq_free_rsrc(struct irdma_pci_f *rf, struct irdma_cq *iwcq)
1685 {
1686         struct irdma_sc_cq *cq = &iwcq->sc_cq;
1687
1688         if (!iwcq->user_mode) {
1689                 dma_free_coherent(rf->sc_dev.hw->device, iwcq->kmem.size,
1690                                   iwcq->kmem.va, iwcq->kmem.pa);
1691                 iwcq->kmem.va = NULL;
1692                 dma_free_coherent(rf->sc_dev.hw->device,
1693                                   iwcq->kmem_shadow.size,
1694                                   iwcq->kmem_shadow.va, iwcq->kmem_shadow.pa);
1695                 iwcq->kmem_shadow.va = NULL;
1696         }
1697
1698         irdma_free_rsrc(rf, rf->allocated_cqs, cq->cq_uk.cq_id);
1699 }
1700
1701 /**
1702  * irdma_free_cqbuf - worker to free a cq buffer
1703  * @work: provides access to the cq buffer to free
1704  */
1705 static void irdma_free_cqbuf(struct work_struct *work)
1706 {
1707         struct irdma_cq_buf *cq_buf = container_of(work, struct irdma_cq_buf, work);
1708
1709         dma_free_coherent(cq_buf->hw->device, cq_buf->kmem_buf.size,
1710                           cq_buf->kmem_buf.va, cq_buf->kmem_buf.pa);
1711         cq_buf->kmem_buf.va = NULL;
1712         kfree(cq_buf);
1713 }
1714
1715 /**
1716  * irdma_process_resize_list - remove resized cq buffers from the resize_list
1717  * @iwcq: cq which owns the resize_list
1718  * @iwdev: irdma device
1719  * @lcqe_buf: the buffer where the last cqe is received
1720  */
1721 static int irdma_process_resize_list(struct irdma_cq *iwcq,
1722                                      struct irdma_device *iwdev,
1723                                      struct irdma_cq_buf *lcqe_buf)
1724 {
1725         struct list_head *tmp_node, *list_node;
1726         struct irdma_cq_buf *cq_buf;
1727         int cnt = 0;
1728
1729         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
1730                 cq_buf = list_entry(list_node, struct irdma_cq_buf, list);
1731                 if (cq_buf == lcqe_buf)
1732                         return cnt;
1733
1734                 list_del(&cq_buf->list);
1735                 queue_work(iwdev->cleanup_wq, &cq_buf->work);
1736                 cnt++;
1737         }
1738
1739         return cnt;
1740 }
1741
1742 /**
1743  * irdma_destroy_cq - destroy cq
1744  * @ib_cq: cq pointer
1745  * @udata: user data
1746  */
1747 static int irdma_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata)
1748 {
1749         struct irdma_device *iwdev = to_iwdev(ib_cq->device);
1750         struct irdma_cq *iwcq = to_iwcq(ib_cq);
1751         struct irdma_sc_cq *cq = &iwcq->sc_cq;
1752         struct irdma_sc_dev *dev = cq->dev;
1753         struct irdma_sc_ceq *ceq = dev->ceq[cq->ceq_id];
1754         struct irdma_ceq *iwceq = container_of(ceq, struct irdma_ceq, sc_ceq);
1755         unsigned long flags;
1756
1757         spin_lock_irqsave(&iwcq->lock, flags);
1758         if (!list_empty(&iwcq->resize_list))
1759                 irdma_process_resize_list(iwcq, iwdev, NULL);
1760         spin_unlock_irqrestore(&iwcq->lock, flags);
1761
1762         irdma_cq_wq_destroy(iwdev->rf, cq);
1763         irdma_cq_free_rsrc(iwdev->rf, iwcq);
1764
1765         spin_lock_irqsave(&iwceq->ce_lock, flags);
1766         irdma_sc_cleanup_ceqes(cq, ceq);
1767         spin_unlock_irqrestore(&iwceq->ce_lock, flags);
1768
1769         return 0;
1770 }
1771
1772 /**
1773  * irdma_resize_cq - resize cq
1774  * @ibcq: cq to be resized
1775  * @entries: desired cq size
1776  * @udata: user data
1777  */
1778 static int irdma_resize_cq(struct ib_cq *ibcq, int entries,
1779                            struct ib_udata *udata)
1780 {
1781         struct irdma_cq *iwcq = to_iwcq(ibcq);
1782         struct irdma_sc_dev *dev = iwcq->sc_cq.dev;
1783         struct irdma_cqp_request *cqp_request;
1784         struct cqp_cmds_info *cqp_info;
1785         struct irdma_modify_cq_info *m_info;
1786         struct irdma_modify_cq_info info = {};
1787         struct irdma_dma_mem kmem_buf;
1788         struct irdma_cq_mr *cqmr_buf;
1789         struct irdma_pbl *iwpbl_buf;
1790         struct irdma_device *iwdev;
1791         struct irdma_pci_f *rf;
1792         struct irdma_cq_buf *cq_buf = NULL;
1793         unsigned long flags;
1794         int ret;
1795
1796         iwdev = to_iwdev(ibcq->device);
1797         rf = iwdev->rf;
1798
1799         if (!(rf->sc_dev.hw_attrs.uk_attrs.feature_flags &
1800             IRDMA_FEATURE_CQ_RESIZE))
1801                 return -EOPNOTSUPP;
1802
1803         if (entries > rf->max_cqe)
1804                 return -EINVAL;
1805
1806         if (!iwcq->user_mode) {
1807                 entries++;
1808                 if (rf->sc_dev.hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
1809                         entries *= 2;
1810         }
1811
1812         info.cq_size = max(entries, 4);
1813
1814         if (info.cq_size == iwcq->sc_cq.cq_uk.cq_size - 1)
1815                 return 0;
1816
1817         if (udata) {
1818                 struct irdma_resize_cq_req req = {};
1819                 struct irdma_ucontext *ucontext =
1820                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
1821                                                   ibucontext);
1822
1823                 /* CQ resize not supported with legacy GEN_1 libi40iw */
1824                 if (ucontext->legacy_mode)
1825                         return -EOPNOTSUPP;
1826
1827                 if (ib_copy_from_udata(&req, udata,
1828                                        min(sizeof(req), udata->inlen)))
1829                         return -EINVAL;
1830
1831                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
1832                 iwpbl_buf = irdma_get_pbl((unsigned long)req.user_cq_buffer,
1833                                           &ucontext->cq_reg_mem_list);
1834                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
1835
1836                 if (!iwpbl_buf)
1837                         return -ENOMEM;
1838
1839                 cqmr_buf = &iwpbl_buf->cq_mr;
1840                 if (iwpbl_buf->pbl_allocated) {
1841                         info.virtual_map = true;
1842                         info.pbl_chunk_size = 1;
1843                         info.first_pm_pbl_idx = cqmr_buf->cq_pbl.idx;
1844                 } else {
1845                         info.cq_pa = cqmr_buf->cq_pbl.addr;
1846                 }
1847         } else {
1848                 /* Kmode CQ resize */
1849                 int rsize;
1850
1851                 rsize = info.cq_size * sizeof(struct irdma_cqe);
1852                 kmem_buf.size = ALIGN(round_up(rsize, 256), 256);
1853                 kmem_buf.va = dma_alloc_coherent(dev->hw->device,
1854                                                  kmem_buf.size, &kmem_buf.pa,
1855                                                  GFP_KERNEL);
1856                 if (!kmem_buf.va)
1857                         return -ENOMEM;
1858
1859                 info.cq_base = kmem_buf.va;
1860                 info.cq_pa = kmem_buf.pa;
1861                 cq_buf = kzalloc(sizeof(*cq_buf), GFP_KERNEL);
1862                 if (!cq_buf) {
1863                         ret = -ENOMEM;
1864                         goto error;
1865                 }
1866         }
1867
1868         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
1869         if (!cqp_request) {
1870                 ret = -ENOMEM;
1871                 goto error;
1872         }
1873
1874         info.shadow_read_threshold = iwcq->sc_cq.shadow_read_threshold;
1875         info.cq_resize = true;
1876
1877         cqp_info = &cqp_request->info;
1878         m_info = &cqp_info->in.u.cq_modify.info;
1879         memcpy(m_info, &info, sizeof(*m_info));
1880
1881         cqp_info->cqp_cmd = IRDMA_OP_CQ_MODIFY;
1882         cqp_info->in.u.cq_modify.cq = &iwcq->sc_cq;
1883         cqp_info->in.u.cq_modify.scratch = (uintptr_t)cqp_request;
1884         cqp_info->post_sq = 1;
1885         ret = irdma_handle_cqp_op(rf, cqp_request);
1886         irdma_put_cqp_request(&rf->cqp, cqp_request);
1887         if (ret)
1888                 goto error;
1889
1890         spin_lock_irqsave(&iwcq->lock, flags);
1891         if (cq_buf) {
1892                 cq_buf->kmem_buf = iwcq->kmem;
1893                 cq_buf->hw = dev->hw;
1894                 memcpy(&cq_buf->cq_uk, &iwcq->sc_cq.cq_uk, sizeof(cq_buf->cq_uk));
1895                 INIT_WORK(&cq_buf->work, irdma_free_cqbuf);
1896                 list_add_tail(&cq_buf->list, &iwcq->resize_list);
1897                 iwcq->kmem = kmem_buf;
1898         }
1899
1900         irdma_sc_cq_resize(&iwcq->sc_cq, &info);
1901         ibcq->cqe = info.cq_size - 1;
1902         spin_unlock_irqrestore(&iwcq->lock, flags);
1903
1904         return 0;
1905 error:
1906         if (!udata) {
1907                 dma_free_coherent(dev->hw->device, kmem_buf.size, kmem_buf.va,
1908                                   kmem_buf.pa);
1909                 kmem_buf.va = NULL;
1910         }
1911         kfree(cq_buf);
1912
1913         return ret;
1914 }
1915
1916 static inline int cq_validate_flags(u32 flags, u8 hw_rev)
1917 {
1918         /* GEN1 does not support CQ create flags */
1919         if (hw_rev == IRDMA_GEN_1)
1920                 return flags ? -EOPNOTSUPP : 0;
1921
1922         return flags & ~IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION ? -EOPNOTSUPP : 0;
1923 }
1924
1925 /**
1926  * irdma_create_cq - create cq
1927  * @ibcq: CQ allocated
1928  * @attr: attributes for cq
1929  * @udata: user data
1930  */
1931 static int irdma_create_cq(struct ib_cq *ibcq,
1932                            const struct ib_cq_init_attr *attr,
1933                            struct ib_udata *udata)
1934 {
1935         struct ib_device *ibdev = ibcq->device;
1936         struct irdma_device *iwdev = to_iwdev(ibdev);
1937         struct irdma_pci_f *rf = iwdev->rf;
1938         struct irdma_cq *iwcq = to_iwcq(ibcq);
1939         u32 cq_num = 0;
1940         struct irdma_sc_cq *cq;
1941         struct irdma_sc_dev *dev = &rf->sc_dev;
1942         struct irdma_cq_init_info info = {};
1943         struct irdma_cqp_request *cqp_request;
1944         struct cqp_cmds_info *cqp_info;
1945         struct irdma_cq_uk_init_info *ukinfo = &info.cq_uk_init_info;
1946         unsigned long flags;
1947         int err_code;
1948         int entries = attr->cqe;
1949
1950         err_code = cq_validate_flags(attr->flags, dev->hw_attrs.uk_attrs.hw_rev);
1951         if (err_code)
1952                 return err_code;
1953         err_code = irdma_alloc_rsrc(rf, rf->allocated_cqs, rf->max_cq, &cq_num,
1954                                     &rf->next_cq);
1955         if (err_code)
1956                 return err_code;
1957
1958         cq = &iwcq->sc_cq;
1959         cq->back_cq = iwcq;
1960         spin_lock_init(&iwcq->lock);
1961         INIT_LIST_HEAD(&iwcq->resize_list);
1962         info.dev = dev;
1963         ukinfo->cq_size = max(entries, 4);
1964         ukinfo->cq_id = cq_num;
1965         iwcq->ibcq.cqe = info.cq_uk_init_info.cq_size;
1966         if (attr->comp_vector < rf->ceqs_count)
1967                 info.ceq_id = attr->comp_vector;
1968         info.ceq_id_valid = true;
1969         info.ceqe_mask = 1;
1970         info.type = IRDMA_CQ_TYPE_IWARP;
1971         info.vsi = &iwdev->vsi;
1972
1973         if (udata) {
1974                 struct irdma_ucontext *ucontext;
1975                 struct irdma_create_cq_req req = {};
1976                 struct irdma_cq_mr *cqmr;
1977                 struct irdma_pbl *iwpbl;
1978                 struct irdma_pbl *iwpbl_shadow;
1979                 struct irdma_cq_mr *cqmr_shadow;
1980
1981                 iwcq->user_mode = true;
1982                 ucontext =
1983                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
1984                                                   ibucontext);
1985                 if (ib_copy_from_udata(&req, udata,
1986                                        min(sizeof(req), udata->inlen))) {
1987                         err_code = -EFAULT;
1988                         goto cq_free_rsrc;
1989                 }
1990
1991                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
1992                 iwpbl = irdma_get_pbl((unsigned long)req.user_cq_buf,
1993                                       &ucontext->cq_reg_mem_list);
1994                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
1995                 if (!iwpbl) {
1996                         err_code = -EPROTO;
1997                         goto cq_free_rsrc;
1998                 }
1999
2000                 iwcq->iwpbl = iwpbl;
2001                 iwcq->cq_mem_size = 0;
2002                 cqmr = &iwpbl->cq_mr;
2003
2004                 if (rf->sc_dev.hw_attrs.uk_attrs.feature_flags &
2005                     IRDMA_FEATURE_CQ_RESIZE && !ucontext->legacy_mode) {
2006                         spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2007                         iwpbl_shadow = irdma_get_pbl(
2008                                         (unsigned long)req.user_shadow_area,
2009                                         &ucontext->cq_reg_mem_list);
2010                         spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2011
2012                         if (!iwpbl_shadow) {
2013                                 err_code = -EPROTO;
2014                                 goto cq_free_rsrc;
2015                         }
2016                         iwcq->iwpbl_shadow = iwpbl_shadow;
2017                         cqmr_shadow = &iwpbl_shadow->cq_mr;
2018                         info.shadow_area_pa = cqmr_shadow->cq_pbl.addr;
2019                         cqmr->split = true;
2020                 } else {
2021                         info.shadow_area_pa = cqmr->shadow;
2022                 }
2023                 if (iwpbl->pbl_allocated) {
2024                         info.virtual_map = true;
2025                         info.pbl_chunk_size = 1;
2026                         info.first_pm_pbl_idx = cqmr->cq_pbl.idx;
2027                 } else {
2028                         info.cq_base_pa = cqmr->cq_pbl.addr;
2029                 }
2030         } else {
2031                 /* Kmode allocations */
2032                 int rsize;
2033
2034                 if (entries < 1 || entries > rf->max_cqe) {
2035                         err_code = -EINVAL;
2036                         goto cq_free_rsrc;
2037                 }
2038
2039                 entries++;
2040                 if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
2041                         entries *= 2;
2042                 ukinfo->cq_size = entries;
2043
2044                 rsize = info.cq_uk_init_info.cq_size * sizeof(struct irdma_cqe);
2045                 iwcq->kmem.size = ALIGN(round_up(rsize, 256), 256);
2046                 iwcq->kmem.va = dma_alloc_coherent(dev->hw->device,
2047                                                    iwcq->kmem.size,
2048                                                    &iwcq->kmem.pa, GFP_KERNEL);
2049                 if (!iwcq->kmem.va) {
2050                         err_code = -ENOMEM;
2051                         goto cq_free_rsrc;
2052                 }
2053
2054                 iwcq->kmem_shadow.size = ALIGN(IRDMA_SHADOW_AREA_SIZE << 3,
2055                                                64);
2056                 iwcq->kmem_shadow.va = dma_alloc_coherent(dev->hw->device,
2057                                                           iwcq->kmem_shadow.size,
2058                                                           &iwcq->kmem_shadow.pa,
2059                                                           GFP_KERNEL);
2060                 if (!iwcq->kmem_shadow.va) {
2061                         err_code = -ENOMEM;
2062                         goto cq_free_rsrc;
2063                 }
2064                 info.shadow_area_pa = iwcq->kmem_shadow.pa;
2065                 ukinfo->shadow_area = iwcq->kmem_shadow.va;
2066                 ukinfo->cq_base = iwcq->kmem.va;
2067                 info.cq_base_pa = iwcq->kmem.pa;
2068         }
2069
2070         if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
2071                 info.shadow_read_threshold = min(info.cq_uk_init_info.cq_size / 2,
2072                                                  (u32)IRDMA_MAX_CQ_READ_THRESH);
2073
2074         if (irdma_sc_cq_init(cq, &info)) {
2075                 ibdev_dbg(&iwdev->ibdev, "VERBS: init cq fail\n");
2076                 err_code = -EPROTO;
2077                 goto cq_free_rsrc;
2078         }
2079
2080         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
2081         if (!cqp_request) {
2082                 err_code = -ENOMEM;
2083                 goto cq_free_rsrc;
2084         }
2085
2086         cqp_info = &cqp_request->info;
2087         cqp_info->cqp_cmd = IRDMA_OP_CQ_CREATE;
2088         cqp_info->post_sq = 1;
2089         cqp_info->in.u.cq_create.cq = cq;
2090         cqp_info->in.u.cq_create.check_overflow = true;
2091         cqp_info->in.u.cq_create.scratch = (uintptr_t)cqp_request;
2092         err_code = irdma_handle_cqp_op(rf, cqp_request);
2093         irdma_put_cqp_request(&rf->cqp, cqp_request);
2094         if (err_code)
2095                 goto cq_free_rsrc;
2096
2097         if (udata) {
2098                 struct irdma_create_cq_resp resp = {};
2099
2100                 resp.cq_id = info.cq_uk_init_info.cq_id;
2101                 resp.cq_size = info.cq_uk_init_info.cq_size;
2102                 if (ib_copy_to_udata(udata, &resp,
2103                                      min(sizeof(resp), udata->outlen))) {
2104                         ibdev_dbg(&iwdev->ibdev,
2105                                   "VERBS: copy to user data\n");
2106                         err_code = -EPROTO;
2107                         goto cq_destroy;
2108                 }
2109         }
2110         return 0;
2111 cq_destroy:
2112         irdma_cq_wq_destroy(rf, cq);
2113 cq_free_rsrc:
2114         irdma_cq_free_rsrc(rf, iwcq);
2115
2116         return err_code;
2117 }
2118
2119 /**
2120  * irdma_get_mr_access - get hw MR access permissions from IB access flags
2121  * @access: IB access flags
2122  */
2123 static inline u16 irdma_get_mr_access(int access)
2124 {
2125         u16 hw_access = 0;
2126
2127         hw_access |= (access & IB_ACCESS_LOCAL_WRITE) ?
2128                      IRDMA_ACCESS_FLAGS_LOCALWRITE : 0;
2129         hw_access |= (access & IB_ACCESS_REMOTE_WRITE) ?
2130                      IRDMA_ACCESS_FLAGS_REMOTEWRITE : 0;
2131         hw_access |= (access & IB_ACCESS_REMOTE_READ) ?
2132                      IRDMA_ACCESS_FLAGS_REMOTEREAD : 0;
2133         hw_access |= (access & IB_ACCESS_MW_BIND) ?
2134                      IRDMA_ACCESS_FLAGS_BIND_WINDOW : 0;
2135         hw_access |= (access & IB_ZERO_BASED) ?
2136                      IRDMA_ACCESS_FLAGS_ZERO_BASED : 0;
2137         hw_access |= IRDMA_ACCESS_FLAGS_LOCALREAD;
2138
2139         return hw_access;
2140 }
2141
2142 /**
2143  * irdma_free_stag - free stag resource
2144  * @iwdev: irdma device
2145  * @stag: stag to free
2146  */
2147 static void irdma_free_stag(struct irdma_device *iwdev, u32 stag)
2148 {
2149         u32 stag_idx;
2150
2151         stag_idx = (stag & iwdev->rf->mr_stagmask) >> IRDMA_CQPSQ_STAG_IDX_S;
2152         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_mrs, stag_idx);
2153 }
2154
2155 /**
2156  * irdma_create_stag - create random stag
2157  * @iwdev: irdma device
2158  */
2159 static u32 irdma_create_stag(struct irdma_device *iwdev)
2160 {
2161         u32 stag = 0;
2162         u32 stag_index = 0;
2163         u32 next_stag_index;
2164         u32 driver_key;
2165         u32 random;
2166         u8 consumer_key;
2167         int ret;
2168
2169         get_random_bytes(&random, sizeof(random));
2170         consumer_key = (u8)random;
2171
2172         driver_key = random & ~iwdev->rf->mr_stagmask;
2173         next_stag_index = (random & iwdev->rf->mr_stagmask) >> 8;
2174         next_stag_index %= iwdev->rf->max_mr;
2175
2176         ret = irdma_alloc_rsrc(iwdev->rf, iwdev->rf->allocated_mrs,
2177                                iwdev->rf->max_mr, &stag_index,
2178                                &next_stag_index);
2179         if (ret)
2180                 return stag;
2181         stag = stag_index << IRDMA_CQPSQ_STAG_IDX_S;
2182         stag |= driver_key;
2183         stag += (u32)consumer_key;
2184
2185         return stag;
2186 }
2187
2188 /**
2189  * irdma_next_pbl_addr - Get next pbl address
2190  * @pbl: pointer to a pble
2191  * @pinfo: info pointer
2192  * @idx: index
2193  */
2194 static inline u64 *irdma_next_pbl_addr(u64 *pbl, struct irdma_pble_info **pinfo,
2195                                        u32 *idx)
2196 {
2197         *idx += 1;
2198         if (!(*pinfo) || *idx != (*pinfo)->cnt)
2199                 return ++pbl;
2200         *idx = 0;
2201         (*pinfo)++;
2202
2203         return (*pinfo)->addr;
2204 }
2205
2206 /**
2207  * irdma_copy_user_pgaddrs - copy user page address to pble's os locally
2208  * @iwmr: iwmr for IB's user page addresses
2209  * @pbl: ple pointer to save 1 level or 0 level pble
2210  * @level: indicated level 0, 1 or 2
2211  */
2212 static void irdma_copy_user_pgaddrs(struct irdma_mr *iwmr, u64 *pbl,
2213                                     enum irdma_pble_level level)
2214 {
2215         struct ib_umem *region = iwmr->region;
2216         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2217         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2218         struct irdma_pble_info *pinfo;
2219         struct ib_block_iter biter;
2220         u32 idx = 0;
2221         u32 pbl_cnt = 0;
2222
2223         pinfo = (level == PBLE_LEVEL_1) ? NULL : palloc->level2.leaf;
2224
2225         if (iwmr->type == IRDMA_MEMREG_TYPE_QP)
2226                 iwpbl->qp_mr.sq_page = sg_page(region->sgt_append.sgt.sgl);
2227
2228         rdma_umem_for_each_dma_block(region, &biter, iwmr->page_size) {
2229                 *pbl = rdma_block_iter_dma_address(&biter);
2230                 if (++pbl_cnt == palloc->total_cnt)
2231                         break;
2232                 pbl = irdma_next_pbl_addr(pbl, &pinfo, &idx);
2233         }
2234 }
2235
2236 /**
2237  * irdma_check_mem_contiguous - check if pbls stored in arr are contiguous
2238  * @arr: lvl1 pbl array
2239  * @npages: page count
2240  * @pg_size: page size
2241  *
2242  */
2243 static bool irdma_check_mem_contiguous(u64 *arr, u32 npages, u32 pg_size)
2244 {
2245         u32 pg_idx;
2246
2247         for (pg_idx = 0; pg_idx < npages; pg_idx++) {
2248                 if ((*arr + (pg_size * pg_idx)) != arr[pg_idx])
2249                         return false;
2250         }
2251
2252         return true;
2253 }
2254
2255 /**
2256  * irdma_check_mr_contiguous - check if MR is physically contiguous
2257  * @palloc: pbl allocation struct
2258  * @pg_size: page size
2259  */
2260 static bool irdma_check_mr_contiguous(struct irdma_pble_alloc *palloc,
2261                                       u32 pg_size)
2262 {
2263         struct irdma_pble_level2 *lvl2 = &palloc->level2;
2264         struct irdma_pble_info *leaf = lvl2->leaf;
2265         u64 *arr = NULL;
2266         u64 *start_addr = NULL;
2267         int i;
2268         bool ret;
2269
2270         if (palloc->level == PBLE_LEVEL_1) {
2271                 arr = palloc->level1.addr;
2272                 ret = irdma_check_mem_contiguous(arr, palloc->total_cnt,
2273                                                  pg_size);
2274                 return ret;
2275         }
2276
2277         start_addr = leaf->addr;
2278
2279         for (i = 0; i < lvl2->leaf_cnt; i++, leaf++) {
2280                 arr = leaf->addr;
2281                 if ((*start_addr + (i * pg_size * PBLE_PER_PAGE)) != *arr)
2282                         return false;
2283                 ret = irdma_check_mem_contiguous(arr, leaf->cnt, pg_size);
2284                 if (!ret)
2285                         return false;
2286         }
2287
2288         return true;
2289 }
2290
2291 /**
2292  * irdma_setup_pbles - copy user pg address to pble's
2293  * @rf: RDMA PCI function
2294  * @iwmr: mr pointer for this memory registration
2295  * @use_pbles: flag if to use pble's
2296  */
2297 static int irdma_setup_pbles(struct irdma_pci_f *rf, struct irdma_mr *iwmr,
2298                              bool use_pbles)
2299 {
2300         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2301         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2302         struct irdma_pble_info *pinfo;
2303         u64 *pbl;
2304         int status;
2305         enum irdma_pble_level level = PBLE_LEVEL_1;
2306
2307         if (use_pbles) {
2308                 status = irdma_get_pble(rf->pble_rsrc, palloc, iwmr->page_cnt,
2309                                         false);
2310                 if (status)
2311                         return status;
2312
2313                 iwpbl->pbl_allocated = true;
2314                 level = palloc->level;
2315                 pinfo = (level == PBLE_LEVEL_1) ? &palloc->level1 :
2316                                                   palloc->level2.leaf;
2317                 pbl = pinfo->addr;
2318         } else {
2319                 pbl = iwmr->pgaddrmem;
2320         }
2321
2322         irdma_copy_user_pgaddrs(iwmr, pbl, level);
2323
2324         if (use_pbles)
2325                 iwmr->pgaddrmem[0] = *pbl;
2326
2327         return 0;
2328 }
2329
2330 /**
2331  * irdma_handle_q_mem - handle memory for qp and cq
2332  * @iwdev: irdma device
2333  * @req: information for q memory management
2334  * @iwpbl: pble struct
2335  * @use_pbles: flag to use pble
2336  */
2337 static int irdma_handle_q_mem(struct irdma_device *iwdev,
2338                               struct irdma_mem_reg_req *req,
2339                               struct irdma_pbl *iwpbl, bool use_pbles)
2340 {
2341         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2342         struct irdma_mr *iwmr = iwpbl->iwmr;
2343         struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;
2344         struct irdma_cq_mr *cqmr = &iwpbl->cq_mr;
2345         struct irdma_hmc_pble *hmc_p;
2346         u64 *arr = iwmr->pgaddrmem;
2347         u32 pg_size, total;
2348         int err = 0;
2349         bool ret = true;
2350
2351         pg_size = iwmr->page_size;
2352         err = irdma_setup_pbles(iwdev->rf, iwmr, use_pbles);
2353         if (err)
2354                 return err;
2355
2356         if (use_pbles && palloc->level != PBLE_LEVEL_1) {
2357                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2358                 iwpbl->pbl_allocated = false;
2359                 return -ENOMEM;
2360         }
2361
2362         if (use_pbles)
2363                 arr = palloc->level1.addr;
2364
2365         switch (iwmr->type) {
2366         case IRDMA_MEMREG_TYPE_QP:
2367                 total = req->sq_pages + req->rq_pages;
2368                 hmc_p = &qpmr->sq_pbl;
2369                 qpmr->shadow = (dma_addr_t)arr[total];
2370
2371                 if (use_pbles) {
2372                         ret = irdma_check_mem_contiguous(arr, req->sq_pages,
2373                                                          pg_size);
2374                         if (ret)
2375                                 ret = irdma_check_mem_contiguous(&arr[req->sq_pages],
2376                                                                  req->rq_pages,
2377                                                                  pg_size);
2378                 }
2379
2380                 if (!ret) {
2381                         hmc_p->idx = palloc->level1.idx;
2382                         hmc_p = &qpmr->rq_pbl;
2383                         hmc_p->idx = palloc->level1.idx + req->sq_pages;
2384                 } else {
2385                         hmc_p->addr = arr[0];
2386                         hmc_p = &qpmr->rq_pbl;
2387                         hmc_p->addr = arr[req->sq_pages];
2388                 }
2389                 break;
2390         case IRDMA_MEMREG_TYPE_CQ:
2391                 hmc_p = &cqmr->cq_pbl;
2392
2393                 if (!cqmr->split)
2394                         cqmr->shadow = (dma_addr_t)arr[req->cq_pages];
2395
2396                 if (use_pbles)
2397                         ret = irdma_check_mem_contiguous(arr, req->cq_pages,
2398                                                          pg_size);
2399
2400                 if (!ret)
2401                         hmc_p->idx = palloc->level1.idx;
2402                 else
2403                         hmc_p->addr = arr[0];
2404         break;
2405         default:
2406                 ibdev_dbg(&iwdev->ibdev, "VERBS: MR type error\n");
2407                 err = -EINVAL;
2408         }
2409
2410         if (use_pbles && ret) {
2411                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2412                 iwpbl->pbl_allocated = false;
2413         }
2414
2415         return err;
2416 }
2417
2418 /**
2419  * irdma_hw_alloc_mw - create the hw memory window
2420  * @iwdev: irdma device
2421  * @iwmr: pointer to memory window info
2422  */
2423 static int irdma_hw_alloc_mw(struct irdma_device *iwdev, struct irdma_mr *iwmr)
2424 {
2425         struct irdma_mw_alloc_info *info;
2426         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2427         struct irdma_cqp_request *cqp_request;
2428         struct cqp_cmds_info *cqp_info;
2429         int status;
2430
2431         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2432         if (!cqp_request)
2433                 return -ENOMEM;
2434
2435         cqp_info = &cqp_request->info;
2436         info = &cqp_info->in.u.mw_alloc.info;
2437         memset(info, 0, sizeof(*info));
2438         if (iwmr->ibmw.type == IB_MW_TYPE_1)
2439                 info->mw_wide = true;
2440
2441         info->page_size = PAGE_SIZE;
2442         info->mw_stag_index = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2443         info->pd_id = iwpd->sc_pd.pd_id;
2444         info->remote_access = true;
2445         cqp_info->cqp_cmd = IRDMA_OP_MW_ALLOC;
2446         cqp_info->post_sq = 1;
2447         cqp_info->in.u.mw_alloc.dev = &iwdev->rf->sc_dev;
2448         cqp_info->in.u.mw_alloc.scratch = (uintptr_t)cqp_request;
2449         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2450         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2451
2452         return status;
2453 }
2454
2455 /**
2456  * irdma_alloc_mw - Allocate memory window
2457  * @ibmw: Memory Window
2458  * @udata: user data pointer
2459  */
2460 static int irdma_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata)
2461 {
2462         struct irdma_device *iwdev = to_iwdev(ibmw->device);
2463         struct irdma_mr *iwmr = to_iwmw(ibmw);
2464         int err_code;
2465         u32 stag;
2466
2467         stag = irdma_create_stag(iwdev);
2468         if (!stag)
2469                 return -ENOMEM;
2470
2471         iwmr->stag = stag;
2472         ibmw->rkey = stag;
2473
2474         err_code = irdma_hw_alloc_mw(iwdev, iwmr);
2475         if (err_code) {
2476                 irdma_free_stag(iwdev, stag);
2477                 return err_code;
2478         }
2479
2480         return 0;
2481 }
2482
2483 /**
2484  * irdma_dealloc_mw - Dealloc memory window
2485  * @ibmw: memory window structure.
2486  */
2487 static int irdma_dealloc_mw(struct ib_mw *ibmw)
2488 {
2489         struct ib_pd *ibpd = ibmw->pd;
2490         struct irdma_pd *iwpd = to_iwpd(ibpd);
2491         struct irdma_mr *iwmr = to_iwmr((struct ib_mr *)ibmw);
2492         struct irdma_device *iwdev = to_iwdev(ibmw->device);
2493         struct irdma_cqp_request *cqp_request;
2494         struct cqp_cmds_info *cqp_info;
2495         struct irdma_dealloc_stag_info *info;
2496
2497         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2498         if (!cqp_request)
2499                 return -ENOMEM;
2500
2501         cqp_info = &cqp_request->info;
2502         info = &cqp_info->in.u.dealloc_stag.info;
2503         memset(info, 0, sizeof(*info));
2504         info->pd_id = iwpd->sc_pd.pd_id;
2505         info->stag_idx = ibmw->rkey >> IRDMA_CQPSQ_STAG_IDX_S;
2506         info->mr = false;
2507         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
2508         cqp_info->post_sq = 1;
2509         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
2510         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
2511         irdma_handle_cqp_op(iwdev->rf, cqp_request);
2512         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2513         irdma_free_stag(iwdev, iwmr->stag);
2514
2515         return 0;
2516 }
2517
2518 /**
2519  * irdma_hw_alloc_stag - cqp command to allocate stag
2520  * @iwdev: irdma device
2521  * @iwmr: irdma mr pointer
2522  */
2523 static int irdma_hw_alloc_stag(struct irdma_device *iwdev,
2524                                struct irdma_mr *iwmr)
2525 {
2526         struct irdma_allocate_stag_info *info;
2527         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2528         int status;
2529         struct irdma_cqp_request *cqp_request;
2530         struct cqp_cmds_info *cqp_info;
2531
2532         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2533         if (!cqp_request)
2534                 return -ENOMEM;
2535
2536         cqp_info = &cqp_request->info;
2537         info = &cqp_info->in.u.alloc_stag.info;
2538         memset(info, 0, sizeof(*info));
2539         info->page_size = PAGE_SIZE;
2540         info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2541         info->pd_id = iwpd->sc_pd.pd_id;
2542         info->total_len = iwmr->len;
2543         info->remote_access = true;
2544         cqp_info->cqp_cmd = IRDMA_OP_ALLOC_STAG;
2545         cqp_info->post_sq = 1;
2546         cqp_info->in.u.alloc_stag.dev = &iwdev->rf->sc_dev;
2547         cqp_info->in.u.alloc_stag.scratch = (uintptr_t)cqp_request;
2548         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2549         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2550
2551         return status;
2552 }
2553
2554 /**
2555  * irdma_alloc_mr - register stag for fast memory registration
2556  * @pd: ibpd pointer
2557  * @mr_type: memory for stag registrion
2558  * @max_num_sg: man number of pages
2559  */
2560 static struct ib_mr *irdma_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
2561                                     u32 max_num_sg)
2562 {
2563         struct irdma_device *iwdev = to_iwdev(pd->device);
2564         struct irdma_pble_alloc *palloc;
2565         struct irdma_pbl *iwpbl;
2566         struct irdma_mr *iwmr;
2567         u32 stag;
2568         int err_code;
2569
2570         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2571         if (!iwmr)
2572                 return ERR_PTR(-ENOMEM);
2573
2574         stag = irdma_create_stag(iwdev);
2575         if (!stag) {
2576                 err_code = -ENOMEM;
2577                 goto err;
2578         }
2579
2580         iwmr->stag = stag;
2581         iwmr->ibmr.rkey = stag;
2582         iwmr->ibmr.lkey = stag;
2583         iwmr->ibmr.pd = pd;
2584         iwmr->ibmr.device = pd->device;
2585         iwpbl = &iwmr->iwpbl;
2586         iwpbl->iwmr = iwmr;
2587         iwmr->type = IRDMA_MEMREG_TYPE_MEM;
2588         palloc = &iwpbl->pble_alloc;
2589         iwmr->page_cnt = max_num_sg;
2590         err_code = irdma_get_pble(iwdev->rf->pble_rsrc, palloc, iwmr->page_cnt,
2591                                   true);
2592         if (err_code)
2593                 goto err_get_pble;
2594
2595         err_code = irdma_hw_alloc_stag(iwdev, iwmr);
2596         if (err_code)
2597                 goto err_alloc_stag;
2598
2599         iwpbl->pbl_allocated = true;
2600
2601         return &iwmr->ibmr;
2602 err_alloc_stag:
2603         irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2604 err_get_pble:
2605         irdma_free_stag(iwdev, stag);
2606 err:
2607         kfree(iwmr);
2608
2609         return ERR_PTR(err_code);
2610 }
2611
2612 /**
2613  * irdma_set_page - populate pbl list for fmr
2614  * @ibmr: ib mem to access iwarp mr pointer
2615  * @addr: page dma address fro pbl list
2616  */
2617 static int irdma_set_page(struct ib_mr *ibmr, u64 addr)
2618 {
2619         struct irdma_mr *iwmr = to_iwmr(ibmr);
2620         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2621         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2622         u64 *pbl;
2623
2624         if (unlikely(iwmr->npages == iwmr->page_cnt))
2625                 return -ENOMEM;
2626
2627         pbl = palloc->level1.addr;
2628         pbl[iwmr->npages++] = addr;
2629
2630         return 0;
2631 }
2632
2633 /**
2634  * irdma_map_mr_sg - map of sg list for fmr
2635  * @ibmr: ib mem to access iwarp mr pointer
2636  * @sg: scatter gather list
2637  * @sg_nents: number of sg pages
2638  * @sg_offset: scatter gather list for fmr
2639  */
2640 static int irdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
2641                            int sg_nents, unsigned int *sg_offset)
2642 {
2643         struct irdma_mr *iwmr = to_iwmr(ibmr);
2644
2645         iwmr->npages = 0;
2646
2647         return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, irdma_set_page);
2648 }
2649
2650 /**
2651  * irdma_hwreg_mr - send cqp command for memory registration
2652  * @iwdev: irdma device
2653  * @iwmr: irdma mr pointer
2654  * @access: access for MR
2655  */
2656 static int irdma_hwreg_mr(struct irdma_device *iwdev, struct irdma_mr *iwmr,
2657                           u16 access)
2658 {
2659         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2660         struct irdma_reg_ns_stag_info *stag_info;
2661         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2662         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2663         struct irdma_cqp_request *cqp_request;
2664         struct cqp_cmds_info *cqp_info;
2665         int ret;
2666
2667         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2668         if (!cqp_request)
2669                 return -ENOMEM;
2670
2671         cqp_info = &cqp_request->info;
2672         stag_info = &cqp_info->in.u.mr_reg_non_shared.info;
2673         memset(stag_info, 0, sizeof(*stag_info));
2674         stag_info->va = iwpbl->user_base;
2675         stag_info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2676         stag_info->stag_key = (u8)iwmr->stag;
2677         stag_info->total_len = iwmr->len;
2678         stag_info->access_rights = irdma_get_mr_access(access);
2679         stag_info->pd_id = iwpd->sc_pd.pd_id;
2680         if (stag_info->access_rights & IRDMA_ACCESS_FLAGS_ZERO_BASED)
2681                 stag_info->addr_type = IRDMA_ADDR_TYPE_ZERO_BASED;
2682         else
2683                 stag_info->addr_type = IRDMA_ADDR_TYPE_VA_BASED;
2684         stag_info->page_size = iwmr->page_size;
2685
2686         if (iwpbl->pbl_allocated) {
2687                 if (palloc->level == PBLE_LEVEL_1) {
2688                         stag_info->first_pm_pbl_index = palloc->level1.idx;
2689                         stag_info->chunk_size = 1;
2690                 } else {
2691                         stag_info->first_pm_pbl_index = palloc->level2.root.idx;
2692                         stag_info->chunk_size = 3;
2693                 }
2694         } else {
2695                 stag_info->reg_addr_pa = iwmr->pgaddrmem[0];
2696         }
2697
2698         cqp_info->cqp_cmd = IRDMA_OP_MR_REG_NON_SHARED;
2699         cqp_info->post_sq = 1;
2700         cqp_info->in.u.mr_reg_non_shared.dev = &iwdev->rf->sc_dev;
2701         cqp_info->in.u.mr_reg_non_shared.scratch = (uintptr_t)cqp_request;
2702         ret = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2703         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2704
2705         return ret;
2706 }
2707
2708 /**
2709  * irdma_reg_user_mr - Register a user memory region
2710  * @pd: ptr of pd
2711  * @start: virtual start address
2712  * @len: length of mr
2713  * @virt: virtual address
2714  * @access: access of mr
2715  * @udata: user data
2716  */
2717 static struct ib_mr *irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,
2718                                        u64 virt, int access,
2719                                        struct ib_udata *udata)
2720 {
2721         struct irdma_device *iwdev = to_iwdev(pd->device);
2722         struct irdma_ucontext *ucontext;
2723         struct irdma_pble_alloc *palloc;
2724         struct irdma_pbl *iwpbl;
2725         struct irdma_mr *iwmr;
2726         struct ib_umem *region;
2727         struct irdma_mem_reg_req req;
2728         u32 total, stag = 0;
2729         u8 shadow_pgcnt = 1;
2730         bool use_pbles = false;
2731         unsigned long flags;
2732         int err = -EINVAL;
2733         int ret;
2734
2735         if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size)
2736                 return ERR_PTR(-EINVAL);
2737
2738         region = ib_umem_get(pd->device, start, len, access);
2739
2740         if (IS_ERR(region)) {
2741                 ibdev_dbg(&iwdev->ibdev,
2742                           "VERBS: Failed to create ib_umem region\n");
2743                 return (struct ib_mr *)region;
2744         }
2745
2746         if (ib_copy_from_udata(&req, udata, min(sizeof(req), udata->inlen))) {
2747                 ib_umem_release(region);
2748                 return ERR_PTR(-EFAULT);
2749         }
2750
2751         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2752         if (!iwmr) {
2753                 ib_umem_release(region);
2754                 return ERR_PTR(-ENOMEM);
2755         }
2756
2757         iwpbl = &iwmr->iwpbl;
2758         iwpbl->iwmr = iwmr;
2759         iwmr->region = region;
2760         iwmr->ibmr.pd = pd;
2761         iwmr->ibmr.device = pd->device;
2762         iwmr->ibmr.iova = virt;
2763         iwmr->page_size = PAGE_SIZE;
2764
2765         if (req.reg_type == IRDMA_MEMREG_TYPE_MEM) {
2766                 iwmr->page_size = ib_umem_find_best_pgsz(region,
2767                                                          SZ_4K | SZ_2M | SZ_1G,
2768                                                          virt);
2769                 if (unlikely(!iwmr->page_size)) {
2770                         kfree(iwmr);
2771                         ib_umem_release(region);
2772                         return ERR_PTR(-EOPNOTSUPP);
2773                 }
2774         }
2775         iwmr->len = region->length;
2776         iwpbl->user_base = virt;
2777         palloc = &iwpbl->pble_alloc;
2778         iwmr->type = req.reg_type;
2779         iwmr->page_cnt = ib_umem_num_dma_blocks(region, iwmr->page_size);
2780
2781         switch (req.reg_type) {
2782         case IRDMA_MEMREG_TYPE_QP:
2783                 total = req.sq_pages + req.rq_pages + shadow_pgcnt;
2784                 if (total > iwmr->page_cnt) {
2785                         err = -EINVAL;
2786                         goto error;
2787                 }
2788                 total = req.sq_pages + req.rq_pages;
2789                 use_pbles = (total > 2);
2790                 err = irdma_handle_q_mem(iwdev, &req, iwpbl, use_pbles);
2791                 if (err)
2792                         goto error;
2793
2794                 ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext,
2795                                                      ibucontext);
2796                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
2797                 list_add_tail(&iwpbl->list, &ucontext->qp_reg_mem_list);
2798                 iwpbl->on_list = true;
2799                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
2800                 break;
2801         case IRDMA_MEMREG_TYPE_CQ:
2802                 if (iwdev->rf->sc_dev.hw_attrs.uk_attrs.feature_flags & IRDMA_FEATURE_CQ_RESIZE)
2803                         shadow_pgcnt = 0;
2804                 total = req.cq_pages + shadow_pgcnt;
2805                 if (total > iwmr->page_cnt) {
2806                         err = -EINVAL;
2807                         goto error;
2808                 }
2809
2810                 use_pbles = (req.cq_pages > 1);
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->cq_reg_mem_list_lock, flags);
2818                 list_add_tail(&iwpbl->list, &ucontext->cq_reg_mem_list);
2819                 iwpbl->on_list = true;
2820                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2821                 break;
2822         case IRDMA_MEMREG_TYPE_MEM:
2823                 use_pbles = (iwmr->page_cnt != 1);
2824
2825                 err = irdma_setup_pbles(iwdev->rf, iwmr, use_pbles);
2826                 if (err)
2827                         goto error;
2828
2829                 if (use_pbles) {
2830                         ret = irdma_check_mr_contiguous(palloc,
2831                                                         iwmr->page_size);
2832                         if (ret) {
2833                                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2834                                 iwpbl->pbl_allocated = false;
2835                         }
2836                 }
2837
2838                 stag = irdma_create_stag(iwdev);
2839                 if (!stag) {
2840                         err = -ENOMEM;
2841                         goto error;
2842                 }
2843
2844                 iwmr->stag = stag;
2845                 iwmr->ibmr.rkey = stag;
2846                 iwmr->ibmr.lkey = stag;
2847                 err = irdma_hwreg_mr(iwdev, iwmr, access);
2848                 if (err) {
2849                         irdma_free_stag(iwdev, stag);
2850                         goto error;
2851                 }
2852
2853                 break;
2854         default:
2855                 goto error;
2856         }
2857
2858         iwmr->type = req.reg_type;
2859
2860         return &iwmr->ibmr;
2861
2862 error:
2863         if (palloc->level != PBLE_LEVEL_0 && iwpbl->pbl_allocated)
2864                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2865         ib_umem_release(region);
2866         kfree(iwmr);
2867
2868         return ERR_PTR(err);
2869 }
2870
2871 /**
2872  * irdma_reg_phys_mr - register kernel physical memory
2873  * @pd: ibpd pointer
2874  * @addr: physical address of memory to register
2875  * @size: size of memory to register
2876  * @access: Access rights
2877  * @iova_start: start of virtual address for physical buffers
2878  */
2879 struct ib_mr *irdma_reg_phys_mr(struct ib_pd *pd, u64 addr, u64 size, int access,
2880                                 u64 *iova_start)
2881 {
2882         struct irdma_device *iwdev = to_iwdev(pd->device);
2883         struct irdma_pbl *iwpbl;
2884         struct irdma_mr *iwmr;
2885         u32 stag;
2886         int ret;
2887
2888         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2889         if (!iwmr)
2890                 return ERR_PTR(-ENOMEM);
2891
2892         iwmr->ibmr.pd = pd;
2893         iwmr->ibmr.device = pd->device;
2894         iwpbl = &iwmr->iwpbl;
2895         iwpbl->iwmr = iwmr;
2896         iwmr->type = IRDMA_MEMREG_TYPE_MEM;
2897         iwpbl->user_base = *iova_start;
2898         stag = irdma_create_stag(iwdev);
2899         if (!stag) {
2900                 ret = -ENOMEM;
2901                 goto err;
2902         }
2903
2904         iwmr->stag = stag;
2905         iwmr->ibmr.iova = *iova_start;
2906         iwmr->ibmr.rkey = stag;
2907         iwmr->ibmr.lkey = stag;
2908         iwmr->page_cnt = 1;
2909         iwmr->pgaddrmem[0] = addr;
2910         iwmr->len = size;
2911         iwmr->page_size = SZ_4K;
2912         ret = irdma_hwreg_mr(iwdev, iwmr, access);
2913         if (ret) {
2914                 irdma_free_stag(iwdev, stag);
2915                 goto err;
2916         }
2917
2918         return &iwmr->ibmr;
2919
2920 err:
2921         kfree(iwmr);
2922
2923         return ERR_PTR(ret);
2924 }
2925
2926 /**
2927  * irdma_get_dma_mr - register physical mem
2928  * @pd: ptr of pd
2929  * @acc: access for memory
2930  */
2931 static struct ib_mr *irdma_get_dma_mr(struct ib_pd *pd, int acc)
2932 {
2933         u64 kva = 0;
2934
2935         return irdma_reg_phys_mr(pd, 0, 0, acc, &kva);
2936 }
2937
2938 /**
2939  * irdma_del_memlist - Deleting pbl list entries for CQ/QP
2940  * @iwmr: iwmr for IB's user page addresses
2941  * @ucontext: ptr to user context
2942  */
2943 static void irdma_del_memlist(struct irdma_mr *iwmr,
2944                               struct irdma_ucontext *ucontext)
2945 {
2946         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2947         unsigned long flags;
2948
2949         switch (iwmr->type) {
2950         case IRDMA_MEMREG_TYPE_CQ:
2951                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2952                 if (iwpbl->on_list) {
2953                         iwpbl->on_list = false;
2954                         list_del(&iwpbl->list);
2955                 }
2956                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2957                 break;
2958         case IRDMA_MEMREG_TYPE_QP:
2959                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
2960                 if (iwpbl->on_list) {
2961                         iwpbl->on_list = false;
2962                         list_del(&iwpbl->list);
2963                 }
2964                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
2965                 break;
2966         default:
2967                 break;
2968         }
2969 }
2970
2971 /**
2972  * irdma_dereg_mr - deregister mr
2973  * @ib_mr: mr ptr for dereg
2974  * @udata: user data
2975  */
2976 static int irdma_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
2977 {
2978         struct ib_pd *ibpd = ib_mr->pd;
2979         struct irdma_pd *iwpd = to_iwpd(ibpd);
2980         struct irdma_mr *iwmr = to_iwmr(ib_mr);
2981         struct irdma_device *iwdev = to_iwdev(ib_mr->device);
2982         struct irdma_dealloc_stag_info *info;
2983         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2984         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2985         struct irdma_cqp_request *cqp_request;
2986         struct cqp_cmds_info *cqp_info;
2987
2988         if (iwmr->type != IRDMA_MEMREG_TYPE_MEM) {
2989                 if (iwmr->region) {
2990                         struct irdma_ucontext *ucontext;
2991
2992                         ucontext = rdma_udata_to_drv_context(udata,
2993                                                 struct irdma_ucontext,
2994                                                 ibucontext);
2995                         irdma_del_memlist(iwmr, ucontext);
2996                 }
2997                 goto done;
2998         }
2999
3000         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
3001         if (!cqp_request)
3002                 return -ENOMEM;
3003
3004         cqp_info = &cqp_request->info;
3005         info = &cqp_info->in.u.dealloc_stag.info;
3006         memset(info, 0, sizeof(*info));
3007         info->pd_id = iwpd->sc_pd.pd_id;
3008         info->stag_idx = ib_mr->rkey >> IRDMA_CQPSQ_STAG_IDX_S;
3009         info->mr = true;
3010         if (iwpbl->pbl_allocated)
3011                 info->dealloc_pbl = true;
3012
3013         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
3014         cqp_info->post_sq = 1;
3015         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
3016         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
3017         irdma_handle_cqp_op(iwdev->rf, cqp_request);
3018         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
3019         irdma_free_stag(iwdev, iwmr->stag);
3020 done:
3021         if (iwpbl->pbl_allocated)
3022                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
3023         ib_umem_release(iwmr->region);
3024         kfree(iwmr);
3025
3026         return 0;
3027 }
3028
3029 /**
3030  * irdma_post_send -  kernel application wr
3031  * @ibqp: qp ptr for wr
3032  * @ib_wr: work request ptr
3033  * @bad_wr: return of bad wr if err
3034  */
3035 static int irdma_post_send(struct ib_qp *ibqp,
3036                            const struct ib_send_wr *ib_wr,
3037                            const struct ib_send_wr **bad_wr)
3038 {
3039         struct irdma_qp *iwqp;
3040         struct irdma_qp_uk *ukqp;
3041         struct irdma_sc_dev *dev;
3042         struct irdma_post_sq_info info;
3043         int err = 0;
3044         unsigned long flags;
3045         bool inv_stag;
3046         struct irdma_ah *ah;
3047         bool reflush = false;
3048
3049         iwqp = to_iwqp(ibqp);
3050         ukqp = &iwqp->sc_qp.qp_uk;
3051         dev = &iwqp->iwdev->rf->sc_dev;
3052
3053         spin_lock_irqsave(&iwqp->lock, flags);
3054         if (iwqp->flush_issued && ukqp->sq_flush_complete)
3055                 reflush = true;
3056         while (ib_wr) {
3057                 memset(&info, 0, sizeof(info));
3058                 inv_stag = false;
3059                 info.wr_id = (ib_wr->wr_id);
3060                 if ((ib_wr->send_flags & IB_SEND_SIGNALED) || iwqp->sig_all)
3061                         info.signaled = true;
3062                 if (ib_wr->send_flags & IB_SEND_FENCE)
3063                         info.read_fence = true;
3064                 switch (ib_wr->opcode) {
3065                 case IB_WR_SEND_WITH_IMM:
3066                         if (ukqp->qp_caps & IRDMA_SEND_WITH_IMM) {
3067                                 info.imm_data_valid = true;
3068                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
3069                         } else {
3070                                 err = -EINVAL;
3071                                 break;
3072                         }
3073                         fallthrough;
3074                 case IB_WR_SEND:
3075                 case IB_WR_SEND_WITH_INV:
3076                         if (ib_wr->opcode == IB_WR_SEND ||
3077                             ib_wr->opcode == IB_WR_SEND_WITH_IMM) {
3078                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
3079                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL;
3080                                 else
3081                                         info.op_type = IRDMA_OP_TYPE_SEND;
3082                         } else {
3083                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
3084                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL_INV;
3085                                 else
3086                                         info.op_type = IRDMA_OP_TYPE_SEND_INV;
3087                                 info.stag_to_inv = ib_wr->ex.invalidate_rkey;
3088                         }
3089
3090                         if (ib_wr->send_flags & IB_SEND_INLINE) {
3091                                 info.op.inline_send.data = (void *)(unsigned long)
3092                                                            ib_wr->sg_list[0].addr;
3093                                 info.op.inline_send.len = ib_wr->sg_list[0].length;
3094                                 if (iwqp->ibqp.qp_type == IB_QPT_UD ||
3095                                     iwqp->ibqp.qp_type == IB_QPT_GSI) {
3096                                         ah = to_iwah(ud_wr(ib_wr)->ah);
3097                                         info.op.inline_send.ah_id = ah->sc_ah.ah_info.ah_idx;
3098                                         info.op.inline_send.qkey = ud_wr(ib_wr)->remote_qkey;
3099                                         info.op.inline_send.dest_qp = ud_wr(ib_wr)->remote_qpn;
3100                                 }
3101                                 err = irdma_uk_inline_send(ukqp, &info, false);
3102                         } else {
3103                                 info.op.send.num_sges = ib_wr->num_sge;
3104                                 info.op.send.sg_list = ib_wr->sg_list;
3105                                 if (iwqp->ibqp.qp_type == IB_QPT_UD ||
3106                                     iwqp->ibqp.qp_type == IB_QPT_GSI) {
3107                                         ah = to_iwah(ud_wr(ib_wr)->ah);
3108                                         info.op.send.ah_id = ah->sc_ah.ah_info.ah_idx;
3109                                         info.op.send.qkey = ud_wr(ib_wr)->remote_qkey;
3110                                         info.op.send.dest_qp = ud_wr(ib_wr)->remote_qpn;
3111                                 }
3112                                 err = irdma_uk_send(ukqp, &info, false);
3113                         }
3114                         break;
3115                 case IB_WR_RDMA_WRITE_WITH_IMM:
3116                         if (ukqp->qp_caps & IRDMA_WRITE_WITH_IMM) {
3117                                 info.imm_data_valid = true;
3118                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
3119                         } else {
3120                                 err = -EINVAL;
3121                                 break;
3122                         }
3123                         fallthrough;
3124                 case IB_WR_RDMA_WRITE:
3125                         if (ib_wr->send_flags & IB_SEND_SOLICITED)
3126                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE_SOL;
3127                         else
3128                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE;
3129
3130                         if (ib_wr->send_flags & IB_SEND_INLINE) {
3131                                 info.op.inline_rdma_write.data = (void *)(uintptr_t)ib_wr->sg_list[0].addr;
3132                                 info.op.inline_rdma_write.len =
3133                                                 ib_wr->sg_list[0].length;
3134                                 info.op.inline_rdma_write.rem_addr.addr =
3135                                                 rdma_wr(ib_wr)->remote_addr;
3136                                 info.op.inline_rdma_write.rem_addr.lkey =
3137                                                 rdma_wr(ib_wr)->rkey;
3138                                 err = irdma_uk_inline_rdma_write(ukqp, &info, false);
3139                         } else {
3140                                 info.op.rdma_write.lo_sg_list = (void *)ib_wr->sg_list;
3141                                 info.op.rdma_write.num_lo_sges = ib_wr->num_sge;
3142                                 info.op.rdma_write.rem_addr.addr = rdma_wr(ib_wr)->remote_addr;
3143                                 info.op.rdma_write.rem_addr.lkey = rdma_wr(ib_wr)->rkey;
3144                                 err = irdma_uk_rdma_write(ukqp, &info, false);
3145                         }
3146                         break;
3147                 case IB_WR_RDMA_READ_WITH_INV:
3148                         inv_stag = true;
3149                         fallthrough;
3150                 case IB_WR_RDMA_READ:
3151                         if (ib_wr->num_sge >
3152                             dev->hw_attrs.uk_attrs.max_hw_read_sges) {
3153                                 err = -EINVAL;
3154                                 break;
3155                         }
3156                         info.op_type = IRDMA_OP_TYPE_RDMA_READ;
3157                         info.op.rdma_read.rem_addr.addr = rdma_wr(ib_wr)->remote_addr;
3158                         info.op.rdma_read.rem_addr.lkey = rdma_wr(ib_wr)->rkey;
3159                         info.op.rdma_read.lo_sg_list = (void *)ib_wr->sg_list;
3160                         info.op.rdma_read.num_lo_sges = ib_wr->num_sge;
3161                         err = irdma_uk_rdma_read(ukqp, &info, inv_stag, false);
3162                         break;
3163                 case IB_WR_LOCAL_INV:
3164                         info.op_type = IRDMA_OP_TYPE_INV_STAG;
3165                         info.op.inv_local_stag.target_stag = ib_wr->ex.invalidate_rkey;
3166                         err = irdma_uk_stag_local_invalidate(ukqp, &info, true);
3167                         break;
3168                 case IB_WR_REG_MR: {
3169                         struct irdma_mr *iwmr = to_iwmr(reg_wr(ib_wr)->mr);
3170                         struct irdma_pble_alloc *palloc = &iwmr->iwpbl.pble_alloc;
3171                         struct irdma_fast_reg_stag_info stag_info = {};
3172
3173                         stag_info.signaled = info.signaled;
3174                         stag_info.read_fence = info.read_fence;
3175                         stag_info.access_rights = irdma_get_mr_access(reg_wr(ib_wr)->access);
3176                         stag_info.stag_key = reg_wr(ib_wr)->key & 0xff;
3177                         stag_info.stag_idx = reg_wr(ib_wr)->key >> 8;
3178                         stag_info.page_size = reg_wr(ib_wr)->mr->page_size;
3179                         stag_info.wr_id = ib_wr->wr_id;
3180                         stag_info.addr_type = IRDMA_ADDR_TYPE_VA_BASED;
3181                         stag_info.va = (void *)(uintptr_t)iwmr->ibmr.iova;
3182                         stag_info.total_len = iwmr->ibmr.length;
3183                         stag_info.reg_addr_pa = *palloc->level1.addr;
3184                         stag_info.first_pm_pbl_index = palloc->level1.idx;
3185                         stag_info.local_fence = ib_wr->send_flags & IB_SEND_FENCE;
3186                         if (iwmr->npages > IRDMA_MIN_PAGES_PER_FMR)
3187                                 stag_info.chunk_size = 1;
3188                         err = irdma_sc_mr_fast_register(&iwqp->sc_qp, &stag_info,
3189                                                         true);
3190                         break;
3191                 }
3192                 default:
3193                         err = -EINVAL;
3194                         ibdev_dbg(&iwqp->iwdev->ibdev,
3195                                   "VERBS: upost_send bad opcode = 0x%x\n",
3196                                   ib_wr->opcode);
3197                         break;
3198                 }
3199
3200                 if (err)
3201                         break;
3202                 ib_wr = ib_wr->next;
3203         }
3204
3205         if (!iwqp->flush_issued && iwqp->hw_iwarp_state <= IRDMA_QP_STATE_RTS) {
3206                 irdma_uk_qp_post_wr(ukqp);
3207                 spin_unlock_irqrestore(&iwqp->lock, flags);
3208         } else if (reflush) {
3209                 ukqp->sq_flush_complete = false;
3210                 spin_unlock_irqrestore(&iwqp->lock, flags);
3211                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_SQ | IRDMA_REFLUSH);
3212         } else {
3213                 spin_unlock_irqrestore(&iwqp->lock, flags);
3214         }
3215         if (err)
3216                 *bad_wr = ib_wr;
3217
3218         return err;
3219 }
3220
3221 /**
3222  * irdma_post_recv - post receive wr for kernel application
3223  * @ibqp: ib qp pointer
3224  * @ib_wr: work request for receive
3225  * @bad_wr: bad wr caused an error
3226  */
3227 static int irdma_post_recv(struct ib_qp *ibqp,
3228                            const struct ib_recv_wr *ib_wr,
3229                            const struct ib_recv_wr **bad_wr)
3230 {
3231         struct irdma_qp *iwqp;
3232         struct irdma_qp_uk *ukqp;
3233         struct irdma_post_rq_info post_recv = {};
3234         unsigned long flags;
3235         int err = 0;
3236         bool reflush = false;
3237
3238         iwqp = to_iwqp(ibqp);
3239         ukqp = &iwqp->sc_qp.qp_uk;
3240
3241         spin_lock_irqsave(&iwqp->lock, flags);
3242         if (iwqp->flush_issued && ukqp->rq_flush_complete)
3243                 reflush = true;
3244         while (ib_wr) {
3245                 post_recv.num_sges = ib_wr->num_sge;
3246                 post_recv.wr_id = ib_wr->wr_id;
3247                 post_recv.sg_list = ib_wr->sg_list;
3248                 err = irdma_uk_post_receive(ukqp, &post_recv);
3249                 if (err) {
3250                         ibdev_dbg(&iwqp->iwdev->ibdev,
3251                                   "VERBS: post_recv err %d\n", err);
3252                         goto out;
3253                 }
3254
3255                 ib_wr = ib_wr->next;
3256         }
3257
3258 out:
3259         if (reflush) {
3260                 ukqp->rq_flush_complete = false;
3261                 spin_unlock_irqrestore(&iwqp->lock, flags);
3262                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_RQ | IRDMA_REFLUSH);
3263         } else {
3264                 spin_unlock_irqrestore(&iwqp->lock, flags);
3265         }
3266
3267         if (err)
3268                 *bad_wr = ib_wr;
3269
3270         return err;
3271 }
3272
3273 /**
3274  * irdma_flush_err_to_ib_wc_status - return change flush error code to IB status
3275  * @opcode: iwarp flush code
3276  */
3277 static enum ib_wc_status irdma_flush_err_to_ib_wc_status(enum irdma_flush_opcode opcode)
3278 {
3279         switch (opcode) {
3280         case FLUSH_PROT_ERR:
3281                 return IB_WC_LOC_PROT_ERR;
3282         case FLUSH_REM_ACCESS_ERR:
3283                 return IB_WC_REM_ACCESS_ERR;
3284         case FLUSH_LOC_QP_OP_ERR:
3285                 return IB_WC_LOC_QP_OP_ERR;
3286         case FLUSH_REM_OP_ERR:
3287                 return IB_WC_REM_OP_ERR;
3288         case FLUSH_LOC_LEN_ERR:
3289                 return IB_WC_LOC_LEN_ERR;
3290         case FLUSH_GENERAL_ERR:
3291                 return IB_WC_WR_FLUSH_ERR;
3292         case FLUSH_RETRY_EXC_ERR:
3293                 return IB_WC_RETRY_EXC_ERR;
3294         case FLUSH_MW_BIND_ERR:
3295                 return IB_WC_MW_BIND_ERR;
3296         case FLUSH_FATAL_ERR:
3297         default:
3298                 return IB_WC_FATAL_ERR;
3299         }
3300 }
3301
3302 /**
3303  * irdma_process_cqe - process cqe info
3304  * @entry: processed cqe
3305  * @cq_poll_info: cqe info
3306  */
3307 static void irdma_process_cqe(struct ib_wc *entry,
3308                               struct irdma_cq_poll_info *cq_poll_info)
3309 {
3310         struct irdma_qp *iwqp;
3311         struct irdma_sc_qp *qp;
3312
3313         entry->wc_flags = 0;
3314         entry->pkey_index = 0;
3315         entry->wr_id = cq_poll_info->wr_id;
3316
3317         qp = cq_poll_info->qp_handle;
3318         iwqp = qp->qp_uk.back_qp;
3319         entry->qp = qp->qp_uk.back_qp;
3320
3321         if (cq_poll_info->error) {
3322                 entry->status = (cq_poll_info->comp_status == IRDMA_COMPL_STATUS_FLUSHED) ?
3323                                 irdma_flush_err_to_ib_wc_status(cq_poll_info->minor_err) : IB_WC_GENERAL_ERR;
3324
3325                 entry->vendor_err = cq_poll_info->major_err << 16 |
3326                                     cq_poll_info->minor_err;
3327         } else {
3328                 entry->status = IB_WC_SUCCESS;
3329                 if (cq_poll_info->imm_valid) {
3330                         entry->ex.imm_data = htonl(cq_poll_info->imm_data);
3331                         entry->wc_flags |= IB_WC_WITH_IMM;
3332                 }
3333                 if (cq_poll_info->ud_smac_valid) {
3334                         ether_addr_copy(entry->smac, cq_poll_info->ud_smac);
3335                         entry->wc_flags |= IB_WC_WITH_SMAC;
3336                 }
3337
3338                 if (cq_poll_info->ud_vlan_valid) {
3339                         u16 vlan = cq_poll_info->ud_vlan & VLAN_VID_MASK;
3340
3341                         entry->sl = cq_poll_info->ud_vlan >> VLAN_PRIO_SHIFT;
3342                         if (vlan) {
3343                                 entry->vlan_id = vlan;
3344                                 entry->wc_flags |= IB_WC_WITH_VLAN;
3345                         }
3346                 } else {
3347                         entry->sl = 0;
3348                 }
3349         }
3350
3351         switch (cq_poll_info->op_type) {
3352         case IRDMA_OP_TYPE_RDMA_WRITE:
3353         case IRDMA_OP_TYPE_RDMA_WRITE_SOL:
3354                 entry->opcode = IB_WC_RDMA_WRITE;
3355                 break;
3356         case IRDMA_OP_TYPE_RDMA_READ_INV_STAG:
3357         case IRDMA_OP_TYPE_RDMA_READ:
3358                 entry->opcode = IB_WC_RDMA_READ;
3359                 break;
3360         case IRDMA_OP_TYPE_SEND_INV:
3361         case IRDMA_OP_TYPE_SEND_SOL:
3362         case IRDMA_OP_TYPE_SEND_SOL_INV:
3363         case IRDMA_OP_TYPE_SEND:
3364                 entry->opcode = IB_WC_SEND;
3365                 break;
3366         case IRDMA_OP_TYPE_FAST_REG_NSMR:
3367                 entry->opcode = IB_WC_REG_MR;
3368                 break;
3369         case IRDMA_OP_TYPE_INV_STAG:
3370                 entry->opcode = IB_WC_LOCAL_INV;
3371                 break;
3372         case IRDMA_OP_TYPE_REC_IMM:
3373         case IRDMA_OP_TYPE_REC:
3374                 entry->opcode = cq_poll_info->op_type == IRDMA_OP_TYPE_REC_IMM ?
3375                         IB_WC_RECV_RDMA_WITH_IMM : IB_WC_RECV;
3376                 if (qp->qp_uk.qp_type != IRDMA_QP_TYPE_ROCE_UD &&
3377                     cq_poll_info->stag_invalid_set) {
3378                         entry->ex.invalidate_rkey = cq_poll_info->inv_stag;
3379                         entry->wc_flags |= IB_WC_WITH_INVALIDATE;
3380                 }
3381                 break;
3382         default:
3383                 ibdev_err(&iwqp->iwdev->ibdev,
3384                           "Invalid opcode = %d in CQE\n", cq_poll_info->op_type);
3385                 entry->status = IB_WC_GENERAL_ERR;
3386                 return;
3387         }
3388
3389         if (qp->qp_uk.qp_type == IRDMA_QP_TYPE_ROCE_UD) {
3390                 entry->src_qp = cq_poll_info->ud_src_qpn;
3391                 entry->slid = 0;
3392                 entry->wc_flags |=
3393                         (IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE);
3394                 entry->network_hdr_type = cq_poll_info->ipv4 ?
3395                                                   RDMA_NETWORK_IPV4 :
3396                                                   RDMA_NETWORK_IPV6;
3397         } else {
3398                 entry->src_qp = cq_poll_info->qp_id;
3399         }
3400
3401         entry->byte_len = cq_poll_info->bytes_xfered;
3402 }
3403
3404 /**
3405  * irdma_poll_one - poll one entry of the CQ
3406  * @ukcq: ukcq to poll
3407  * @cur_cqe: current CQE info to be filled in
3408  * @entry: ibv_wc object to be filled for non-extended CQ or NULL for extended CQ
3409  *
3410  * Returns the internal irdma device error code or 0 on success
3411  */
3412 static inline int irdma_poll_one(struct irdma_cq_uk *ukcq,
3413                                  struct irdma_cq_poll_info *cur_cqe,
3414                                  struct ib_wc *entry)
3415 {
3416         int ret = irdma_uk_cq_poll_cmpl(ukcq, cur_cqe);
3417
3418         if (ret)
3419                 return ret;
3420
3421         irdma_process_cqe(entry, cur_cqe);
3422
3423         return 0;
3424 }
3425
3426 /**
3427  * __irdma_poll_cq - poll cq for completion (kernel apps)
3428  * @iwcq: cq to poll
3429  * @num_entries: number of entries to poll
3430  * @entry: wr of a completed entry
3431  */
3432 static int __irdma_poll_cq(struct irdma_cq *iwcq, int num_entries, struct ib_wc *entry)
3433 {
3434         struct list_head *tmp_node, *list_node;
3435         struct irdma_cq_buf *last_buf = NULL;
3436         struct irdma_cq_poll_info *cur_cqe = &iwcq->cur_cqe;
3437         struct irdma_cq_buf *cq_buf;
3438         int ret;
3439         struct irdma_device *iwdev;
3440         struct irdma_cq_uk *ukcq;
3441         bool cq_new_cqe = false;
3442         int resized_bufs = 0;
3443         int npolled = 0;
3444
3445         iwdev = to_iwdev(iwcq->ibcq.device);
3446         ukcq = &iwcq->sc_cq.cq_uk;
3447
3448         /* go through the list of previously resized CQ buffers */
3449         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
3450                 cq_buf = container_of(list_node, struct irdma_cq_buf, list);
3451                 while (npolled < num_entries) {
3452                         ret = irdma_poll_one(&cq_buf->cq_uk, cur_cqe, entry + npolled);
3453                         if (!ret) {
3454                                 ++npolled;
3455                                 cq_new_cqe = true;
3456                                 continue;
3457                         }
3458                         if (ret == -ENOENT)
3459                                 break;
3460                          /* QP using the CQ is destroyed. Skip reporting this CQE */
3461                         if (ret == -EFAULT) {
3462                                 cq_new_cqe = true;
3463                                 continue;
3464                         }
3465                         goto error;
3466                 }
3467
3468                 /* save the resized CQ buffer which received the last cqe */
3469                 if (cq_new_cqe)
3470                         last_buf = cq_buf;
3471                 cq_new_cqe = false;
3472         }
3473
3474         /* check the current CQ for new cqes */
3475         while (npolled < num_entries) {
3476                 ret = irdma_poll_one(ukcq, cur_cqe, entry + npolled);
3477                 if (!ret) {
3478                         ++npolled;
3479                         cq_new_cqe = true;
3480                         continue;
3481                 }
3482
3483                 if (ret == -ENOENT)
3484                         break;
3485                 /* QP using the CQ is destroyed. Skip reporting this CQE */
3486                 if (ret == -EFAULT) {
3487                         cq_new_cqe = true;
3488                         continue;
3489                 }
3490                 goto error;
3491         }
3492
3493         if (cq_new_cqe)
3494                 /* all previous CQ resizes are complete */
3495                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, NULL);
3496         else if (last_buf)
3497                 /* only CQ resizes up to the last_buf are complete */
3498                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, last_buf);
3499         if (resized_bufs)
3500                 /* report to the HW the number of complete CQ resizes */
3501                 irdma_uk_cq_set_resized_cnt(ukcq, resized_bufs);
3502
3503         return npolled;
3504 error:
3505         ibdev_dbg(&iwdev->ibdev, "%s: Error polling CQ, irdma_err: %d\n",
3506                   __func__, ret);
3507
3508         return ret;
3509 }
3510
3511 /**
3512  * irdma_poll_cq - poll cq for completion (kernel apps)
3513  * @ibcq: cq to poll
3514  * @num_entries: number of entries to poll
3515  * @entry: wr of a completed entry
3516  */
3517 static int irdma_poll_cq(struct ib_cq *ibcq, int num_entries,
3518                          struct ib_wc *entry)
3519 {
3520         struct irdma_cq *iwcq;
3521         unsigned long flags;
3522         int ret;
3523
3524         iwcq = to_iwcq(ibcq);
3525
3526         spin_lock_irqsave(&iwcq->lock, flags);
3527         ret = __irdma_poll_cq(iwcq, num_entries, entry);
3528         spin_unlock_irqrestore(&iwcq->lock, flags);
3529
3530         return ret;
3531 }
3532
3533 /**
3534  * irdma_req_notify_cq - arm cq kernel application
3535  * @ibcq: cq to arm
3536  * @notify_flags: notofication flags
3537  */
3538 static int irdma_req_notify_cq(struct ib_cq *ibcq,
3539                                enum ib_cq_notify_flags notify_flags)
3540 {
3541         struct irdma_cq *iwcq;
3542         struct irdma_cq_uk *ukcq;
3543         unsigned long flags;
3544         enum irdma_cmpl_notify cq_notify;
3545         bool promo_event = false;
3546         int ret = 0;
3547
3548         cq_notify = notify_flags == IB_CQ_SOLICITED ?
3549                     IRDMA_CQ_COMPL_SOLICITED : IRDMA_CQ_COMPL_EVENT;
3550         iwcq = to_iwcq(ibcq);
3551         ukcq = &iwcq->sc_cq.cq_uk;
3552
3553         spin_lock_irqsave(&iwcq->lock, flags);
3554         /* Only promote to arm the CQ for any event if the last arm event was solicited. */
3555         if (iwcq->last_notify == IRDMA_CQ_COMPL_SOLICITED && notify_flags != IB_CQ_SOLICITED)
3556                 promo_event = true;
3557
3558         if (!iwcq->armed || promo_event) {
3559                 iwcq->armed = true;
3560                 iwcq->last_notify = cq_notify;
3561                 irdma_uk_cq_request_notification(ukcq, cq_notify);
3562         }
3563
3564         if ((notify_flags & IB_CQ_REPORT_MISSED_EVENTS) && !irdma_cq_empty(iwcq))
3565                 ret = 1;
3566         spin_unlock_irqrestore(&iwcq->lock, flags);
3567
3568         return ret;
3569 }
3570
3571 static int irdma_roce_port_immutable(struct ib_device *ibdev, u32 port_num,
3572                                      struct ib_port_immutable *immutable)
3573 {
3574         struct ib_port_attr attr;
3575         int err;
3576
3577         immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
3578         err = ib_query_port(ibdev, port_num, &attr);
3579         if (err)
3580                 return err;
3581
3582         immutable->max_mad_size = IB_MGMT_MAD_SIZE;
3583         immutable->pkey_tbl_len = attr.pkey_tbl_len;
3584         immutable->gid_tbl_len = attr.gid_tbl_len;
3585
3586         return 0;
3587 }
3588
3589 static int irdma_iw_port_immutable(struct ib_device *ibdev, u32 port_num,
3590                                    struct ib_port_immutable *immutable)
3591 {
3592         struct ib_port_attr attr;
3593         int err;
3594
3595         immutable->core_cap_flags = RDMA_CORE_PORT_IWARP;
3596         err = ib_query_port(ibdev, port_num, &attr);
3597         if (err)
3598                 return err;
3599         immutable->gid_tbl_len = attr.gid_tbl_len;
3600
3601         return 0;
3602 }
3603
3604 static const struct rdma_stat_desc irdma_hw_stat_descs[] = {
3605         /* 32bit names */
3606         [IRDMA_HW_STAT_INDEX_RXVLANERR].name = "rxVlanErrors",
3607         [IRDMA_HW_STAT_INDEX_IP4RXDISCARD].name = "ip4InDiscards",
3608         [IRDMA_HW_STAT_INDEX_IP4RXTRUNC].name = "ip4InTruncatedPkts",
3609         [IRDMA_HW_STAT_INDEX_IP4TXNOROUTE].name = "ip4OutNoRoutes",
3610         [IRDMA_HW_STAT_INDEX_IP6RXDISCARD].name = "ip6InDiscards",
3611         [IRDMA_HW_STAT_INDEX_IP6RXTRUNC].name = "ip6InTruncatedPkts",
3612         [IRDMA_HW_STAT_INDEX_IP6TXNOROUTE].name = "ip6OutNoRoutes",
3613         [IRDMA_HW_STAT_INDEX_TCPRTXSEG].name = "tcpRetransSegs",
3614         [IRDMA_HW_STAT_INDEX_TCPRXOPTERR].name = "tcpInOptErrors",
3615         [IRDMA_HW_STAT_INDEX_TCPRXPROTOERR].name = "tcpInProtoErrors",
3616         [IRDMA_HW_STAT_INDEX_RXRPCNPHANDLED].name = "cnpHandled",
3617         [IRDMA_HW_STAT_INDEX_RXRPCNPIGNORED].name = "cnpIgnored",
3618         [IRDMA_HW_STAT_INDEX_TXNPCNPSENT].name = "cnpSent",
3619
3620         /* 64bit names */
3621         [IRDMA_HW_STAT_INDEX_IP4RXOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3622                 "ip4InOctets",
3623         [IRDMA_HW_STAT_INDEX_IP4RXPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3624                 "ip4InPkts",
3625         [IRDMA_HW_STAT_INDEX_IP4RXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3626                 "ip4InReasmRqd",
3627         [IRDMA_HW_STAT_INDEX_IP4RXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3628                 "ip4InMcastOctets",
3629         [IRDMA_HW_STAT_INDEX_IP4RXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3630                 "ip4InMcastPkts",
3631         [IRDMA_HW_STAT_INDEX_IP4TXOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3632                 "ip4OutOctets",
3633         [IRDMA_HW_STAT_INDEX_IP4TXPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3634                 "ip4OutPkts",
3635         [IRDMA_HW_STAT_INDEX_IP4TXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3636                 "ip4OutSegRqd",
3637         [IRDMA_HW_STAT_INDEX_IP4TXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3638                 "ip4OutMcastOctets",
3639         [IRDMA_HW_STAT_INDEX_IP4TXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3640                 "ip4OutMcastPkts",
3641         [IRDMA_HW_STAT_INDEX_IP6RXOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3642                 "ip6InOctets",
3643         [IRDMA_HW_STAT_INDEX_IP6RXPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3644                 "ip6InPkts",
3645         [IRDMA_HW_STAT_INDEX_IP6RXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3646                 "ip6InReasmRqd",
3647         [IRDMA_HW_STAT_INDEX_IP6RXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3648                 "ip6InMcastOctets",
3649         [IRDMA_HW_STAT_INDEX_IP6RXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3650                 "ip6InMcastPkts",
3651         [IRDMA_HW_STAT_INDEX_IP6TXOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3652                 "ip6OutOctets",
3653         [IRDMA_HW_STAT_INDEX_IP6TXPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3654                 "ip6OutPkts",
3655         [IRDMA_HW_STAT_INDEX_IP6TXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3656                 "ip6OutSegRqd",
3657         [IRDMA_HW_STAT_INDEX_IP6TXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3658                 "ip6OutMcastOctets",
3659         [IRDMA_HW_STAT_INDEX_IP6TXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3660                 "ip6OutMcastPkts",
3661         [IRDMA_HW_STAT_INDEX_TCPRXSEGS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3662                 "tcpInSegs",
3663         [IRDMA_HW_STAT_INDEX_TCPTXSEG + IRDMA_HW_STAT_INDEX_MAX_32].name =
3664                 "tcpOutSegs",
3665         [IRDMA_HW_STAT_INDEX_RDMARXRDS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3666                 "iwInRdmaReads",
3667         [IRDMA_HW_STAT_INDEX_RDMARXSNDS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3668                 "iwInRdmaSends",
3669         [IRDMA_HW_STAT_INDEX_RDMARXWRS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3670                 "iwInRdmaWrites",
3671         [IRDMA_HW_STAT_INDEX_RDMATXRDS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3672                 "iwOutRdmaReads",
3673         [IRDMA_HW_STAT_INDEX_RDMATXSNDS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3674                 "iwOutRdmaSends",
3675         [IRDMA_HW_STAT_INDEX_RDMATXWRS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3676                 "iwOutRdmaWrites",
3677         [IRDMA_HW_STAT_INDEX_RDMAVBND + IRDMA_HW_STAT_INDEX_MAX_32].name =
3678                 "iwRdmaBnd",
3679         [IRDMA_HW_STAT_INDEX_RDMAVINV + IRDMA_HW_STAT_INDEX_MAX_32].name =
3680                 "iwRdmaInv",
3681         [IRDMA_HW_STAT_INDEX_UDPRXPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3682                 "RxUDP",
3683         [IRDMA_HW_STAT_INDEX_UDPTXPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3684                 "TxUDP",
3685         [IRDMA_HW_STAT_INDEX_RXNPECNMARKEDPKTS + IRDMA_HW_STAT_INDEX_MAX_32]
3686                 .name = "RxECNMrkd",
3687 };
3688
3689 static void irdma_get_dev_fw_str(struct ib_device *dev, char *str)
3690 {
3691         struct irdma_device *iwdev = to_iwdev(dev);
3692
3693         snprintf(str, IB_FW_VERSION_NAME_MAX, "%u.%u",
3694                  irdma_fw_major_ver(&iwdev->rf->sc_dev),
3695                  irdma_fw_minor_ver(&iwdev->rf->sc_dev));
3696 }
3697
3698 /**
3699  * irdma_alloc_hw_port_stats - Allocate a hw stats structure
3700  * @ibdev: device pointer from stack
3701  * @port_num: port number
3702  */
3703 static struct rdma_hw_stats *irdma_alloc_hw_port_stats(struct ib_device *ibdev,
3704                                                        u32 port_num)
3705 {
3706         int num_counters = IRDMA_HW_STAT_INDEX_MAX_32 +
3707                            IRDMA_HW_STAT_INDEX_MAX_64;
3708         unsigned long lifespan = RDMA_HW_STATS_DEFAULT_LIFESPAN;
3709
3710         BUILD_BUG_ON(ARRAY_SIZE(irdma_hw_stat_descs) !=
3711                      (IRDMA_HW_STAT_INDEX_MAX_32 + IRDMA_HW_STAT_INDEX_MAX_64));
3712
3713         return rdma_alloc_hw_stats_struct(irdma_hw_stat_descs, num_counters,
3714                                           lifespan);
3715 }
3716
3717 /**
3718  * irdma_get_hw_stats - Populates the rdma_hw_stats structure
3719  * @ibdev: device pointer from stack
3720  * @stats: stats pointer from stack
3721  * @port_num: port number
3722  * @index: which hw counter the stack is requesting we update
3723  */
3724 static int irdma_get_hw_stats(struct ib_device *ibdev,
3725                               struct rdma_hw_stats *stats, u32 port_num,
3726                               int index)
3727 {
3728         struct irdma_device *iwdev = to_iwdev(ibdev);
3729         struct irdma_dev_hw_stats *hw_stats = &iwdev->vsi.pestat->hw_stats;
3730
3731         if (iwdev->rf->rdma_ver >= IRDMA_GEN_2)
3732                 irdma_cqp_gather_stats_cmd(&iwdev->rf->sc_dev, iwdev->vsi.pestat, true);
3733         else
3734                 irdma_cqp_gather_stats_gen1(&iwdev->rf->sc_dev, iwdev->vsi.pestat);
3735
3736         memcpy(&stats->value[0], hw_stats, sizeof(*hw_stats));
3737
3738         return stats->num_counters;
3739 }
3740
3741 /**
3742  * irdma_query_gid - Query port GID
3743  * @ibdev: device pointer from stack
3744  * @port: port number
3745  * @index: Entry index
3746  * @gid: Global ID
3747  */
3748 static int irdma_query_gid(struct ib_device *ibdev, u32 port, int index,
3749                            union ib_gid *gid)
3750 {
3751         struct irdma_device *iwdev = to_iwdev(ibdev);
3752
3753         memset(gid->raw, 0, sizeof(gid->raw));
3754         ether_addr_copy(gid->raw, iwdev->netdev->dev_addr);
3755
3756         return 0;
3757 }
3758
3759 /**
3760  * mcast_list_add -  Add a new mcast item to list
3761  * @rf: RDMA PCI function
3762  * @new_elem: pointer to element to add
3763  */
3764 static void mcast_list_add(struct irdma_pci_f *rf,
3765                            struct mc_table_list *new_elem)
3766 {
3767         list_add(&new_elem->list, &rf->mc_qht_list.list);
3768 }
3769
3770 /**
3771  * mcast_list_del - Remove an mcast item from list
3772  * @mc_qht_elem: pointer to mcast table list element
3773  */
3774 static void mcast_list_del(struct mc_table_list *mc_qht_elem)
3775 {
3776         if (mc_qht_elem)
3777                 list_del(&mc_qht_elem->list);
3778 }
3779
3780 /**
3781  * mcast_list_lookup_ip - Search mcast list for address
3782  * @rf: RDMA PCI function
3783  * @ip_mcast: pointer to mcast IP address
3784  */
3785 static struct mc_table_list *mcast_list_lookup_ip(struct irdma_pci_f *rf,
3786                                                   u32 *ip_mcast)
3787 {
3788         struct mc_table_list *mc_qht_el;
3789         struct list_head *pos, *q;
3790
3791         list_for_each_safe (pos, q, &rf->mc_qht_list.list) {
3792                 mc_qht_el = list_entry(pos, struct mc_table_list, list);
3793                 if (!memcmp(mc_qht_el->mc_info.dest_ip, ip_mcast,
3794                             sizeof(mc_qht_el->mc_info.dest_ip)))
3795                         return mc_qht_el;
3796         }
3797
3798         return NULL;
3799 }
3800
3801 /**
3802  * irdma_mcast_cqp_op - perform a mcast cqp operation
3803  * @iwdev: irdma device
3804  * @mc_grp_ctx: mcast group info
3805  * @op: operation
3806  *
3807  * returns error status
3808  */
3809 static int irdma_mcast_cqp_op(struct irdma_device *iwdev,
3810                               struct irdma_mcast_grp_info *mc_grp_ctx, u8 op)
3811 {
3812         struct cqp_cmds_info *cqp_info;
3813         struct irdma_cqp_request *cqp_request;
3814         int status;
3815
3816         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
3817         if (!cqp_request)
3818                 return -ENOMEM;
3819
3820         cqp_request->info.in.u.mc_create.info = *mc_grp_ctx;
3821         cqp_info = &cqp_request->info;
3822         cqp_info->cqp_cmd = op;
3823         cqp_info->post_sq = 1;
3824         cqp_info->in.u.mc_create.scratch = (uintptr_t)cqp_request;
3825         cqp_info->in.u.mc_create.cqp = &iwdev->rf->cqp.sc_cqp;
3826         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
3827         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
3828
3829         return status;
3830 }
3831
3832 /**
3833  * irdma_mcast_mac - Get the multicast MAC for an IP address
3834  * @ip_addr: IPv4 or IPv6 address
3835  * @mac: pointer to result MAC address
3836  * @ipv4: flag indicating IPv4 or IPv6
3837  *
3838  */
3839 void irdma_mcast_mac(u32 *ip_addr, u8 *mac, bool ipv4)
3840 {
3841         u8 *ip = (u8 *)ip_addr;
3842
3843         if (ipv4) {
3844                 unsigned char mac4[ETH_ALEN] = {0x01, 0x00, 0x5E, 0x00,
3845                                                 0x00, 0x00};
3846
3847                 mac4[3] = ip[2] & 0x7F;
3848                 mac4[4] = ip[1];
3849                 mac4[5] = ip[0];
3850                 ether_addr_copy(mac, mac4);
3851         } else {
3852                 unsigned char mac6[ETH_ALEN] = {0x33, 0x33, 0x00, 0x00,
3853                                                 0x00, 0x00};
3854
3855                 mac6[2] = ip[3];
3856                 mac6[3] = ip[2];
3857                 mac6[4] = ip[1];
3858                 mac6[5] = ip[0];
3859                 ether_addr_copy(mac, mac6);
3860         }
3861 }
3862
3863 /**
3864  * irdma_attach_mcast - attach a qp to a multicast group
3865  * @ibqp: ptr to qp
3866  * @ibgid: pointer to global ID
3867  * @lid: local ID
3868  *
3869  * returns error status
3870  */
3871 static int irdma_attach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
3872 {
3873         struct irdma_qp *iwqp = to_iwqp(ibqp);
3874         struct irdma_device *iwdev = iwqp->iwdev;
3875         struct irdma_pci_f *rf = iwdev->rf;
3876         struct mc_table_list *mc_qht_elem;
3877         struct irdma_mcast_grp_ctx_entry_info mcg_info = {};
3878         unsigned long flags;
3879         u32 ip_addr[4] = {};
3880         u32 mgn;
3881         u32 no_mgs;
3882         int ret = 0;
3883         bool ipv4;
3884         u16 vlan_id;
3885         union irdma_sockaddr sgid_addr;
3886         unsigned char dmac[ETH_ALEN];
3887
3888         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
3889
3890         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid)) {
3891                 irdma_copy_ip_ntohl(ip_addr,
3892                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
3893                 irdma_netdev_vlan_ipv6(ip_addr, &vlan_id, NULL);
3894                 ipv4 = false;
3895                 ibdev_dbg(&iwdev->ibdev,
3896                           "VERBS: qp_id=%d, IP6address=%pI6\n", ibqp->qp_num,
3897                           ip_addr);
3898                 irdma_mcast_mac(ip_addr, dmac, false);
3899         } else {
3900                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
3901                 ipv4 = true;
3902                 vlan_id = irdma_get_vlan_ipv4(ip_addr);
3903                 irdma_mcast_mac(ip_addr, dmac, true);
3904                 ibdev_dbg(&iwdev->ibdev,
3905                           "VERBS: qp_id=%d, IP4address=%pI4, MAC=%pM\n",
3906                           ibqp->qp_num, ip_addr, dmac);
3907         }
3908
3909         spin_lock_irqsave(&rf->qh_list_lock, flags);
3910         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
3911         if (!mc_qht_elem) {
3912                 struct irdma_dma_mem *dma_mem_mc;
3913
3914                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3915                 mc_qht_elem = kzalloc(sizeof(*mc_qht_elem), GFP_KERNEL);
3916                 if (!mc_qht_elem)
3917                         return -ENOMEM;
3918
3919                 mc_qht_elem->mc_info.ipv4_valid = ipv4;
3920                 memcpy(mc_qht_elem->mc_info.dest_ip, ip_addr,
3921                        sizeof(mc_qht_elem->mc_info.dest_ip));
3922                 ret = irdma_alloc_rsrc(rf, rf->allocated_mcgs, rf->max_mcg,
3923                                        &mgn, &rf->next_mcg);
3924                 if (ret) {
3925                         kfree(mc_qht_elem);
3926                         return -ENOMEM;
3927                 }
3928
3929                 mc_qht_elem->mc_info.mgn = mgn;
3930                 dma_mem_mc = &mc_qht_elem->mc_grp_ctx.dma_mem_mc;
3931                 dma_mem_mc->size = ALIGN(sizeof(u64) * IRDMA_MAX_MGS_PER_CTX,
3932                                          IRDMA_HW_PAGE_SIZE);
3933                 dma_mem_mc->va = dma_alloc_coherent(rf->hw.device,
3934                                                     dma_mem_mc->size,
3935                                                     &dma_mem_mc->pa,
3936                                                     GFP_KERNEL);
3937                 if (!dma_mem_mc->va) {
3938                         irdma_free_rsrc(rf, rf->allocated_mcgs, mgn);
3939                         kfree(mc_qht_elem);
3940                         return -ENOMEM;
3941                 }
3942
3943                 mc_qht_elem->mc_grp_ctx.mg_id = (u16)mgn;
3944                 memcpy(mc_qht_elem->mc_grp_ctx.dest_ip_addr, ip_addr,
3945                        sizeof(mc_qht_elem->mc_grp_ctx.dest_ip_addr));
3946                 mc_qht_elem->mc_grp_ctx.ipv4_valid = ipv4;
3947                 mc_qht_elem->mc_grp_ctx.vlan_id = vlan_id;
3948                 if (vlan_id < VLAN_N_VID)
3949                         mc_qht_elem->mc_grp_ctx.vlan_valid = true;
3950                 mc_qht_elem->mc_grp_ctx.hmc_fcn_id = iwdev->vsi.fcn_id;
3951                 mc_qht_elem->mc_grp_ctx.qs_handle =
3952                         iwqp->sc_qp.vsi->qos[iwqp->sc_qp.user_pri].qs_handle;
3953                 ether_addr_copy(mc_qht_elem->mc_grp_ctx.dest_mac_addr, dmac);
3954
3955                 spin_lock_irqsave(&rf->qh_list_lock, flags);
3956                 mcast_list_add(rf, mc_qht_elem);
3957         } else {
3958                 if (mc_qht_elem->mc_grp_ctx.no_of_mgs ==
3959                     IRDMA_MAX_MGS_PER_CTX) {
3960                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3961                         return -ENOMEM;
3962                 }
3963         }
3964
3965         mcg_info.qp_id = iwqp->ibqp.qp_num;
3966         no_mgs = mc_qht_elem->mc_grp_ctx.no_of_mgs;
3967         irdma_sc_add_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
3968         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3969
3970         /* Only if there is a change do we need to modify or create */
3971         if (!no_mgs) {
3972                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
3973                                          IRDMA_OP_MC_CREATE);
3974         } else if (no_mgs != mc_qht_elem->mc_grp_ctx.no_of_mgs) {
3975                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
3976                                          IRDMA_OP_MC_MODIFY);
3977         } else {
3978                 return 0;
3979         }
3980
3981         if (ret)
3982                 goto error;
3983
3984         return 0;
3985
3986 error:
3987         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
3988         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
3989                 mcast_list_del(mc_qht_elem);
3990                 dma_free_coherent(rf->hw.device,
3991                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.size,
3992                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.va,
3993                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.pa);
3994                 mc_qht_elem->mc_grp_ctx.dma_mem_mc.va = NULL;
3995                 irdma_free_rsrc(rf, rf->allocated_mcgs,
3996                                 mc_qht_elem->mc_grp_ctx.mg_id);
3997                 kfree(mc_qht_elem);
3998         }
3999
4000         return ret;
4001 }
4002
4003 /**
4004  * irdma_detach_mcast - detach a qp from a multicast group
4005  * @ibqp: ptr to qp
4006  * @ibgid: pointer to global ID
4007  * @lid: local ID
4008  *
4009  * returns error status
4010  */
4011 static int irdma_detach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
4012 {
4013         struct irdma_qp *iwqp = to_iwqp(ibqp);
4014         struct irdma_device *iwdev = iwqp->iwdev;
4015         struct irdma_pci_f *rf = iwdev->rf;
4016         u32 ip_addr[4] = {};
4017         struct mc_table_list *mc_qht_elem;
4018         struct irdma_mcast_grp_ctx_entry_info mcg_info = {};
4019         int ret;
4020         unsigned long flags;
4021         union irdma_sockaddr sgid_addr;
4022
4023         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
4024         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid))
4025                 irdma_copy_ip_ntohl(ip_addr,
4026                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4027         else
4028                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
4029
4030         spin_lock_irqsave(&rf->qh_list_lock, flags);
4031         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
4032         if (!mc_qht_elem) {
4033                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4034                 ibdev_dbg(&iwdev->ibdev,
4035                           "VERBS: address not found MCG\n");
4036                 return 0;
4037         }
4038
4039         mcg_info.qp_id = iwqp->ibqp.qp_num;
4040         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
4041         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
4042                 mcast_list_del(mc_qht_elem);
4043                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4044                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4045                                          IRDMA_OP_MC_DESTROY);
4046                 if (ret) {
4047                         ibdev_dbg(&iwdev->ibdev,
4048                                   "VERBS: failed MC_DESTROY MCG\n");
4049                         spin_lock_irqsave(&rf->qh_list_lock, flags);
4050                         mcast_list_add(rf, mc_qht_elem);
4051                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4052                         return -EAGAIN;
4053                 }
4054
4055                 dma_free_coherent(rf->hw.device,
4056                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.size,
4057                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.va,
4058                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.pa);
4059                 mc_qht_elem->mc_grp_ctx.dma_mem_mc.va = NULL;
4060                 irdma_free_rsrc(rf, rf->allocated_mcgs,
4061                                 mc_qht_elem->mc_grp_ctx.mg_id);
4062                 kfree(mc_qht_elem);
4063         } else {
4064                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4065                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4066                                          IRDMA_OP_MC_MODIFY);
4067                 if (ret) {
4068                         ibdev_dbg(&iwdev->ibdev,
4069                                   "VERBS: failed Modify MCG\n");
4070                         return ret;
4071                 }
4072         }
4073
4074         return 0;
4075 }
4076
4077 static int irdma_create_hw_ah(struct irdma_device *iwdev, struct irdma_ah *ah, bool sleep)
4078 {
4079         struct irdma_pci_f *rf = iwdev->rf;
4080         int err;
4081
4082         err = irdma_alloc_rsrc(rf, rf->allocated_ahs, rf->max_ah, &ah->sc_ah.ah_info.ah_idx,
4083                                &rf->next_ah);
4084         if (err)
4085                 return err;
4086
4087         err = irdma_ah_cqp_op(rf, &ah->sc_ah, IRDMA_OP_AH_CREATE, sleep,
4088                               irdma_gsi_ud_qp_ah_cb, &ah->sc_ah);
4089
4090         if (err) {
4091                 ibdev_dbg(&iwdev->ibdev, "VERBS: CQP-OP Create AH fail");
4092                 goto err_ah_create;
4093         }
4094
4095         if (!sleep) {
4096                 int cnt = CQP_COMPL_WAIT_TIME_MS * CQP_TIMEOUT_THRESHOLD;
4097
4098                 do {
4099                         irdma_cqp_ce_handler(rf, &rf->ccq.sc_cq);
4100                         mdelay(1);
4101                 } while (!ah->sc_ah.ah_info.ah_valid && --cnt);
4102
4103                 if (!cnt) {
4104                         ibdev_dbg(&iwdev->ibdev, "VERBS: CQP create AH timed out");
4105                         err = -ETIMEDOUT;
4106                         goto err_ah_create;
4107                 }
4108         }
4109         return 0;
4110
4111 err_ah_create:
4112         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_ahs, ah->sc_ah.ah_info.ah_idx);
4113
4114         return err;
4115 }
4116
4117 static int irdma_setup_ah(struct ib_ah *ibah, struct rdma_ah_init_attr *attr)
4118 {
4119         struct irdma_pd *pd = to_iwpd(ibah->pd);
4120         struct irdma_ah *ah = container_of(ibah, struct irdma_ah, ibah);
4121         struct rdma_ah_attr *ah_attr = attr->ah_attr;
4122         const struct ib_gid_attr *sgid_attr;
4123         struct irdma_device *iwdev = to_iwdev(ibah->pd->device);
4124         struct irdma_pci_f *rf = iwdev->rf;
4125         struct irdma_sc_ah *sc_ah;
4126         struct irdma_ah_info *ah_info;
4127         union irdma_sockaddr sgid_addr, dgid_addr;
4128         int err;
4129         u8 dmac[ETH_ALEN];
4130
4131         ah->pd = pd;
4132         sc_ah = &ah->sc_ah;
4133         sc_ah->ah_info.vsi = &iwdev->vsi;
4134         irdma_sc_init_ah(&rf->sc_dev, sc_ah);
4135         ah->sgid_index = ah_attr->grh.sgid_index;
4136         sgid_attr = ah_attr->grh.sgid_attr;
4137         memcpy(&ah->dgid, &ah_attr->grh.dgid, sizeof(ah->dgid));
4138         rdma_gid2ip((struct sockaddr *)&sgid_addr, &sgid_attr->gid);
4139         rdma_gid2ip((struct sockaddr *)&dgid_addr, &ah_attr->grh.dgid);
4140         ah->av.attrs = *ah_attr;
4141         ah->av.net_type = rdma_gid_attr_network_type(sgid_attr);
4142         ah_info = &sc_ah->ah_info;
4143         ah_info->pd_idx = pd->sc_pd.pd_id;
4144         if (ah_attr->ah_flags & IB_AH_GRH) {
4145                 ah_info->flow_label = ah_attr->grh.flow_label;
4146                 ah_info->hop_ttl = ah_attr->grh.hop_limit;
4147                 ah_info->tc_tos = ah_attr->grh.traffic_class;
4148         }
4149
4150         ether_addr_copy(dmac, ah_attr->roce.dmac);
4151         if (ah->av.net_type == RDMA_NETWORK_IPV4) {
4152                 ah_info->ipv4_valid = true;
4153                 ah_info->dest_ip_addr[0] =
4154                         ntohl(dgid_addr.saddr_in.sin_addr.s_addr);
4155                 ah_info->src_ip_addr[0] =
4156                         ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
4157                 ah_info->do_lpbk = irdma_ipv4_is_lpb(ah_info->src_ip_addr[0],
4158                                                      ah_info->dest_ip_addr[0]);
4159                 if (ipv4_is_multicast(dgid_addr.saddr_in.sin_addr.s_addr)) {
4160                         ah_info->do_lpbk = true;
4161                         irdma_mcast_mac(ah_info->dest_ip_addr, dmac, true);
4162                 }
4163         } else {
4164                 irdma_copy_ip_ntohl(ah_info->dest_ip_addr,
4165                                     dgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4166                 irdma_copy_ip_ntohl(ah_info->src_ip_addr,
4167                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4168                 ah_info->do_lpbk = irdma_ipv6_is_lpb(ah_info->src_ip_addr,
4169                                                      ah_info->dest_ip_addr);
4170                 if (rdma_is_multicast_addr(&dgid_addr.saddr_in6.sin6_addr)) {
4171                         ah_info->do_lpbk = true;
4172                         irdma_mcast_mac(ah_info->dest_ip_addr, dmac, false);
4173                 }
4174         }
4175
4176         err = rdma_read_gid_l2_fields(sgid_attr, &ah_info->vlan_tag,
4177                                       ah_info->mac_addr);
4178         if (err)
4179                 return err;
4180
4181         ah_info->dst_arpindex = irdma_add_arp(iwdev->rf, ah_info->dest_ip_addr,
4182                                               ah_info->ipv4_valid, dmac);
4183
4184         if (ah_info->dst_arpindex == -1)
4185                 return -EINVAL;
4186
4187         if (ah_info->vlan_tag >= VLAN_N_VID && iwdev->dcb_vlan_mode)
4188                 ah_info->vlan_tag = 0;
4189
4190         if (ah_info->vlan_tag < VLAN_N_VID) {
4191                 ah_info->insert_vlan_tag = true;
4192                 ah_info->vlan_tag |=
4193                         rt_tos2priority(ah_info->tc_tos) << VLAN_PRIO_SHIFT;
4194         }
4195
4196         return 0;
4197 }
4198
4199 /**
4200  * irdma_ah_exists - Check for existing identical AH
4201  * @iwdev: irdma device
4202  * @new_ah: AH to check for
4203  *
4204  * returns true if AH is found, false if not found.
4205  */
4206 static bool irdma_ah_exists(struct irdma_device *iwdev,
4207                             struct irdma_ah *new_ah)
4208 {
4209         struct irdma_ah *ah;
4210         u32 key = new_ah->sc_ah.ah_info.dest_ip_addr[0] ^
4211                   new_ah->sc_ah.ah_info.dest_ip_addr[1] ^
4212                   new_ah->sc_ah.ah_info.dest_ip_addr[2] ^
4213                   new_ah->sc_ah.ah_info.dest_ip_addr[3];
4214
4215         hash_for_each_possible(iwdev->ah_hash_tbl, ah, list, key) {
4216                 /* Set ah_valid and ah_id the same so memcmp can work */
4217                 new_ah->sc_ah.ah_info.ah_idx = ah->sc_ah.ah_info.ah_idx;
4218                 new_ah->sc_ah.ah_info.ah_valid = ah->sc_ah.ah_info.ah_valid;
4219                 if (!memcmp(&ah->sc_ah.ah_info, &new_ah->sc_ah.ah_info,
4220                             sizeof(ah->sc_ah.ah_info))) {
4221                         refcount_inc(&ah->refcnt);
4222                         new_ah->parent_ah = ah;
4223                         return true;
4224                 }
4225         }
4226
4227         return false;
4228 }
4229
4230 /**
4231  * irdma_destroy_ah - Destroy address handle
4232  * @ibah: pointer to address handle
4233  * @ah_flags: flags for sleepable
4234  */
4235 static int irdma_destroy_ah(struct ib_ah *ibah, u32 ah_flags)
4236 {
4237         struct irdma_device *iwdev = to_iwdev(ibah->device);
4238         struct irdma_ah *ah = to_iwah(ibah);
4239
4240         if ((ah_flags & RDMA_DESTROY_AH_SLEEPABLE) && ah->parent_ah) {
4241                 mutex_lock(&iwdev->ah_tbl_lock);
4242                 if (!refcount_dec_and_test(&ah->parent_ah->refcnt)) {
4243                         mutex_unlock(&iwdev->ah_tbl_lock);
4244                         return 0;
4245                 }
4246                 hash_del(&ah->parent_ah->list);
4247                 kfree(ah->parent_ah);
4248                 mutex_unlock(&iwdev->ah_tbl_lock);
4249         }
4250
4251         irdma_ah_cqp_op(iwdev->rf, &ah->sc_ah, IRDMA_OP_AH_DESTROY,
4252                         false, NULL, ah);
4253
4254         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_ahs,
4255                         ah->sc_ah.ah_info.ah_idx);
4256
4257         return 0;
4258 }
4259
4260 /**
4261  * irdma_create_user_ah - create user address handle
4262  * @ibah: address handle
4263  * @attr: address handle attributes
4264  * @udata: User data
4265  *
4266  * returns 0 on success, error otherwise
4267  */
4268 static int irdma_create_user_ah(struct ib_ah *ibah,
4269                                 struct rdma_ah_init_attr *attr,
4270                                 struct ib_udata *udata)
4271 {
4272         struct irdma_ah *ah = container_of(ibah, struct irdma_ah, ibah);
4273         struct irdma_device *iwdev = to_iwdev(ibah->pd->device);
4274         struct irdma_create_ah_resp uresp;
4275         struct irdma_ah *parent_ah;
4276         int err;
4277
4278         err = irdma_setup_ah(ibah, attr);
4279         if (err)
4280                 return err;
4281         mutex_lock(&iwdev->ah_tbl_lock);
4282         if (!irdma_ah_exists(iwdev, ah)) {
4283                 err = irdma_create_hw_ah(iwdev, ah, true);
4284                 if (err) {
4285                         mutex_unlock(&iwdev->ah_tbl_lock);
4286                         return err;
4287                 }
4288                 /* Add new AH to list */
4289                 parent_ah = kmemdup(ah, sizeof(*ah), GFP_KERNEL);
4290                 if (parent_ah) {
4291                         u32 key = parent_ah->sc_ah.ah_info.dest_ip_addr[0] ^
4292                                   parent_ah->sc_ah.ah_info.dest_ip_addr[1] ^
4293                                   parent_ah->sc_ah.ah_info.dest_ip_addr[2] ^
4294                                   parent_ah->sc_ah.ah_info.dest_ip_addr[3];
4295
4296                         ah->parent_ah = parent_ah;
4297                         hash_add(iwdev->ah_hash_tbl, &parent_ah->list, key);
4298                         refcount_set(&parent_ah->refcnt, 1);
4299                 }
4300         }
4301         mutex_unlock(&iwdev->ah_tbl_lock);
4302
4303         uresp.ah_id = ah->sc_ah.ah_info.ah_idx;
4304         err = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp), udata->outlen));
4305         if (err)
4306                 irdma_destroy_ah(ibah, attr->flags);
4307
4308         return err;
4309 }
4310
4311 /**
4312  * irdma_create_ah - create address handle
4313  * @ibah: address handle
4314  * @attr: address handle attributes
4315  * @udata: NULL
4316  *
4317  * returns 0 on success, error otherwise
4318  */
4319 static int irdma_create_ah(struct ib_ah *ibah, struct rdma_ah_init_attr *attr,
4320                            struct ib_udata *udata)
4321 {
4322         struct irdma_ah *ah = container_of(ibah, struct irdma_ah, ibah);
4323         struct irdma_device *iwdev = to_iwdev(ibah->pd->device);
4324         int err;
4325
4326         err = irdma_setup_ah(ibah, attr);
4327         if (err)
4328                 return err;
4329         err = irdma_create_hw_ah(iwdev, ah, attr->flags & RDMA_CREATE_AH_SLEEPABLE);
4330
4331         return err;
4332 }
4333
4334 /**
4335  * irdma_query_ah - Query address handle
4336  * @ibah: pointer to address handle
4337  * @ah_attr: address handle attributes
4338  */
4339 static int irdma_query_ah(struct ib_ah *ibah, struct rdma_ah_attr *ah_attr)
4340 {
4341         struct irdma_ah *ah = to_iwah(ibah);
4342
4343         memset(ah_attr, 0, sizeof(*ah_attr));
4344         if (ah->av.attrs.ah_flags & IB_AH_GRH) {
4345                 ah_attr->ah_flags = IB_AH_GRH;
4346                 ah_attr->grh.flow_label = ah->sc_ah.ah_info.flow_label;
4347                 ah_attr->grh.traffic_class = ah->sc_ah.ah_info.tc_tos;
4348                 ah_attr->grh.hop_limit = ah->sc_ah.ah_info.hop_ttl;
4349                 ah_attr->grh.sgid_index = ah->sgid_index;
4350                 ah_attr->grh.sgid_index = ah->sgid_index;
4351                 memcpy(&ah_attr->grh.dgid, &ah->dgid,
4352                        sizeof(ah_attr->grh.dgid));
4353         }
4354
4355         return 0;
4356 }
4357
4358 static enum rdma_link_layer irdma_get_link_layer(struct ib_device *ibdev,
4359                                                  u32 port_num)
4360 {
4361         return IB_LINK_LAYER_ETHERNET;
4362 }
4363
4364 static const struct ib_device_ops irdma_roce_dev_ops = {
4365         .attach_mcast = irdma_attach_mcast,
4366         .create_ah = irdma_create_ah,
4367         .create_user_ah = irdma_create_user_ah,
4368         .destroy_ah = irdma_destroy_ah,
4369         .detach_mcast = irdma_detach_mcast,
4370         .get_link_layer = irdma_get_link_layer,
4371         .get_port_immutable = irdma_roce_port_immutable,
4372         .modify_qp = irdma_modify_qp_roce,
4373         .query_ah = irdma_query_ah,
4374         .query_pkey = irdma_query_pkey,
4375 };
4376
4377 static const struct ib_device_ops irdma_iw_dev_ops = {
4378         .modify_qp = irdma_modify_qp,
4379         .get_port_immutable = irdma_iw_port_immutable,
4380         .query_gid = irdma_query_gid,
4381 };
4382
4383 static const struct ib_device_ops irdma_dev_ops = {
4384         .owner = THIS_MODULE,
4385         .driver_id = RDMA_DRIVER_IRDMA,
4386         .uverbs_abi_ver = IRDMA_ABI_VER,
4387
4388         .alloc_hw_port_stats = irdma_alloc_hw_port_stats,
4389         .alloc_mr = irdma_alloc_mr,
4390         .alloc_mw = irdma_alloc_mw,
4391         .alloc_pd = irdma_alloc_pd,
4392         .alloc_ucontext = irdma_alloc_ucontext,
4393         .create_cq = irdma_create_cq,
4394         .create_qp = irdma_create_qp,
4395         .dealloc_driver = irdma_ib_dealloc_device,
4396         .dealloc_mw = irdma_dealloc_mw,
4397         .dealloc_pd = irdma_dealloc_pd,
4398         .dealloc_ucontext = irdma_dealloc_ucontext,
4399         .dereg_mr = irdma_dereg_mr,
4400         .destroy_cq = irdma_destroy_cq,
4401         .destroy_qp = irdma_destroy_qp,
4402         .disassociate_ucontext = irdma_disassociate_ucontext,
4403         .get_dev_fw_str = irdma_get_dev_fw_str,
4404         .get_dma_mr = irdma_get_dma_mr,
4405         .get_hw_stats = irdma_get_hw_stats,
4406         .map_mr_sg = irdma_map_mr_sg,
4407         .mmap = irdma_mmap,
4408         .mmap_free = irdma_mmap_free,
4409         .poll_cq = irdma_poll_cq,
4410         .post_recv = irdma_post_recv,
4411         .post_send = irdma_post_send,
4412         .query_device = irdma_query_device,
4413         .query_port = irdma_query_port,
4414         .query_qp = irdma_query_qp,
4415         .reg_user_mr = irdma_reg_user_mr,
4416         .req_notify_cq = irdma_req_notify_cq,
4417         .resize_cq = irdma_resize_cq,
4418         INIT_RDMA_OBJ_SIZE(ib_pd, irdma_pd, ibpd),
4419         INIT_RDMA_OBJ_SIZE(ib_ucontext, irdma_ucontext, ibucontext),
4420         INIT_RDMA_OBJ_SIZE(ib_ah, irdma_ah, ibah),
4421         INIT_RDMA_OBJ_SIZE(ib_cq, irdma_cq, ibcq),
4422         INIT_RDMA_OBJ_SIZE(ib_mw, irdma_mr, ibmw),
4423         INIT_RDMA_OBJ_SIZE(ib_qp, irdma_qp, ibqp),
4424 };
4425
4426 /**
4427  * irdma_init_roce_device - initialization of roce rdma device
4428  * @iwdev: irdma device
4429  */
4430 static void irdma_init_roce_device(struct irdma_device *iwdev)
4431 {
4432         iwdev->ibdev.node_type = RDMA_NODE_IB_CA;
4433         addrconf_addr_eui48((u8 *)&iwdev->ibdev.node_guid,
4434                             iwdev->netdev->dev_addr);
4435         ib_set_device_ops(&iwdev->ibdev, &irdma_roce_dev_ops);
4436 }
4437
4438 /**
4439  * irdma_init_iw_device - initialization of iwarp rdma device
4440  * @iwdev: irdma device
4441  */
4442 static int irdma_init_iw_device(struct irdma_device *iwdev)
4443 {
4444         struct net_device *netdev = iwdev->netdev;
4445
4446         iwdev->ibdev.node_type = RDMA_NODE_RNIC;
4447         addrconf_addr_eui48((u8 *)&iwdev->ibdev.node_guid,
4448                             netdev->dev_addr);
4449         iwdev->ibdev.ops.iw_add_ref = irdma_qp_add_ref;
4450         iwdev->ibdev.ops.iw_rem_ref = irdma_qp_rem_ref;
4451         iwdev->ibdev.ops.iw_get_qp = irdma_get_qp;
4452         iwdev->ibdev.ops.iw_connect = irdma_connect;
4453         iwdev->ibdev.ops.iw_accept = irdma_accept;
4454         iwdev->ibdev.ops.iw_reject = irdma_reject;
4455         iwdev->ibdev.ops.iw_create_listen = irdma_create_listen;
4456         iwdev->ibdev.ops.iw_destroy_listen = irdma_destroy_listen;
4457         memcpy(iwdev->ibdev.iw_ifname, netdev->name,
4458                sizeof(iwdev->ibdev.iw_ifname));
4459         ib_set_device_ops(&iwdev->ibdev, &irdma_iw_dev_ops);
4460
4461         return 0;
4462 }
4463
4464 /**
4465  * irdma_init_rdma_device - initialization of rdma device
4466  * @iwdev: irdma device
4467  */
4468 static int irdma_init_rdma_device(struct irdma_device *iwdev)
4469 {
4470         struct pci_dev *pcidev = iwdev->rf->pcidev;
4471         int ret;
4472
4473         if (iwdev->roce_mode) {
4474                 irdma_init_roce_device(iwdev);
4475         } else {
4476                 ret = irdma_init_iw_device(iwdev);
4477                 if (ret)
4478                         return ret;
4479         }
4480         iwdev->ibdev.phys_port_cnt = 1;
4481         iwdev->ibdev.num_comp_vectors = iwdev->rf->ceqs_count;
4482         iwdev->ibdev.dev.parent = &pcidev->dev;
4483         ib_set_device_ops(&iwdev->ibdev, &irdma_dev_ops);
4484
4485         return 0;
4486 }
4487
4488 /**
4489  * irdma_port_ibevent - indicate port event
4490  * @iwdev: irdma device
4491  */
4492 void irdma_port_ibevent(struct irdma_device *iwdev)
4493 {
4494         struct ib_event event;
4495
4496         event.device = &iwdev->ibdev;
4497         event.element.port_num = 1;
4498         event.event =
4499                 iwdev->iw_status ? IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
4500         ib_dispatch_event(&event);
4501 }
4502
4503 /**
4504  * irdma_ib_unregister_device - unregister rdma device from IB
4505  * core
4506  * @iwdev: irdma device
4507  */
4508 void irdma_ib_unregister_device(struct irdma_device *iwdev)
4509 {
4510         iwdev->iw_status = 0;
4511         irdma_port_ibevent(iwdev);
4512         ib_unregister_device(&iwdev->ibdev);
4513 }
4514
4515 /**
4516  * irdma_ib_register_device - register irdma device to IB core
4517  * @iwdev: irdma device
4518  */
4519 int irdma_ib_register_device(struct irdma_device *iwdev)
4520 {
4521         int ret;
4522
4523         ret = irdma_init_rdma_device(iwdev);
4524         if (ret)
4525                 return ret;
4526
4527         ret = ib_device_set_netdev(&iwdev->ibdev, iwdev->netdev, 1);
4528         if (ret)
4529                 goto error;
4530         dma_set_max_seg_size(iwdev->rf->hw.device, UINT_MAX);
4531         ret = ib_register_device(&iwdev->ibdev, "irdma%d", iwdev->rf->hw.device);
4532         if (ret)
4533                 goto error;
4534
4535         iwdev->iw_status = 1;
4536         irdma_port_ibevent(iwdev);
4537
4538         return 0;
4539
4540 error:
4541         if (ret)
4542                 ibdev_dbg(&iwdev->ibdev, "VERBS: Register RDMA device fail\n");
4543
4544         return ret;
4545 }
4546
4547 /**
4548  * irdma_ib_dealloc_device
4549  * @ibdev: ib device
4550  *
4551  * callback from ibdev dealloc_driver to deallocate resources
4552  * unber irdma device
4553  */
4554 void irdma_ib_dealloc_device(struct ib_device *ibdev)
4555 {
4556         struct irdma_device *iwdev = to_iwdev(ibdev);
4557
4558         irdma_rt_deinit_hw(iwdev);
4559         irdma_ctrl_deinit_hw(iwdev->rf);
4560         kfree(iwdev->rf);
4561 }