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 = IB_DEVICE_MEM_WINDOW |
  29                                   IB_DEVICE_MEM_MGT_EXTENSIONS;
  30         props->kernel_cap_flags = IBK_LOCAL_DMA_LKEY;
  31         props->vendor_id = pcidev->vendor;
  32         props->vendor_part_id = pcidev->device;
  33
  34         props->hw_ver = rf->pcidev->revision;
  35         props->page_size_cap = hw_attrs->page_size_cap;
  36         props->max_mr_size = hw_attrs->max_mr_size;
  37         props->max_qp = rf->max_qp - rf->used_qps;
  38         props->max_qp_wr = hw_attrs->max_qp_wr;
  39         props->max_send_sge = hw_attrs->uk_attrs.max_hw_wq_frags;
  40         props->max_recv_sge = hw_attrs->uk_attrs.max_hw_wq_frags;
  41         props->max_cq = rf->max_cq - rf->used_cqs;
  42         props->max_cqe = rf->max_cqe - 1;
  43         props->max_mr = rf->max_mr - rf->used_mrs;
  44         props->max_mw = props->max_mr;
  45         props->max_pd = rf->max_pd - rf->used_pds;
  46         props->max_sge_rd = hw_attrs->uk_attrs.max_hw_read_sges;
  47         props->max_qp_rd_atom = hw_attrs->max_hw_ird;
  48         props->max_qp_init_rd_atom = hw_attrs->max_hw_ord;
  49         if (rdma_protocol_roce(ibdev, 1)) {
  50                 props->device_cap_flags |= IB_DEVICE_RC_RNR_NAK_GEN;
  51                 props->max_pkeys = IRDMA_PKEY_TBL_SZ;
  52         }
  53
  54         props->max_ah = rf->max_ah;
  55         props->max_mcast_grp = rf->max_mcg;
  56         props->max_mcast_qp_attach = IRDMA_MAX_MGS_PER_CTX;
  57         props->max_total_mcast_qp_attach = rf->max_qp * IRDMA_MAX_MGS_PER_CTX;
  58         props->max_fast_reg_page_list_len = IRDMA_MAX_PAGES_PER_FMR;
  59 #define HCA_CLOCK_TIMESTAMP_MASK 0x1ffff
  60         if (hw_attrs->uk_attrs.hw_rev >= IRDMA_GEN_2)
  61                 props->timestamp_mask = HCA_CLOCK_TIMESTAMP_MASK;
  62
  63         return 0;
  64 }
  65
  66 /**
  67  * irdma_query_port - get port attributes
  68  * @ibdev: device pointer from stack
  69  * @port: port number for query
  70  * @props: returning device attributes
  71  */
  72 static int irdma_query_port(struct ib_device *ibdev, u32 port,
  73                             struct ib_port_attr *props)
  74 {
  75         struct irdma_device *iwdev = to_iwdev(ibdev);
  76         struct net_device *netdev = iwdev->netdev;
  77
  78         /* no need to zero out pros here. done by caller */
  79
  80         props->max_mtu = IB_MTU_4096;
  81         props->active_mtu = ib_mtu_int_to_enum(netdev->mtu);
  82         props->lid = 1;
  83         props->lmc = 0;
  84         props->sm_lid = 0;
  85         props->sm_sl = 0;
  86         if (netif_carrier_ok(netdev) && netif_running(netdev)) {
  87                 props->state = IB_PORT_ACTIVE;
  88                 props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
  89         } else {
  90                 props->state = IB_PORT_DOWN;
  91                 props->phys_state = IB_PORT_PHYS_STATE_DISABLED;
  92         }
  93
  94         ib_get_eth_speed(ibdev, port, &props->active_speed,
  95                          &props->active_width);
  96
  97         if (rdma_protocol_roce(ibdev, 1)) {
  98                 props->gid_tbl_len = 32;
  99                 props->ip_gids = true;
 100                 props->pkey_tbl_len = IRDMA_PKEY_TBL_SZ;
 101         } else {
 102                 props->gid_tbl_len = 1;
 103         }
 104         props->qkey_viol_cntr = 0;
 105         props->port_cap_flags |= IB_PORT_CM_SUP | IB_PORT_REINIT_SUP;
 106         props->max_msg_sz = iwdev->rf->sc_dev.hw_attrs.max_hw_outbound_msg_size;
 107
 108         return 0;
 109 }
 110
 111 /**
 112  * irdma_disassociate_ucontext - Disassociate user context
 113  * @context: ib user context
 114  */
 115 static void irdma_disassociate_ucontext(struct ib_ucontext *context)
 116 {
 117 }
 118
 119 static int irdma_mmap_legacy(struct irdma_ucontext *ucontext,
 120                              struct vm_area_struct *vma)
 121 {
 122         u64 pfn;
 123
 124         if (vma->vm_pgoff || vma->vm_end - vma->vm_start != PAGE_SIZE)
 125                 return -EINVAL;
 126
 127         vma->vm_private_data = ucontext;
 128         pfn = ((uintptr_t)ucontext->iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET] +
 129                pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;
 130
 131         return rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn, PAGE_SIZE,
 132                                  pgprot_noncached(vma->vm_page_prot), NULL);
 133 }
 134
 135 static void irdma_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
 136 {
 137         struct irdma_user_mmap_entry *entry = to_irdma_mmap_entry(rdma_entry);
 138
 139         kfree(entry);
 140 }
 141
 142 static struct rdma_user_mmap_entry*
 143 irdma_user_mmap_entry_insert(struct irdma_ucontext *ucontext, u64 bar_offset,
 144                              enum irdma_mmap_flag mmap_flag, u64 *mmap_offset)
 145 {
 146         struct irdma_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 147         int ret;
 148
 149         if (!entry)
 150                 return NULL;
 151
 152         entry->bar_offset = bar_offset;
 153         entry->mmap_flag = mmap_flag;
 154
 155         ret = rdma_user_mmap_entry_insert(&ucontext->ibucontext,
 156                                           &entry->rdma_entry, PAGE_SIZE);
 157         if (ret) {
 158                 kfree(entry);
 159                 return NULL;
 160         }
 161         *mmap_offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
 162
 163         return &entry->rdma_entry;
 164 }
 165
 166 /**
 167  * irdma_mmap - user memory map
 168  * @context: context created during alloc
 169  * @vma: kernel info for user memory map
 170  */
 171 static int irdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
 172 {
 173         struct rdma_user_mmap_entry *rdma_entry;
 174         struct irdma_user_mmap_entry *entry;
 175         struct irdma_ucontext *ucontext;
 176         u64 pfn;
 177         int ret;
 178
 179         ucontext = to_ucontext(context);
 180
 181         /* Legacy support for libi40iw with hard-coded mmap key */
 182         if (ucontext->legacy_mode)
 183                 return irdma_mmap_legacy(ucontext, vma);
 184
 185         rdma_entry = rdma_user_mmap_entry_get(&ucontext->ibucontext, vma);
 186         if (!rdma_entry) {
 187                 ibdev_dbg(&ucontext->iwdev->ibdev,
 188                           "VERBS: pgoff[0x%lx] does not have valid entry\n",
 189                           vma->vm_pgoff);
 190                 return -EINVAL;
 191         }
 192
 193         entry = to_irdma_mmap_entry(rdma_entry);
 194         ibdev_dbg(&ucontext->iwdev->ibdev,
 195                   "VERBS: bar_offset [0x%llx] mmap_flag [%d]\n",
 196                   entry->bar_offset, entry->mmap_flag);
 197
 198         pfn = (entry->bar_offset +
 199               pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;
 200
 201         switch (entry->mmap_flag) {
 202         case IRDMA_MMAP_IO_NC:
 203                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
 204                                         pgprot_noncached(vma->vm_page_prot),
 205                                         rdma_entry);
 206                 break;
 207         case IRDMA_MMAP_IO_WC:
 208                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
 209                                         pgprot_writecombine(vma->vm_page_prot),
 210                                         rdma_entry);
 211                 break;
 212         default:
 213                 ret = -EINVAL;
 214         }
 215
 216         if (ret)
 217                 ibdev_dbg(&ucontext->iwdev->ibdev,
 218                           "VERBS: bar_offset [0x%llx] mmap_flag[%d] err[%d]\n",
 219                           entry->bar_offset, entry->mmap_flag, ret);
 220         rdma_user_mmap_entry_put(rdma_entry);
 221
 222         return ret;
 223 }
 224
 225 /**
 226  * irdma_alloc_push_page - allocate a push page for qp
 227  * @iwqp: qp pointer
 228  */
 229 static void irdma_alloc_push_page(struct irdma_qp *iwqp)
 230 {
 231         struct irdma_cqp_request *cqp_request;
 232         struct cqp_cmds_info *cqp_info;
 233         struct irdma_device *iwdev = iwqp->iwdev;
 234         struct irdma_sc_qp *qp = &iwqp->sc_qp;
 235         int status;
 236
 237         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
 238         if (!cqp_request)
 239                 return;
 240
 241         cqp_info = &cqp_request->info;
 242         cqp_info->cqp_cmd = IRDMA_OP_MANAGE_PUSH_PAGE;
 243         cqp_info->post_sq = 1;
 244         cqp_info->in.u.manage_push_page.info.push_idx = 0;
 245         cqp_info->in.u.manage_push_page.info.qs_handle =
 246                 qp->vsi->qos[qp->user_pri].qs_handle;
 247         cqp_info->in.u.manage_push_page.info.free_page = 0;
 248         cqp_info->in.u.manage_push_page.info.push_page_type = 0;
 249         cqp_info->in.u.manage_push_page.cqp = &iwdev->rf->cqp.sc_cqp;
 250         cqp_info->in.u.manage_push_page.scratch = (uintptr_t)cqp_request;
 251
 252         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
 253         if (!status && cqp_request->compl_info.op_ret_val <
 254             iwdev->rf->sc_dev.hw_attrs.max_hw_device_pages) {
 255                 qp->push_idx = cqp_request->compl_info.op_ret_val;
 256                 qp->push_offset = 0;
 257         }
 258
 259         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
 260 }
 261
 262 /**
 263  * irdma_alloc_ucontext - Allocate the user context data structure
 264  * @uctx: uverbs context pointer
 265  * @udata: user data
 266  *
 267  * This keeps track of all objects associated with a particular
 268  * user-mode client.
 269  */
 270 static int irdma_alloc_ucontext(struct ib_ucontext *uctx,
 271                                 struct ib_udata *udata)
 272 {
 273 #define IRDMA_ALLOC_UCTX_MIN_REQ_LEN offsetofend(struct irdma_alloc_ucontext_req, rsvd8)
 274 #define IRDMA_ALLOC_UCTX_MIN_RESP_LEN offsetofend(struct irdma_alloc_ucontext_resp, rsvd)
 275         struct ib_device *ibdev = uctx->device;
 276         struct irdma_device *iwdev = to_iwdev(ibdev);
 277         struct irdma_alloc_ucontext_req req = {};
 278         struct irdma_alloc_ucontext_resp uresp = {};
 279         struct irdma_ucontext *ucontext = to_ucontext(uctx);
 280         struct irdma_uk_attrs *uk_attrs;
 281
 282         if (udata->inlen < IRDMA_ALLOC_UCTX_MIN_REQ_LEN ||
 283             udata->outlen < IRDMA_ALLOC_UCTX_MIN_RESP_LEN)
 284                 return -EINVAL;
 285
 286         if (ib_copy_from_udata(&req, udata, min(sizeof(req), udata->inlen)))
 287                 return -EINVAL;
 288
 289         if (req.userspace_ver < 4 || req.userspace_ver > IRDMA_ABI_VER)
 290                 goto ver_error;
 291
 292         ucontext->iwdev = iwdev;
 293         ucontext->abi_ver = req.userspace_ver;
 294
 295         uk_attrs = &iwdev->rf->sc_dev.hw_attrs.uk_attrs;
 296         /* GEN_1 legacy support with libi40iw */
 297         if (udata->outlen == IRDMA_ALLOC_UCTX_MIN_RESP_LEN) {
 298                 if (uk_attrs->hw_rev != IRDMA_GEN_1)
 299                         return -EOPNOTSUPP;
 300
 301                 ucontext->legacy_mode = true;
 302                 uresp.max_qps = iwdev->rf->max_qp;
 303                 uresp.max_pds = iwdev->rf->sc_dev.hw_attrs.max_hw_pds;
 304                 uresp.wq_size = iwdev->rf->sc_dev.hw_attrs.max_qp_wr * 2;
 305                 uresp.kernel_ver = req.userspace_ver;
 306                 if (ib_copy_to_udata(udata, &uresp,
 307                                      min(sizeof(uresp), udata->outlen)))
 308                         return -EFAULT;
 309         } else {
 310                 u64 bar_off = (uintptr_t)iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET];
 311
 312                 ucontext->db_mmap_entry =
 313                         irdma_user_mmap_entry_insert(ucontext, bar_off,
 314                                                      IRDMA_MMAP_IO_NC,
 315                                                      &uresp.db_mmap_key);
 316                 if (!ucontext->db_mmap_entry)
 317                         return -ENOMEM;
 318
 319                 uresp.kernel_ver = IRDMA_ABI_VER;
 320                 uresp.feature_flags = uk_attrs->feature_flags;
 321                 uresp.max_hw_wq_frags = uk_attrs->max_hw_wq_frags;
 322                 uresp.max_hw_read_sges = uk_attrs->max_hw_read_sges;
 323                 uresp.max_hw_inline = uk_attrs->max_hw_inline;
 324                 uresp.max_hw_rq_quanta = uk_attrs->max_hw_rq_quanta;
 325                 uresp.max_hw_wq_quanta = uk_attrs->max_hw_wq_quanta;
 326                 uresp.max_hw_sq_chunk = uk_attrs->max_hw_sq_chunk;
 327                 uresp.max_hw_cq_size = uk_attrs->max_hw_cq_size;
 328                 uresp.min_hw_cq_size = uk_attrs->min_hw_cq_size;
 329                 uresp.hw_rev = uk_attrs->hw_rev;
 330                 if (ib_copy_to_udata(udata, &uresp,
 331                                      min(sizeof(uresp), udata->outlen))) {
 332                         rdma_user_mmap_entry_remove(ucontext->db_mmap_entry);
 333                         return -EFAULT;
 334                 }
 335         }
 336
 337         INIT_LIST_HEAD(&ucontext->cq_reg_mem_list);
 338         spin_lock_init(&ucontext->cq_reg_mem_list_lock);
 339         INIT_LIST_HEAD(&ucontext->qp_reg_mem_list);
 340         spin_lock_init(&ucontext->qp_reg_mem_list_lock);
 341
 342         return 0;
 343
 344 ver_error:
 345         ibdev_err(&iwdev->ibdev,
 346                   "Invalid userspace driver version detected. Detected version %d, should be %d\n",
 347                   req.userspace_ver, IRDMA_ABI_VER);
 348         return -EINVAL;
 349 }
 350
 351 /**
 352  * irdma_dealloc_ucontext - deallocate the user context data structure
 353  * @context: user context created during alloc
 354  */
 355 static void irdma_dealloc_ucontext(struct ib_ucontext *context)
 356 {
 357         struct irdma_ucontext *ucontext = to_ucontext(context);
 358
 359         rdma_user_mmap_entry_remove(ucontext->db_mmap_entry);
 360 }
 361
 362 /**
 363  * irdma_alloc_pd - allocate protection domain
 364  * @pd: PD pointer
 365  * @udata: user data
 366  */
 367 static int irdma_alloc_pd(struct ib_pd *pd, struct ib_udata *udata)
 368 {
 369 #define IRDMA_ALLOC_PD_MIN_RESP_LEN offsetofend(struct irdma_alloc_pd_resp, rsvd)
 370         struct irdma_pd *iwpd = to_iwpd(pd);
 371         struct irdma_device *iwdev = to_iwdev(pd->device);
 372         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 373         struct irdma_pci_f *rf = iwdev->rf;
 374         struct irdma_alloc_pd_resp uresp = {};
 375         struct irdma_sc_pd *sc_pd;
 376         u32 pd_id = 0;
 377         int err;
 378
 379         if (udata && udata->outlen < IRDMA_ALLOC_PD_MIN_RESP_LEN)
 380                 return -EINVAL;
 381
 382         err = irdma_alloc_rsrc(rf, rf->allocated_pds, rf->max_pd, &pd_id,
 383                                &rf->next_pd);
 384         if (err)
 385                 return err;
 386
 387         sc_pd = &iwpd->sc_pd;
 388         if (udata) {
 389                 struct irdma_ucontext *ucontext =
 390                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
 391                                                   ibucontext);
 392                 irdma_sc_pd_init(dev, sc_pd, pd_id, ucontext->abi_ver);
 393                 uresp.pd_id = pd_id;
 394                 if (ib_copy_to_udata(udata, &uresp,
 395                                      min(sizeof(uresp), udata->outlen))) {
 396                         err = -EFAULT;
 397                         goto error;
 398                 }
 399         } else {
 400                 irdma_sc_pd_init(dev, sc_pd, pd_id, IRDMA_ABI_VER);
 401         }
 402
 403         return 0;
 404 error:
 405         irdma_free_rsrc(rf, rf->allocated_pds, pd_id);
 406
 407         return err;
 408 }
 409
 410 /**
 411  * irdma_dealloc_pd - deallocate pd
 412  * @ibpd: ptr of pd to be deallocated
 413  * @udata: user data
 414  */
 415 static int irdma_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 416 {
 417         struct irdma_pd *iwpd = to_iwpd(ibpd);
 418         struct irdma_device *iwdev = to_iwdev(ibpd->device);
 419
 420         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_pds, iwpd->sc_pd.pd_id);
 421
 422         return 0;
 423 }
 424
 425 /**
 426  * irdma_get_pbl - Retrieve pbl from a list given a virtual
 427  * address
 428  * @va: user virtual address
 429  * @pbl_list: pbl list to search in (QP's or CQ's)
 430  */
 431 static struct irdma_pbl *irdma_get_pbl(unsigned long va,
 432                                        struct list_head *pbl_list)
 433 {
 434         struct irdma_pbl *iwpbl;
 435
 436         list_for_each_entry (iwpbl, pbl_list, list) {
 437                 if (iwpbl->user_base == va) {
 438                         list_del(&iwpbl->list);
 439                         iwpbl->on_list = false;
 440                         return iwpbl;
 441                 }
 442         }
 443
 444         return NULL;
 445 }
 446
 447 /**
 448  * irdma_clean_cqes - clean cq entries for qp
 449  * @iwqp: qp ptr (user or kernel)
 450  * @iwcq: cq ptr
 451  */
 452 static void irdma_clean_cqes(struct irdma_qp *iwqp, struct irdma_cq *iwcq)
 453 {
 454         struct irdma_cq_uk *ukcq = &iwcq->sc_cq.cq_uk;
 455         unsigned long flags;
 456
 457         spin_lock_irqsave(&iwcq->lock, flags);
 458         irdma_uk_clean_cq(&iwqp->sc_qp.qp_uk, ukcq);
 459         spin_unlock_irqrestore(&iwcq->lock, flags);
 460 }
 461
 462 static void irdma_remove_push_mmap_entries(struct irdma_qp *iwqp)
 463 {
 464         if (iwqp->push_db_mmap_entry) {
 465                 rdma_user_mmap_entry_remove(iwqp->push_db_mmap_entry);
 466                 iwqp->push_db_mmap_entry = NULL;
 467         }
 468         if (iwqp->push_wqe_mmap_entry) {
 469                 rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
 470                 iwqp->push_wqe_mmap_entry = NULL;
 471         }
 472 }
 473
 474 static int irdma_setup_push_mmap_entries(struct irdma_ucontext *ucontext,
 475                                          struct irdma_qp *iwqp,
 476                                          u64 *push_wqe_mmap_key,
 477                                          u64 *push_db_mmap_key)
 478 {
 479         struct irdma_device *iwdev = ucontext->iwdev;
 480         u64 rsvd, bar_off;
 481
 482         rsvd = IRDMA_PF_BAR_RSVD;
 483         bar_off = (uintptr_t)iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET];
 484         /* skip over db page */
 485         bar_off += IRDMA_HW_PAGE_SIZE;
 486         /* push wqe page */
 487         bar_off += rsvd + iwqp->sc_qp.push_idx * IRDMA_HW_PAGE_SIZE;
 488         iwqp->push_wqe_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
 489                                         bar_off, IRDMA_MMAP_IO_WC,
 490                                         push_wqe_mmap_key);
 491         if (!iwqp->push_wqe_mmap_entry)
 492                 return -ENOMEM;
 493
 494         /* push doorbell page */
 495         bar_off += IRDMA_HW_PAGE_SIZE;
 496         iwqp->push_db_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
 497                                         bar_off, IRDMA_MMAP_IO_NC,
 498                                         push_db_mmap_key);
 499         if (!iwqp->push_db_mmap_entry) {
 500                 rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
 501                 return -ENOMEM;
 502         }
 503
 504         return 0;
 505 }
 506
 507 /**
 508  * irdma_destroy_qp - destroy qp
 509  * @ibqp: qp's ib pointer also to get to device's qp address
 510  * @udata: user data
 511  */
 512 static int irdma_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
 513 {
 514         struct irdma_qp *iwqp = to_iwqp(ibqp);
 515         struct irdma_device *iwdev = iwqp->iwdev;
 516
 517         iwqp->sc_qp.qp_uk.destroy_pending = true;
 518
 519         if (iwqp->iwarp_state == IRDMA_QP_STATE_RTS)
 520                 irdma_modify_qp_to_err(&iwqp->sc_qp);
 521
 522         if (!iwqp->user_mode)
 523                 cancel_delayed_work_sync(&iwqp->dwork_flush);
 524
 525         irdma_qp_rem_ref(&iwqp->ibqp);
 526         wait_for_completion(&iwqp->free_qp);
 527         irdma_free_lsmm_rsrc(iwqp);
 528         irdma_cqp_qp_destroy_cmd(&iwdev->rf->sc_dev, &iwqp->sc_qp);
 529
 530         if (!iwqp->user_mode) {
 531                 if (iwqp->iwscq) {
 532                         irdma_clean_cqes(iwqp, iwqp->iwscq);
 533                         if (iwqp->iwrcq != iwqp->iwscq)
 534                                 irdma_clean_cqes(iwqp, iwqp->iwrcq);
 535                 }
 536         }
 537         irdma_remove_push_mmap_entries(iwqp);
 538         irdma_free_qp_rsrc(iwqp);
 539
 540         return 0;
 541 }
 542
 543 /**
 544  * irdma_setup_virt_qp - setup for allocation of virtual qp
 545  * @iwdev: irdma device
 546  * @iwqp: qp ptr
 547  * @init_info: initialize info to return
 548  */
 549 static void irdma_setup_virt_qp(struct irdma_device *iwdev,
 550                                struct irdma_qp *iwqp,
 551                                struct irdma_qp_init_info *init_info)
 552 {
 553         struct irdma_pbl *iwpbl = iwqp->iwpbl;
 554         struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;
 555
 556         iwqp->page = qpmr->sq_page;
 557         init_info->shadow_area_pa = qpmr->shadow;
 558         if (iwpbl->pbl_allocated) {
 559                 init_info->virtual_map = true;
 560                 init_info->sq_pa = qpmr->sq_pbl.idx;
 561                 init_info->rq_pa = qpmr->rq_pbl.idx;
 562         } else {
 563                 init_info->sq_pa = qpmr->sq_pbl.addr;
 564                 init_info->rq_pa = qpmr->rq_pbl.addr;
 565         }
 566 }
 567
 568 /**
 569  * irdma_setup_kmode_qp - setup initialization for kernel mode qp
 570  * @iwdev: iwarp device
 571  * @iwqp: qp ptr (user or kernel)
 572  * @info: initialize info to return
 573  * @init_attr: Initial QP create attributes
 574  */
 575 static int irdma_setup_kmode_qp(struct irdma_device *iwdev,
 576                                 struct irdma_qp *iwqp,
 577                                 struct irdma_qp_init_info *info,
 578                                 struct ib_qp_init_attr *init_attr)
 579 {
 580         struct irdma_dma_mem *mem = &iwqp->kqp.dma_mem;
 581         u32 sqdepth, rqdepth;
 582         u8 sqshift, rqshift;
 583         u32 size;
 584         int status;
 585         struct irdma_qp_uk_init_info *ukinfo = &info->qp_uk_init_info;
 586         struct irdma_uk_attrs *uk_attrs = &iwdev->rf->sc_dev.hw_attrs.uk_attrs;
 587
 588         irdma_get_wqe_shift(uk_attrs,
 589                 uk_attrs->hw_rev >= IRDMA_GEN_2 ? ukinfo->max_sq_frag_cnt + 1 :
 590                                                   ukinfo->max_sq_frag_cnt,
 591                 ukinfo->max_inline_data, &sqshift);
 592         status = irdma_get_sqdepth(uk_attrs, ukinfo->sq_size, sqshift,
 593                                    &sqdepth);
 594         if (status)
 595                 return status;
 596
 597         if (uk_attrs->hw_rev == IRDMA_GEN_1)
 598                 rqshift = IRDMA_MAX_RQ_WQE_SHIFT_GEN1;
 599         else
 600                 irdma_get_wqe_shift(uk_attrs, ukinfo->max_rq_frag_cnt, 0,
 601                                     &rqshift);
 602
 603         status = irdma_get_rqdepth(uk_attrs, ukinfo->rq_size, rqshift,
 604                                    &rqdepth);
 605         if (status)
 606                 return status;
 607
 608         iwqp->kqp.sq_wrid_mem =
 609                 kcalloc(sqdepth, sizeof(*iwqp->kqp.sq_wrid_mem), GFP_KERNEL);
 610         if (!iwqp->kqp.sq_wrid_mem)
 611                 return -ENOMEM;
 612
 613         iwqp->kqp.rq_wrid_mem =
 614                 kcalloc(rqdepth, sizeof(*iwqp->kqp.rq_wrid_mem), GFP_KERNEL);
 615         if (!iwqp->kqp.rq_wrid_mem) {
 616                 kfree(iwqp->kqp.sq_wrid_mem);
 617                 iwqp->kqp.sq_wrid_mem = NULL;
 618                 return -ENOMEM;
 619         }
 620
 621         ukinfo->sq_wrtrk_array = iwqp->kqp.sq_wrid_mem;
 622         ukinfo->rq_wrid_array = iwqp->kqp.rq_wrid_mem;
 623
 624         size = (sqdepth + rqdepth) * IRDMA_QP_WQE_MIN_SIZE;
 625         size += (IRDMA_SHADOW_AREA_SIZE << 3);
 626
 627         mem->size = ALIGN(size, 256);
 628         mem->va = dma_alloc_coherent(iwdev->rf->hw.device, mem->size,
 629                                      &mem->pa, GFP_KERNEL);
 630         if (!mem->va) {
 631                 kfree(iwqp->kqp.sq_wrid_mem);
 632                 iwqp->kqp.sq_wrid_mem = NULL;
 633                 kfree(iwqp->kqp.rq_wrid_mem);
 634                 iwqp->kqp.rq_wrid_mem = NULL;
 635                 return -ENOMEM;
 636         }
 637
 638         ukinfo->sq = mem->va;
 639         info->sq_pa = mem->pa;
 640         ukinfo->rq = &ukinfo->sq[sqdepth];
 641         info->rq_pa = info->sq_pa + (sqdepth * IRDMA_QP_WQE_MIN_SIZE);
 642         ukinfo->shadow_area = ukinfo->rq[rqdepth].elem;
 643         info->shadow_area_pa = info->rq_pa + (rqdepth * IRDMA_QP_WQE_MIN_SIZE);
 644         ukinfo->sq_size = sqdepth >> sqshift;
 645         ukinfo->rq_size = rqdepth >> rqshift;
 646         ukinfo->qp_id = iwqp->ibqp.qp_num;
 647
 648         init_attr->cap.max_send_wr = (sqdepth - IRDMA_SQ_RSVD) >> sqshift;
 649         init_attr->cap.max_recv_wr = (rqdepth - IRDMA_RQ_RSVD) >> rqshift;
 650
 651         return 0;
 652 }
 653
 654 static int irdma_cqp_create_qp_cmd(struct irdma_qp *iwqp)
 655 {
 656         struct irdma_pci_f *rf = iwqp->iwdev->rf;
 657         struct irdma_cqp_request *cqp_request;
 658         struct cqp_cmds_info *cqp_info;
 659         struct irdma_create_qp_info *qp_info;
 660         int status;
 661
 662         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
 663         if (!cqp_request)
 664                 return -ENOMEM;
 665
 666         cqp_info = &cqp_request->info;
 667         qp_info = &cqp_request->info.in.u.qp_create.info;
 668         memset(qp_info, 0, sizeof(*qp_info));
 669         qp_info->mac_valid = true;
 670         qp_info->cq_num_valid = true;
 671         qp_info->next_iwarp_state = IRDMA_QP_STATE_IDLE;
 672
 673         cqp_info->cqp_cmd = IRDMA_OP_QP_CREATE;
 674         cqp_info->post_sq = 1;
 675         cqp_info->in.u.qp_create.qp = &iwqp->sc_qp;
 676         cqp_info->in.u.qp_create.scratch = (uintptr_t)cqp_request;
 677         status = irdma_handle_cqp_op(rf, cqp_request);
 678         irdma_put_cqp_request(&rf->cqp, cqp_request);
 679
 680         return status;
 681 }
 682
 683 static void irdma_roce_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
 684                                                struct irdma_qp_host_ctx_info *ctx_info)
 685 {
 686         struct irdma_device *iwdev = iwqp->iwdev;
 687         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 688         struct irdma_roce_offload_info *roce_info;
 689         struct irdma_udp_offload_info *udp_info;
 690
 691         udp_info = &iwqp->udp_info;
 692         udp_info->snd_mss = ib_mtu_enum_to_int(ib_mtu_int_to_enum(iwdev->vsi.mtu));
 693         udp_info->cwnd = iwdev->roce_cwnd;
 694         udp_info->rexmit_thresh = 2;
 695         udp_info->rnr_nak_thresh = 2;
 696         udp_info->src_port = 0xc000;
 697         udp_info->dst_port = ROCE_V2_UDP_DPORT;
 698         roce_info = &iwqp->roce_info;
 699         ether_addr_copy(roce_info->mac_addr, iwdev->netdev->dev_addr);
 700
 701         roce_info->rd_en = true;
 702         roce_info->wr_rdresp_en = true;
 703         roce_info->bind_en = true;
 704         roce_info->dcqcn_en = false;
 705         roce_info->rtomin = 5;
 706
 707         roce_info->ack_credits = iwdev->roce_ackcreds;
 708         roce_info->ird_size = dev->hw_attrs.max_hw_ird;
 709         roce_info->ord_size = dev->hw_attrs.max_hw_ord;
 710
 711         if (!iwqp->user_mode) {
 712                 roce_info->priv_mode_en = true;
 713                 roce_info->fast_reg_en = true;
 714                 roce_info->udprivcq_en = true;
 715         }
 716         roce_info->roce_tver = 0;
 717
 718         ctx_info->roce_info = &iwqp->roce_info;
 719         ctx_info->udp_info = &iwqp->udp_info;
 720         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
 721 }
 722
 723 static void irdma_iw_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
 724                                              struct irdma_qp_host_ctx_info *ctx_info)
 725 {
 726         struct irdma_device *iwdev = iwqp->iwdev;
 727         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 728         struct irdma_iwarp_offload_info *iwarp_info;
 729
 730         iwarp_info = &iwqp->iwarp_info;
 731         ether_addr_copy(iwarp_info->mac_addr, iwdev->netdev->dev_addr);
 732         iwarp_info->rd_en = true;
 733         iwarp_info->wr_rdresp_en = true;
 734         iwarp_info->bind_en = true;
 735         iwarp_info->ecn_en = true;
 736         iwarp_info->rtomin = 5;
 737
 738         if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
 739                 iwarp_info->ib_rd_en = true;
 740         if (!iwqp->user_mode) {
 741                 iwarp_info->priv_mode_en = true;
 742                 iwarp_info->fast_reg_en = true;
 743         }
 744         iwarp_info->ddp_ver = 1;
 745         iwarp_info->rdmap_ver = 1;
 746
 747         ctx_info->iwarp_info = &iwqp->iwarp_info;
 748         ctx_info->iwarp_info_valid = true;
 749         irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
 750         ctx_info->iwarp_info_valid = false;
 751 }
 752
 753 static int irdma_validate_qp_attrs(struct ib_qp_init_attr *init_attr,
 754                                    struct irdma_device *iwdev)
 755 {
 756         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 757         struct irdma_uk_attrs *uk_attrs = &dev->hw_attrs.uk_attrs;
 758
 759         if (init_attr->create_flags)
 760                 return -EOPNOTSUPP;
 761
 762         if (init_attr->cap.max_inline_data > uk_attrs->max_hw_inline ||
 763             init_attr->cap.max_send_sge > uk_attrs->max_hw_wq_frags ||
 764             init_attr->cap.max_recv_sge > uk_attrs->max_hw_wq_frags)
 765                 return -EINVAL;
 766
 767         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 768                 if (init_attr->qp_type != IB_QPT_RC &&
 769                     init_attr->qp_type != IB_QPT_UD &&
 770                     init_attr->qp_type != IB_QPT_GSI)
 771                         return -EOPNOTSUPP;
 772         } else {
 773                 if (init_attr->qp_type != IB_QPT_RC)
 774                         return -EOPNOTSUPP;
 775         }
 776
 777         return 0;
 778 }
 779
 780 static void irdma_flush_worker(struct work_struct *work)
 781 {
 782         struct delayed_work *dwork = to_delayed_work(work);
 783         struct irdma_qp *iwqp = container_of(dwork, struct irdma_qp, dwork_flush);
 784
 785         irdma_generate_flush_completions(iwqp);
 786 }
 787
 788 /**
 789  * irdma_create_qp - create qp
 790  * @ibqp: ptr of qp
 791  * @init_attr: attributes for qp
 792  * @udata: user data for create qp
 793  */
 794 static int irdma_create_qp(struct ib_qp *ibqp,
 795                            struct ib_qp_init_attr *init_attr,
 796                            struct ib_udata *udata)
 797 {
 798 #define IRDMA_CREATE_QP_MIN_REQ_LEN offsetofend(struct irdma_create_qp_req, user_compl_ctx)
 799 #define IRDMA_CREATE_QP_MIN_RESP_LEN offsetofend(struct irdma_create_qp_resp, rsvd)
 800         struct ib_pd *ibpd = ibqp->pd;
 801         struct irdma_pd *iwpd = to_iwpd(ibpd);
 802         struct irdma_device *iwdev = to_iwdev(ibpd->device);
 803         struct irdma_pci_f *rf = iwdev->rf;
 804         struct irdma_qp *iwqp = to_iwqp(ibqp);
 805         struct irdma_create_qp_req req = {};
 806         struct irdma_create_qp_resp uresp = {};
 807         u32 qp_num = 0;
 808         int err_code;
 809         int sq_size;
 810         int rq_size;
 811         struct irdma_sc_qp *qp;
 812         struct irdma_sc_dev *dev = &rf->sc_dev;
 813         struct irdma_uk_attrs *uk_attrs = &dev->hw_attrs.uk_attrs;
 814         struct irdma_qp_init_info init_info = {};
 815         struct irdma_qp_host_ctx_info *ctx_info;
 816         unsigned long flags;
 817
 818         err_code = irdma_validate_qp_attrs(init_attr, iwdev);
 819         if (err_code)
 820                 return err_code;
 821
 822         if (udata && (udata->inlen < IRDMA_CREATE_QP_MIN_REQ_LEN ||
 823                       udata->outlen < IRDMA_CREATE_QP_MIN_RESP_LEN))
 824                 return -EINVAL;
 825
 826         sq_size = init_attr->cap.max_send_wr;
 827         rq_size = init_attr->cap.max_recv_wr;
 828
 829         init_info.vsi = &iwdev->vsi;
 830         init_info.qp_uk_init_info.uk_attrs = uk_attrs;
 831         init_info.qp_uk_init_info.sq_size = sq_size;
 832         init_info.qp_uk_init_info.rq_size = rq_size;
 833         init_info.qp_uk_init_info.max_sq_frag_cnt = init_attr->cap.max_send_sge;
 834         init_info.qp_uk_init_info.max_rq_frag_cnt = init_attr->cap.max_recv_sge;
 835         init_info.qp_uk_init_info.max_inline_data = init_attr->cap.max_inline_data;
 836
 837         qp = &iwqp->sc_qp;
 838         qp->qp_uk.back_qp = iwqp;
 839         qp->push_idx = IRDMA_INVALID_PUSH_PAGE_INDEX;
 840
 841         iwqp->iwdev = iwdev;
 842         iwqp->q2_ctx_mem.size = ALIGN(IRDMA_Q2_BUF_SIZE + IRDMA_QP_CTX_SIZE,
 843                                       256);
 844         iwqp->q2_ctx_mem.va = dma_alloc_coherent(dev->hw->device,
 845                                                  iwqp->q2_ctx_mem.size,
 846                                                  &iwqp->q2_ctx_mem.pa,
 847                                                  GFP_KERNEL);
 848         if (!iwqp->q2_ctx_mem.va)
 849                 return -ENOMEM;
 850
 851         init_info.q2 = iwqp->q2_ctx_mem.va;
 852         init_info.q2_pa = iwqp->q2_ctx_mem.pa;
 853         init_info.host_ctx = (__le64 *)(init_info.q2 + IRDMA_Q2_BUF_SIZE);
 854         init_info.host_ctx_pa = init_info.q2_pa + IRDMA_Q2_BUF_SIZE;
 855
 856         if (init_attr->qp_type == IB_QPT_GSI)
 857                 qp_num = 1;
 858         else
 859                 err_code = irdma_alloc_rsrc(rf, rf->allocated_qps, rf->max_qp,
 860                                             &qp_num, &rf->next_qp);
 861         if (err_code)
 862                 goto error;
 863
 864         iwqp->iwpd = iwpd;
 865         iwqp->ibqp.qp_num = qp_num;
 866         qp = &iwqp->sc_qp;
 867         iwqp->iwscq = to_iwcq(init_attr->send_cq);
 868         iwqp->iwrcq = to_iwcq(init_attr->recv_cq);
 869         iwqp->host_ctx.va = init_info.host_ctx;
 870         iwqp->host_ctx.pa = init_info.host_ctx_pa;
 871         iwqp->host_ctx.size = IRDMA_QP_CTX_SIZE;
 872
 873         init_info.pd = &iwpd->sc_pd;
 874         init_info.qp_uk_init_info.qp_id = iwqp->ibqp.qp_num;
 875         if (!rdma_protocol_roce(&iwdev->ibdev, 1))
 876                 init_info.qp_uk_init_info.first_sq_wq = 1;
 877         iwqp->ctx_info.qp_compl_ctx = (uintptr_t)qp;
 878         init_waitqueue_head(&iwqp->waitq);
 879         init_waitqueue_head(&iwqp->mod_qp_waitq);
 880
 881         if (udata) {
 882                 err_code = ib_copy_from_udata(&req, udata,
 883                                               min(sizeof(req), udata->inlen));
 884                 if (err_code) {
 885                         ibdev_dbg(&iwdev->ibdev,
 886                                   "VERBS: ib_copy_from_data fail\n");
 887                         goto error;
 888                 }
 889
 890                 iwqp->ctx_info.qp_compl_ctx = req.user_compl_ctx;
 891                 iwqp->user_mode = 1;
 892                 if (req.user_wqe_bufs) {
 893                         struct irdma_ucontext *ucontext =
 894                                 rdma_udata_to_drv_context(udata,
 895                                                           struct irdma_ucontext,
 896                                                           ibucontext);
 897
 898                         init_info.qp_uk_init_info.legacy_mode = ucontext->legacy_mode;
 899                         spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
 900                         iwqp->iwpbl = irdma_get_pbl((unsigned long)req.user_wqe_bufs,
 901                                                     &ucontext->qp_reg_mem_list);
 902                         spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
 903
 904                         if (!iwqp->iwpbl) {
 905                                 err_code = -ENODATA;
 906                                 ibdev_dbg(&iwdev->ibdev, "VERBS: no pbl info\n");
 907                                 goto error;
 908                         }
 909                 }
 910                 init_info.qp_uk_init_info.abi_ver = iwpd->sc_pd.abi_ver;
 911                 irdma_setup_virt_qp(iwdev, iwqp, &init_info);
 912         } else {
 913                 INIT_DELAYED_WORK(&iwqp->dwork_flush, irdma_flush_worker);
 914                 init_info.qp_uk_init_info.abi_ver = IRDMA_ABI_VER;
 915                 err_code = irdma_setup_kmode_qp(iwdev, iwqp, &init_info, init_attr);
 916         }
 917
 918         if (err_code) {
 919                 ibdev_dbg(&iwdev->ibdev, "VERBS: setup qp failed\n");
 920                 goto error;
 921         }
 922
 923         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 924                 if (init_attr->qp_type == IB_QPT_RC) {
 925                         init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_ROCE_RC;
 926                         init_info.qp_uk_init_info.qp_caps = IRDMA_SEND_WITH_IMM |
 927                                                             IRDMA_WRITE_WITH_IMM |
 928                                                             IRDMA_ROCE;
 929                 } else {
 930                         init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_ROCE_UD;
 931                         init_info.qp_uk_init_info.qp_caps = IRDMA_SEND_WITH_IMM |
 932                                                             IRDMA_ROCE;
 933                 }
 934         } else {
 935                 init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_IWARP;
 936                 init_info.qp_uk_init_info.qp_caps = IRDMA_WRITE_WITH_IMM;
 937         }
 938
 939         if (dev->hw_attrs.uk_attrs.hw_rev > IRDMA_GEN_1)
 940                 init_info.qp_uk_init_info.qp_caps |= IRDMA_PUSH_MODE;
 941
 942         err_code = irdma_sc_qp_init(qp, &init_info);
 943         if (err_code) {
 944                 ibdev_dbg(&iwdev->ibdev, "VERBS: qp_init fail\n");
 945                 goto error;
 946         }
 947
 948         ctx_info = &iwqp->ctx_info;
 949         ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
 950         ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
 951
 952         if (rdma_protocol_roce(&iwdev->ibdev, 1))
 953                 irdma_roce_fill_and_set_qpctx_info(iwqp, ctx_info);
 954         else
 955                 irdma_iw_fill_and_set_qpctx_info(iwqp, ctx_info);
 956
 957         err_code = irdma_cqp_create_qp_cmd(iwqp);
 958         if (err_code)
 959                 goto error;
 960
 961         refcount_set(&iwqp->refcnt, 1);
 962         spin_lock_init(&iwqp->lock);
 963         spin_lock_init(&iwqp->sc_qp.pfpdu.lock);
 964         iwqp->sig_all = (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ? 1 : 0;
 965         rf->qp_table[qp_num] = iwqp;
 966         iwqp->max_send_wr = sq_size;
 967         iwqp->max_recv_wr = rq_size;
 968
 969         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 970                 if (dev->ws_add(&iwdev->vsi, 0)) {
 971                         irdma_cqp_qp_destroy_cmd(&rf->sc_dev, &iwqp->sc_qp);
 972                         err_code = -EINVAL;
 973                         goto error;
 974                 }
 975
 976                 irdma_qp_add_qos(&iwqp->sc_qp);
 977         }
 978
 979         if (udata) {
 980                 /* GEN_1 legacy support with libi40iw does not have expanded uresp struct */
 981                 if (udata->outlen < sizeof(uresp)) {
 982                         uresp.lsmm = 1;
 983                         uresp.push_idx = IRDMA_INVALID_PUSH_PAGE_INDEX_GEN_1;
 984                 } else {
 985                         if (rdma_protocol_iwarp(&iwdev->ibdev, 1))
 986                                 uresp.lsmm = 1;
 987                 }
 988                 uresp.actual_sq_size = sq_size;
 989                 uresp.actual_rq_size = rq_size;
 990                 uresp.qp_id = qp_num;
 991                 uresp.qp_caps = qp->qp_uk.qp_caps;
 992
 993                 err_code = ib_copy_to_udata(udata, &uresp,
 994                                             min(sizeof(uresp), udata->outlen));
 995                 if (err_code) {
 996                         ibdev_dbg(&iwdev->ibdev, "VERBS: copy_to_udata failed\n");
 997                         irdma_destroy_qp(&iwqp->ibqp, udata);
 998                         return err_code;
 999                 }
1000         }
1001
1002         init_completion(&iwqp->free_qp);
1003         return 0;
1004
1005 error:
1006         irdma_free_qp_rsrc(iwqp);
1007         return err_code;
1008 }
1009
1010 static int irdma_get_ib_acc_flags(struct irdma_qp *iwqp)
1011 {
1012         int acc_flags = 0;
1013
1014         if (rdma_protocol_roce(iwqp->ibqp.device, 1)) {
1015                 if (iwqp->roce_info.wr_rdresp_en) {
1016                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
1017                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
1018                 }
1019                 if (iwqp->roce_info.rd_en)
1020                         acc_flags |= IB_ACCESS_REMOTE_READ;
1021                 if (iwqp->roce_info.bind_en)
1022                         acc_flags |= IB_ACCESS_MW_BIND;
1023         } else {
1024                 if (iwqp->iwarp_info.wr_rdresp_en) {
1025                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
1026                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
1027                 }
1028                 if (iwqp->iwarp_info.rd_en)
1029                         acc_flags |= IB_ACCESS_REMOTE_READ;
1030                 if (iwqp->iwarp_info.bind_en)
1031                         acc_flags |= IB_ACCESS_MW_BIND;
1032         }
1033         return acc_flags;
1034 }
1035
1036 /**
1037  * irdma_query_qp - query qp attributes
1038  * @ibqp: qp pointer
1039  * @attr: attributes pointer
1040  * @attr_mask: Not used
1041  * @init_attr: qp attributes to return
1042  */
1043 static int irdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1044                           int attr_mask, struct ib_qp_init_attr *init_attr)
1045 {
1046         struct irdma_qp *iwqp = to_iwqp(ibqp);
1047         struct irdma_sc_qp *qp = &iwqp->sc_qp;
1048
1049         memset(attr, 0, sizeof(*attr));
1050         memset(init_attr, 0, sizeof(*init_attr));
1051
1052         attr->qp_state = iwqp->ibqp_state;
1053         attr->cur_qp_state = iwqp->ibqp_state;
1054         attr->cap.max_send_wr = iwqp->max_send_wr;
1055         attr->cap.max_recv_wr = iwqp->max_recv_wr;
1056         attr->cap.max_inline_data = qp->qp_uk.max_inline_data;
1057         attr->cap.max_send_sge = qp->qp_uk.max_sq_frag_cnt;
1058         attr->cap.max_recv_sge = qp->qp_uk.max_rq_frag_cnt;
1059         attr->qp_access_flags = irdma_get_ib_acc_flags(iwqp);
1060         attr->port_num = 1;
1061         if (rdma_protocol_roce(ibqp->device, 1)) {
1062                 attr->path_mtu = ib_mtu_int_to_enum(iwqp->udp_info.snd_mss);
1063                 attr->qkey = iwqp->roce_info.qkey;
1064                 attr->rq_psn = iwqp->udp_info.epsn;
1065                 attr->sq_psn = iwqp->udp_info.psn_nxt;
1066                 attr->dest_qp_num = iwqp->roce_info.dest_qp;
1067                 attr->pkey_index = iwqp->roce_info.p_key;
1068                 attr->retry_cnt = iwqp->udp_info.rexmit_thresh;
1069                 attr->rnr_retry = iwqp->udp_info.rnr_nak_thresh;
1070                 attr->max_rd_atomic = iwqp->roce_info.ord_size;
1071                 attr->max_dest_rd_atomic = iwqp->roce_info.ird_size;
1072         }
1073
1074         init_attr->event_handler = iwqp->ibqp.event_handler;
1075         init_attr->qp_context = iwqp->ibqp.qp_context;
1076         init_attr->send_cq = iwqp->ibqp.send_cq;
1077         init_attr->recv_cq = iwqp->ibqp.recv_cq;
1078         init_attr->cap = attr->cap;
1079
1080         return 0;
1081 }
1082
1083 /**
1084  * irdma_query_pkey - Query partition key
1085  * @ibdev: device pointer from stack
1086  * @port: port number
1087  * @index: index of pkey
1088  * @pkey: pointer to store the pkey
1089  */
1090 static int irdma_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1091                             u16 *pkey)
1092 {
1093         if (index >= IRDMA_PKEY_TBL_SZ)
1094                 return -EINVAL;
1095
1096         *pkey = IRDMA_DEFAULT_PKEY;
1097         return 0;
1098 }
1099
1100 /**
1101  * irdma_modify_qp_roce - modify qp request
1102  * @ibqp: qp's pointer for modify
1103  * @attr: access attributes
1104  * @attr_mask: state mask
1105  * @udata: user data
1106  */
1107 int irdma_modify_qp_roce(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1108                          int attr_mask, struct ib_udata *udata)
1109 {
1110 #define IRDMA_MODIFY_QP_MIN_REQ_LEN offsetofend(struct irdma_modify_qp_req, rq_flush)
1111 #define IRDMA_MODIFY_QP_MIN_RESP_LEN offsetofend(struct irdma_modify_qp_resp, push_valid)
1112         struct irdma_pd *iwpd = to_iwpd(ibqp->pd);
1113         struct irdma_qp *iwqp = to_iwqp(ibqp);
1114         struct irdma_device *iwdev = iwqp->iwdev;
1115         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
1116         struct irdma_qp_host_ctx_info *ctx_info;
1117         struct irdma_roce_offload_info *roce_info;
1118         struct irdma_udp_offload_info *udp_info;
1119         struct irdma_modify_qp_info info = {};
1120         struct irdma_modify_qp_resp uresp = {};
1121         struct irdma_modify_qp_req ureq = {};
1122         unsigned long flags;
1123         u8 issue_modify_qp = 0;
1124         int ret = 0;
1125
1126         ctx_info = &iwqp->ctx_info;
1127         roce_info = &iwqp->roce_info;
1128         udp_info = &iwqp->udp_info;
1129
1130         if (udata) {
1131                 /* udata inlen/outlen can be 0 when supporting legacy libi40iw */
1132                 if ((udata->inlen && udata->inlen < IRDMA_MODIFY_QP_MIN_REQ_LEN) ||
1133                     (udata->outlen && udata->outlen < IRDMA_MODIFY_QP_MIN_RESP_LEN))
1134                         return -EINVAL;
1135         }
1136
1137         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1138                 return -EOPNOTSUPP;
1139
1140         if (attr_mask & IB_QP_DEST_QPN)
1141                 roce_info->dest_qp = attr->dest_qp_num;
1142
1143         if (attr_mask & IB_QP_PKEY_INDEX) {
1144                 ret = irdma_query_pkey(ibqp->device, 0, attr->pkey_index,
1145                                        &roce_info->p_key);
1146                 if (ret)
1147                         return ret;
1148         }
1149
1150         if (attr_mask & IB_QP_QKEY)
1151                 roce_info->qkey = attr->qkey;
1152
1153         if (attr_mask & IB_QP_PATH_MTU)
1154                 udp_info->snd_mss = ib_mtu_enum_to_int(attr->path_mtu);
1155
1156         if (attr_mask & IB_QP_SQ_PSN) {
1157                 udp_info->psn_nxt = attr->sq_psn;
1158                 udp_info->lsn =  0xffff;
1159                 udp_info->psn_una = attr->sq_psn;
1160                 udp_info->psn_max = attr->sq_psn;
1161         }
1162
1163         if (attr_mask & IB_QP_RQ_PSN)
1164                 udp_info->epsn = attr->rq_psn;
1165
1166         if (attr_mask & IB_QP_RNR_RETRY)
1167                 udp_info->rnr_nak_thresh = attr->rnr_retry;
1168
1169         if (attr_mask & IB_QP_RETRY_CNT)
1170                 udp_info->rexmit_thresh = attr->retry_cnt;
1171
1172         ctx_info->roce_info->pd_id = iwpd->sc_pd.pd_id;
1173
1174         if (attr_mask & IB_QP_AV) {
1175                 struct irdma_av *av = &iwqp->roce_ah.av;
1176                 const struct ib_gid_attr *sgid_attr;
1177                 u16 vlan_id = VLAN_N_VID;
1178                 u32 local_ip[4];
1179
1180                 memset(&iwqp->roce_ah, 0, sizeof(iwqp->roce_ah));
1181                 if (attr->ah_attr.ah_flags & IB_AH_GRH) {
1182                         udp_info->ttl = attr->ah_attr.grh.hop_limit;
1183                         udp_info->flow_label = attr->ah_attr.grh.flow_label;
1184                         udp_info->tos = attr->ah_attr.grh.traffic_class;
1185                         udp_info->src_port =
1186                                 rdma_get_udp_sport(udp_info->flow_label,
1187                                                    ibqp->qp_num,
1188                                                    roce_info->dest_qp);
1189                         irdma_qp_rem_qos(&iwqp->sc_qp);
1190                         dev->ws_remove(iwqp->sc_qp.vsi, ctx_info->user_pri);
1191                         ctx_info->user_pri = rt_tos2priority(udp_info->tos);
1192                         iwqp->sc_qp.user_pri = ctx_info->user_pri;
1193                         if (dev->ws_add(iwqp->sc_qp.vsi, ctx_info->user_pri))
1194                                 return -ENOMEM;
1195                         irdma_qp_add_qos(&iwqp->sc_qp);
1196                 }
1197                 sgid_attr = attr->ah_attr.grh.sgid_attr;
1198                 ret = rdma_read_gid_l2_fields(sgid_attr, &vlan_id,
1199                                               ctx_info->roce_info->mac_addr);
1200                 if (ret)
1201                         return ret;
1202
1203                 if (vlan_id >= VLAN_N_VID && iwdev->dcb_vlan_mode)
1204                         vlan_id = 0;
1205                 if (vlan_id < VLAN_N_VID) {
1206                         udp_info->insert_vlan_tag = true;
1207                         udp_info->vlan_tag = vlan_id |
1208                                 ctx_info->user_pri << VLAN_PRIO_SHIFT;
1209                 } else {
1210                         udp_info->insert_vlan_tag = false;
1211                 }
1212
1213                 av->attrs = attr->ah_attr;
1214                 rdma_gid2ip((struct sockaddr *)&av->sgid_addr, &sgid_attr->gid);
1215                 rdma_gid2ip((struct sockaddr *)&av->dgid_addr, &attr->ah_attr.grh.dgid);
1216                 av->net_type = rdma_gid_attr_network_type(sgid_attr);
1217                 if (av->net_type == RDMA_NETWORK_IPV6) {
1218                         __be32 *daddr =
1219                                 av->dgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32;
1220                         __be32 *saddr =
1221                                 av->sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32;
1222
1223                         irdma_copy_ip_ntohl(&udp_info->dest_ip_addr[0], daddr);
1224                         irdma_copy_ip_ntohl(&udp_info->local_ipaddr[0], saddr);
1225
1226                         udp_info->ipv4 = false;
1227                         irdma_copy_ip_ntohl(local_ip, daddr);
1228
1229                 } else if (av->net_type == RDMA_NETWORK_IPV4) {
1230                         __be32 saddr = av->sgid_addr.saddr_in.sin_addr.s_addr;
1231                         __be32 daddr = av->dgid_addr.saddr_in.sin_addr.s_addr;
1232
1233                         local_ip[0] = ntohl(daddr);
1234
1235                         udp_info->ipv4 = true;
1236                         udp_info->dest_ip_addr[0] = 0;
1237                         udp_info->dest_ip_addr[1] = 0;
1238                         udp_info->dest_ip_addr[2] = 0;
1239                         udp_info->dest_ip_addr[3] = local_ip[0];
1240
1241                         udp_info->local_ipaddr[0] = 0;
1242                         udp_info->local_ipaddr[1] = 0;
1243                         udp_info->local_ipaddr[2] = 0;
1244                         udp_info->local_ipaddr[3] = ntohl(saddr);
1245                 }
1246                 udp_info->arp_idx =
1247                         irdma_add_arp(iwdev->rf, local_ip, udp_info->ipv4,
1248                                       attr->ah_attr.roce.dmac);
1249         }
1250
1251         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1252                 if (attr->max_rd_atomic > dev->hw_attrs.max_hw_ord) {
1253                         ibdev_err(&iwdev->ibdev,
1254                                   "rd_atomic = %d, above max_hw_ord=%d\n",
1255                                   attr->max_rd_atomic,
1256                                   dev->hw_attrs.max_hw_ord);
1257                         return -EINVAL;
1258                 }
1259                 if (attr->max_rd_atomic)
1260                         roce_info->ord_size = attr->max_rd_atomic;
1261                 info.ord_valid = true;
1262         }
1263
1264         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1265                 if (attr->max_dest_rd_atomic > dev->hw_attrs.max_hw_ird) {
1266                         ibdev_err(&iwdev->ibdev,
1267                                   "rd_atomic = %d, above max_hw_ird=%d\n",
1268                                    attr->max_rd_atomic,
1269                                    dev->hw_attrs.max_hw_ird);
1270                         return -EINVAL;
1271                 }
1272                 if (attr->max_dest_rd_atomic)
1273                         roce_info->ird_size = attr->max_dest_rd_atomic;
1274         }
1275
1276         if (attr_mask & IB_QP_ACCESS_FLAGS) {
1277                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
1278                         roce_info->wr_rdresp_en = true;
1279                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
1280                         roce_info->wr_rdresp_en = true;
1281                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
1282                         roce_info->rd_en = true;
1283         }
1284
1285         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
1286
1287         ibdev_dbg(&iwdev->ibdev,
1288                   "VERBS: caller: %pS qp_id=%d to_ibqpstate=%d ibqpstate=%d irdma_qpstate=%d attr_mask=0x%x\n",
1289                   __builtin_return_address(0), ibqp->qp_num, attr->qp_state,
1290                   iwqp->ibqp_state, iwqp->iwarp_state, attr_mask);
1291
1292         spin_lock_irqsave(&iwqp->lock, flags);
1293         if (attr_mask & IB_QP_STATE) {
1294                 if (!ib_modify_qp_is_ok(iwqp->ibqp_state, attr->qp_state,
1295                                         iwqp->ibqp.qp_type, attr_mask)) {
1296                         ibdev_warn(&iwdev->ibdev, "modify_qp invalid for qp_id=%d, old_state=0x%x, new_state=0x%x\n",
1297                                    iwqp->ibqp.qp_num, iwqp->ibqp_state,
1298                                    attr->qp_state);
1299                         ret = -EINVAL;
1300                         goto exit;
1301                 }
1302                 info.curr_iwarp_state = iwqp->iwarp_state;
1303
1304                 switch (attr->qp_state) {
1305                 case IB_QPS_INIT:
1306                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1307                                 ret = -EINVAL;
1308                                 goto exit;
1309                         }
1310
1311                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
1312                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
1313                                 issue_modify_qp = 1;
1314                         }
1315                         break;
1316                 case IB_QPS_RTR:
1317                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1318                                 ret = -EINVAL;
1319                                 goto exit;
1320                         }
1321                         info.arp_cache_idx_valid = true;
1322                         info.cq_num_valid = true;
1323                         info.next_iwarp_state = IRDMA_QP_STATE_RTR;
1324                         issue_modify_qp = 1;
1325                         break;
1326                 case IB_QPS_RTS:
1327                         if (iwqp->ibqp_state < IB_QPS_RTR ||
1328                             iwqp->ibqp_state == IB_QPS_ERR) {
1329                                 ret = -EINVAL;
1330                                 goto exit;
1331                         }
1332
1333                         info.arp_cache_idx_valid = true;
1334                         info.cq_num_valid = true;
1335                         info.ord_valid = true;
1336                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
1337                         issue_modify_qp = 1;
1338                         if (iwdev->push_mode && udata &&
1339                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1340                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1341                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1342                                 irdma_alloc_push_page(iwqp);
1343                                 spin_lock_irqsave(&iwqp->lock, flags);
1344                         }
1345                         break;
1346                 case IB_QPS_SQD:
1347                         if (iwqp->iwarp_state == IRDMA_QP_STATE_SQD)
1348                                 goto exit;
1349
1350                         if (iwqp->iwarp_state != IRDMA_QP_STATE_RTS) {
1351                                 ret = -EINVAL;
1352                                 goto exit;
1353                         }
1354
1355                         info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1356                         issue_modify_qp = 1;
1357                         break;
1358                 case IB_QPS_SQE:
1359                 case IB_QPS_ERR:
1360                 case IB_QPS_RESET:
1361                         if (iwqp->iwarp_state == IRDMA_QP_STATE_RTS) {
1362                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1363                                 info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1364                                 irdma_hw_modify_qp(iwdev, iwqp, &info, true);
1365                                 spin_lock_irqsave(&iwqp->lock, flags);
1366                         }
1367
1368                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1369                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1370                                 if (udata && udata->inlen) {
1371                                         if (ib_copy_from_udata(&ureq, udata,
1372                                             min(sizeof(ureq), udata->inlen)))
1373                                                 return -EINVAL;
1374
1375                                         irdma_flush_wqes(iwqp,
1376                                             (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1377                                             (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1378                                             IRDMA_REFLUSH);
1379                                 }
1380                                 return 0;
1381                         }
1382
1383                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1384                         issue_modify_qp = 1;
1385                         break;
1386                 default:
1387                         ret = -EINVAL;
1388                         goto exit;
1389                 }
1390
1391                 iwqp->ibqp_state = attr->qp_state;
1392         }
1393
1394         ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1395         ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1396         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1397         spin_unlock_irqrestore(&iwqp->lock, flags);
1398
1399         if (attr_mask & IB_QP_STATE) {
1400                 if (issue_modify_qp) {
1401                         ctx_info->rem_endpoint_idx = udp_info->arp_idx;
1402                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1403                                 return -EINVAL;
1404                         spin_lock_irqsave(&iwqp->lock, flags);
1405                         if (iwqp->iwarp_state == info.curr_iwarp_state) {
1406                                 iwqp->iwarp_state = info.next_iwarp_state;
1407                                 iwqp->ibqp_state = attr->qp_state;
1408                         }
1409                         if (iwqp->ibqp_state > IB_QPS_RTS &&
1410                             !iwqp->flush_issued) {
1411                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1412                                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_SQ |
1413                                                        IRDMA_FLUSH_RQ |
1414                                                        IRDMA_FLUSH_WAIT);
1415                                 iwqp->flush_issued = 1;
1416                         } else {
1417                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1418                         }
1419                 } else {
1420                         iwqp->ibqp_state = attr->qp_state;
1421                 }
1422                 if (udata && udata->outlen && dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1423                         struct irdma_ucontext *ucontext;
1424
1425                         ucontext = rdma_udata_to_drv_context(udata,
1426                                         struct irdma_ucontext, ibucontext);
1427                         if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1428                             !iwqp->push_wqe_mmap_entry &&
1429                             !irdma_setup_push_mmap_entries(ucontext, iwqp,
1430                                 &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1431                                 uresp.push_valid = 1;
1432                                 uresp.push_offset = iwqp->sc_qp.push_offset;
1433                         }
1434                         ret = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1435                                                udata->outlen));
1436                         if (ret) {
1437                                 irdma_remove_push_mmap_entries(iwqp);
1438                                 ibdev_dbg(&iwdev->ibdev,
1439                                           "VERBS: copy_to_udata failed\n");
1440                                 return ret;
1441                         }
1442                 }
1443         }
1444
1445         return 0;
1446 exit:
1447         spin_unlock_irqrestore(&iwqp->lock, flags);
1448
1449         return ret;
1450 }
1451
1452 /**
1453  * irdma_modify_qp - modify qp request
1454  * @ibqp: qp's pointer for modify
1455  * @attr: access attributes
1456  * @attr_mask: state mask
1457  * @udata: user data
1458  */
1459 int irdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
1460                     struct ib_udata *udata)
1461 {
1462 #define IRDMA_MODIFY_QP_MIN_REQ_LEN offsetofend(struct irdma_modify_qp_req, rq_flush)
1463 #define IRDMA_MODIFY_QP_MIN_RESP_LEN offsetofend(struct irdma_modify_qp_resp, push_valid)
1464         struct irdma_qp *iwqp = to_iwqp(ibqp);
1465         struct irdma_device *iwdev = iwqp->iwdev;
1466         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
1467         struct irdma_qp_host_ctx_info *ctx_info;
1468         struct irdma_tcp_offload_info *tcp_info;
1469         struct irdma_iwarp_offload_info *offload_info;
1470         struct irdma_modify_qp_info info = {};
1471         struct irdma_modify_qp_resp uresp = {};
1472         struct irdma_modify_qp_req ureq = {};
1473         u8 issue_modify_qp = 0;
1474         u8 dont_wait = 0;
1475         int err;
1476         unsigned long flags;
1477
1478         if (udata) {
1479                 /* udata inlen/outlen can be 0 when supporting legacy libi40iw */
1480                 if ((udata->inlen && udata->inlen < IRDMA_MODIFY_QP_MIN_REQ_LEN) ||
1481                     (udata->outlen && udata->outlen < IRDMA_MODIFY_QP_MIN_RESP_LEN))
1482                         return -EINVAL;
1483         }
1484
1485         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1486                 return -EOPNOTSUPP;
1487
1488         ctx_info = &iwqp->ctx_info;
1489         offload_info = &iwqp->iwarp_info;
1490         tcp_info = &iwqp->tcp_info;
1491         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
1492         ibdev_dbg(&iwdev->ibdev,
1493                   "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",
1494                   __builtin_return_address(0), ibqp->qp_num, attr->qp_state,
1495                   iwqp->ibqp_state, iwqp->iwarp_state, iwqp->last_aeq,
1496                   iwqp->hw_tcp_state, iwqp->hw_iwarp_state, attr_mask);
1497
1498         spin_lock_irqsave(&iwqp->lock, flags);
1499         if (attr_mask & IB_QP_STATE) {
1500                 info.curr_iwarp_state = iwqp->iwarp_state;
1501                 switch (attr->qp_state) {
1502                 case IB_QPS_INIT:
1503                 case IB_QPS_RTR:
1504                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1505                                 err = -EINVAL;
1506                                 goto exit;
1507                         }
1508
1509                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
1510                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
1511                                 issue_modify_qp = 1;
1512                         }
1513                         if (iwdev->push_mode && udata &&
1514                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1515                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1516                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1517                                 irdma_alloc_push_page(iwqp);
1518                                 spin_lock_irqsave(&iwqp->lock, flags);
1519                         }
1520                         break;
1521                 case IB_QPS_RTS:
1522                         if (iwqp->iwarp_state > IRDMA_QP_STATE_RTS ||
1523                             !iwqp->cm_id) {
1524                                 err = -EINVAL;
1525                                 goto exit;
1526                         }
1527
1528                         issue_modify_qp = 1;
1529                         iwqp->hw_tcp_state = IRDMA_TCP_STATE_ESTABLISHED;
1530                         iwqp->hte_added = 1;
1531                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
1532                         info.tcp_ctx_valid = true;
1533                         info.ord_valid = true;
1534                         info.arp_cache_idx_valid = true;
1535                         info.cq_num_valid = true;
1536                         break;
1537                 case IB_QPS_SQD:
1538                         if (iwqp->hw_iwarp_state > IRDMA_QP_STATE_RTS) {
1539                                 err = 0;
1540                                 goto exit;
1541                         }
1542
1543                         if (iwqp->iwarp_state == IRDMA_QP_STATE_CLOSING ||
1544                             iwqp->iwarp_state < IRDMA_QP_STATE_RTS) {
1545                                 err = 0;
1546                                 goto exit;
1547                         }
1548
1549                         if (iwqp->iwarp_state > IRDMA_QP_STATE_CLOSING) {
1550                                 err = -EINVAL;
1551                                 goto exit;
1552                         }
1553
1554                         info.next_iwarp_state = IRDMA_QP_STATE_CLOSING;
1555                         issue_modify_qp = 1;
1556                         break;
1557                 case IB_QPS_SQE:
1558                         if (iwqp->iwarp_state >= IRDMA_QP_STATE_TERMINATE) {
1559                                 err = -EINVAL;
1560                                 goto exit;
1561                         }
1562
1563                         info.next_iwarp_state = IRDMA_QP_STATE_TERMINATE;
1564                         issue_modify_qp = 1;
1565                         break;
1566                 case IB_QPS_ERR:
1567                 case IB_QPS_RESET:
1568                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1569                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1570                                 if (udata && udata->inlen) {
1571                                         if (ib_copy_from_udata(&ureq, udata,
1572                                             min(sizeof(ureq), udata->inlen)))
1573                                                 return -EINVAL;
1574
1575                                         irdma_flush_wqes(iwqp,
1576                                             (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1577                                             (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1578                                             IRDMA_REFLUSH);
1579                                 }
1580                                 return 0;
1581                         }
1582
1583                         if (iwqp->sc_qp.term_flags) {
1584                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1585                                 irdma_terminate_del_timer(&iwqp->sc_qp);
1586                                 spin_lock_irqsave(&iwqp->lock, flags);
1587                         }
1588                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1589                         if (iwqp->hw_tcp_state > IRDMA_TCP_STATE_CLOSED &&
1590                             iwdev->iw_status &&
1591                             iwqp->hw_tcp_state != IRDMA_TCP_STATE_TIME_WAIT)
1592                                 info.reset_tcp_conn = true;
1593                         else
1594                                 dont_wait = 1;
1595
1596                         issue_modify_qp = 1;
1597                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1598                         break;
1599                 default:
1600                         err = -EINVAL;
1601                         goto exit;
1602                 }
1603
1604                 iwqp->ibqp_state = attr->qp_state;
1605         }
1606         if (attr_mask & IB_QP_ACCESS_FLAGS) {
1607                 ctx_info->iwarp_info_valid = true;
1608                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
1609                         offload_info->wr_rdresp_en = true;
1610                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
1611                         offload_info->wr_rdresp_en = true;
1612                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
1613                         offload_info->rd_en = true;
1614         }
1615
1616         if (ctx_info->iwarp_info_valid) {
1617                 ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1618                 ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1619                 irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1620         }
1621         spin_unlock_irqrestore(&iwqp->lock, flags);
1622
1623         if (attr_mask & IB_QP_STATE) {
1624                 if (issue_modify_qp) {
1625                         ctx_info->rem_endpoint_idx = tcp_info->arp_idx;
1626                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1627                                 return -EINVAL;
1628                 }
1629
1630                 spin_lock_irqsave(&iwqp->lock, flags);
1631                 if (iwqp->iwarp_state == info.curr_iwarp_state) {
1632                         iwqp->iwarp_state = info.next_iwarp_state;
1633                         iwqp->ibqp_state = attr->qp_state;
1634                 }
1635                 spin_unlock_irqrestore(&iwqp->lock, flags);
1636         }
1637
1638         if (issue_modify_qp && iwqp->ibqp_state > IB_QPS_RTS) {
1639                 if (dont_wait) {
1640                         if (iwqp->hw_tcp_state) {
1641                                 spin_lock_irqsave(&iwqp->lock, flags);
1642                                 iwqp->hw_tcp_state = IRDMA_TCP_STATE_CLOSED;
1643                                 iwqp->last_aeq = IRDMA_AE_RESET_SENT;
1644                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1645                         }
1646                         irdma_cm_disconn(iwqp);
1647                 } else {
1648                         int close_timer_started;
1649
1650                         spin_lock_irqsave(&iwdev->cm_core.ht_lock, flags);
1651
1652                         if (iwqp->cm_node) {
1653                                 refcount_inc(&iwqp->cm_node->refcnt);
1654                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1655                                 close_timer_started = atomic_inc_return(&iwqp->close_timer_started);
1656                                 if (iwqp->cm_id && close_timer_started == 1)
1657                                         irdma_schedule_cm_timer(iwqp->cm_node,
1658                                                 (struct irdma_puda_buf *)iwqp,
1659                                                 IRDMA_TIMER_TYPE_CLOSE, 1, 0);
1660
1661                                 irdma_rem_ref_cm_node(iwqp->cm_node);
1662                         } else {
1663                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1664                         }
1665                 }
1666         }
1667         if (attr_mask & IB_QP_STATE && udata && udata->outlen &&
1668             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1669                 struct irdma_ucontext *ucontext;
1670
1671                 ucontext = rdma_udata_to_drv_context(udata,
1672                                         struct irdma_ucontext, ibucontext);
1673                 if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1674                     !iwqp->push_wqe_mmap_entry &&
1675                     !irdma_setup_push_mmap_entries(ucontext, iwqp,
1676                         &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1677                         uresp.push_valid = 1;
1678                         uresp.push_offset = iwqp->sc_qp.push_offset;
1679                 }
1680
1681                 err = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1682                                        udata->outlen));
1683                 if (err) {
1684                         irdma_remove_push_mmap_entries(iwqp);
1685                         ibdev_dbg(&iwdev->ibdev,
1686                                   "VERBS: copy_to_udata failed\n");
1687                         return err;
1688                 }
1689         }
1690
1691         return 0;
1692 exit:
1693         spin_unlock_irqrestore(&iwqp->lock, flags);
1694
1695         return err;
1696 }
1697
1698 /**
1699  * irdma_cq_free_rsrc - free up resources for cq
1700  * @rf: RDMA PCI function
1701  * @iwcq: cq ptr
1702  */
1703 static void irdma_cq_free_rsrc(struct irdma_pci_f *rf, struct irdma_cq *iwcq)
1704 {
1705         struct irdma_sc_cq *cq = &iwcq->sc_cq;
1706
1707         if (!iwcq->user_mode) {
1708                 dma_free_coherent(rf->sc_dev.hw->device, iwcq->kmem.size,
1709                                   iwcq->kmem.va, iwcq->kmem.pa);
1710                 iwcq->kmem.va = NULL;
1711                 dma_free_coherent(rf->sc_dev.hw->device,
1712                                   iwcq->kmem_shadow.size,
1713                                   iwcq->kmem_shadow.va, iwcq->kmem_shadow.pa);
1714                 iwcq->kmem_shadow.va = NULL;
1715         }
1716
1717         irdma_free_rsrc(rf, rf->allocated_cqs, cq->cq_uk.cq_id);
1718 }
1719
1720 /**
1721  * irdma_free_cqbuf - worker to free a cq buffer
1722  * @work: provides access to the cq buffer to free
1723  */
1724 static void irdma_free_cqbuf(struct work_struct *work)
1725 {
1726         struct irdma_cq_buf *cq_buf = container_of(work, struct irdma_cq_buf, work);
1727
1728         dma_free_coherent(cq_buf->hw->device, cq_buf->kmem_buf.size,
1729                           cq_buf->kmem_buf.va, cq_buf->kmem_buf.pa);
1730         cq_buf->kmem_buf.va = NULL;
1731         kfree(cq_buf);
1732 }
1733
1734 /**
1735  * irdma_process_resize_list - remove resized cq buffers from the resize_list
1736  * @iwcq: cq which owns the resize_list
1737  * @iwdev: irdma device
1738  * @lcqe_buf: the buffer where the last cqe is received
1739  */
1740 static int irdma_process_resize_list(struct irdma_cq *iwcq,
1741                                      struct irdma_device *iwdev,
1742                                      struct irdma_cq_buf *lcqe_buf)
1743 {
1744         struct list_head *tmp_node, *list_node;
1745         struct irdma_cq_buf *cq_buf;
1746         int cnt = 0;
1747
1748         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
1749                 cq_buf = list_entry(list_node, struct irdma_cq_buf, list);
1750                 if (cq_buf == lcqe_buf)
1751                         return cnt;
1752
1753                 list_del(&cq_buf->list);
1754                 queue_work(iwdev->cleanup_wq, &cq_buf->work);
1755                 cnt++;
1756         }
1757
1758         return cnt;
1759 }
1760
1761 /**
1762  * irdma_destroy_cq - destroy cq
1763  * @ib_cq: cq pointer
1764  * @udata: user data
1765  */
1766 static int irdma_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata)
1767 {
1768         struct irdma_device *iwdev = to_iwdev(ib_cq->device);
1769         struct irdma_cq *iwcq = to_iwcq(ib_cq);
1770         struct irdma_sc_cq *cq = &iwcq->sc_cq;
1771         struct irdma_sc_dev *dev = cq->dev;
1772         struct irdma_sc_ceq *ceq = dev->ceq[cq->ceq_id];
1773         struct irdma_ceq *iwceq = container_of(ceq, struct irdma_ceq, sc_ceq);
1774         unsigned long flags;
1775
1776         spin_lock_irqsave(&iwcq->lock, flags);
1777         if (!list_empty(&iwcq->cmpl_generated))
1778                 irdma_remove_cmpls_list(iwcq);
1779         if (!list_empty(&iwcq->resize_list))
1780                 irdma_process_resize_list(iwcq, iwdev, NULL);
1781         spin_unlock_irqrestore(&iwcq->lock, flags);
1782
1783         irdma_cq_wq_destroy(iwdev->rf, cq);
1784
1785         spin_lock_irqsave(&iwceq->ce_lock, flags);
1786         irdma_sc_cleanup_ceqes(cq, ceq);
1787         spin_unlock_irqrestore(&iwceq->ce_lock, flags);
1788         irdma_cq_free_rsrc(iwdev->rf, iwcq);
1789
1790         return 0;
1791 }
1792
1793 /**
1794  * irdma_resize_cq - resize cq
1795  * @ibcq: cq to be resized
1796  * @entries: desired cq size
1797  * @udata: user data
1798  */
1799 static int irdma_resize_cq(struct ib_cq *ibcq, int entries,
1800                            struct ib_udata *udata)
1801 {
1802 #define IRDMA_RESIZE_CQ_MIN_REQ_LEN offsetofend(struct irdma_resize_cq_req, user_cq_buffer)
1803         struct irdma_cq *iwcq = to_iwcq(ibcq);
1804         struct irdma_sc_dev *dev = iwcq->sc_cq.dev;
1805         struct irdma_cqp_request *cqp_request;
1806         struct cqp_cmds_info *cqp_info;
1807         struct irdma_modify_cq_info *m_info;
1808         struct irdma_modify_cq_info info = {};
1809         struct irdma_dma_mem kmem_buf;
1810         struct irdma_cq_mr *cqmr_buf;
1811         struct irdma_pbl *iwpbl_buf;
1812         struct irdma_device *iwdev;
1813         struct irdma_pci_f *rf;
1814         struct irdma_cq_buf *cq_buf = NULL;
1815         unsigned long flags;
1816         int ret;
1817
1818         iwdev = to_iwdev(ibcq->device);
1819         rf = iwdev->rf;
1820
1821         if (!(rf->sc_dev.hw_attrs.uk_attrs.feature_flags &
1822             IRDMA_FEATURE_CQ_RESIZE))
1823                 return -EOPNOTSUPP;
1824
1825         if (udata && udata->inlen < IRDMA_RESIZE_CQ_MIN_REQ_LEN)
1826                 return -EINVAL;
1827
1828         if (entries > rf->max_cqe)
1829                 return -EINVAL;
1830
1831         if (!iwcq->user_mode) {
1832                 entries++;
1833                 if (rf->sc_dev.hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
1834                         entries *= 2;
1835         }
1836
1837         info.cq_size = max(entries, 4);
1838
1839         if (info.cq_size == iwcq->sc_cq.cq_uk.cq_size - 1)
1840                 return 0;
1841
1842         if (udata) {
1843                 struct irdma_resize_cq_req req = {};
1844                 struct irdma_ucontext *ucontext =
1845                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
1846                                                   ibucontext);
1847
1848                 /* CQ resize not supported with legacy GEN_1 libi40iw */
1849                 if (ucontext->legacy_mode)
1850                         return -EOPNOTSUPP;
1851
1852                 if (ib_copy_from_udata(&req, udata,
1853                                        min(sizeof(req), udata->inlen)))
1854                         return -EINVAL;
1855
1856                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
1857                 iwpbl_buf = irdma_get_pbl((unsigned long)req.user_cq_buffer,
1858                                           &ucontext->cq_reg_mem_list);
1859                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
1860
1861                 if (!iwpbl_buf)
1862                         return -ENOMEM;
1863
1864                 cqmr_buf = &iwpbl_buf->cq_mr;
1865                 if (iwpbl_buf->pbl_allocated) {
1866                         info.virtual_map = true;
1867                         info.pbl_chunk_size = 1;
1868                         info.first_pm_pbl_idx = cqmr_buf->cq_pbl.idx;
1869                 } else {
1870                         info.cq_pa = cqmr_buf->cq_pbl.addr;
1871                 }
1872         } else {
1873                 /* Kmode CQ resize */
1874                 int rsize;
1875
1876                 rsize = info.cq_size * sizeof(struct irdma_cqe);
1877                 kmem_buf.size = ALIGN(round_up(rsize, 256), 256);
1878                 kmem_buf.va = dma_alloc_coherent(dev->hw->device,
1879                                                  kmem_buf.size, &kmem_buf.pa,
1880                                                  GFP_KERNEL);
1881                 if (!kmem_buf.va)
1882                         return -ENOMEM;
1883
1884                 info.cq_base = kmem_buf.va;
1885                 info.cq_pa = kmem_buf.pa;
1886                 cq_buf = kzalloc(sizeof(*cq_buf), GFP_KERNEL);
1887                 if (!cq_buf) {
1888                         ret = -ENOMEM;
1889                         goto error;
1890                 }
1891         }
1892
1893         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
1894         if (!cqp_request) {
1895                 ret = -ENOMEM;
1896                 goto error;
1897         }
1898
1899         info.shadow_read_threshold = iwcq->sc_cq.shadow_read_threshold;
1900         info.cq_resize = true;
1901
1902         cqp_info = &cqp_request->info;
1903         m_info = &cqp_info->in.u.cq_modify.info;
1904         memcpy(m_info, &info, sizeof(*m_info));
1905
1906         cqp_info->cqp_cmd = IRDMA_OP_CQ_MODIFY;
1907         cqp_info->in.u.cq_modify.cq = &iwcq->sc_cq;
1908         cqp_info->in.u.cq_modify.scratch = (uintptr_t)cqp_request;
1909         cqp_info->post_sq = 1;
1910         ret = irdma_handle_cqp_op(rf, cqp_request);
1911         irdma_put_cqp_request(&rf->cqp, cqp_request);
1912         if (ret)
1913                 goto error;
1914
1915         spin_lock_irqsave(&iwcq->lock, flags);
1916         if (cq_buf) {
1917                 cq_buf->kmem_buf = iwcq->kmem;
1918                 cq_buf->hw = dev->hw;
1919                 memcpy(&cq_buf->cq_uk, &iwcq->sc_cq.cq_uk, sizeof(cq_buf->cq_uk));
1920                 INIT_WORK(&cq_buf->work, irdma_free_cqbuf);
1921                 list_add_tail(&cq_buf->list, &iwcq->resize_list);
1922                 iwcq->kmem = kmem_buf;
1923         }
1924
1925         irdma_sc_cq_resize(&iwcq->sc_cq, &info);
1926         ibcq->cqe = info.cq_size - 1;
1927         spin_unlock_irqrestore(&iwcq->lock, flags);
1928
1929         return 0;
1930 error:
1931         if (!udata) {
1932                 dma_free_coherent(dev->hw->device, kmem_buf.size, kmem_buf.va,
1933                                   kmem_buf.pa);
1934                 kmem_buf.va = NULL;
1935         }
1936         kfree(cq_buf);
1937
1938         return ret;
1939 }
1940
1941 static inline int cq_validate_flags(u32 flags, u8 hw_rev)
1942 {
1943         /* GEN1 does not support CQ create flags */
1944         if (hw_rev == IRDMA_GEN_1)
1945                 return flags ? -EOPNOTSUPP : 0;
1946
1947         return flags & ~IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION ? -EOPNOTSUPP : 0;
1948 }
1949
1950 /**
1951  * irdma_create_cq - create cq
1952  * @ibcq: CQ allocated
1953  * @attr: attributes for cq
1954  * @udata: user data
1955  */
1956 static int irdma_create_cq(struct ib_cq *ibcq,
1957                            const struct ib_cq_init_attr *attr,
1958                            struct ib_udata *udata)
1959 {
1960 #define IRDMA_CREATE_CQ_MIN_REQ_LEN offsetofend(struct irdma_create_cq_req, user_cq_buf)
1961 #define IRDMA_CREATE_CQ_MIN_RESP_LEN offsetofend(struct irdma_create_cq_resp, cq_size)
1962         struct ib_device *ibdev = ibcq->device;
1963         struct irdma_device *iwdev = to_iwdev(ibdev);
1964         struct irdma_pci_f *rf = iwdev->rf;
1965         struct irdma_cq *iwcq = to_iwcq(ibcq);
1966         u32 cq_num = 0;
1967         struct irdma_sc_cq *cq;
1968         struct irdma_sc_dev *dev = &rf->sc_dev;
1969         struct irdma_cq_init_info info = {};
1970         struct irdma_cqp_request *cqp_request;
1971         struct cqp_cmds_info *cqp_info;
1972         struct irdma_cq_uk_init_info *ukinfo = &info.cq_uk_init_info;
1973         unsigned long flags;
1974         int err_code;
1975         int entries = attr->cqe;
1976
1977         err_code = cq_validate_flags(attr->flags, dev->hw_attrs.uk_attrs.hw_rev);
1978         if (err_code)
1979                 return err_code;
1980
1981         if (udata && (udata->inlen < IRDMA_CREATE_CQ_MIN_REQ_LEN ||
1982                       udata->outlen < IRDMA_CREATE_CQ_MIN_RESP_LEN))
1983                 return -EINVAL;
1984
1985         err_code = irdma_alloc_rsrc(rf, rf->allocated_cqs, rf->max_cq, &cq_num,
1986                                     &rf->next_cq);
1987         if (err_code)
1988                 return err_code;
1989
1990         cq = &iwcq->sc_cq;
1991         cq->back_cq = iwcq;
1992         spin_lock_init(&iwcq->lock);
1993         INIT_LIST_HEAD(&iwcq->resize_list);
1994         INIT_LIST_HEAD(&iwcq->cmpl_generated);
1995         info.dev = dev;
1996         ukinfo->cq_size = max(entries, 4);
1997         ukinfo->cq_id = cq_num;
1998         iwcq->ibcq.cqe = info.cq_uk_init_info.cq_size;
1999         if (attr->comp_vector < rf->ceqs_count)
2000                 info.ceq_id = attr->comp_vector;
2001         info.ceq_id_valid = true;
2002         info.ceqe_mask = 1;
2003         info.type = IRDMA_CQ_TYPE_IWARP;
2004         info.vsi = &iwdev->vsi;
2005
2006         if (udata) {
2007                 struct irdma_ucontext *ucontext;
2008                 struct irdma_create_cq_req req = {};
2009                 struct irdma_cq_mr *cqmr;
2010                 struct irdma_pbl *iwpbl;
2011                 struct irdma_pbl *iwpbl_shadow;
2012                 struct irdma_cq_mr *cqmr_shadow;
2013
2014                 iwcq->user_mode = true;
2015                 ucontext =
2016                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
2017                                                   ibucontext);
2018                 if (ib_copy_from_udata(&req, udata,
2019                                        min(sizeof(req), udata->inlen))) {
2020                         err_code = -EFAULT;
2021                         goto cq_free_rsrc;
2022                 }
2023
2024                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2025                 iwpbl = irdma_get_pbl((unsigned long)req.user_cq_buf,
2026                                       &ucontext->cq_reg_mem_list);
2027                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2028                 if (!iwpbl) {
2029                         err_code = -EPROTO;
2030                         goto cq_free_rsrc;
2031                 }
2032
2033                 iwcq->iwpbl = iwpbl;
2034                 iwcq->cq_mem_size = 0;
2035                 cqmr = &iwpbl->cq_mr;
2036
2037                 if (rf->sc_dev.hw_attrs.uk_attrs.feature_flags &
2038                     IRDMA_FEATURE_CQ_RESIZE && !ucontext->legacy_mode) {
2039                         spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2040                         iwpbl_shadow = irdma_get_pbl(
2041                                         (unsigned long)req.user_shadow_area,
2042                                         &ucontext->cq_reg_mem_list);
2043                         spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2044
2045                         if (!iwpbl_shadow) {
2046                                 err_code = -EPROTO;
2047                                 goto cq_free_rsrc;
2048                         }
2049                         iwcq->iwpbl_shadow = iwpbl_shadow;
2050                         cqmr_shadow = &iwpbl_shadow->cq_mr;
2051                         info.shadow_area_pa = cqmr_shadow->cq_pbl.addr;
2052                         cqmr->split = true;
2053                 } else {
2054                         info.shadow_area_pa = cqmr->shadow;
2055                 }
2056                 if (iwpbl->pbl_allocated) {
2057                         info.virtual_map = true;
2058                         info.pbl_chunk_size = 1;
2059                         info.first_pm_pbl_idx = cqmr->cq_pbl.idx;
2060                 } else {
2061                         info.cq_base_pa = cqmr->cq_pbl.addr;
2062                 }
2063         } else {
2064                 /* Kmode allocations */
2065                 int rsize;
2066
2067                 if (entries < 1 || entries > rf->max_cqe) {
2068                         err_code = -EINVAL;
2069                         goto cq_free_rsrc;
2070                 }
2071
2072                 entries++;
2073                 if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
2074                         entries *= 2;
2075                 ukinfo->cq_size = entries;
2076
2077                 rsize = info.cq_uk_init_info.cq_size * sizeof(struct irdma_cqe);
2078                 iwcq->kmem.size = ALIGN(round_up(rsize, 256), 256);
2079                 iwcq->kmem.va = dma_alloc_coherent(dev->hw->device,
2080                                                    iwcq->kmem.size,
2081                                                    &iwcq->kmem.pa, GFP_KERNEL);
2082                 if (!iwcq->kmem.va) {
2083                         err_code = -ENOMEM;
2084                         goto cq_free_rsrc;
2085                 }
2086
2087                 iwcq->kmem_shadow.size = ALIGN(IRDMA_SHADOW_AREA_SIZE << 3,
2088                                                64);
2089                 iwcq->kmem_shadow.va = dma_alloc_coherent(dev->hw->device,
2090                                                           iwcq->kmem_shadow.size,
2091                                                           &iwcq->kmem_shadow.pa,
2092                                                           GFP_KERNEL);
2093                 if (!iwcq->kmem_shadow.va) {
2094                         err_code = -ENOMEM;
2095                         goto cq_free_rsrc;
2096                 }
2097                 info.shadow_area_pa = iwcq->kmem_shadow.pa;
2098                 ukinfo->shadow_area = iwcq->kmem_shadow.va;
2099                 ukinfo->cq_base = iwcq->kmem.va;
2100                 info.cq_base_pa = iwcq->kmem.pa;
2101         }
2102
2103         if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
2104                 info.shadow_read_threshold = min(info.cq_uk_init_info.cq_size / 2,
2105                                                  (u32)IRDMA_MAX_CQ_READ_THRESH);
2106
2107         if (irdma_sc_cq_init(cq, &info)) {
2108                 ibdev_dbg(&iwdev->ibdev, "VERBS: init cq fail\n");
2109                 err_code = -EPROTO;
2110                 goto cq_free_rsrc;
2111         }
2112
2113         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
2114         if (!cqp_request) {
2115                 err_code = -ENOMEM;
2116                 goto cq_free_rsrc;
2117         }
2118
2119         cqp_info = &cqp_request->info;
2120         cqp_info->cqp_cmd = IRDMA_OP_CQ_CREATE;
2121         cqp_info->post_sq = 1;
2122         cqp_info->in.u.cq_create.cq = cq;
2123         cqp_info->in.u.cq_create.check_overflow = true;
2124         cqp_info->in.u.cq_create.scratch = (uintptr_t)cqp_request;
2125         err_code = irdma_handle_cqp_op(rf, cqp_request);
2126         irdma_put_cqp_request(&rf->cqp, cqp_request);
2127         if (err_code)
2128                 goto cq_free_rsrc;
2129
2130         if (udata) {
2131                 struct irdma_create_cq_resp resp = {};
2132
2133                 resp.cq_id = info.cq_uk_init_info.cq_id;
2134                 resp.cq_size = info.cq_uk_init_info.cq_size;
2135                 if (ib_copy_to_udata(udata, &resp,
2136                                      min(sizeof(resp), udata->outlen))) {
2137                         ibdev_dbg(&iwdev->ibdev,
2138                                   "VERBS: copy to user data\n");
2139                         err_code = -EPROTO;
2140                         goto cq_destroy;
2141                 }
2142         }
2143         return 0;
2144 cq_destroy:
2145         irdma_cq_wq_destroy(rf, cq);
2146 cq_free_rsrc:
2147         irdma_cq_free_rsrc(rf, iwcq);
2148
2149         return err_code;
2150 }
2151
2152 /**
2153  * irdma_get_mr_access - get hw MR access permissions from IB access flags
2154  * @access: IB access flags
2155  */
2156 static inline u16 irdma_get_mr_access(int access)
2157 {
2158         u16 hw_access = 0;
2159
2160         hw_access |= (access & IB_ACCESS_LOCAL_WRITE) ?
2161                      IRDMA_ACCESS_FLAGS_LOCALWRITE : 0;
2162         hw_access |= (access & IB_ACCESS_REMOTE_WRITE) ?
2163                      IRDMA_ACCESS_FLAGS_REMOTEWRITE : 0;
2164         hw_access |= (access & IB_ACCESS_REMOTE_READ) ?
2165                      IRDMA_ACCESS_FLAGS_REMOTEREAD : 0;
2166         hw_access |= (access & IB_ACCESS_MW_BIND) ?
2167                      IRDMA_ACCESS_FLAGS_BIND_WINDOW : 0;
2168         hw_access |= (access & IB_ZERO_BASED) ?
2169                      IRDMA_ACCESS_FLAGS_ZERO_BASED : 0;
2170         hw_access |= IRDMA_ACCESS_FLAGS_LOCALREAD;
2171
2172         return hw_access;
2173 }
2174
2175 /**
2176  * irdma_free_stag - free stag resource
2177  * @iwdev: irdma device
2178  * @stag: stag to free
2179  */
2180 static void irdma_free_stag(struct irdma_device *iwdev, u32 stag)
2181 {
2182         u32 stag_idx;
2183
2184         stag_idx = (stag & iwdev->rf->mr_stagmask) >> IRDMA_CQPSQ_STAG_IDX_S;
2185         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_mrs, stag_idx);
2186 }
2187
2188 /**
2189  * irdma_create_stag - create random stag
2190  * @iwdev: irdma device
2191  */
2192 static u32 irdma_create_stag(struct irdma_device *iwdev)
2193 {
2194         u32 stag = 0;
2195         u32 stag_index = 0;
2196         u32 next_stag_index;
2197         u32 driver_key;
2198         u32 random;
2199         u8 consumer_key;
2200         int ret;
2201
2202         get_random_bytes(&random, sizeof(random));
2203         consumer_key = (u8)random;
2204
2205         driver_key = random & ~iwdev->rf->mr_stagmask;
2206         next_stag_index = (random & iwdev->rf->mr_stagmask) >> 8;
2207         next_stag_index %= iwdev->rf->max_mr;
2208
2209         ret = irdma_alloc_rsrc(iwdev->rf, iwdev->rf->allocated_mrs,
2210                                iwdev->rf->max_mr, &stag_index,
2211                                &next_stag_index);
2212         if (ret)
2213                 return stag;
2214         stag = stag_index << IRDMA_CQPSQ_STAG_IDX_S;
2215         stag |= driver_key;
2216         stag += (u32)consumer_key;
2217
2218         return stag;
2219 }
2220
2221 /**
2222  * irdma_next_pbl_addr - Get next pbl address
2223  * @pbl: pointer to a pble
2224  * @pinfo: info pointer
2225  * @idx: index
2226  */
2227 static inline u64 *irdma_next_pbl_addr(u64 *pbl, struct irdma_pble_info **pinfo,
2228                                        u32 *idx)
2229 {
2230         *idx += 1;
2231         if (!(*pinfo) || *idx != (*pinfo)->cnt)
2232                 return ++pbl;
2233         *idx = 0;
2234         (*pinfo)++;
2235
2236         return (*pinfo)->addr;
2237 }
2238
2239 /**
2240  * irdma_copy_user_pgaddrs - copy user page address to pble's os locally
2241  * @iwmr: iwmr for IB's user page addresses
2242  * @pbl: ple pointer to save 1 level or 0 level pble
2243  * @level: indicated level 0, 1 or 2
2244  */
2245 static void irdma_copy_user_pgaddrs(struct irdma_mr *iwmr, u64 *pbl,
2246                                     enum irdma_pble_level level)
2247 {
2248         struct ib_umem *region = iwmr->region;
2249         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2250         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2251         struct irdma_pble_info *pinfo;
2252         struct ib_block_iter biter;
2253         u32 idx = 0;
2254         u32 pbl_cnt = 0;
2255
2256         pinfo = (level == PBLE_LEVEL_1) ? NULL : palloc->level2.leaf;
2257
2258         if (iwmr->type == IRDMA_MEMREG_TYPE_QP)
2259                 iwpbl->qp_mr.sq_page = sg_page(region->sgt_append.sgt.sgl);
2260
2261         rdma_umem_for_each_dma_block(region, &biter, iwmr->page_size) {
2262                 *pbl = rdma_block_iter_dma_address(&biter);
2263                 if (++pbl_cnt == palloc->total_cnt)
2264                         break;
2265                 pbl = irdma_next_pbl_addr(pbl, &pinfo, &idx);
2266         }
2267 }
2268
2269 /**
2270  * irdma_check_mem_contiguous - check if pbls stored in arr are contiguous
2271  * @arr: lvl1 pbl array
2272  * @npages: page count
2273  * @pg_size: page size
2274  *
2275  */
2276 static bool irdma_check_mem_contiguous(u64 *arr, u32 npages, u32 pg_size)
2277 {
2278         u32 pg_idx;
2279
2280         for (pg_idx = 0; pg_idx < npages; pg_idx++) {
2281                 if ((*arr + (pg_size * pg_idx)) != arr[pg_idx])
2282                         return false;
2283         }
2284
2285         return true;
2286 }
2287
2288 /**
2289  * irdma_check_mr_contiguous - check if MR is physically contiguous
2290  * @palloc: pbl allocation struct
2291  * @pg_size: page size
2292  */
2293 static bool irdma_check_mr_contiguous(struct irdma_pble_alloc *palloc,
2294                                       u32 pg_size)
2295 {
2296         struct irdma_pble_level2 *lvl2 = &palloc->level2;
2297         struct irdma_pble_info *leaf = lvl2->leaf;
2298         u64 *arr = NULL;
2299         u64 *start_addr = NULL;
2300         int i;
2301         bool ret;
2302
2303         if (palloc->level == PBLE_LEVEL_1) {
2304                 arr = palloc->level1.addr;
2305                 ret = irdma_check_mem_contiguous(arr, palloc->total_cnt,
2306                                                  pg_size);
2307                 return ret;
2308         }
2309
2310         start_addr = leaf->addr;
2311
2312         for (i = 0; i < lvl2->leaf_cnt; i++, leaf++) {
2313                 arr = leaf->addr;
2314                 if ((*start_addr + (i * pg_size * PBLE_PER_PAGE)) != *arr)
2315                         return false;
2316                 ret = irdma_check_mem_contiguous(arr, leaf->cnt, pg_size);
2317                 if (!ret)
2318                         return false;
2319         }
2320
2321         return true;
2322 }
2323
2324 /**
2325  * irdma_setup_pbles - copy user pg address to pble's
2326  * @rf: RDMA PCI function
2327  * @iwmr: mr pointer for this memory registration
2328  * @lvl: requested pble levels
2329  */
2330 static int irdma_setup_pbles(struct irdma_pci_f *rf, struct irdma_mr *iwmr,
2331                              u8 lvl)
2332 {
2333         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2334         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2335         struct irdma_pble_info *pinfo;
2336         u64 *pbl;
2337         int status;
2338         enum irdma_pble_level level = PBLE_LEVEL_1;
2339
2340         if (lvl) {
2341                 status = irdma_get_pble(rf->pble_rsrc, palloc, iwmr->page_cnt,
2342                                         lvl);
2343                 if (status)
2344                         return status;
2345
2346                 iwpbl->pbl_allocated = true;
2347                 level = palloc->level;
2348                 pinfo = (level == PBLE_LEVEL_1) ? &palloc->level1 :
2349                                                   palloc->level2.leaf;
2350                 pbl = pinfo->addr;
2351         } else {
2352                 pbl = iwmr->pgaddrmem;
2353         }
2354
2355         irdma_copy_user_pgaddrs(iwmr, pbl, level);
2356
2357         if (lvl)
2358                 iwmr->pgaddrmem[0] = *pbl;
2359
2360         return 0;
2361 }
2362
2363 /**
2364  * irdma_handle_q_mem - handle memory for qp and cq
2365  * @iwdev: irdma device
2366  * @req: information for q memory management
2367  * @iwpbl: pble struct
2368  * @lvl: pble level mask
2369  */
2370 static int irdma_handle_q_mem(struct irdma_device *iwdev,
2371                               struct irdma_mem_reg_req *req,
2372                               struct irdma_pbl *iwpbl, u8 lvl)
2373 {
2374         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2375         struct irdma_mr *iwmr = iwpbl->iwmr;
2376         struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;
2377         struct irdma_cq_mr *cqmr = &iwpbl->cq_mr;
2378         struct irdma_hmc_pble *hmc_p;
2379         u64 *arr = iwmr->pgaddrmem;
2380         u32 pg_size, total;
2381         int err = 0;
2382         bool ret = true;
2383
2384         pg_size = iwmr->page_size;
2385         err = irdma_setup_pbles(iwdev->rf, iwmr, lvl);
2386         if (err)
2387                 return err;
2388
2389         if (lvl)
2390                 arr = palloc->level1.addr;
2391
2392         switch (iwmr->type) {
2393         case IRDMA_MEMREG_TYPE_QP:
2394                 total = req->sq_pages + req->rq_pages;
2395                 hmc_p = &qpmr->sq_pbl;
2396                 qpmr->shadow = (dma_addr_t)arr[total];
2397
2398                 if (lvl) {
2399                         ret = irdma_check_mem_contiguous(arr, req->sq_pages,
2400                                                          pg_size);
2401                         if (ret)
2402                                 ret = irdma_check_mem_contiguous(&arr[req->sq_pages],
2403                                                                  req->rq_pages,
2404                                                                  pg_size);
2405                 }
2406
2407                 if (!ret) {
2408                         hmc_p->idx = palloc->level1.idx;
2409                         hmc_p = &qpmr->rq_pbl;
2410                         hmc_p->idx = palloc->level1.idx + req->sq_pages;
2411                 } else {
2412                         hmc_p->addr = arr[0];
2413                         hmc_p = &qpmr->rq_pbl;
2414                         hmc_p->addr = arr[req->sq_pages];
2415                 }
2416                 break;
2417         case IRDMA_MEMREG_TYPE_CQ:
2418                 hmc_p = &cqmr->cq_pbl;
2419
2420                 if (!cqmr->split)
2421                         cqmr->shadow = (dma_addr_t)arr[req->cq_pages];
2422
2423                 if (lvl)
2424                         ret = irdma_check_mem_contiguous(arr, req->cq_pages,
2425                                                          pg_size);
2426
2427                 if (!ret)
2428                         hmc_p->idx = palloc->level1.idx;
2429                 else
2430                         hmc_p->addr = arr[0];
2431         break;
2432         default:
2433                 ibdev_dbg(&iwdev->ibdev, "VERBS: MR type error\n");
2434                 err = -EINVAL;
2435         }
2436
2437         if (lvl && ret) {
2438                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2439                 iwpbl->pbl_allocated = false;
2440         }
2441
2442         return err;
2443 }
2444
2445 /**
2446  * irdma_hw_alloc_mw - create the hw memory window
2447  * @iwdev: irdma device
2448  * @iwmr: pointer to memory window info
2449  */
2450 static int irdma_hw_alloc_mw(struct irdma_device *iwdev, struct irdma_mr *iwmr)
2451 {
2452         struct irdma_mw_alloc_info *info;
2453         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2454         struct irdma_cqp_request *cqp_request;
2455         struct cqp_cmds_info *cqp_info;
2456         int status;
2457
2458         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2459         if (!cqp_request)
2460                 return -ENOMEM;
2461
2462         cqp_info = &cqp_request->info;
2463         info = &cqp_info->in.u.mw_alloc.info;
2464         memset(info, 0, sizeof(*info));
2465         if (iwmr->ibmw.type == IB_MW_TYPE_1)
2466                 info->mw_wide = true;
2467
2468         info->page_size = PAGE_SIZE;
2469         info->mw_stag_index = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2470         info->pd_id = iwpd->sc_pd.pd_id;
2471         info->remote_access = true;
2472         cqp_info->cqp_cmd = IRDMA_OP_MW_ALLOC;
2473         cqp_info->post_sq = 1;
2474         cqp_info->in.u.mw_alloc.dev = &iwdev->rf->sc_dev;
2475         cqp_info->in.u.mw_alloc.scratch = (uintptr_t)cqp_request;
2476         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2477         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2478
2479         return status;
2480 }
2481
2482 /**
2483  * irdma_alloc_mw - Allocate memory window
2484  * @ibmw: Memory Window
2485  * @udata: user data pointer
2486  */
2487 static int irdma_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata)
2488 {
2489         struct irdma_device *iwdev = to_iwdev(ibmw->device);
2490         struct irdma_mr *iwmr = to_iwmw(ibmw);
2491         int err_code;
2492         u32 stag;
2493
2494         stag = irdma_create_stag(iwdev);
2495         if (!stag)
2496                 return -ENOMEM;
2497
2498         iwmr->stag = stag;
2499         ibmw->rkey = stag;
2500
2501         err_code = irdma_hw_alloc_mw(iwdev, iwmr);
2502         if (err_code) {
2503                 irdma_free_stag(iwdev, stag);
2504                 return err_code;
2505         }
2506
2507         return 0;
2508 }
2509
2510 /**
2511  * irdma_dealloc_mw - Dealloc memory window
2512  * @ibmw: memory window structure.
2513  */
2514 static int irdma_dealloc_mw(struct ib_mw *ibmw)
2515 {
2516         struct ib_pd *ibpd = ibmw->pd;
2517         struct irdma_pd *iwpd = to_iwpd(ibpd);
2518         struct irdma_mr *iwmr = to_iwmr((struct ib_mr *)ibmw);
2519         struct irdma_device *iwdev = to_iwdev(ibmw->device);
2520         struct irdma_cqp_request *cqp_request;
2521         struct cqp_cmds_info *cqp_info;
2522         struct irdma_dealloc_stag_info *info;
2523
2524         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2525         if (!cqp_request)
2526                 return -ENOMEM;
2527
2528         cqp_info = &cqp_request->info;
2529         info = &cqp_info->in.u.dealloc_stag.info;
2530         memset(info, 0, sizeof(*info));
2531         info->pd_id = iwpd->sc_pd.pd_id;
2532         info->stag_idx = ibmw->rkey >> IRDMA_CQPSQ_STAG_IDX_S;
2533         info->mr = false;
2534         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
2535         cqp_info->post_sq = 1;
2536         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
2537         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
2538         irdma_handle_cqp_op(iwdev->rf, cqp_request);
2539         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2540         irdma_free_stag(iwdev, iwmr->stag);
2541
2542         return 0;
2543 }
2544
2545 /**
2546  * irdma_hw_alloc_stag - cqp command to allocate stag
2547  * @iwdev: irdma device
2548  * @iwmr: irdma mr pointer
2549  */
2550 static int irdma_hw_alloc_stag(struct irdma_device *iwdev,
2551                                struct irdma_mr *iwmr)
2552 {
2553         struct irdma_allocate_stag_info *info;
2554         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2555         int status;
2556         struct irdma_cqp_request *cqp_request;
2557         struct cqp_cmds_info *cqp_info;
2558
2559         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2560         if (!cqp_request)
2561                 return -ENOMEM;
2562
2563         cqp_info = &cqp_request->info;
2564         info = &cqp_info->in.u.alloc_stag.info;
2565         memset(info, 0, sizeof(*info));
2566         info->page_size = PAGE_SIZE;
2567         info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2568         info->pd_id = iwpd->sc_pd.pd_id;
2569         info->total_len = iwmr->len;
2570         info->remote_access = true;
2571         cqp_info->cqp_cmd = IRDMA_OP_ALLOC_STAG;
2572         cqp_info->post_sq = 1;
2573         cqp_info->in.u.alloc_stag.dev = &iwdev->rf->sc_dev;
2574         cqp_info->in.u.alloc_stag.scratch = (uintptr_t)cqp_request;
2575         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2576         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2577
2578         return status;
2579 }
2580
2581 /**
2582  * irdma_alloc_mr - register stag for fast memory registration
2583  * @pd: ibpd pointer
2584  * @mr_type: memory for stag registrion
2585  * @max_num_sg: man number of pages
2586  */
2587 static struct ib_mr *irdma_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
2588                                     u32 max_num_sg)
2589 {
2590         struct irdma_device *iwdev = to_iwdev(pd->device);
2591         struct irdma_pble_alloc *palloc;
2592         struct irdma_pbl *iwpbl;
2593         struct irdma_mr *iwmr;
2594         u32 stag;
2595         int err_code;
2596
2597         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2598         if (!iwmr)
2599                 return ERR_PTR(-ENOMEM);
2600
2601         stag = irdma_create_stag(iwdev);
2602         if (!stag) {
2603                 err_code = -ENOMEM;
2604                 goto err;
2605         }
2606
2607         iwmr->stag = stag;
2608         iwmr->ibmr.rkey = stag;
2609         iwmr->ibmr.lkey = stag;
2610         iwmr->ibmr.pd = pd;
2611         iwmr->ibmr.device = pd->device;
2612         iwpbl = &iwmr->iwpbl;
2613         iwpbl->iwmr = iwmr;
2614         iwmr->type = IRDMA_MEMREG_TYPE_MEM;
2615         palloc = &iwpbl->pble_alloc;
2616         iwmr->page_cnt = max_num_sg;
2617         err_code = irdma_get_pble(iwdev->rf->pble_rsrc, palloc, iwmr->page_cnt,
2618                                   false);
2619         if (err_code)
2620                 goto err_get_pble;
2621
2622         err_code = irdma_hw_alloc_stag(iwdev, iwmr);
2623         if (err_code)
2624                 goto err_alloc_stag;
2625
2626         iwpbl->pbl_allocated = true;
2627
2628         return &iwmr->ibmr;
2629 err_alloc_stag:
2630         irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2631 err_get_pble:
2632         irdma_free_stag(iwdev, stag);
2633 err:
2634         kfree(iwmr);
2635
2636         return ERR_PTR(err_code);
2637 }
2638
2639 /**
2640  * irdma_set_page - populate pbl list for fmr
2641  * @ibmr: ib mem to access iwarp mr pointer
2642  * @addr: page dma address fro pbl list
2643  */
2644 static int irdma_set_page(struct ib_mr *ibmr, u64 addr)
2645 {
2646         struct irdma_mr *iwmr = to_iwmr(ibmr);
2647         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2648         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2649         u64 *pbl;
2650
2651         if (unlikely(iwmr->npages == iwmr->page_cnt))
2652                 return -ENOMEM;
2653
2654         if (palloc->level == PBLE_LEVEL_2) {
2655                 struct irdma_pble_info *palloc_info =
2656                         palloc->level2.leaf + (iwmr->npages >> PBLE_512_SHIFT);
2657
2658                 palloc_info->addr[iwmr->npages & (PBLE_PER_PAGE - 1)] = addr;
2659         } else {
2660                 pbl = palloc->level1.addr;
2661                 pbl[iwmr->npages] = addr;
2662         }
2663         iwmr->npages++;
2664
2665         return 0;
2666 }
2667
2668 /**
2669  * irdma_map_mr_sg - map of sg list for fmr
2670  * @ibmr: ib mem to access iwarp mr pointer
2671  * @sg: scatter gather list
2672  * @sg_nents: number of sg pages
2673  * @sg_offset: scatter gather list for fmr
2674  */
2675 static int irdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
2676                            int sg_nents, unsigned int *sg_offset)
2677 {
2678         struct irdma_mr *iwmr = to_iwmr(ibmr);
2679
2680         iwmr->npages = 0;
2681
2682         return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, irdma_set_page);
2683 }
2684
2685 /**
2686  * irdma_hwreg_mr - send cqp command for memory registration
2687  * @iwdev: irdma device
2688  * @iwmr: irdma mr pointer
2689  * @access: access for MR
2690  */
2691 static int irdma_hwreg_mr(struct irdma_device *iwdev, struct irdma_mr *iwmr,
2692                           u16 access)
2693 {
2694         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2695         struct irdma_reg_ns_stag_info *stag_info;
2696         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2697         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2698         struct irdma_cqp_request *cqp_request;
2699         struct cqp_cmds_info *cqp_info;
2700         int ret;
2701
2702         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2703         if (!cqp_request)
2704                 return -ENOMEM;
2705
2706         cqp_info = &cqp_request->info;
2707         stag_info = &cqp_info->in.u.mr_reg_non_shared.info;
2708         memset(stag_info, 0, sizeof(*stag_info));
2709         stag_info->va = iwpbl->user_base;
2710         stag_info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2711         stag_info->stag_key = (u8)iwmr->stag;
2712         stag_info->total_len = iwmr->len;
2713         stag_info->access_rights = irdma_get_mr_access(access);
2714         stag_info->pd_id = iwpd->sc_pd.pd_id;
2715         if (stag_info->access_rights & IRDMA_ACCESS_FLAGS_ZERO_BASED)
2716                 stag_info->addr_type = IRDMA_ADDR_TYPE_ZERO_BASED;
2717         else
2718                 stag_info->addr_type = IRDMA_ADDR_TYPE_VA_BASED;
2719         stag_info->page_size = iwmr->page_size;
2720
2721         if (iwpbl->pbl_allocated) {
2722                 if (palloc->level == PBLE_LEVEL_1) {
2723                         stag_info->first_pm_pbl_index = palloc->level1.idx;
2724                         stag_info->chunk_size = 1;
2725                 } else {
2726                         stag_info->first_pm_pbl_index = palloc->level2.root.idx;
2727                         stag_info->chunk_size = 3;
2728                 }
2729         } else {
2730                 stag_info->reg_addr_pa = iwmr->pgaddrmem[0];
2731         }
2732
2733         cqp_info->cqp_cmd = IRDMA_OP_MR_REG_NON_SHARED;
2734         cqp_info->post_sq = 1;
2735         cqp_info->in.u.mr_reg_non_shared.dev = &iwdev->rf->sc_dev;
2736         cqp_info->in.u.mr_reg_non_shared.scratch = (uintptr_t)cqp_request;
2737         ret = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2738         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2739
2740         return ret;
2741 }
2742
2743 static int irdma_reg_user_mr_type_mem(struct irdma_mr *iwmr, int access)
2744 {
2745         struct irdma_device *iwdev = to_iwdev(iwmr->ibmr.device);
2746         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2747         u32 stag;
2748         u8 lvl;
2749         int err;
2750
2751         lvl = iwmr->page_cnt != 1 ? PBLE_LEVEL_1 | PBLE_LEVEL_2 : PBLE_LEVEL_0;
2752
2753         err = irdma_setup_pbles(iwdev->rf, iwmr, lvl);
2754         if (err)
2755                 return err;
2756
2757         if (lvl) {
2758                 err = irdma_check_mr_contiguous(&iwpbl->pble_alloc,
2759                                                 iwmr->page_size);
2760                 if (err) {
2761                         irdma_free_pble(iwdev->rf->pble_rsrc, &iwpbl->pble_alloc);
2762                         iwpbl->pbl_allocated = false;
2763                 }
2764         }
2765
2766         stag = irdma_create_stag(iwdev);
2767         if (!stag) {
2768                 err = -ENOMEM;
2769                 goto free_pble;
2770         }
2771
2772         iwmr->stag = stag;
2773         iwmr->ibmr.rkey = stag;
2774         iwmr->ibmr.lkey = stag;
2775         err = irdma_hwreg_mr(iwdev, iwmr, access);
2776         if (err)
2777                 goto err_hwreg;
2778
2779         return 0;
2780
2781 err_hwreg:
2782         irdma_free_stag(iwdev, stag);
2783
2784 free_pble:
2785         if (iwpbl->pble_alloc.level != PBLE_LEVEL_0 && iwpbl->pbl_allocated)
2786                 irdma_free_pble(iwdev->rf->pble_rsrc, &iwpbl->pble_alloc);
2787
2788         return err;
2789 }
2790
2791 static struct irdma_mr *irdma_alloc_iwmr(struct ib_umem *region,
2792                                          struct ib_pd *pd, u64 virt,
2793                                          enum irdma_memreg_type reg_type)
2794 {
2795         struct irdma_device *iwdev = to_iwdev(pd->device);
2796         struct irdma_pbl *iwpbl = NULL;
2797         struct irdma_mr *iwmr = NULL;
2798         unsigned long pgsz_bitmap;
2799
2800         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2801         if (!iwmr)
2802                 return ERR_PTR(-ENOMEM);
2803
2804         iwpbl = &iwmr->iwpbl;
2805         iwpbl->iwmr = iwmr;
2806         iwmr->region = region;
2807         iwmr->ibmr.pd = pd;
2808         iwmr->ibmr.device = pd->device;
2809         iwmr->ibmr.iova = virt;
2810         iwmr->type = reg_type;
2811
2812         pgsz_bitmap = (reg_type == IRDMA_MEMREG_TYPE_MEM) ?
2813                 iwdev->rf->sc_dev.hw_attrs.page_size_cap : PAGE_SIZE;
2814
2815         iwmr->page_size = ib_umem_find_best_pgsz(region, pgsz_bitmap, virt);
2816         if (unlikely(!iwmr->page_size)) {
2817                 kfree(iwmr);
2818                 return ERR_PTR(-EOPNOTSUPP);
2819         }
2820
2821         iwmr->len = region->length;
2822         iwpbl->user_base = virt;
2823         iwmr->page_cnt = ib_umem_num_dma_blocks(region, iwmr->page_size);
2824
2825         return iwmr;
2826 }
2827
2828 static void irdma_free_iwmr(struct irdma_mr *iwmr)
2829 {
2830         kfree(iwmr);
2831 }
2832
2833 static int irdma_reg_user_mr_type_qp(struct irdma_mem_reg_req req,
2834                                      struct ib_udata *udata,
2835                                      struct irdma_mr *iwmr)
2836 {
2837         struct irdma_device *iwdev = to_iwdev(iwmr->ibmr.device);
2838         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2839         struct irdma_ucontext *ucontext = NULL;
2840         unsigned long flags;
2841         u32 total;
2842         int err;
2843         u8 lvl;
2844
2845         total = req.sq_pages + req.rq_pages + 1;
2846         if (total > iwmr->page_cnt)
2847                 return -EINVAL;
2848
2849         total = req.sq_pages + req.rq_pages;
2850         lvl = total > 2 ? PBLE_LEVEL_1 : PBLE_LEVEL_0;
2851         err = irdma_handle_q_mem(iwdev, &req, iwpbl, lvl);
2852         if (err)
2853                 return err;
2854
2855         ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext,
2856                                              ibucontext);
2857         spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
2858         list_add_tail(&iwpbl->list, &ucontext->qp_reg_mem_list);
2859         iwpbl->on_list = true;
2860         spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
2861
2862         return 0;
2863 }
2864
2865 static int irdma_reg_user_mr_type_cq(struct irdma_mem_reg_req req,
2866                                      struct ib_udata *udata,
2867                                      struct irdma_mr *iwmr)
2868 {
2869         struct irdma_device *iwdev = to_iwdev(iwmr->ibmr.device);
2870         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2871         struct irdma_ucontext *ucontext = NULL;
2872         u8 shadow_pgcnt = 1;
2873         unsigned long flags;
2874         u32 total;
2875         int err;
2876         u8 lvl;
2877
2878         if (iwdev->rf->sc_dev.hw_attrs.uk_attrs.feature_flags & IRDMA_FEATURE_CQ_RESIZE)
2879                 shadow_pgcnt = 0;
2880         total = req.cq_pages + shadow_pgcnt;
2881         if (total > iwmr->page_cnt)
2882                 return -EINVAL;
2883
2884         lvl = req.cq_pages > 1 ? PBLE_LEVEL_1 : PBLE_LEVEL_0;
2885         err = irdma_handle_q_mem(iwdev, &req, iwpbl, lvl);
2886         if (err)
2887                 return err;
2888
2889         ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext,
2890                                              ibucontext);
2891         spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2892         list_add_tail(&iwpbl->list, &ucontext->cq_reg_mem_list);
2893         iwpbl->on_list = true;
2894         spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2895
2896         return 0;
2897 }
2898
2899 /**
2900  * irdma_reg_user_mr - Register a user memory region
2901  * @pd: ptr of pd
2902  * @start: virtual start address
2903  * @len: length of mr
2904  * @virt: virtual address
2905  * @access: access of mr
2906  * @udata: user data
2907  */
2908 static struct ib_mr *irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,
2909                                        u64 virt, int access,
2910                                        struct ib_udata *udata)
2911 {
2912 #define IRDMA_MEM_REG_MIN_REQ_LEN offsetofend(struct irdma_mem_reg_req, sq_pages)
2913         struct irdma_device *iwdev = to_iwdev(pd->device);
2914         struct irdma_mem_reg_req req = {};
2915         struct ib_umem *region = NULL;
2916         struct irdma_mr *iwmr = NULL;
2917         int err;
2918
2919         if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size)
2920                 return ERR_PTR(-EINVAL);
2921
2922         if (udata->inlen < IRDMA_MEM_REG_MIN_REQ_LEN)
2923                 return ERR_PTR(-EINVAL);
2924
2925         region = ib_umem_get(pd->device, start, len, access);
2926
2927         if (IS_ERR(region)) {
2928                 ibdev_dbg(&iwdev->ibdev,
2929                           "VERBS: Failed to create ib_umem region\n");
2930                 return (struct ib_mr *)region;
2931         }
2932
2933         if (ib_copy_from_udata(&req, udata, min(sizeof(req), udata->inlen))) {
2934                 ib_umem_release(region);
2935                 return ERR_PTR(-EFAULT);
2936         }
2937
2938         iwmr = irdma_alloc_iwmr(region, pd, virt, req.reg_type);
2939         if (IS_ERR(iwmr)) {
2940                 ib_umem_release(region);
2941                 return (struct ib_mr *)iwmr;
2942         }
2943
2944         switch (req.reg_type) {
2945         case IRDMA_MEMREG_TYPE_QP:
2946                 err = irdma_reg_user_mr_type_qp(req, udata, iwmr);
2947                 if (err)
2948                         goto error;
2949
2950                 break;
2951         case IRDMA_MEMREG_TYPE_CQ:
2952                 err = irdma_reg_user_mr_type_cq(req, udata, iwmr);
2953                 if (err)
2954                         goto error;
2955                 break;
2956         case IRDMA_MEMREG_TYPE_MEM:
2957                 err = irdma_reg_user_mr_type_mem(iwmr, access);
2958                 if (err)
2959                         goto error;
2960
2961                 break;
2962         default:
2963                 err = -EINVAL;
2964                 goto error;
2965         }
2966
2967         return &iwmr->ibmr;
2968 error:
2969         ib_umem_release(region);
2970         irdma_free_iwmr(iwmr);
2971
2972         return ERR_PTR(err);
2973 }
2974
2975 static struct ib_mr *irdma_reg_user_mr_dmabuf(struct ib_pd *pd, u64 start,
2976                                               u64 len, u64 virt,
2977                                               int fd, int access,
2978                                               struct ib_udata *udata)
2979 {
2980         struct irdma_device *iwdev = to_iwdev(pd->device);
2981         struct ib_umem_dmabuf *umem_dmabuf;
2982         struct irdma_mr *iwmr;
2983         int err;
2984
2985         if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size)
2986                 return ERR_PTR(-EINVAL);
2987
2988         umem_dmabuf = ib_umem_dmabuf_get_pinned(pd->device, start, len, fd, access);
2989         if (IS_ERR(umem_dmabuf)) {
2990                 err = PTR_ERR(umem_dmabuf);
2991                 ibdev_dbg(&iwdev->ibdev, "Failed to get dmabuf umem[%d]\n", err);
2992                 return ERR_PTR(err);
2993         }
2994
2995         iwmr = irdma_alloc_iwmr(&umem_dmabuf->umem, pd, virt, IRDMA_MEMREG_TYPE_MEM);
2996         if (IS_ERR(iwmr)) {
2997                 err = PTR_ERR(iwmr);
2998                 goto err_release;
2999         }
3000
3001         err = irdma_reg_user_mr_type_mem(iwmr, access);
3002         if (err)
3003                 goto err_iwmr;
3004
3005         return &iwmr->ibmr;
3006
3007 err_iwmr:
3008         irdma_free_iwmr(iwmr);
3009
3010 err_release:
3011         ib_umem_release(&umem_dmabuf->umem);
3012
3013         return ERR_PTR(err);
3014 }
3015
3016 /**
3017  * irdma_reg_phys_mr - register kernel physical memory
3018  * @pd: ibpd pointer
3019  * @addr: physical address of memory to register
3020  * @size: size of memory to register
3021  * @access: Access rights
3022  * @iova_start: start of virtual address for physical buffers
3023  */
3024 struct ib_mr *irdma_reg_phys_mr(struct ib_pd *pd, u64 addr, u64 size, int access,
3025                                 u64 *iova_start)
3026 {
3027         struct irdma_device *iwdev = to_iwdev(pd->device);
3028         struct irdma_pbl *iwpbl;
3029         struct irdma_mr *iwmr;
3030         u32 stag;
3031         int ret;
3032
3033         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
3034         if (!iwmr)
3035                 return ERR_PTR(-ENOMEM);
3036
3037         iwmr->ibmr.pd = pd;
3038         iwmr->ibmr.device = pd->device;
3039         iwpbl = &iwmr->iwpbl;
3040         iwpbl->iwmr = iwmr;
3041         iwmr->type = IRDMA_MEMREG_TYPE_MEM;
3042         iwpbl->user_base = *iova_start;
3043         stag = irdma_create_stag(iwdev);
3044         if (!stag) {
3045                 ret = -ENOMEM;
3046                 goto err;
3047         }
3048
3049         iwmr->stag = stag;
3050         iwmr->ibmr.iova = *iova_start;
3051         iwmr->ibmr.rkey = stag;
3052         iwmr->ibmr.lkey = stag;
3053         iwmr->page_cnt = 1;
3054         iwmr->pgaddrmem[0] = addr;
3055         iwmr->len = size;
3056         iwmr->page_size = SZ_4K;
3057         ret = irdma_hwreg_mr(iwdev, iwmr, access);
3058         if (ret) {
3059                 irdma_free_stag(iwdev, stag);
3060                 goto err;
3061         }
3062
3063         return &iwmr->ibmr;
3064
3065 err:
3066         kfree(iwmr);
3067
3068         return ERR_PTR(ret);
3069 }
3070
3071 /**
3072  * irdma_get_dma_mr - register physical mem
3073  * @pd: ptr of pd
3074  * @acc: access for memory
3075  */
3076 static struct ib_mr *irdma_get_dma_mr(struct ib_pd *pd, int acc)
3077 {
3078         u64 kva = 0;
3079
3080         return irdma_reg_phys_mr(pd, 0, 0, acc, &kva);
3081 }
3082
3083 /**
3084  * irdma_del_memlist - Deleting pbl list entries for CQ/QP
3085  * @iwmr: iwmr for IB's user page addresses
3086  * @ucontext: ptr to user context
3087  */
3088 static void irdma_del_memlist(struct irdma_mr *iwmr,
3089                               struct irdma_ucontext *ucontext)
3090 {
3091         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
3092         unsigned long flags;
3093
3094         switch (iwmr->type) {
3095         case IRDMA_MEMREG_TYPE_CQ:
3096                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
3097                 if (iwpbl->on_list) {
3098                         iwpbl->on_list = false;
3099                         list_del(&iwpbl->list);
3100                 }
3101                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
3102                 break;
3103         case IRDMA_MEMREG_TYPE_QP:
3104                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
3105                 if (iwpbl->on_list) {
3106                         iwpbl->on_list = false;
3107                         list_del(&iwpbl->list);
3108                 }
3109                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
3110                 break;
3111         default:
3112                 break;
3113         }
3114 }
3115
3116 /**
3117  * irdma_dereg_mr - deregister mr
3118  * @ib_mr: mr ptr for dereg
3119  * @udata: user data
3120  */
3121 static int irdma_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
3122 {
3123         struct ib_pd *ibpd = ib_mr->pd;
3124         struct irdma_pd *iwpd = to_iwpd(ibpd);
3125         struct irdma_mr *iwmr = to_iwmr(ib_mr);
3126         struct irdma_device *iwdev = to_iwdev(ib_mr->device);
3127         struct irdma_dealloc_stag_info *info;
3128         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
3129         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
3130         struct irdma_cqp_request *cqp_request;
3131         struct cqp_cmds_info *cqp_info;
3132         int status;
3133
3134         if (iwmr->type != IRDMA_MEMREG_TYPE_MEM) {
3135                 if (iwmr->region) {
3136                         struct irdma_ucontext *ucontext;
3137
3138                         ucontext = rdma_udata_to_drv_context(udata,
3139                                                 struct irdma_ucontext,
3140                                                 ibucontext);
3141                         irdma_del_memlist(iwmr, ucontext);
3142                 }
3143                 goto done;
3144         }
3145
3146         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
3147         if (!cqp_request)
3148                 return -ENOMEM;
3149
3150         cqp_info = &cqp_request->info;
3151         info = &cqp_info->in.u.dealloc_stag.info;
3152         memset(info, 0, sizeof(*info));
3153         info->pd_id = iwpd->sc_pd.pd_id;
3154         info->stag_idx = ib_mr->rkey >> IRDMA_CQPSQ_STAG_IDX_S;
3155         info->mr = true;
3156         if (iwpbl->pbl_allocated)
3157                 info->dealloc_pbl = true;
3158
3159         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
3160         cqp_info->post_sq = 1;
3161         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
3162         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
3163         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
3164         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
3165         if (status)
3166                 return status;
3167
3168         irdma_free_stag(iwdev, iwmr->stag);
3169 done:
3170         if (iwpbl->pbl_allocated)
3171                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
3172         ib_umem_release(iwmr->region);
3173         kfree(iwmr);
3174
3175         return 0;
3176 }
3177
3178 /**
3179  * irdma_post_send -  kernel application wr
3180  * @ibqp: qp ptr for wr
3181  * @ib_wr: work request ptr
3182  * @bad_wr: return of bad wr if err
3183  */
3184 static int irdma_post_send(struct ib_qp *ibqp,
3185                            const struct ib_send_wr *ib_wr,
3186                            const struct ib_send_wr **bad_wr)
3187 {
3188         struct irdma_qp *iwqp;
3189         struct irdma_qp_uk *ukqp;
3190         struct irdma_sc_dev *dev;
3191         struct irdma_post_sq_info info;
3192         int err = 0;
3193         unsigned long flags;
3194         bool inv_stag;
3195         struct irdma_ah *ah;
3196
3197         iwqp = to_iwqp(ibqp);
3198         ukqp = &iwqp->sc_qp.qp_uk;
3199         dev = &iwqp->iwdev->rf->sc_dev;
3200
3201         spin_lock_irqsave(&iwqp->lock, flags);
3202         while (ib_wr) {
3203                 memset(&info, 0, sizeof(info));
3204                 inv_stag = false;
3205                 info.wr_id = (ib_wr->wr_id);
3206                 if ((ib_wr->send_flags & IB_SEND_SIGNALED) || iwqp->sig_all)
3207                         info.signaled = true;
3208                 if (ib_wr->send_flags & IB_SEND_FENCE)
3209                         info.read_fence = true;
3210                 switch (ib_wr->opcode) {
3211                 case IB_WR_SEND_WITH_IMM:
3212                         if (ukqp->qp_caps & IRDMA_SEND_WITH_IMM) {
3213                                 info.imm_data_valid = true;
3214                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
3215                         } else {
3216                                 err = -EINVAL;
3217                                 break;
3218                         }
3219                         fallthrough;
3220                 case IB_WR_SEND:
3221                 case IB_WR_SEND_WITH_INV:
3222                         if (ib_wr->opcode == IB_WR_SEND ||
3223                             ib_wr->opcode == IB_WR_SEND_WITH_IMM) {
3224                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
3225                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL;
3226                                 else
3227                                         info.op_type = IRDMA_OP_TYPE_SEND;
3228                         } else {
3229                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
3230                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL_INV;
3231                                 else
3232                                         info.op_type = IRDMA_OP_TYPE_SEND_INV;
3233                                 info.stag_to_inv = ib_wr->ex.invalidate_rkey;
3234                         }
3235
3236                         info.op.send.num_sges = ib_wr->num_sge;
3237                         info.op.send.sg_list = ib_wr->sg_list;
3238                         if (iwqp->ibqp.qp_type == IB_QPT_UD ||
3239                             iwqp->ibqp.qp_type == IB_QPT_GSI) {
3240                                 ah = to_iwah(ud_wr(ib_wr)->ah);
3241                                 info.op.send.ah_id = ah->sc_ah.ah_info.ah_idx;
3242                                 info.op.send.qkey = ud_wr(ib_wr)->remote_qkey;
3243                                 info.op.send.dest_qp = ud_wr(ib_wr)->remote_qpn;
3244                         }
3245
3246                         if (ib_wr->send_flags & IB_SEND_INLINE)
3247                                 err = irdma_uk_inline_send(ukqp, &info, false);
3248                         else
3249                                 err = irdma_uk_send(ukqp, &info, false);
3250                         break;
3251                 case IB_WR_RDMA_WRITE_WITH_IMM:
3252                         if (ukqp->qp_caps & IRDMA_WRITE_WITH_IMM) {
3253                                 info.imm_data_valid = true;
3254                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
3255                         } else {
3256                                 err = -EINVAL;
3257                                 break;
3258                         }
3259                         fallthrough;
3260                 case IB_WR_RDMA_WRITE:
3261                         if (ib_wr->send_flags & IB_SEND_SOLICITED)
3262                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE_SOL;
3263                         else
3264                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE;
3265
3266                         info.op.rdma_write.num_lo_sges = ib_wr->num_sge;
3267                         info.op.rdma_write.lo_sg_list = ib_wr->sg_list;
3268                         info.op.rdma_write.rem_addr.addr =
3269                                 rdma_wr(ib_wr)->remote_addr;
3270                         info.op.rdma_write.rem_addr.lkey = rdma_wr(ib_wr)->rkey;
3271                         if (ib_wr->send_flags & IB_SEND_INLINE)
3272                                 err = irdma_uk_inline_rdma_write(ukqp, &info, false);
3273                         else
3274                                 err = irdma_uk_rdma_write(ukqp, &info, false);
3275                         break;
3276                 case IB_WR_RDMA_READ_WITH_INV:
3277                         inv_stag = true;
3278                         fallthrough;
3279                 case IB_WR_RDMA_READ:
3280                         if (ib_wr->num_sge >
3281                             dev->hw_attrs.uk_attrs.max_hw_read_sges) {
3282                                 err = -EINVAL;
3283                                 break;
3284                         }
3285                         info.op_type = IRDMA_OP_TYPE_RDMA_READ;
3286                         info.op.rdma_read.rem_addr.addr = rdma_wr(ib_wr)->remote_addr;
3287                         info.op.rdma_read.rem_addr.lkey = rdma_wr(ib_wr)->rkey;
3288                         info.op.rdma_read.lo_sg_list = (void *)ib_wr->sg_list;
3289                         info.op.rdma_read.num_lo_sges = ib_wr->num_sge;
3290                         err = irdma_uk_rdma_read(ukqp, &info, inv_stag, false);
3291                         break;
3292                 case IB_WR_LOCAL_INV:
3293                         info.op_type = IRDMA_OP_TYPE_INV_STAG;
3294                         info.op.inv_local_stag.target_stag = ib_wr->ex.invalidate_rkey;
3295                         err = irdma_uk_stag_local_invalidate(ukqp, &info, true);
3296                         break;
3297                 case IB_WR_REG_MR: {
3298                         struct irdma_mr *iwmr = to_iwmr(reg_wr(ib_wr)->mr);
3299                         struct irdma_pble_alloc *palloc = &iwmr->iwpbl.pble_alloc;
3300                         struct irdma_fast_reg_stag_info stag_info = {};
3301
3302                         stag_info.signaled = info.signaled;
3303                         stag_info.read_fence = info.read_fence;
3304                         stag_info.access_rights = irdma_get_mr_access(reg_wr(ib_wr)->access);
3305                         stag_info.stag_key = reg_wr(ib_wr)->key & 0xff;
3306                         stag_info.stag_idx = reg_wr(ib_wr)->key >> 8;
3307                         stag_info.page_size = reg_wr(ib_wr)->mr->page_size;
3308                         stag_info.wr_id = ib_wr->wr_id;
3309                         stag_info.addr_type = IRDMA_ADDR_TYPE_VA_BASED;
3310                         stag_info.va = (void *)(uintptr_t)iwmr->ibmr.iova;
3311                         stag_info.total_len = iwmr->ibmr.length;
3312                         stag_info.reg_addr_pa = *palloc->level1.addr;
3313                         stag_info.first_pm_pbl_index = palloc->level1.idx;
3314                         stag_info.local_fence = ib_wr->send_flags & IB_SEND_FENCE;
3315                         if (iwmr->npages > IRDMA_MIN_PAGES_PER_FMR)
3316                                 stag_info.chunk_size = 1;
3317                         err = irdma_sc_mr_fast_register(&iwqp->sc_qp, &stag_info,
3318                                                         true);
3319                         break;
3320                 }
3321                 default:
3322                         err = -EINVAL;
3323                         ibdev_dbg(&iwqp->iwdev->ibdev,
3324                                   "VERBS: upost_send bad opcode = 0x%x\n",
3325                                   ib_wr->opcode);
3326                         break;
3327                 }
3328
3329                 if (err)
3330                         break;
3331                 ib_wr = ib_wr->next;
3332         }
3333
3334         if (!iwqp->flush_issued) {
3335                 if (iwqp->hw_iwarp_state <= IRDMA_QP_STATE_RTS)
3336                         irdma_uk_qp_post_wr(ukqp);
3337                 spin_unlock_irqrestore(&iwqp->lock, flags);
3338         } else {
3339                 spin_unlock_irqrestore(&iwqp->lock, flags);
3340                 mod_delayed_work(iwqp->iwdev->cleanup_wq, &iwqp->dwork_flush,
3341                                  msecs_to_jiffies(IRDMA_FLUSH_DELAY_MS));
3342         }
3343         if (err)
3344                 *bad_wr = ib_wr;
3345
3346         return err;
3347 }
3348
3349 /**
3350  * irdma_post_recv - post receive wr for kernel application
3351  * @ibqp: ib qp pointer
3352  * @ib_wr: work request for receive
3353  * @bad_wr: bad wr caused an error
3354  */
3355 static int irdma_post_recv(struct ib_qp *ibqp,
3356                            const struct ib_recv_wr *ib_wr,
3357                            const struct ib_recv_wr **bad_wr)
3358 {
3359         struct irdma_qp *iwqp;
3360         struct irdma_qp_uk *ukqp;
3361         struct irdma_post_rq_info post_recv = {};
3362         unsigned long flags;
3363         int err = 0;
3364
3365         iwqp = to_iwqp(ibqp);
3366         ukqp = &iwqp->sc_qp.qp_uk;
3367
3368         spin_lock_irqsave(&iwqp->lock, flags);
3369         while (ib_wr) {
3370                 post_recv.num_sges = ib_wr->num_sge;
3371                 post_recv.wr_id = ib_wr->wr_id;
3372                 post_recv.sg_list = ib_wr->sg_list;
3373                 err = irdma_uk_post_receive(ukqp, &post_recv);
3374                 if (err) {
3375                         ibdev_dbg(&iwqp->iwdev->ibdev,
3376                                   "VERBS: post_recv err %d\n", err);
3377                         goto out;
3378                 }
3379
3380                 ib_wr = ib_wr->next;
3381         }
3382
3383 out:
3384         spin_unlock_irqrestore(&iwqp->lock, flags);
3385         if (iwqp->flush_issued)
3386                 mod_delayed_work(iwqp->iwdev->cleanup_wq, &iwqp->dwork_flush,
3387                                  msecs_to_jiffies(IRDMA_FLUSH_DELAY_MS));
3388
3389         if (err)
3390                 *bad_wr = ib_wr;
3391
3392         return err;
3393 }
3394
3395 /**
3396  * irdma_flush_err_to_ib_wc_status - return change flush error code to IB status
3397  * @opcode: iwarp flush code
3398  */
3399 static enum ib_wc_status irdma_flush_err_to_ib_wc_status(enum irdma_flush_opcode opcode)
3400 {
3401         switch (opcode) {
3402         case FLUSH_PROT_ERR:
3403                 return IB_WC_LOC_PROT_ERR;
3404         case FLUSH_REM_ACCESS_ERR:
3405                 return IB_WC_REM_ACCESS_ERR;
3406         case FLUSH_LOC_QP_OP_ERR:
3407                 return IB_WC_LOC_QP_OP_ERR;
3408         case FLUSH_REM_OP_ERR:
3409                 return IB_WC_REM_OP_ERR;
3410         case FLUSH_LOC_LEN_ERR:
3411                 return IB_WC_LOC_LEN_ERR;
3412         case FLUSH_GENERAL_ERR:
3413                 return IB_WC_WR_FLUSH_ERR;
3414         case FLUSH_RETRY_EXC_ERR:
3415                 return IB_WC_RETRY_EXC_ERR;
3416         case FLUSH_MW_BIND_ERR:
3417                 return IB_WC_MW_BIND_ERR;
3418         case FLUSH_REM_INV_REQ_ERR:
3419                 return IB_WC_REM_INV_REQ_ERR;
3420         case FLUSH_FATAL_ERR:
3421         default:
3422                 return IB_WC_FATAL_ERR;
3423         }
3424 }
3425
3426 /**
3427  * irdma_process_cqe - process cqe info
3428  * @entry: processed cqe
3429  * @cq_poll_info: cqe info
3430  */
3431 static void irdma_process_cqe(struct ib_wc *entry,
3432                               struct irdma_cq_poll_info *cq_poll_info)
3433 {
3434         struct irdma_sc_qp *qp;
3435
3436         entry->wc_flags = 0;
3437         entry->pkey_index = 0;
3438         entry->wr_id = cq_poll_info->wr_id;
3439
3440         qp = cq_poll_info->qp_handle;
3441         entry->qp = qp->qp_uk.back_qp;
3442
3443         if (cq_poll_info->error) {
3444                 entry->status = (cq_poll_info->comp_status == IRDMA_COMPL_STATUS_FLUSHED) ?
3445                                 irdma_flush_err_to_ib_wc_status(cq_poll_info->minor_err) : IB_WC_GENERAL_ERR;
3446
3447                 entry->vendor_err = cq_poll_info->major_err << 16 |
3448                                     cq_poll_info->minor_err;
3449         } else {
3450                 entry->status = IB_WC_SUCCESS;
3451                 if (cq_poll_info->imm_valid) {
3452                         entry->ex.imm_data = htonl(cq_poll_info->imm_data);
3453                         entry->wc_flags |= IB_WC_WITH_IMM;
3454                 }
3455                 if (cq_poll_info->ud_smac_valid) {
3456                         ether_addr_copy(entry->smac, cq_poll_info->ud_smac);
3457                         entry->wc_flags |= IB_WC_WITH_SMAC;
3458                 }
3459
3460                 if (cq_poll_info->ud_vlan_valid) {
3461                         u16 vlan = cq_poll_info->ud_vlan & VLAN_VID_MASK;
3462
3463                         entry->sl = cq_poll_info->ud_vlan >> VLAN_PRIO_SHIFT;
3464                         if (vlan) {
3465                                 entry->vlan_id = vlan;
3466                                 entry->wc_flags |= IB_WC_WITH_VLAN;
3467                         }
3468                 } else {
3469                         entry->sl = 0;
3470                 }
3471         }
3472
3473         if (cq_poll_info->q_type == IRDMA_CQE_QTYPE_SQ) {
3474                 set_ib_wc_op_sq(cq_poll_info, entry);
3475         } else {
3476                 set_ib_wc_op_rq(cq_poll_info, entry,
3477                                 qp->qp_uk.qp_caps & IRDMA_SEND_WITH_IMM ?
3478                                 true : false);
3479                 if (qp->qp_uk.qp_type != IRDMA_QP_TYPE_ROCE_UD &&
3480                     cq_poll_info->stag_invalid_set) {
3481                         entry->ex.invalidate_rkey = cq_poll_info->inv_stag;
3482                         entry->wc_flags |= IB_WC_WITH_INVALIDATE;
3483                 }
3484         }
3485
3486         if (qp->qp_uk.qp_type == IRDMA_QP_TYPE_ROCE_UD) {
3487                 entry->src_qp = cq_poll_info->ud_src_qpn;
3488                 entry->slid = 0;
3489                 entry->wc_flags |=
3490                         (IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE);
3491                 entry->network_hdr_type = cq_poll_info->ipv4 ?
3492                                                   RDMA_NETWORK_IPV4 :
3493                                                   RDMA_NETWORK_IPV6;
3494         } else {
3495                 entry->src_qp = cq_poll_info->qp_id;
3496         }
3497
3498         entry->byte_len = cq_poll_info->bytes_xfered;
3499 }
3500
3501 /**
3502  * irdma_poll_one - poll one entry of the CQ
3503  * @ukcq: ukcq to poll
3504  * @cur_cqe: current CQE info to be filled in
3505  * @entry: ibv_wc object to be filled for non-extended CQ or NULL for extended CQ
3506  *
3507  * Returns the internal irdma device error code or 0 on success
3508  */
3509 static inline int irdma_poll_one(struct irdma_cq_uk *ukcq,
3510                                  struct irdma_cq_poll_info *cur_cqe,
3511                                  struct ib_wc *entry)
3512 {
3513         int ret = irdma_uk_cq_poll_cmpl(ukcq, cur_cqe);
3514
3515         if (ret)
3516                 return ret;
3517
3518         irdma_process_cqe(entry, cur_cqe);
3519
3520         return 0;
3521 }
3522
3523 /**
3524  * __irdma_poll_cq - poll cq for completion (kernel apps)
3525  * @iwcq: cq to poll
3526  * @num_entries: number of entries to poll
3527  * @entry: wr of a completed entry
3528  */
3529 static int __irdma_poll_cq(struct irdma_cq *iwcq, int num_entries, struct ib_wc *entry)
3530 {
3531         struct list_head *tmp_node, *list_node;
3532         struct irdma_cq_buf *last_buf = NULL;
3533         struct irdma_cq_poll_info *cur_cqe = &iwcq->cur_cqe;
3534         struct irdma_cq_buf *cq_buf;
3535         int ret;
3536         struct irdma_device *iwdev;
3537         struct irdma_cq_uk *ukcq;
3538         bool cq_new_cqe = false;
3539         int resized_bufs = 0;
3540         int npolled = 0;
3541
3542         iwdev = to_iwdev(iwcq->ibcq.device);
3543         ukcq = &iwcq->sc_cq.cq_uk;
3544
3545         /* go through the list of previously resized CQ buffers */
3546         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
3547                 cq_buf = container_of(list_node, struct irdma_cq_buf, list);
3548                 while (npolled < num_entries) {
3549                         ret = irdma_poll_one(&cq_buf->cq_uk, cur_cqe, entry + npolled);
3550                         if (!ret) {
3551                                 ++npolled;
3552                                 cq_new_cqe = true;
3553                                 continue;
3554                         }
3555                         if (ret == -ENOENT)
3556                                 break;
3557                          /* QP using the CQ is destroyed. Skip reporting this CQE */
3558                         if (ret == -EFAULT) {
3559                                 cq_new_cqe = true;
3560                                 continue;
3561                         }
3562                         goto error;
3563                 }
3564
3565                 /* save the resized CQ buffer which received the last cqe */
3566                 if (cq_new_cqe)
3567                         last_buf = cq_buf;
3568                 cq_new_cqe = false;
3569         }
3570
3571         /* check the current CQ for new cqes */
3572         while (npolled < num_entries) {
3573                 ret = irdma_poll_one(ukcq, cur_cqe, entry + npolled);
3574                 if (ret == -ENOENT) {
3575                         ret = irdma_generated_cmpls(iwcq, cur_cqe);
3576                         if (!ret)
3577                                 irdma_process_cqe(entry + npolled, cur_cqe);
3578                 }
3579                 if (!ret) {
3580                         ++npolled;
3581                         cq_new_cqe = true;
3582                         continue;
3583                 }
3584
3585                 if (ret == -ENOENT)
3586                         break;
3587                 /* QP using the CQ is destroyed. Skip reporting this CQE */
3588                 if (ret == -EFAULT) {
3589                         cq_new_cqe = true;
3590                         continue;
3591                 }
3592                 goto error;
3593         }
3594
3595         if (cq_new_cqe)
3596                 /* all previous CQ resizes are complete */
3597                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, NULL);
3598         else if (last_buf)
3599                 /* only CQ resizes up to the last_buf are complete */
3600                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, last_buf);
3601         if (resized_bufs)
3602                 /* report to the HW the number of complete CQ resizes */
3603                 irdma_uk_cq_set_resized_cnt(ukcq, resized_bufs);
3604
3605         return npolled;
3606 error:
3607         ibdev_dbg(&iwdev->ibdev, "%s: Error polling CQ, irdma_err: %d\n",
3608                   __func__, ret);
3609
3610         return ret;
3611 }
3612
3613 /**
3614  * irdma_poll_cq - poll cq for completion (kernel apps)
3615  * @ibcq: cq to poll
3616  * @num_entries: number of entries to poll
3617  * @entry: wr of a completed entry
3618  */
3619 static int irdma_poll_cq(struct ib_cq *ibcq, int num_entries,
3620                          struct ib_wc *entry)
3621 {
3622         struct irdma_cq *iwcq;
3623         unsigned long flags;
3624         int ret;
3625
3626         iwcq = to_iwcq(ibcq);
3627
3628         spin_lock_irqsave(&iwcq->lock, flags);
3629         ret = __irdma_poll_cq(iwcq, num_entries, entry);
3630         spin_unlock_irqrestore(&iwcq->lock, flags);
3631
3632         return ret;
3633 }
3634
3635 /**
3636  * irdma_req_notify_cq - arm cq kernel application
3637  * @ibcq: cq to arm
3638  * @notify_flags: notofication flags
3639  */
3640 static int irdma_req_notify_cq(struct ib_cq *ibcq,
3641                                enum ib_cq_notify_flags notify_flags)
3642 {
3643         struct irdma_cq *iwcq;
3644         struct irdma_cq_uk *ukcq;
3645         unsigned long flags;
3646         enum irdma_cmpl_notify cq_notify;
3647         bool promo_event = false;
3648         int ret = 0;
3649
3650         cq_notify = notify_flags == IB_CQ_SOLICITED ?
3651                     IRDMA_CQ_COMPL_SOLICITED : IRDMA_CQ_COMPL_EVENT;
3652         iwcq = to_iwcq(ibcq);
3653         ukcq = &iwcq->sc_cq.cq_uk;
3654
3655         spin_lock_irqsave(&iwcq->lock, flags);
3656         /* Only promote to arm the CQ for any event if the last arm event was solicited. */
3657         if (iwcq->last_notify == IRDMA_CQ_COMPL_SOLICITED && notify_flags != IB_CQ_SOLICITED)
3658                 promo_event = true;
3659
3660         if (!atomic_cmpxchg(&iwcq->armed, 0, 1) || promo_event) {
3661                 iwcq->last_notify = cq_notify;
3662                 irdma_uk_cq_request_notification(ukcq, cq_notify);
3663         }
3664
3665         if ((notify_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
3666             (!irdma_cq_empty(iwcq) || !list_empty(&iwcq->cmpl_generated)))
3667                 ret = 1;
3668         spin_unlock_irqrestore(&iwcq->lock, flags);
3669
3670         return ret;
3671 }
3672
3673 static int irdma_roce_port_immutable(struct ib_device *ibdev, u32 port_num,
3674                                      struct ib_port_immutable *immutable)
3675 {
3676         struct ib_port_attr attr;
3677         int err;
3678
3679         immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
3680         err = ib_query_port(ibdev, port_num, &attr);
3681         if (err)
3682                 return err;
3683
3684         immutable->max_mad_size = IB_MGMT_MAD_SIZE;
3685         immutable->pkey_tbl_len = attr.pkey_tbl_len;
3686         immutable->gid_tbl_len = attr.gid_tbl_len;
3687
3688         return 0;
3689 }
3690
3691 static int irdma_iw_port_immutable(struct ib_device *ibdev, u32 port_num,
3692                                    struct ib_port_immutable *immutable)
3693 {
3694         struct ib_port_attr attr;
3695         int err;
3696
3697         immutable->core_cap_flags = RDMA_CORE_PORT_IWARP;
3698         err = ib_query_port(ibdev, port_num, &attr);
3699         if (err)
3700                 return err;
3701         immutable->gid_tbl_len = attr.gid_tbl_len;
3702
3703         return 0;
3704 }
3705
3706 static const struct rdma_stat_desc irdma_hw_stat_names[] = {
3707         /* gen1 - 32-bit */
3708         [IRDMA_HW_STAT_INDEX_IP4RXDISCARD].name         = "ip4InDiscards",
3709         [IRDMA_HW_STAT_INDEX_IP4RXTRUNC].name           = "ip4InTruncatedPkts",
3710         [IRDMA_HW_STAT_INDEX_IP4TXNOROUTE].name         = "ip4OutNoRoutes",
3711         [IRDMA_HW_STAT_INDEX_IP6RXDISCARD].name         = "ip6InDiscards",
3712         [IRDMA_HW_STAT_INDEX_IP6RXTRUNC].name           = "ip6InTruncatedPkts",
3713         [IRDMA_HW_STAT_INDEX_IP6TXNOROUTE].name         = "ip6OutNoRoutes",
3714         [IRDMA_HW_STAT_INDEX_TCPRTXSEG].name            = "tcpRetransSegs",
3715         [IRDMA_HW_STAT_INDEX_TCPRXOPTERR].name          = "tcpInOptErrors",
3716         [IRDMA_HW_STAT_INDEX_TCPRXPROTOERR].name        = "tcpInProtoErrors",
3717         [IRDMA_HW_STAT_INDEX_RXVLANERR].name            = "rxVlanErrors",
3718         /* gen1 - 64-bit */
3719         [IRDMA_HW_STAT_INDEX_IP4RXOCTS].name            = "ip4InOctets",
3720         [IRDMA_HW_STAT_INDEX_IP4RXPKTS].name            = "ip4InPkts",
3721         [IRDMA_HW_STAT_INDEX_IP4RXFRAGS].name           = "ip4InReasmRqd",
3722         [IRDMA_HW_STAT_INDEX_IP4RXMCPKTS].name          = "ip4InMcastPkts",
3723         [IRDMA_HW_STAT_INDEX_IP4TXOCTS].name            = "ip4OutOctets",
3724         [IRDMA_HW_STAT_INDEX_IP4TXPKTS].name            = "ip4OutPkts",
3725         [IRDMA_HW_STAT_INDEX_IP4TXFRAGS].name           = "ip4OutSegRqd",
3726         [IRDMA_HW_STAT_INDEX_IP4TXMCPKTS].name          = "ip4OutMcastPkts",
3727         [IRDMA_HW_STAT_INDEX_IP6RXOCTS].name            = "ip6InOctets",
3728         [IRDMA_HW_STAT_INDEX_IP6RXPKTS].name            = "ip6InPkts",
3729         [IRDMA_HW_STAT_INDEX_IP6RXFRAGS].name           = "ip6InReasmRqd",
3730         [IRDMA_HW_STAT_INDEX_IP6RXMCPKTS].name          = "ip6InMcastPkts",
3731         [IRDMA_HW_STAT_INDEX_IP6TXOCTS].name            = "ip6OutOctets",
3732         [IRDMA_HW_STAT_INDEX_IP6TXPKTS].name            = "ip6OutPkts",
3733         [IRDMA_HW_STAT_INDEX_IP6TXFRAGS].name           = "ip6OutSegRqd",
3734         [IRDMA_HW_STAT_INDEX_IP6TXMCPKTS].name          = "ip6OutMcastPkts",
3735         [IRDMA_HW_STAT_INDEX_TCPRXSEGS].name            = "tcpInSegs",
3736         [IRDMA_HW_STAT_INDEX_TCPTXSEG].name             = "tcpOutSegs",
3737         [IRDMA_HW_STAT_INDEX_RDMARXRDS].name            = "iwInRdmaReads",
3738         [IRDMA_HW_STAT_INDEX_RDMARXSNDS].name           = "iwInRdmaSends",
3739         [IRDMA_HW_STAT_INDEX_RDMARXWRS].name            = "iwInRdmaWrites",
3740         [IRDMA_HW_STAT_INDEX_RDMATXRDS].name            = "iwOutRdmaReads",
3741         [IRDMA_HW_STAT_INDEX_RDMATXSNDS].name           = "iwOutRdmaSends",
3742         [IRDMA_HW_STAT_INDEX_RDMATXWRS].name            = "iwOutRdmaWrites",
3743         [IRDMA_HW_STAT_INDEX_RDMAVBND].name             = "iwRdmaBnd",
3744         [IRDMA_HW_STAT_INDEX_RDMAVINV].name             = "iwRdmaInv",
3745
3746         /* gen2 - 32-bit */
3747         [IRDMA_HW_STAT_INDEX_RXRPCNPHANDLED].name       = "cnpHandled",
3748         [IRDMA_HW_STAT_INDEX_RXRPCNPIGNORED].name       = "cnpIgnored",
3749         [IRDMA_HW_STAT_INDEX_TXNPCNPSENT].name          = "cnpSent",
3750         /* gen2 - 64-bit */
3751         [IRDMA_HW_STAT_INDEX_IP4RXMCOCTS].name          = "ip4InMcastOctets",
3752         [IRDMA_HW_STAT_INDEX_IP4TXMCOCTS].name          = "ip4OutMcastOctets",
3753         [IRDMA_HW_STAT_INDEX_IP6RXMCOCTS].name          = "ip6InMcastOctets",
3754         [IRDMA_HW_STAT_INDEX_IP6TXMCOCTS].name          = "ip6OutMcastOctets",
3755         [IRDMA_HW_STAT_INDEX_UDPRXPKTS].name            = "RxUDP",
3756         [IRDMA_HW_STAT_INDEX_UDPTXPKTS].name            = "TxUDP",
3757         [IRDMA_HW_STAT_INDEX_RXNPECNMARKEDPKTS].name    = "RxECNMrkd",
3758
3759 };
3760
3761 static void irdma_get_dev_fw_str(struct ib_device *dev, char *str)
3762 {
3763         struct irdma_device *iwdev = to_iwdev(dev);
3764
3765         snprintf(str, IB_FW_VERSION_NAME_MAX, "%u.%u",
3766                  irdma_fw_major_ver(&iwdev->rf->sc_dev),
3767                  irdma_fw_minor_ver(&iwdev->rf->sc_dev));
3768 }
3769
3770 /**
3771  * irdma_alloc_hw_port_stats - Allocate a hw stats structure
3772  * @ibdev: device pointer from stack
3773  * @port_num: port number
3774  */
3775 static struct rdma_hw_stats *irdma_alloc_hw_port_stats(struct ib_device *ibdev,
3776                                                        u32 port_num)
3777 {
3778         struct irdma_device *iwdev = to_iwdev(ibdev);
3779         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
3780
3781         int num_counters = dev->hw_attrs.max_stat_idx;
3782         unsigned long lifespan = RDMA_HW_STATS_DEFAULT_LIFESPAN;
3783
3784         return rdma_alloc_hw_stats_struct(irdma_hw_stat_names, num_counters,
3785                                           lifespan);
3786 }
3787
3788 /**
3789  * irdma_get_hw_stats - Populates the rdma_hw_stats structure
3790  * @ibdev: device pointer from stack
3791  * @stats: stats pointer from stack
3792  * @port_num: port number
3793  * @index: which hw counter the stack is requesting we update
3794  */
3795 static int irdma_get_hw_stats(struct ib_device *ibdev,
3796                               struct rdma_hw_stats *stats, u32 port_num,
3797                               int index)
3798 {
3799         struct irdma_device *iwdev = to_iwdev(ibdev);
3800         struct irdma_dev_hw_stats *hw_stats = &iwdev->vsi.pestat->hw_stats;
3801
3802         if (iwdev->rf->rdma_ver >= IRDMA_GEN_2)
3803                 irdma_cqp_gather_stats_cmd(&iwdev->rf->sc_dev, iwdev->vsi.pestat, true);
3804         else
3805                 irdma_cqp_gather_stats_gen1(&iwdev->rf->sc_dev, iwdev->vsi.pestat);
3806
3807         memcpy(&stats->value[0], hw_stats, sizeof(u64) * stats->num_counters);
3808
3809         return stats->num_counters;
3810 }
3811
3812 /**
3813  * irdma_query_gid - Query port GID
3814  * @ibdev: device pointer from stack
3815  * @port: port number
3816  * @index: Entry index
3817  * @gid: Global ID
3818  */
3819 static int irdma_query_gid(struct ib_device *ibdev, u32 port, int index,
3820                            union ib_gid *gid)
3821 {
3822         struct irdma_device *iwdev = to_iwdev(ibdev);
3823
3824         memset(gid->raw, 0, sizeof(gid->raw));
3825         ether_addr_copy(gid->raw, iwdev->netdev->dev_addr);
3826
3827         return 0;
3828 }
3829
3830 /**
3831  * mcast_list_add -  Add a new mcast item to list
3832  * @rf: RDMA PCI function
3833  * @new_elem: pointer to element to add
3834  */
3835 static void mcast_list_add(struct irdma_pci_f *rf,
3836                            struct mc_table_list *new_elem)
3837 {
3838         list_add(&new_elem->list, &rf->mc_qht_list.list);
3839 }
3840
3841 /**
3842  * mcast_list_del - Remove an mcast item from list
3843  * @mc_qht_elem: pointer to mcast table list element
3844  */
3845 static void mcast_list_del(struct mc_table_list *mc_qht_elem)
3846 {
3847         if (mc_qht_elem)
3848                 list_del(&mc_qht_elem->list);
3849 }
3850
3851 /**
3852  * mcast_list_lookup_ip - Search mcast list for address
3853  * @rf: RDMA PCI function
3854  * @ip_mcast: pointer to mcast IP address
3855  */
3856 static struct mc_table_list *mcast_list_lookup_ip(struct irdma_pci_f *rf,
3857                                                   u32 *ip_mcast)
3858 {
3859         struct mc_table_list *mc_qht_el;
3860         struct list_head *pos, *q;
3861
3862         list_for_each_safe (pos, q, &rf->mc_qht_list.list) {
3863                 mc_qht_el = list_entry(pos, struct mc_table_list, list);
3864                 if (!memcmp(mc_qht_el->mc_info.dest_ip, ip_mcast,
3865                             sizeof(mc_qht_el->mc_info.dest_ip)))
3866                         return mc_qht_el;
3867         }
3868
3869         return NULL;
3870 }
3871
3872 /**
3873  * irdma_mcast_cqp_op - perform a mcast cqp operation
3874  * @iwdev: irdma device
3875  * @mc_grp_ctx: mcast group info
3876  * @op: operation
3877  *
3878  * returns error status
3879  */
3880 static int irdma_mcast_cqp_op(struct irdma_device *iwdev,
3881                               struct irdma_mcast_grp_info *mc_grp_ctx, u8 op)
3882 {
3883         struct cqp_cmds_info *cqp_info;
3884         struct irdma_cqp_request *cqp_request;
3885         int status;
3886
3887         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
3888         if (!cqp_request)
3889                 return -ENOMEM;
3890
3891         cqp_request->info.in.u.mc_create.info = *mc_grp_ctx;
3892         cqp_info = &cqp_request->info;
3893         cqp_info->cqp_cmd = op;
3894         cqp_info->post_sq = 1;
3895         cqp_info->in.u.mc_create.scratch = (uintptr_t)cqp_request;
3896         cqp_info->in.u.mc_create.cqp = &iwdev->rf->cqp.sc_cqp;
3897         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
3898         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
3899
3900         return status;
3901 }
3902
3903 /**
3904  * irdma_mcast_mac - Get the multicast MAC for an IP address
3905  * @ip_addr: IPv4 or IPv6 address
3906  * @mac: pointer to result MAC address
3907  * @ipv4: flag indicating IPv4 or IPv6
3908  *
3909  */
3910 void irdma_mcast_mac(u32 *ip_addr, u8 *mac, bool ipv4)
3911 {
3912         u8 *ip = (u8 *)ip_addr;
3913
3914         if (ipv4) {
3915                 unsigned char mac4[ETH_ALEN] = {0x01, 0x00, 0x5E, 0x00,
3916                                                 0x00, 0x00};
3917
3918                 mac4[3] = ip[2] & 0x7F;
3919                 mac4[4] = ip[1];
3920                 mac4[5] = ip[0];
3921                 ether_addr_copy(mac, mac4);
3922         } else {
3923                 unsigned char mac6[ETH_ALEN] = {0x33, 0x33, 0x00, 0x00,
3924                                                 0x00, 0x00};
3925
3926                 mac6[2] = ip[3];
3927                 mac6[3] = ip[2];
3928                 mac6[4] = ip[1];
3929                 mac6[5] = ip[0];
3930                 ether_addr_copy(mac, mac6);
3931         }
3932 }
3933
3934 /**
3935  * irdma_attach_mcast - attach a qp to a multicast group
3936  * @ibqp: ptr to qp
3937  * @ibgid: pointer to global ID
3938  * @lid: local ID
3939  *
3940  * returns error status
3941  */
3942 static int irdma_attach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
3943 {
3944         struct irdma_qp *iwqp = to_iwqp(ibqp);
3945         struct irdma_device *iwdev = iwqp->iwdev;
3946         struct irdma_pci_f *rf = iwdev->rf;
3947         struct mc_table_list *mc_qht_elem;
3948         struct irdma_mcast_grp_ctx_entry_info mcg_info = {};
3949         unsigned long flags;
3950         u32 ip_addr[4] = {};
3951         u32 mgn;
3952         u32 no_mgs;
3953         int ret = 0;
3954         bool ipv4;
3955         u16 vlan_id;
3956         union irdma_sockaddr sgid_addr;
3957         unsigned char dmac[ETH_ALEN];
3958
3959         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
3960
3961         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid)) {
3962                 irdma_copy_ip_ntohl(ip_addr,
3963                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
3964                 irdma_netdev_vlan_ipv6(ip_addr, &vlan_id, NULL);
3965                 ipv4 = false;
3966                 ibdev_dbg(&iwdev->ibdev,
3967                           "VERBS: qp_id=%d, IP6address=%pI6\n", ibqp->qp_num,
3968                           ip_addr);
3969                 irdma_mcast_mac(ip_addr, dmac, false);
3970         } else {
3971                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
3972                 ipv4 = true;
3973                 vlan_id = irdma_get_vlan_ipv4(ip_addr);
3974                 irdma_mcast_mac(ip_addr, dmac, true);
3975                 ibdev_dbg(&iwdev->ibdev,
3976                           "VERBS: qp_id=%d, IP4address=%pI4, MAC=%pM\n",
3977                           ibqp->qp_num, ip_addr, dmac);
3978         }
3979
3980         spin_lock_irqsave(&rf->qh_list_lock, flags);
3981         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
3982         if (!mc_qht_elem) {
3983                 struct irdma_dma_mem *dma_mem_mc;
3984
3985                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3986                 mc_qht_elem = kzalloc(sizeof(*mc_qht_elem), GFP_KERNEL);
3987                 if (!mc_qht_elem)
3988                         return -ENOMEM;
3989
3990                 mc_qht_elem->mc_info.ipv4_valid = ipv4;
3991                 memcpy(mc_qht_elem->mc_info.dest_ip, ip_addr,
3992                        sizeof(mc_qht_elem->mc_info.dest_ip));
3993                 ret = irdma_alloc_rsrc(rf, rf->allocated_mcgs, rf->max_mcg,
3994                                        &mgn, &rf->next_mcg);
3995                 if (ret) {
3996                         kfree(mc_qht_elem);
3997                         return -ENOMEM;
3998                 }
3999
4000                 mc_qht_elem->mc_info.mgn = mgn;
4001                 dma_mem_mc = &mc_qht_elem->mc_grp_ctx.dma_mem_mc;
4002                 dma_mem_mc->size = ALIGN(sizeof(u64) * IRDMA_MAX_MGS_PER_CTX,
4003                                          IRDMA_HW_PAGE_SIZE);
4004                 dma_mem_mc->va = dma_alloc_coherent(rf->hw.device,
4005                                                     dma_mem_mc->size,
4006                                                     &dma_mem_mc->pa,
4007                                                     GFP_KERNEL);
4008                 if (!dma_mem_mc->va) {
4009                         irdma_free_rsrc(rf, rf->allocated_mcgs, mgn);
4010                         kfree(mc_qht_elem);
4011                         return -ENOMEM;
4012                 }
4013
4014                 mc_qht_elem->mc_grp_ctx.mg_id = (u16)mgn;
4015                 memcpy(mc_qht_elem->mc_grp_ctx.dest_ip_addr, ip_addr,
4016                        sizeof(mc_qht_elem->mc_grp_ctx.dest_ip_addr));
4017                 mc_qht_elem->mc_grp_ctx.ipv4_valid = ipv4;
4018                 mc_qht_elem->mc_grp_ctx.vlan_id = vlan_id;
4019                 if (vlan_id < VLAN_N_VID)
4020                         mc_qht_elem->mc_grp_ctx.vlan_valid = true;
4021                 mc_qht_elem->mc_grp_ctx.hmc_fcn_id = iwdev->rf->sc_dev.hmc_fn_id;
4022                 mc_qht_elem->mc_grp_ctx.qs_handle =
4023                         iwqp->sc_qp.vsi->qos[iwqp->sc_qp.user_pri].qs_handle;
4024                 ether_addr_copy(mc_qht_elem->mc_grp_ctx.dest_mac_addr, dmac);
4025
4026                 spin_lock_irqsave(&rf->qh_list_lock, flags);
4027                 mcast_list_add(rf, mc_qht_elem);
4028         } else {
4029                 if (mc_qht_elem->mc_grp_ctx.no_of_mgs ==
4030                     IRDMA_MAX_MGS_PER_CTX) {
4031                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4032                         return -ENOMEM;
4033                 }
4034         }
4035
4036         mcg_info.qp_id = iwqp->ibqp.qp_num;
4037         no_mgs = mc_qht_elem->mc_grp_ctx.no_of_mgs;
4038         irdma_sc_add_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
4039         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4040
4041         /* Only if there is a change do we need to modify or create */
4042         if (!no_mgs) {
4043                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4044                                          IRDMA_OP_MC_CREATE);
4045         } else if (no_mgs != mc_qht_elem->mc_grp_ctx.no_of_mgs) {
4046                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4047                                          IRDMA_OP_MC_MODIFY);
4048         } else {
4049                 return 0;
4050         }
4051
4052         if (ret)
4053                 goto error;
4054
4055         return 0;
4056
4057 error:
4058         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
4059         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
4060                 mcast_list_del(mc_qht_elem);
4061                 dma_free_coherent(rf->hw.device,
4062                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.size,
4063                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.va,
4064                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.pa);
4065                 mc_qht_elem->mc_grp_ctx.dma_mem_mc.va = NULL;
4066                 irdma_free_rsrc(rf, rf->allocated_mcgs,
4067                                 mc_qht_elem->mc_grp_ctx.mg_id);
4068                 kfree(mc_qht_elem);
4069         }
4070
4071         return ret;
4072 }
4073
4074 /**
4075  * irdma_detach_mcast - detach a qp from a multicast group
4076  * @ibqp: ptr to qp
4077  * @ibgid: pointer to global ID
4078  * @lid: local ID
4079  *
4080  * returns error status
4081  */
4082 static int irdma_detach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
4083 {
4084         struct irdma_qp *iwqp = to_iwqp(ibqp);
4085         struct irdma_device *iwdev = iwqp->iwdev;
4086         struct irdma_pci_f *rf = iwdev->rf;
4087         u32 ip_addr[4] = {};
4088         struct mc_table_list *mc_qht_elem;
4089         struct irdma_mcast_grp_ctx_entry_info mcg_info = {};
4090         int ret;
4091         unsigned long flags;
4092         union irdma_sockaddr sgid_addr;
4093
4094         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
4095         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid))
4096                 irdma_copy_ip_ntohl(ip_addr,
4097                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4098         else
4099                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
4100
4101         spin_lock_irqsave(&rf->qh_list_lock, flags);
4102         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
4103         if (!mc_qht_elem) {
4104                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4105                 ibdev_dbg(&iwdev->ibdev,
4106                           "VERBS: address not found MCG\n");
4107                 return 0;
4108         }
4109
4110         mcg_info.qp_id = iwqp->ibqp.qp_num;
4111         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
4112         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
4113                 mcast_list_del(mc_qht_elem);
4114                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4115                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4116                                          IRDMA_OP_MC_DESTROY);
4117                 if (ret) {
4118                         ibdev_dbg(&iwdev->ibdev,
4119                                   "VERBS: failed MC_DESTROY MCG\n");
4120                         spin_lock_irqsave(&rf->qh_list_lock, flags);
4121                         mcast_list_add(rf, mc_qht_elem);
4122                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4123                         return -EAGAIN;
4124                 }
4125
4126                 dma_free_coherent(rf->hw.device,
4127                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.size,
4128                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.va,
4129                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.pa);
4130                 mc_qht_elem->mc_grp_ctx.dma_mem_mc.va = NULL;
4131                 irdma_free_rsrc(rf, rf->allocated_mcgs,
4132                                 mc_qht_elem->mc_grp_ctx.mg_id);
4133                 kfree(mc_qht_elem);
4134         } else {
4135                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4136                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4137                                          IRDMA_OP_MC_MODIFY);
4138                 if (ret) {
4139                         ibdev_dbg(&iwdev->ibdev,
4140                                   "VERBS: failed Modify MCG\n");
4141                         return ret;
4142                 }
4143         }
4144
4145         return 0;
4146 }
4147
4148 static int irdma_create_hw_ah(struct irdma_device *iwdev, struct irdma_ah *ah, bool sleep)
4149 {
4150         struct irdma_pci_f *rf = iwdev->rf;
4151         int err;
4152
4153         err = irdma_alloc_rsrc(rf, rf->allocated_ahs, rf->max_ah, &ah->sc_ah.ah_info.ah_idx,
4154                                &rf->next_ah);
4155         if (err)
4156                 return err;
4157
4158         err = irdma_ah_cqp_op(rf, &ah->sc_ah, IRDMA_OP_AH_CREATE, sleep,
4159                               irdma_gsi_ud_qp_ah_cb, &ah->sc_ah);
4160
4161         if (err) {
4162                 ibdev_dbg(&iwdev->ibdev, "VERBS: CQP-OP Create AH fail");
4163                 goto err_ah_create;
4164         }
4165
4166         if (!sleep) {
4167                 int cnt = CQP_COMPL_WAIT_TIME_MS * CQP_TIMEOUT_THRESHOLD;
4168
4169                 do {
4170                         irdma_cqp_ce_handler(rf, &rf->ccq.sc_cq);
4171                         mdelay(1);
4172                 } while (!ah->sc_ah.ah_info.ah_valid && --cnt);
4173
4174                 if (!cnt) {
4175                         ibdev_dbg(&iwdev->ibdev, "VERBS: CQP create AH timed out");
4176                         err = -ETIMEDOUT;
4177                         goto err_ah_create;
4178                 }
4179         }
4180         return 0;
4181
4182 err_ah_create:
4183         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_ahs, ah->sc_ah.ah_info.ah_idx);
4184
4185         return err;
4186 }
4187
4188 static int irdma_setup_ah(struct ib_ah *ibah, struct rdma_ah_init_attr *attr)
4189 {
4190         struct irdma_pd *pd = to_iwpd(ibah->pd);
4191         struct irdma_ah *ah = container_of(ibah, struct irdma_ah, ibah);
4192         struct rdma_ah_attr *ah_attr = attr->ah_attr;
4193         const struct ib_gid_attr *sgid_attr;
4194         struct irdma_device *iwdev = to_iwdev(ibah->pd->device);
4195         struct irdma_pci_f *rf = iwdev->rf;
4196         struct irdma_sc_ah *sc_ah;
4197         struct irdma_ah_info *ah_info;
4198         union irdma_sockaddr sgid_addr, dgid_addr;
4199         int err;
4200         u8 dmac[ETH_ALEN];
4201
4202         ah->pd = pd;
4203         sc_ah = &ah->sc_ah;
4204         sc_ah->ah_info.vsi = &iwdev->vsi;
4205         irdma_sc_init_ah(&rf->sc_dev, sc_ah);
4206         ah->sgid_index = ah_attr->grh.sgid_index;
4207         sgid_attr = ah_attr->grh.sgid_attr;
4208         memcpy(&ah->dgid, &ah_attr->grh.dgid, sizeof(ah->dgid));
4209         rdma_gid2ip((struct sockaddr *)&sgid_addr, &sgid_attr->gid);
4210         rdma_gid2ip((struct sockaddr *)&dgid_addr, &ah_attr->grh.dgid);
4211         ah->av.attrs = *ah_attr;
4212         ah->av.net_type = rdma_gid_attr_network_type(sgid_attr);
4213         ah_info = &sc_ah->ah_info;
4214         ah_info->pd_idx = pd->sc_pd.pd_id;
4215         if (ah_attr->ah_flags & IB_AH_GRH) {
4216                 ah_info->flow_label = ah_attr->grh.flow_label;
4217                 ah_info->hop_ttl = ah_attr->grh.hop_limit;
4218                 ah_info->tc_tos = ah_attr->grh.traffic_class;
4219         }
4220
4221         ether_addr_copy(dmac, ah_attr->roce.dmac);
4222         if (ah->av.net_type == RDMA_NETWORK_IPV4) {
4223                 ah_info->ipv4_valid = true;
4224                 ah_info->dest_ip_addr[0] =
4225                         ntohl(dgid_addr.saddr_in.sin_addr.s_addr);
4226                 ah_info->src_ip_addr[0] =
4227                         ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
4228                 ah_info->do_lpbk = irdma_ipv4_is_lpb(ah_info->src_ip_addr[0],
4229                                                      ah_info->dest_ip_addr[0]);
4230                 if (ipv4_is_multicast(dgid_addr.saddr_in.sin_addr.s_addr)) {
4231                         ah_info->do_lpbk = true;
4232                         irdma_mcast_mac(ah_info->dest_ip_addr, dmac, true);
4233                 }
4234         } else {
4235                 irdma_copy_ip_ntohl(ah_info->dest_ip_addr,
4236                                     dgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4237                 irdma_copy_ip_ntohl(ah_info->src_ip_addr,
4238                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4239                 ah_info->do_lpbk = irdma_ipv6_is_lpb(ah_info->src_ip_addr,
4240                                                      ah_info->dest_ip_addr);
4241                 if (rdma_is_multicast_addr(&dgid_addr.saddr_in6.sin6_addr)) {
4242                         ah_info->do_lpbk = true;
4243                         irdma_mcast_mac(ah_info->dest_ip_addr, dmac, false);
4244                 }
4245         }
4246
4247         err = rdma_read_gid_l2_fields(sgid_attr, &ah_info->vlan_tag,
4248                                       ah_info->mac_addr);
4249         if (err)
4250                 return err;
4251
4252         ah_info->dst_arpindex = irdma_add_arp(iwdev->rf, ah_info->dest_ip_addr,
4253                                               ah_info->ipv4_valid, dmac);
4254
4255         if (ah_info->dst_arpindex == -1)
4256                 return -EINVAL;
4257
4258         if (ah_info->vlan_tag >= VLAN_N_VID && iwdev->dcb_vlan_mode)
4259                 ah_info->vlan_tag = 0;
4260
4261         if (ah_info->vlan_tag < VLAN_N_VID) {
4262                 ah_info->insert_vlan_tag = true;
4263                 ah_info->vlan_tag |=
4264                         rt_tos2priority(ah_info->tc_tos) << VLAN_PRIO_SHIFT;
4265         }
4266
4267         return 0;
4268 }
4269
4270 /**
4271  * irdma_ah_exists - Check for existing identical AH
4272  * @iwdev: irdma device
4273  * @new_ah: AH to check for
4274  *
4275  * returns true if AH is found, false if not found.
4276  */
4277 static bool irdma_ah_exists(struct irdma_device *iwdev,
4278                             struct irdma_ah *new_ah)
4279 {
4280         struct irdma_ah *ah;
4281         u32 key = new_ah->sc_ah.ah_info.dest_ip_addr[0] ^
4282                   new_ah->sc_ah.ah_info.dest_ip_addr[1] ^
4283                   new_ah->sc_ah.ah_info.dest_ip_addr[2] ^
4284                   new_ah->sc_ah.ah_info.dest_ip_addr[3];
4285
4286         hash_for_each_possible(iwdev->ah_hash_tbl, ah, list, key) {
4287                 /* Set ah_valid and ah_id the same so memcmp can work */
4288                 new_ah->sc_ah.ah_info.ah_idx = ah->sc_ah.ah_info.ah_idx;
4289                 new_ah->sc_ah.ah_info.ah_valid = ah->sc_ah.ah_info.ah_valid;
4290                 if (!memcmp(&ah->sc_ah.ah_info, &new_ah->sc_ah.ah_info,
4291                             sizeof(ah->sc_ah.ah_info))) {
4292                         refcount_inc(&ah->refcnt);
4293                         new_ah->parent_ah = ah;
4294                         return true;
4295                 }
4296         }
4297
4298         return false;
4299 }
4300
4301 /**
4302  * irdma_destroy_ah - Destroy address handle
4303  * @ibah: pointer to address handle
4304  * @ah_flags: flags for sleepable
4305  */
4306 static int irdma_destroy_ah(struct ib_ah *ibah, u32 ah_flags)
4307 {
4308         struct irdma_device *iwdev = to_iwdev(ibah->device);
4309         struct irdma_ah *ah = to_iwah(ibah);
4310
4311         if ((ah_flags & RDMA_DESTROY_AH_SLEEPABLE) && ah->parent_ah) {
4312                 mutex_lock(&iwdev->ah_tbl_lock);
4313                 if (!refcount_dec_and_test(&ah->parent_ah->refcnt)) {
4314                         mutex_unlock(&iwdev->ah_tbl_lock);
4315                         return 0;
4316                 }
4317                 hash_del(&ah->parent_ah->list);
4318                 kfree(ah->parent_ah);
4319                 mutex_unlock(&iwdev->ah_tbl_lock);
4320         }
4321
4322         irdma_ah_cqp_op(iwdev->rf, &ah->sc_ah, IRDMA_OP_AH_DESTROY,
4323                         false, NULL, ah);
4324
4325         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_ahs,
4326                         ah->sc_ah.ah_info.ah_idx);
4327
4328         return 0;
4329 }
4330
4331 /**
4332  * irdma_create_user_ah - create user address handle
4333  * @ibah: address handle
4334  * @attr: address handle attributes
4335  * @udata: User data
4336  *
4337  * returns 0 on success, error otherwise
4338  */
4339 static int irdma_create_user_ah(struct ib_ah *ibah,
4340                                 struct rdma_ah_init_attr *attr,
4341                                 struct ib_udata *udata)
4342 {
4343 #define IRDMA_CREATE_AH_MIN_RESP_LEN offsetofend(struct irdma_create_ah_resp, rsvd)
4344         struct irdma_ah *ah = container_of(ibah, struct irdma_ah, ibah);
4345         struct irdma_device *iwdev = to_iwdev(ibah->pd->device);
4346         struct irdma_create_ah_resp uresp;
4347         struct irdma_ah *parent_ah;
4348         int err;
4349
4350         if (udata && udata->outlen < IRDMA_CREATE_AH_MIN_RESP_LEN)
4351                 return -EINVAL;
4352
4353         err = irdma_setup_ah(ibah, attr);
4354         if (err)
4355                 return err;
4356         mutex_lock(&iwdev->ah_tbl_lock);
4357         if (!irdma_ah_exists(iwdev, ah)) {
4358                 err = irdma_create_hw_ah(iwdev, ah, true);
4359                 if (err) {
4360                         mutex_unlock(&iwdev->ah_tbl_lock);
4361                         return err;
4362                 }
4363                 /* Add new AH to list */
4364                 parent_ah = kmemdup(ah, sizeof(*ah), GFP_KERNEL);
4365                 if (parent_ah) {
4366                         u32 key = parent_ah->sc_ah.ah_info.dest_ip_addr[0] ^
4367                                   parent_ah->sc_ah.ah_info.dest_ip_addr[1] ^
4368                                   parent_ah->sc_ah.ah_info.dest_ip_addr[2] ^
4369                                   parent_ah->sc_ah.ah_info.dest_ip_addr[3];
4370
4371                         ah->parent_ah = parent_ah;
4372                         hash_add(iwdev->ah_hash_tbl, &parent_ah->list, key);
4373                         refcount_set(&parent_ah->refcnt, 1);
4374                 }
4375         }
4376         mutex_unlock(&iwdev->ah_tbl_lock);
4377
4378         uresp.ah_id = ah->sc_ah.ah_info.ah_idx;
4379         err = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp), udata->outlen));
4380         if (err)
4381                 irdma_destroy_ah(ibah, attr->flags);
4382
4383         return err;
4384 }
4385
4386 /**
4387  * irdma_create_ah - create address handle
4388  * @ibah: address handle
4389  * @attr: address handle attributes
4390  * @udata: NULL
4391  *
4392  * returns 0 on success, error otherwise
4393  */
4394 static int irdma_create_ah(struct ib_ah *ibah, struct rdma_ah_init_attr *attr,
4395                            struct ib_udata *udata)
4396 {
4397         struct irdma_ah *ah = container_of(ibah, struct irdma_ah, ibah);
4398         struct irdma_device *iwdev = to_iwdev(ibah->pd->device);
4399         int err;
4400
4401         err = irdma_setup_ah(ibah, attr);
4402         if (err)
4403                 return err;
4404         err = irdma_create_hw_ah(iwdev, ah, attr->flags & RDMA_CREATE_AH_SLEEPABLE);
4405
4406         return err;
4407 }
4408
4409 /**
4410  * irdma_query_ah - Query address handle
4411  * @ibah: pointer to address handle
4412  * @ah_attr: address handle attributes
4413  */
4414 static int irdma_query_ah(struct ib_ah *ibah, struct rdma_ah_attr *ah_attr)
4415 {
4416         struct irdma_ah *ah = to_iwah(ibah);
4417
4418         memset(ah_attr, 0, sizeof(*ah_attr));
4419         if (ah->av.attrs.ah_flags & IB_AH_GRH) {
4420                 ah_attr->ah_flags = IB_AH_GRH;
4421                 ah_attr->grh.flow_label = ah->sc_ah.ah_info.flow_label;
4422                 ah_attr->grh.traffic_class = ah->sc_ah.ah_info.tc_tos;
4423                 ah_attr->grh.hop_limit = ah->sc_ah.ah_info.hop_ttl;
4424                 ah_attr->grh.sgid_index = ah->sgid_index;
4425                 ah_attr->grh.sgid_index = ah->sgid_index;
4426                 memcpy(&ah_attr->grh.dgid, &ah->dgid,
4427                        sizeof(ah_attr->grh.dgid));
4428         }
4429
4430         return 0;
4431 }
4432
4433 static enum rdma_link_layer irdma_get_link_layer(struct ib_device *ibdev,
4434                                                  u32 port_num)
4435 {
4436         return IB_LINK_LAYER_ETHERNET;
4437 }
4438
4439 static const struct ib_device_ops irdma_roce_dev_ops = {
4440         .attach_mcast = irdma_attach_mcast,
4441         .create_ah = irdma_create_ah,
4442         .create_user_ah = irdma_create_user_ah,
4443         .destroy_ah = irdma_destroy_ah,
4444         .detach_mcast = irdma_detach_mcast,
4445         .get_link_layer = irdma_get_link_layer,
4446         .get_port_immutable = irdma_roce_port_immutable,
4447         .modify_qp = irdma_modify_qp_roce,
4448         .query_ah = irdma_query_ah,
4449         .query_pkey = irdma_query_pkey,
4450 };
4451
4452 static const struct ib_device_ops irdma_iw_dev_ops = {
4453         .modify_qp = irdma_modify_qp,
4454         .get_port_immutable = irdma_iw_port_immutable,
4455         .query_gid = irdma_query_gid,
4456 };
4457
4458 static const struct ib_device_ops irdma_dev_ops = {
4459         .owner = THIS_MODULE,
4460         .driver_id = RDMA_DRIVER_IRDMA,
4461         .uverbs_abi_ver = IRDMA_ABI_VER,
4462
4463         .alloc_hw_port_stats = irdma_alloc_hw_port_stats,
4464         .alloc_mr = irdma_alloc_mr,
4465         .alloc_mw = irdma_alloc_mw,
4466         .alloc_pd = irdma_alloc_pd,
4467         .alloc_ucontext = irdma_alloc_ucontext,
4468         .create_cq = irdma_create_cq,
4469         .create_qp = irdma_create_qp,
4470         .dealloc_driver = irdma_ib_dealloc_device,
4471         .dealloc_mw = irdma_dealloc_mw,
4472         .dealloc_pd = irdma_dealloc_pd,
4473         .dealloc_ucontext = irdma_dealloc_ucontext,
4474         .dereg_mr = irdma_dereg_mr,
4475         .destroy_cq = irdma_destroy_cq,
4476         .destroy_qp = irdma_destroy_qp,
4477         .disassociate_ucontext = irdma_disassociate_ucontext,
4478         .get_dev_fw_str = irdma_get_dev_fw_str,
4479         .get_dma_mr = irdma_get_dma_mr,
4480         .get_hw_stats = irdma_get_hw_stats,
4481         .map_mr_sg = irdma_map_mr_sg,
4482         .mmap = irdma_mmap,
4483         .mmap_free = irdma_mmap_free,
4484         .poll_cq = irdma_poll_cq,
4485         .post_recv = irdma_post_recv,
4486         .post_send = irdma_post_send,
4487         .query_device = irdma_query_device,
4488         .query_port = irdma_query_port,
4489         .query_qp = irdma_query_qp,
4490         .reg_user_mr = irdma_reg_user_mr,
4491         .reg_user_mr_dmabuf = irdma_reg_user_mr_dmabuf,
4492         .req_notify_cq = irdma_req_notify_cq,
4493         .resize_cq = irdma_resize_cq,
4494         INIT_RDMA_OBJ_SIZE(ib_pd, irdma_pd, ibpd),
4495         INIT_RDMA_OBJ_SIZE(ib_ucontext, irdma_ucontext, ibucontext),
4496         INIT_RDMA_OBJ_SIZE(ib_ah, irdma_ah, ibah),
4497         INIT_RDMA_OBJ_SIZE(ib_cq, irdma_cq, ibcq),
4498         INIT_RDMA_OBJ_SIZE(ib_mw, irdma_mr, ibmw),
4499         INIT_RDMA_OBJ_SIZE(ib_qp, irdma_qp, ibqp),
4500 };
4501
4502 /**
4503  * irdma_init_roce_device - initialization of roce rdma device
4504  * @iwdev: irdma device
4505  */
4506 static void irdma_init_roce_device(struct irdma_device *iwdev)
4507 {
4508         iwdev->ibdev.node_type = RDMA_NODE_IB_CA;
4509         addrconf_addr_eui48((u8 *)&iwdev->ibdev.node_guid,
4510                             iwdev->netdev->dev_addr);
4511         ib_set_device_ops(&iwdev->ibdev, &irdma_roce_dev_ops);
4512 }
4513
4514 /**
4515  * irdma_init_iw_device - initialization of iwarp rdma device
4516  * @iwdev: irdma device
4517  */
4518 static int irdma_init_iw_device(struct irdma_device *iwdev)
4519 {
4520         struct net_device *netdev = iwdev->netdev;
4521
4522         iwdev->ibdev.node_type = RDMA_NODE_RNIC;
4523         addrconf_addr_eui48((u8 *)&iwdev->ibdev.node_guid,
4524                             netdev->dev_addr);
4525         iwdev->ibdev.ops.iw_add_ref = irdma_qp_add_ref;
4526         iwdev->ibdev.ops.iw_rem_ref = irdma_qp_rem_ref;
4527         iwdev->ibdev.ops.iw_get_qp = irdma_get_qp;
4528         iwdev->ibdev.ops.iw_connect = irdma_connect;
4529         iwdev->ibdev.ops.iw_accept = irdma_accept;
4530         iwdev->ibdev.ops.iw_reject = irdma_reject;
4531         iwdev->ibdev.ops.iw_create_listen = irdma_create_listen;
4532         iwdev->ibdev.ops.iw_destroy_listen = irdma_destroy_listen;
4533         memcpy(iwdev->ibdev.iw_ifname, netdev->name,
4534                sizeof(iwdev->ibdev.iw_ifname));
4535         ib_set_device_ops(&iwdev->ibdev, &irdma_iw_dev_ops);
4536
4537         return 0;
4538 }
4539
4540 /**
4541  * irdma_init_rdma_device - initialization of rdma device
4542  * @iwdev: irdma device
4543  */
4544 static int irdma_init_rdma_device(struct irdma_device *iwdev)
4545 {
4546         struct pci_dev *pcidev = iwdev->rf->pcidev;
4547         int ret;
4548
4549         if (iwdev->roce_mode) {
4550                 irdma_init_roce_device(iwdev);
4551         } else {
4552                 ret = irdma_init_iw_device(iwdev);
4553                 if (ret)
4554                         return ret;
4555         }
4556         iwdev->ibdev.phys_port_cnt = 1;
4557         iwdev->ibdev.num_comp_vectors = iwdev->rf->ceqs_count;
4558         iwdev->ibdev.dev.parent = &pcidev->dev;
4559         ib_set_device_ops(&iwdev->ibdev, &irdma_dev_ops);
4560
4561         return 0;
4562 }
4563
4564 /**
4565  * irdma_port_ibevent - indicate port event
4566  * @iwdev: irdma device
4567  */
4568 void irdma_port_ibevent(struct irdma_device *iwdev)
4569 {
4570         struct ib_event event;
4571
4572         event.device = &iwdev->ibdev;
4573         event.element.port_num = 1;
4574         event.event =
4575                 iwdev->iw_status ? IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
4576         ib_dispatch_event(&event);
4577 }
4578
4579 /**
4580  * irdma_ib_unregister_device - unregister rdma device from IB
4581  * core
4582  * @iwdev: irdma device
4583  */
4584 void irdma_ib_unregister_device(struct irdma_device *iwdev)
4585 {
4586         iwdev->iw_status = 0;
4587         irdma_port_ibevent(iwdev);
4588         ib_unregister_device(&iwdev->ibdev);
4589 }
4590
4591 /**
4592  * irdma_ib_register_device - register irdma device to IB core
4593  * @iwdev: irdma device
4594  */
4595 int irdma_ib_register_device(struct irdma_device *iwdev)
4596 {
4597         int ret;
4598
4599         ret = irdma_init_rdma_device(iwdev);
4600         if (ret)
4601                 return ret;
4602
4603         ret = ib_device_set_netdev(&iwdev->ibdev, iwdev->netdev, 1);
4604         if (ret)
4605                 goto error;
4606         dma_set_max_seg_size(iwdev->rf->hw.device, UINT_MAX);
4607         ret = ib_register_device(&iwdev->ibdev, "irdma%d", iwdev->rf->hw.device);
4608         if (ret)
4609                 goto error;
4610
4611         iwdev->iw_status = 1;
4612         irdma_port_ibevent(iwdev);
4613
4614         return 0;
4615
4616 error:
4617         if (ret)
4618                 ibdev_dbg(&iwdev->ibdev, "VERBS: Register RDMA device fail\n");
4619
4620         return ret;
4621 }
4622
4623 /**
4624  * irdma_ib_dealloc_device
4625  * @ibdev: ib device
4626  *
4627  * callback from ibdev dealloc_driver to deallocate resources
4628  * unber irdma device
4629  */
4630 void irdma_ib_dealloc_device(struct ib_device *ibdev)
4631 {
4632         struct irdma_device *iwdev = to_iwdev(ibdev);
4633
4634         irdma_rt_deinit_hw(iwdev);
4635         irdma_ctrl_deinit_hw(iwdev->rf);
4636         kfree(iwdev->rf);
4637 }