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_get_eth_speed_and_width - Get IB port speed and width from netdev speed
  68  * @link_speed: netdev phy link speed
  69  * @active_speed: IB port speed
  70  * @active_width: IB port width
  71  */
  72 static void irdma_get_eth_speed_and_width(u32 link_speed, u16 *active_speed,
  73                                           u8 *active_width)
  74 {
  75         if (link_speed <= SPEED_1000) {
  76                 *active_width = IB_WIDTH_1X;
  77                 *active_speed = IB_SPEED_SDR;
  78         } else if (link_speed <= SPEED_10000) {
  79                 *active_width = IB_WIDTH_1X;
  80                 *active_speed = IB_SPEED_FDR10;
  81         } else if (link_speed <= SPEED_20000) {
  82                 *active_width = IB_WIDTH_4X;
  83                 *active_speed = IB_SPEED_DDR;
  84         } else if (link_speed <= SPEED_25000) {
  85                 *active_width = IB_WIDTH_1X;
  86                 *active_speed = IB_SPEED_EDR;
  87         } else if (link_speed <= SPEED_40000) {
  88                 *active_width = IB_WIDTH_4X;
  89                 *active_speed = IB_SPEED_FDR10;
  90         } else {
  91                 *active_width = IB_WIDTH_4X;
  92                 *active_speed = IB_SPEED_EDR;
  93         }
  94 }
  95
  96 /**
  97  * irdma_query_port - get port attributes
  98  * @ibdev: device pointer from stack
  99  * @port: port number for query
 100  * @props: returning device attributes
 101  */
 102 static int irdma_query_port(struct ib_device *ibdev, u32 port,
 103                             struct ib_port_attr *props)
 104 {
 105         struct irdma_device *iwdev = to_iwdev(ibdev);
 106         struct net_device *netdev = iwdev->netdev;
 107
 108         /* no need to zero out pros here. done by caller */
 109
 110         props->max_mtu = IB_MTU_4096;
 111         props->active_mtu = ib_mtu_int_to_enum(netdev->mtu);
 112         props->lid = 1;
 113         props->lmc = 0;
 114         props->sm_lid = 0;
 115         props->sm_sl = 0;
 116         if (netif_carrier_ok(netdev) && netif_running(netdev)) {
 117                 props->state = IB_PORT_ACTIVE;
 118                 props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
 119         } else {
 120                 props->state = IB_PORT_DOWN;
 121                 props->phys_state = IB_PORT_PHYS_STATE_DISABLED;
 122         }
 123         irdma_get_eth_speed_and_width(SPEED_100000, &props->active_speed,
 124                                       &props->active_width);
 125
 126         if (rdma_protocol_roce(ibdev, 1)) {
 127                 props->gid_tbl_len = 32;
 128                 props->ip_gids = true;
 129                 props->pkey_tbl_len = IRDMA_PKEY_TBL_SZ;
 130         } else {
 131                 props->gid_tbl_len = 1;
 132         }
 133         props->qkey_viol_cntr = 0;
 134         props->port_cap_flags |= IB_PORT_CM_SUP | IB_PORT_REINIT_SUP;
 135         props->max_msg_sz = iwdev->rf->sc_dev.hw_attrs.max_hw_outbound_msg_size;
 136
 137         return 0;
 138 }
 139
 140 /**
 141  * irdma_disassociate_ucontext - Disassociate user context
 142  * @context: ib user context
 143  */
 144 static void irdma_disassociate_ucontext(struct ib_ucontext *context)
 145 {
 146 }
 147
 148 static int irdma_mmap_legacy(struct irdma_ucontext *ucontext,
 149                              struct vm_area_struct *vma)
 150 {
 151         u64 pfn;
 152
 153         if (vma->vm_pgoff || vma->vm_end - vma->vm_start != PAGE_SIZE)
 154                 return -EINVAL;
 155
 156         vma->vm_private_data = ucontext;
 157         pfn = ((uintptr_t)ucontext->iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET] +
 158                pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;
 159
 160         return rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn, PAGE_SIZE,
 161                                  pgprot_noncached(vma->vm_page_prot), NULL);
 162 }
 163
 164 static void irdma_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
 165 {
 166         struct irdma_user_mmap_entry *entry = to_irdma_mmap_entry(rdma_entry);
 167
 168         kfree(entry);
 169 }
 170
 171 static struct rdma_user_mmap_entry*
 172 irdma_user_mmap_entry_insert(struct irdma_ucontext *ucontext, u64 bar_offset,
 173                              enum irdma_mmap_flag mmap_flag, u64 *mmap_offset)
 174 {
 175         struct irdma_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 176         int ret;
 177
 178         if (!entry)
 179                 return NULL;
 180
 181         entry->bar_offset = bar_offset;
 182         entry->mmap_flag = mmap_flag;
 183
 184         ret = rdma_user_mmap_entry_insert(&ucontext->ibucontext,
 185                                           &entry->rdma_entry, PAGE_SIZE);
 186         if (ret) {
 187                 kfree(entry);
 188                 return NULL;
 189         }
 190         *mmap_offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
 191
 192         return &entry->rdma_entry;
 193 }
 194
 195 /**
 196  * irdma_mmap - user memory map
 197  * @context: context created during alloc
 198  * @vma: kernel info for user memory map
 199  */
 200 static int irdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
 201 {
 202         struct rdma_user_mmap_entry *rdma_entry;
 203         struct irdma_user_mmap_entry *entry;
 204         struct irdma_ucontext *ucontext;
 205         u64 pfn;
 206         int ret;
 207
 208         ucontext = to_ucontext(context);
 209
 210         /* Legacy support for libi40iw with hard-coded mmap key */
 211         if (ucontext->legacy_mode)
 212                 return irdma_mmap_legacy(ucontext, vma);
 213
 214         rdma_entry = rdma_user_mmap_entry_get(&ucontext->ibucontext, vma);
 215         if (!rdma_entry) {
 216                 ibdev_dbg(&ucontext->iwdev->ibdev,
 217                           "VERBS: pgoff[0x%lx] does not have valid entry\n",
 218                           vma->vm_pgoff);
 219                 return -EINVAL;
 220         }
 221
 222         entry = to_irdma_mmap_entry(rdma_entry);
 223         ibdev_dbg(&ucontext->iwdev->ibdev,
 224                   "VERBS: bar_offset [0x%llx] mmap_flag [%d]\n",
 225                   entry->bar_offset, entry->mmap_flag);
 226
 227         pfn = (entry->bar_offset +
 228               pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;
 229
 230         switch (entry->mmap_flag) {
 231         case IRDMA_MMAP_IO_NC:
 232                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
 233                                         pgprot_noncached(vma->vm_page_prot),
 234                                         rdma_entry);
 235                 break;
 236         case IRDMA_MMAP_IO_WC:
 237                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
 238                                         pgprot_writecombine(vma->vm_page_prot),
 239                                         rdma_entry);
 240                 break;
 241         default:
 242                 ret = -EINVAL;
 243         }
 244
 245         if (ret)
 246                 ibdev_dbg(&ucontext->iwdev->ibdev,
 247                           "VERBS: bar_offset [0x%llx] mmap_flag[%d] err[%d]\n",
 248                           entry->bar_offset, entry->mmap_flag, ret);
 249         rdma_user_mmap_entry_put(rdma_entry);
 250
 251         return ret;
 252 }
 253
 254 /**
 255  * irdma_alloc_push_page - allocate a push page for qp
 256  * @iwqp: qp pointer
 257  */
 258 static void irdma_alloc_push_page(struct irdma_qp *iwqp)
 259 {
 260         struct irdma_cqp_request *cqp_request;
 261         struct cqp_cmds_info *cqp_info;
 262         struct irdma_device *iwdev = iwqp->iwdev;
 263         struct irdma_sc_qp *qp = &iwqp->sc_qp;
 264         int status;
 265
 266         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
 267         if (!cqp_request)
 268                 return;
 269
 270         cqp_info = &cqp_request->info;
 271         cqp_info->cqp_cmd = IRDMA_OP_MANAGE_PUSH_PAGE;
 272         cqp_info->post_sq = 1;
 273         cqp_info->in.u.manage_push_page.info.push_idx = 0;
 274         cqp_info->in.u.manage_push_page.info.qs_handle =
 275                 qp->vsi->qos[qp->user_pri].qs_handle;
 276         cqp_info->in.u.manage_push_page.info.free_page = 0;
 277         cqp_info->in.u.manage_push_page.info.push_page_type = 0;
 278         cqp_info->in.u.manage_push_page.cqp = &iwdev->rf->cqp.sc_cqp;
 279         cqp_info->in.u.manage_push_page.scratch = (uintptr_t)cqp_request;
 280
 281         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
 282         if (!status && cqp_request->compl_info.op_ret_val <
 283             iwdev->rf->sc_dev.hw_attrs.max_hw_device_pages) {
 284                 qp->push_idx = cqp_request->compl_info.op_ret_val;
 285                 qp->push_offset = 0;
 286         }
 287
 288         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
 289 }
 290
 291 /**
 292  * irdma_alloc_ucontext - Allocate the user context data structure
 293  * @uctx: uverbs context pointer
 294  * @udata: user data
 295  *
 296  * This keeps track of all objects associated with a particular
 297  * user-mode client.
 298  */
 299 static int irdma_alloc_ucontext(struct ib_ucontext *uctx,
 300                                 struct ib_udata *udata)
 301 {
 302 #define IRDMA_ALLOC_UCTX_MIN_REQ_LEN offsetofend(struct irdma_alloc_ucontext_req, rsvd8)
 303 #define IRDMA_ALLOC_UCTX_MIN_RESP_LEN offsetofend(struct irdma_alloc_ucontext_resp, rsvd)
 304         struct ib_device *ibdev = uctx->device;
 305         struct irdma_device *iwdev = to_iwdev(ibdev);
 306         struct irdma_alloc_ucontext_req req = {};
 307         struct irdma_alloc_ucontext_resp uresp = {};
 308         struct irdma_ucontext *ucontext = to_ucontext(uctx);
 309         struct irdma_uk_attrs *uk_attrs;
 310
 311         if (udata->inlen < IRDMA_ALLOC_UCTX_MIN_REQ_LEN ||
 312             udata->outlen < IRDMA_ALLOC_UCTX_MIN_RESP_LEN)
 313                 return -EINVAL;
 314
 315         if (ib_copy_from_udata(&req, udata, min(sizeof(req), udata->inlen)))
 316                 return -EINVAL;
 317
 318         if (req.userspace_ver < 4 || req.userspace_ver > IRDMA_ABI_VER)
 319                 goto ver_error;
 320
 321         ucontext->iwdev = iwdev;
 322         ucontext->abi_ver = req.userspace_ver;
 323
 324         uk_attrs = &iwdev->rf->sc_dev.hw_attrs.uk_attrs;
 325         /* GEN_1 legacy support with libi40iw */
 326         if (udata->outlen == IRDMA_ALLOC_UCTX_MIN_RESP_LEN) {
 327                 if (uk_attrs->hw_rev != IRDMA_GEN_1)
 328                         return -EOPNOTSUPP;
 329
 330                 ucontext->legacy_mode = true;
 331                 uresp.max_qps = iwdev->rf->max_qp;
 332                 uresp.max_pds = iwdev->rf->sc_dev.hw_attrs.max_hw_pds;
 333                 uresp.wq_size = iwdev->rf->sc_dev.hw_attrs.max_qp_wr * 2;
 334                 uresp.kernel_ver = req.userspace_ver;
 335                 if (ib_copy_to_udata(udata, &uresp,
 336                                      min(sizeof(uresp), udata->outlen)))
 337                         return -EFAULT;
 338         } else {
 339                 u64 bar_off = (uintptr_t)iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET];
 340
 341                 ucontext->db_mmap_entry =
 342                         irdma_user_mmap_entry_insert(ucontext, bar_off,
 343                                                      IRDMA_MMAP_IO_NC,
 344                                                      &uresp.db_mmap_key);
 345                 if (!ucontext->db_mmap_entry)
 346                         return -ENOMEM;
 347
 348                 uresp.kernel_ver = IRDMA_ABI_VER;
 349                 uresp.feature_flags = uk_attrs->feature_flags;
 350                 uresp.max_hw_wq_frags = uk_attrs->max_hw_wq_frags;
 351                 uresp.max_hw_read_sges = uk_attrs->max_hw_read_sges;
 352                 uresp.max_hw_inline = uk_attrs->max_hw_inline;
 353                 uresp.max_hw_rq_quanta = uk_attrs->max_hw_rq_quanta;
 354                 uresp.max_hw_wq_quanta = uk_attrs->max_hw_wq_quanta;
 355                 uresp.max_hw_sq_chunk = uk_attrs->max_hw_sq_chunk;
 356                 uresp.max_hw_cq_size = uk_attrs->max_hw_cq_size;
 357                 uresp.min_hw_cq_size = uk_attrs->min_hw_cq_size;
 358                 uresp.hw_rev = uk_attrs->hw_rev;
 359                 if (ib_copy_to_udata(udata, &uresp,
 360                                      min(sizeof(uresp), udata->outlen))) {
 361                         rdma_user_mmap_entry_remove(ucontext->db_mmap_entry);
 362                         return -EFAULT;
 363                 }
 364         }
 365
 366         INIT_LIST_HEAD(&ucontext->cq_reg_mem_list);
 367         spin_lock_init(&ucontext->cq_reg_mem_list_lock);
 368         INIT_LIST_HEAD(&ucontext->qp_reg_mem_list);
 369         spin_lock_init(&ucontext->qp_reg_mem_list_lock);
 370
 371         return 0;
 372
 373 ver_error:
 374         ibdev_err(&iwdev->ibdev,
 375                   "Invalid userspace driver version detected. Detected version %d, should be %d\n",
 376                   req.userspace_ver, IRDMA_ABI_VER);
 377         return -EINVAL;
 378 }
 379
 380 /**
 381  * irdma_dealloc_ucontext - deallocate the user context data structure
 382  * @context: user context created during alloc
 383  */
 384 static void irdma_dealloc_ucontext(struct ib_ucontext *context)
 385 {
 386         struct irdma_ucontext *ucontext = to_ucontext(context);
 387
 388         rdma_user_mmap_entry_remove(ucontext->db_mmap_entry);
 389 }
 390
 391 /**
 392  * irdma_alloc_pd - allocate protection domain
 393  * @pd: PD pointer
 394  * @udata: user data
 395  */
 396 static int irdma_alloc_pd(struct ib_pd *pd, struct ib_udata *udata)
 397 {
 398 #define IRDMA_ALLOC_PD_MIN_RESP_LEN offsetofend(struct irdma_alloc_pd_resp, rsvd)
 399         struct irdma_pd *iwpd = to_iwpd(pd);
 400         struct irdma_device *iwdev = to_iwdev(pd->device);
 401         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 402         struct irdma_pci_f *rf = iwdev->rf;
 403         struct irdma_alloc_pd_resp uresp = {};
 404         struct irdma_sc_pd *sc_pd;
 405         u32 pd_id = 0;
 406         int err;
 407
 408         if (udata && udata->outlen < IRDMA_ALLOC_PD_MIN_RESP_LEN)
 409                 return -EINVAL;
 410
 411         err = irdma_alloc_rsrc(rf, rf->allocated_pds, rf->max_pd, &pd_id,
 412                                &rf->next_pd);
 413         if (err)
 414                 return err;
 415
 416         sc_pd = &iwpd->sc_pd;
 417         if (udata) {
 418                 struct irdma_ucontext *ucontext =
 419                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
 420                                                   ibucontext);
 421                 irdma_sc_pd_init(dev, sc_pd, pd_id, ucontext->abi_ver);
 422                 uresp.pd_id = pd_id;
 423                 if (ib_copy_to_udata(udata, &uresp,
 424                                      min(sizeof(uresp), udata->outlen))) {
 425                         err = -EFAULT;
 426                         goto error;
 427                 }
 428         } else {
 429                 irdma_sc_pd_init(dev, sc_pd, pd_id, IRDMA_ABI_VER);
 430         }
 431
 432         return 0;
 433 error:
 434         irdma_free_rsrc(rf, rf->allocated_pds, pd_id);
 435
 436         return err;
 437 }
 438
 439 /**
 440  * irdma_dealloc_pd - deallocate pd
 441  * @ibpd: ptr of pd to be deallocated
 442  * @udata: user data
 443  */
 444 static int irdma_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 445 {
 446         struct irdma_pd *iwpd = to_iwpd(ibpd);
 447         struct irdma_device *iwdev = to_iwdev(ibpd->device);
 448
 449         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_pds, iwpd->sc_pd.pd_id);
 450
 451         return 0;
 452 }
 453
 454 /**
 455  * irdma_get_pbl - Retrieve pbl from a list given a virtual
 456  * address
 457  * @va: user virtual address
 458  * @pbl_list: pbl list to search in (QP's or CQ's)
 459  */
 460 static struct irdma_pbl *irdma_get_pbl(unsigned long va,
 461                                        struct list_head *pbl_list)
 462 {
 463         struct irdma_pbl *iwpbl;
 464
 465         list_for_each_entry (iwpbl, pbl_list, list) {
 466                 if (iwpbl->user_base == va) {
 467                         list_del(&iwpbl->list);
 468                         iwpbl->on_list = false;
 469                         return iwpbl;
 470                 }
 471         }
 472
 473         return NULL;
 474 }
 475
 476 /**
 477  * irdma_clean_cqes - clean cq entries for qp
 478  * @iwqp: qp ptr (user or kernel)
 479  * @iwcq: cq ptr
 480  */
 481 static void irdma_clean_cqes(struct irdma_qp *iwqp, struct irdma_cq *iwcq)
 482 {
 483         struct irdma_cq_uk *ukcq = &iwcq->sc_cq.cq_uk;
 484         unsigned long flags;
 485
 486         spin_lock_irqsave(&iwcq->lock, flags);
 487         irdma_uk_clean_cq(&iwqp->sc_qp.qp_uk, ukcq);
 488         spin_unlock_irqrestore(&iwcq->lock, flags);
 489 }
 490
 491 static void irdma_remove_push_mmap_entries(struct irdma_qp *iwqp)
 492 {
 493         if (iwqp->push_db_mmap_entry) {
 494                 rdma_user_mmap_entry_remove(iwqp->push_db_mmap_entry);
 495                 iwqp->push_db_mmap_entry = NULL;
 496         }
 497         if (iwqp->push_wqe_mmap_entry) {
 498                 rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
 499                 iwqp->push_wqe_mmap_entry = NULL;
 500         }
 501 }
 502
 503 static int irdma_setup_push_mmap_entries(struct irdma_ucontext *ucontext,
 504                                          struct irdma_qp *iwqp,
 505                                          u64 *push_wqe_mmap_key,
 506                                          u64 *push_db_mmap_key)
 507 {
 508         struct irdma_device *iwdev = ucontext->iwdev;
 509         u64 rsvd, bar_off;
 510
 511         rsvd = IRDMA_PF_BAR_RSVD;
 512         bar_off = (uintptr_t)iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET];
 513         /* skip over db page */
 514         bar_off += IRDMA_HW_PAGE_SIZE;
 515         /* push wqe page */
 516         bar_off += rsvd + iwqp->sc_qp.push_idx * IRDMA_HW_PAGE_SIZE;
 517         iwqp->push_wqe_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
 518                                         bar_off, IRDMA_MMAP_IO_WC,
 519                                         push_wqe_mmap_key);
 520         if (!iwqp->push_wqe_mmap_entry)
 521                 return -ENOMEM;
 522
 523         /* push doorbell page */
 524         bar_off += IRDMA_HW_PAGE_SIZE;
 525         iwqp->push_db_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
 526                                         bar_off, IRDMA_MMAP_IO_NC,
 527                                         push_db_mmap_key);
 528         if (!iwqp->push_db_mmap_entry) {
 529                 rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
 530                 return -ENOMEM;
 531         }
 532
 533         return 0;
 534 }
 535
 536 /**
 537  * irdma_destroy_qp - destroy qp
 538  * @ibqp: qp's ib pointer also to get to device's qp address
 539  * @udata: user data
 540  */
 541 static int irdma_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
 542 {
 543         struct irdma_qp *iwqp = to_iwqp(ibqp);
 544         struct irdma_device *iwdev = iwqp->iwdev;
 545
 546         iwqp->sc_qp.qp_uk.destroy_pending = true;
 547
 548         if (iwqp->iwarp_state == IRDMA_QP_STATE_RTS)
 549                 irdma_modify_qp_to_err(&iwqp->sc_qp);
 550
 551         if (!iwqp->user_mode)
 552                 cancel_delayed_work_sync(&iwqp->dwork_flush);
 553
 554         irdma_qp_rem_ref(&iwqp->ibqp);
 555         wait_for_completion(&iwqp->free_qp);
 556         irdma_free_lsmm_rsrc(iwqp);
 557         irdma_cqp_qp_destroy_cmd(&iwdev->rf->sc_dev, &iwqp->sc_qp);
 558
 559         if (!iwqp->user_mode) {
 560                 if (iwqp->iwscq) {
 561                         irdma_clean_cqes(iwqp, iwqp->iwscq);
 562                         if (iwqp->iwrcq != iwqp->iwscq)
 563                                 irdma_clean_cqes(iwqp, iwqp->iwrcq);
 564                 }
 565         }
 566         irdma_remove_push_mmap_entries(iwqp);
 567         irdma_free_qp_rsrc(iwqp);
 568
 569         return 0;
 570 }
 571
 572 /**
 573  * irdma_setup_virt_qp - setup for allocation of virtual qp
 574  * @iwdev: irdma device
 575  * @iwqp: qp ptr
 576  * @init_info: initialize info to return
 577  */
 578 static void irdma_setup_virt_qp(struct irdma_device *iwdev,
 579                                struct irdma_qp *iwqp,
 580                                struct irdma_qp_init_info *init_info)
 581 {
 582         struct irdma_pbl *iwpbl = iwqp->iwpbl;
 583         struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;
 584
 585         iwqp->page = qpmr->sq_page;
 586         init_info->shadow_area_pa = qpmr->shadow;
 587         if (iwpbl->pbl_allocated) {
 588                 init_info->virtual_map = true;
 589                 init_info->sq_pa = qpmr->sq_pbl.idx;
 590                 init_info->rq_pa = qpmr->rq_pbl.idx;
 591         } else {
 592                 init_info->sq_pa = qpmr->sq_pbl.addr;
 593                 init_info->rq_pa = qpmr->rq_pbl.addr;
 594         }
 595 }
 596
 597 /**
 598  * irdma_setup_kmode_qp - setup initialization for kernel mode qp
 599  * @iwdev: iwarp device
 600  * @iwqp: qp ptr (user or kernel)
 601  * @info: initialize info to return
 602  * @init_attr: Initial QP create attributes
 603  */
 604 static int irdma_setup_kmode_qp(struct irdma_device *iwdev,
 605                                 struct irdma_qp *iwqp,
 606                                 struct irdma_qp_init_info *info,
 607                                 struct ib_qp_init_attr *init_attr)
 608 {
 609         struct irdma_dma_mem *mem = &iwqp->kqp.dma_mem;
 610         u32 sqdepth, rqdepth;
 611         u8 sqshift, rqshift;
 612         u32 size;
 613         int status;
 614         struct irdma_qp_uk_init_info *ukinfo = &info->qp_uk_init_info;
 615         struct irdma_uk_attrs *uk_attrs = &iwdev->rf->sc_dev.hw_attrs.uk_attrs;
 616
 617         irdma_get_wqe_shift(uk_attrs,
 618                 uk_attrs->hw_rev >= IRDMA_GEN_2 ? ukinfo->max_sq_frag_cnt + 1 :
 619                                                   ukinfo->max_sq_frag_cnt,
 620                 ukinfo->max_inline_data, &sqshift);
 621         status = irdma_get_sqdepth(uk_attrs, ukinfo->sq_size, sqshift,
 622                                    &sqdepth);
 623         if (status)
 624                 return status;
 625
 626         if (uk_attrs->hw_rev == IRDMA_GEN_1)
 627                 rqshift = IRDMA_MAX_RQ_WQE_SHIFT_GEN1;
 628         else
 629                 irdma_get_wqe_shift(uk_attrs, ukinfo->max_rq_frag_cnt, 0,
 630                                     &rqshift);
 631
 632         status = irdma_get_rqdepth(uk_attrs, ukinfo->rq_size, rqshift,
 633                                    &rqdepth);
 634         if (status)
 635                 return status;
 636
 637         iwqp->kqp.sq_wrid_mem =
 638                 kcalloc(sqdepth, sizeof(*iwqp->kqp.sq_wrid_mem), GFP_KERNEL);
 639         if (!iwqp->kqp.sq_wrid_mem)
 640                 return -ENOMEM;
 641
 642         iwqp->kqp.rq_wrid_mem =
 643                 kcalloc(rqdepth, sizeof(*iwqp->kqp.rq_wrid_mem), GFP_KERNEL);
 644         if (!iwqp->kqp.rq_wrid_mem) {
 645                 kfree(iwqp->kqp.sq_wrid_mem);
 646                 iwqp->kqp.sq_wrid_mem = NULL;
 647                 return -ENOMEM;
 648         }
 649
 650         ukinfo->sq_wrtrk_array = iwqp->kqp.sq_wrid_mem;
 651         ukinfo->rq_wrid_array = iwqp->kqp.rq_wrid_mem;
 652
 653         size = (sqdepth + rqdepth) * IRDMA_QP_WQE_MIN_SIZE;
 654         size += (IRDMA_SHADOW_AREA_SIZE << 3);
 655
 656         mem->size = ALIGN(size, 256);
 657         mem->va = dma_alloc_coherent(iwdev->rf->hw.device, mem->size,
 658                                      &mem->pa, GFP_KERNEL);
 659         if (!mem->va) {
 660                 kfree(iwqp->kqp.sq_wrid_mem);
 661                 iwqp->kqp.sq_wrid_mem = NULL;
 662                 kfree(iwqp->kqp.rq_wrid_mem);
 663                 iwqp->kqp.rq_wrid_mem = NULL;
 664                 return -ENOMEM;
 665         }
 666
 667         ukinfo->sq = mem->va;
 668         info->sq_pa = mem->pa;
 669         ukinfo->rq = &ukinfo->sq[sqdepth];
 670         info->rq_pa = info->sq_pa + (sqdepth * IRDMA_QP_WQE_MIN_SIZE);
 671         ukinfo->shadow_area = ukinfo->rq[rqdepth].elem;
 672         info->shadow_area_pa = info->rq_pa + (rqdepth * IRDMA_QP_WQE_MIN_SIZE);
 673         ukinfo->sq_size = sqdepth >> sqshift;
 674         ukinfo->rq_size = rqdepth >> rqshift;
 675         ukinfo->qp_id = iwqp->ibqp.qp_num;
 676
 677         init_attr->cap.max_send_wr = (sqdepth - IRDMA_SQ_RSVD) >> sqshift;
 678         init_attr->cap.max_recv_wr = (rqdepth - IRDMA_RQ_RSVD) >> rqshift;
 679
 680         return 0;
 681 }
 682
 683 static int irdma_cqp_create_qp_cmd(struct irdma_qp *iwqp)
 684 {
 685         struct irdma_pci_f *rf = iwqp->iwdev->rf;
 686         struct irdma_cqp_request *cqp_request;
 687         struct cqp_cmds_info *cqp_info;
 688         struct irdma_create_qp_info *qp_info;
 689         int status;
 690
 691         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
 692         if (!cqp_request)
 693                 return -ENOMEM;
 694
 695         cqp_info = &cqp_request->info;
 696         qp_info = &cqp_request->info.in.u.qp_create.info;
 697         memset(qp_info, 0, sizeof(*qp_info));
 698         qp_info->mac_valid = true;
 699         qp_info->cq_num_valid = true;
 700         qp_info->next_iwarp_state = IRDMA_QP_STATE_IDLE;
 701
 702         cqp_info->cqp_cmd = IRDMA_OP_QP_CREATE;
 703         cqp_info->post_sq = 1;
 704         cqp_info->in.u.qp_create.qp = &iwqp->sc_qp;
 705         cqp_info->in.u.qp_create.scratch = (uintptr_t)cqp_request;
 706         status = irdma_handle_cqp_op(rf, cqp_request);
 707         irdma_put_cqp_request(&rf->cqp, cqp_request);
 708
 709         return status;
 710 }
 711
 712 static void irdma_roce_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
 713                                                struct irdma_qp_host_ctx_info *ctx_info)
 714 {
 715         struct irdma_device *iwdev = iwqp->iwdev;
 716         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 717         struct irdma_roce_offload_info *roce_info;
 718         struct irdma_udp_offload_info *udp_info;
 719
 720         udp_info = &iwqp->udp_info;
 721         udp_info->snd_mss = ib_mtu_enum_to_int(ib_mtu_int_to_enum(iwdev->vsi.mtu));
 722         udp_info->cwnd = iwdev->roce_cwnd;
 723         udp_info->rexmit_thresh = 2;
 724         udp_info->rnr_nak_thresh = 2;
 725         udp_info->src_port = 0xc000;
 726         udp_info->dst_port = ROCE_V2_UDP_DPORT;
 727         roce_info = &iwqp->roce_info;
 728         ether_addr_copy(roce_info->mac_addr, iwdev->netdev->dev_addr);
 729
 730         roce_info->rd_en = true;
 731         roce_info->wr_rdresp_en = true;
 732         roce_info->bind_en = true;
 733         roce_info->dcqcn_en = false;
 734         roce_info->rtomin = 5;
 735
 736         roce_info->ack_credits = iwdev->roce_ackcreds;
 737         roce_info->ird_size = dev->hw_attrs.max_hw_ird;
 738         roce_info->ord_size = dev->hw_attrs.max_hw_ord;
 739
 740         if (!iwqp->user_mode) {
 741                 roce_info->priv_mode_en = true;
 742                 roce_info->fast_reg_en = true;
 743                 roce_info->udprivcq_en = true;
 744         }
 745         roce_info->roce_tver = 0;
 746
 747         ctx_info->roce_info = &iwqp->roce_info;
 748         ctx_info->udp_info = &iwqp->udp_info;
 749         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
 750 }
 751
 752 static void irdma_iw_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
 753                                              struct irdma_qp_host_ctx_info *ctx_info)
 754 {
 755         struct irdma_device *iwdev = iwqp->iwdev;
 756         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 757         struct irdma_iwarp_offload_info *iwarp_info;
 758
 759         iwarp_info = &iwqp->iwarp_info;
 760         ether_addr_copy(iwarp_info->mac_addr, iwdev->netdev->dev_addr);
 761         iwarp_info->rd_en = true;
 762         iwarp_info->wr_rdresp_en = true;
 763         iwarp_info->bind_en = true;
 764         iwarp_info->ecn_en = true;
 765         iwarp_info->rtomin = 5;
 766
 767         if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
 768                 iwarp_info->ib_rd_en = true;
 769         if (!iwqp->user_mode) {
 770                 iwarp_info->priv_mode_en = true;
 771                 iwarp_info->fast_reg_en = true;
 772         }
 773         iwarp_info->ddp_ver = 1;
 774         iwarp_info->rdmap_ver = 1;
 775
 776         ctx_info->iwarp_info = &iwqp->iwarp_info;
 777         ctx_info->iwarp_info_valid = true;
 778         irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
 779         ctx_info->iwarp_info_valid = false;
 780 }
 781
 782 static int irdma_validate_qp_attrs(struct ib_qp_init_attr *init_attr,
 783                                    struct irdma_device *iwdev)
 784 {
 785         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 786         struct irdma_uk_attrs *uk_attrs = &dev->hw_attrs.uk_attrs;
 787
 788         if (init_attr->create_flags)
 789                 return -EOPNOTSUPP;
 790
 791         if (init_attr->cap.max_inline_data > uk_attrs->max_hw_inline ||
 792             init_attr->cap.max_send_sge > uk_attrs->max_hw_wq_frags ||
 793             init_attr->cap.max_recv_sge > uk_attrs->max_hw_wq_frags)
 794                 return -EINVAL;
 795
 796         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 797                 if (init_attr->qp_type != IB_QPT_RC &&
 798                     init_attr->qp_type != IB_QPT_UD &&
 799                     init_attr->qp_type != IB_QPT_GSI)
 800                         return -EOPNOTSUPP;
 801         } else {
 802                 if (init_attr->qp_type != IB_QPT_RC)
 803                         return -EOPNOTSUPP;
 804         }
 805
 806         return 0;
 807 }
 808
 809 static void irdma_flush_worker(struct work_struct *work)
 810 {
 811         struct delayed_work *dwork = to_delayed_work(work);
 812         struct irdma_qp *iwqp = container_of(dwork, struct irdma_qp, dwork_flush);
 813
 814         irdma_generate_flush_completions(iwqp);
 815 }
 816
 817 /**
 818  * irdma_create_qp - create qp
 819  * @ibqp: ptr of qp
 820  * @init_attr: attributes for qp
 821  * @udata: user data for create qp
 822  */
 823 static int irdma_create_qp(struct ib_qp *ibqp,
 824                            struct ib_qp_init_attr *init_attr,
 825                            struct ib_udata *udata)
 826 {
 827 #define IRDMA_CREATE_QP_MIN_REQ_LEN offsetofend(struct irdma_create_qp_req, user_compl_ctx)
 828 #define IRDMA_CREATE_QP_MIN_RESP_LEN offsetofend(struct irdma_create_qp_resp, rsvd)
 829         struct ib_pd *ibpd = ibqp->pd;
 830         struct irdma_pd *iwpd = to_iwpd(ibpd);
 831         struct irdma_device *iwdev = to_iwdev(ibpd->device);
 832         struct irdma_pci_f *rf = iwdev->rf;
 833         struct irdma_qp *iwqp = to_iwqp(ibqp);
 834         struct irdma_create_qp_req req = {};
 835         struct irdma_create_qp_resp uresp = {};
 836         u32 qp_num = 0;
 837         int err_code;
 838         int sq_size;
 839         int rq_size;
 840         struct irdma_sc_qp *qp;
 841         struct irdma_sc_dev *dev = &rf->sc_dev;
 842         struct irdma_uk_attrs *uk_attrs = &dev->hw_attrs.uk_attrs;
 843         struct irdma_qp_init_info init_info = {};
 844         struct irdma_qp_host_ctx_info *ctx_info;
 845         unsigned long flags;
 846
 847         err_code = irdma_validate_qp_attrs(init_attr, iwdev);
 848         if (err_code)
 849                 return err_code;
 850
 851         if (udata && (udata->inlen < IRDMA_CREATE_QP_MIN_REQ_LEN ||
 852                       udata->outlen < IRDMA_CREATE_QP_MIN_RESP_LEN))
 853                 return -EINVAL;
 854
 855         sq_size = init_attr->cap.max_send_wr;
 856         rq_size = init_attr->cap.max_recv_wr;
 857
 858         init_info.vsi = &iwdev->vsi;
 859         init_info.qp_uk_init_info.uk_attrs = uk_attrs;
 860         init_info.qp_uk_init_info.sq_size = sq_size;
 861         init_info.qp_uk_init_info.rq_size = rq_size;
 862         init_info.qp_uk_init_info.max_sq_frag_cnt = init_attr->cap.max_send_sge;
 863         init_info.qp_uk_init_info.max_rq_frag_cnt = init_attr->cap.max_recv_sge;
 864         init_info.qp_uk_init_info.max_inline_data = init_attr->cap.max_inline_data;
 865
 866         qp = &iwqp->sc_qp;
 867         qp->qp_uk.back_qp = iwqp;
 868         qp->push_idx = IRDMA_INVALID_PUSH_PAGE_INDEX;
 869
 870         iwqp->iwdev = iwdev;
 871         iwqp->q2_ctx_mem.size = ALIGN(IRDMA_Q2_BUF_SIZE + IRDMA_QP_CTX_SIZE,
 872                                       256);
 873         iwqp->q2_ctx_mem.va = dma_alloc_coherent(dev->hw->device,
 874                                                  iwqp->q2_ctx_mem.size,
 875                                                  &iwqp->q2_ctx_mem.pa,
 876                                                  GFP_KERNEL);
 877         if (!iwqp->q2_ctx_mem.va)
 878                 return -ENOMEM;
 879
 880         init_info.q2 = iwqp->q2_ctx_mem.va;
 881         init_info.q2_pa = iwqp->q2_ctx_mem.pa;
 882         init_info.host_ctx = (__le64 *)(init_info.q2 + IRDMA_Q2_BUF_SIZE);
 883         init_info.host_ctx_pa = init_info.q2_pa + IRDMA_Q2_BUF_SIZE;
 884
 885         if (init_attr->qp_type == IB_QPT_GSI)
 886                 qp_num = 1;
 887         else
 888                 err_code = irdma_alloc_rsrc(rf, rf->allocated_qps, rf->max_qp,
 889                                             &qp_num, &rf->next_qp);
 890         if (err_code)
 891                 goto error;
 892
 893         iwqp->iwpd = iwpd;
 894         iwqp->ibqp.qp_num = qp_num;
 895         qp = &iwqp->sc_qp;
 896         iwqp->iwscq = to_iwcq(init_attr->send_cq);
 897         iwqp->iwrcq = to_iwcq(init_attr->recv_cq);
 898         iwqp->host_ctx.va = init_info.host_ctx;
 899         iwqp->host_ctx.pa = init_info.host_ctx_pa;
 900         iwqp->host_ctx.size = IRDMA_QP_CTX_SIZE;
 901
 902         init_info.pd = &iwpd->sc_pd;
 903         init_info.qp_uk_init_info.qp_id = iwqp->ibqp.qp_num;
 904         if (!rdma_protocol_roce(&iwdev->ibdev, 1))
 905                 init_info.qp_uk_init_info.first_sq_wq = 1;
 906         iwqp->ctx_info.qp_compl_ctx = (uintptr_t)qp;
 907         init_waitqueue_head(&iwqp->waitq);
 908         init_waitqueue_head(&iwqp->mod_qp_waitq);
 909
 910         if (udata) {
 911                 err_code = ib_copy_from_udata(&req, udata,
 912                                               min(sizeof(req), udata->inlen));
 913                 if (err_code) {
 914                         ibdev_dbg(&iwdev->ibdev,
 915                                   "VERBS: ib_copy_from_data fail\n");
 916                         goto error;
 917                 }
 918
 919                 iwqp->ctx_info.qp_compl_ctx = req.user_compl_ctx;
 920                 iwqp->user_mode = 1;
 921                 if (req.user_wqe_bufs) {
 922                         struct irdma_ucontext *ucontext =
 923                                 rdma_udata_to_drv_context(udata,
 924                                                           struct irdma_ucontext,
 925                                                           ibucontext);
 926
 927                         init_info.qp_uk_init_info.legacy_mode = ucontext->legacy_mode;
 928                         spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
 929                         iwqp->iwpbl = irdma_get_pbl((unsigned long)req.user_wqe_bufs,
 930                                                     &ucontext->qp_reg_mem_list);
 931                         spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
 932
 933                         if (!iwqp->iwpbl) {
 934                                 err_code = -ENODATA;
 935                                 ibdev_dbg(&iwdev->ibdev, "VERBS: no pbl info\n");
 936                                 goto error;
 937                         }
 938                 }
 939                 init_info.qp_uk_init_info.abi_ver = iwpd->sc_pd.abi_ver;
 940                 irdma_setup_virt_qp(iwdev, iwqp, &init_info);
 941         } else {
 942                 INIT_DELAYED_WORK(&iwqp->dwork_flush, irdma_flush_worker);
 943                 init_info.qp_uk_init_info.abi_ver = IRDMA_ABI_VER;
 944                 err_code = irdma_setup_kmode_qp(iwdev, iwqp, &init_info, init_attr);
 945         }
 946
 947         if (err_code) {
 948                 ibdev_dbg(&iwdev->ibdev, "VERBS: setup qp failed\n");
 949                 goto error;
 950         }
 951
 952         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 953                 if (init_attr->qp_type == IB_QPT_RC) {
 954                         init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_ROCE_RC;
 955                         init_info.qp_uk_init_info.qp_caps = IRDMA_SEND_WITH_IMM |
 956                                                             IRDMA_WRITE_WITH_IMM |
 957                                                             IRDMA_ROCE;
 958                 } else {
 959                         init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_ROCE_UD;
 960                         init_info.qp_uk_init_info.qp_caps = IRDMA_SEND_WITH_IMM |
 961                                                             IRDMA_ROCE;
 962                 }
 963         } else {
 964                 init_info.qp_uk_init_info.type = IRDMA_QP_TYPE_IWARP;
 965                 init_info.qp_uk_init_info.qp_caps = IRDMA_WRITE_WITH_IMM;
 966         }
 967
 968         if (dev->hw_attrs.uk_attrs.hw_rev > IRDMA_GEN_1)
 969                 init_info.qp_uk_init_info.qp_caps |= IRDMA_PUSH_MODE;
 970
 971         err_code = irdma_sc_qp_init(qp, &init_info);
 972         if (err_code) {
 973                 ibdev_dbg(&iwdev->ibdev, "VERBS: qp_init fail\n");
 974                 goto error;
 975         }
 976
 977         ctx_info = &iwqp->ctx_info;
 978         ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
 979         ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
 980
 981         if (rdma_protocol_roce(&iwdev->ibdev, 1))
 982                 irdma_roce_fill_and_set_qpctx_info(iwqp, ctx_info);
 983         else
 984                 irdma_iw_fill_and_set_qpctx_info(iwqp, ctx_info);
 985
 986         err_code = irdma_cqp_create_qp_cmd(iwqp);
 987         if (err_code)
 988                 goto error;
 989
 990         refcount_set(&iwqp->refcnt, 1);
 991         spin_lock_init(&iwqp->lock);
 992         spin_lock_init(&iwqp->sc_qp.pfpdu.lock);
 993         iwqp->sig_all = (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ? 1 : 0;
 994         rf->qp_table[qp_num] = iwqp;
 995         iwqp->max_send_wr = sq_size;
 996         iwqp->max_recv_wr = rq_size;
 997
 998         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 999                 if (dev->ws_add(&iwdev->vsi, 0)) {
1000                         irdma_cqp_qp_destroy_cmd(&rf->sc_dev, &iwqp->sc_qp);
1001                         err_code = -EINVAL;
1002                         goto error;
1003                 }
1004
1005                 irdma_qp_add_qos(&iwqp->sc_qp);
1006         }
1007
1008         if (udata) {
1009                 /* GEN_1 legacy support with libi40iw does not have expanded uresp struct */
1010                 if (udata->outlen < sizeof(uresp)) {
1011                         uresp.lsmm = 1;
1012                         uresp.push_idx = IRDMA_INVALID_PUSH_PAGE_INDEX_GEN_1;
1013                 } else {
1014                         if (rdma_protocol_iwarp(&iwdev->ibdev, 1))
1015                                 uresp.lsmm = 1;
1016                 }
1017                 uresp.actual_sq_size = sq_size;
1018                 uresp.actual_rq_size = rq_size;
1019                 uresp.qp_id = qp_num;
1020                 uresp.qp_caps = qp->qp_uk.qp_caps;
1021
1022                 err_code = ib_copy_to_udata(udata, &uresp,
1023                                             min(sizeof(uresp), udata->outlen));
1024                 if (err_code) {
1025                         ibdev_dbg(&iwdev->ibdev, "VERBS: copy_to_udata failed\n");
1026                         irdma_destroy_qp(&iwqp->ibqp, udata);
1027                         return err_code;
1028                 }
1029         }
1030
1031         init_completion(&iwqp->free_qp);
1032         return 0;
1033
1034 error:
1035         irdma_free_qp_rsrc(iwqp);
1036         return err_code;
1037 }
1038
1039 static int irdma_get_ib_acc_flags(struct irdma_qp *iwqp)
1040 {
1041         int acc_flags = 0;
1042
1043         if (rdma_protocol_roce(iwqp->ibqp.device, 1)) {
1044                 if (iwqp->roce_info.wr_rdresp_en) {
1045                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
1046                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
1047                 }
1048                 if (iwqp->roce_info.rd_en)
1049                         acc_flags |= IB_ACCESS_REMOTE_READ;
1050                 if (iwqp->roce_info.bind_en)
1051                         acc_flags |= IB_ACCESS_MW_BIND;
1052         } else {
1053                 if (iwqp->iwarp_info.wr_rdresp_en) {
1054                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
1055                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
1056                 }
1057                 if (iwqp->iwarp_info.rd_en)
1058                         acc_flags |= IB_ACCESS_REMOTE_READ;
1059                 if (iwqp->iwarp_info.bind_en)
1060                         acc_flags |= IB_ACCESS_MW_BIND;
1061         }
1062         return acc_flags;
1063 }
1064
1065 /**
1066  * irdma_query_qp - query qp attributes
1067  * @ibqp: qp pointer
1068  * @attr: attributes pointer
1069  * @attr_mask: Not used
1070  * @init_attr: qp attributes to return
1071  */
1072 static int irdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1073                           int attr_mask, struct ib_qp_init_attr *init_attr)
1074 {
1075         struct irdma_qp *iwqp = to_iwqp(ibqp);
1076         struct irdma_sc_qp *qp = &iwqp->sc_qp;
1077
1078         memset(attr, 0, sizeof(*attr));
1079         memset(init_attr, 0, sizeof(*init_attr));
1080
1081         attr->qp_state = iwqp->ibqp_state;
1082         attr->cur_qp_state = iwqp->ibqp_state;
1083         attr->cap.max_send_wr = iwqp->max_send_wr;
1084         attr->cap.max_recv_wr = iwqp->max_recv_wr;
1085         attr->cap.max_inline_data = qp->qp_uk.max_inline_data;
1086         attr->cap.max_send_sge = qp->qp_uk.max_sq_frag_cnt;
1087         attr->cap.max_recv_sge = qp->qp_uk.max_rq_frag_cnt;
1088         attr->qp_access_flags = irdma_get_ib_acc_flags(iwqp);
1089         attr->port_num = 1;
1090         if (rdma_protocol_roce(ibqp->device, 1)) {
1091                 attr->path_mtu = ib_mtu_int_to_enum(iwqp->udp_info.snd_mss);
1092                 attr->qkey = iwqp->roce_info.qkey;
1093                 attr->rq_psn = iwqp->udp_info.epsn;
1094                 attr->sq_psn = iwqp->udp_info.psn_nxt;
1095                 attr->dest_qp_num = iwqp->roce_info.dest_qp;
1096                 attr->pkey_index = iwqp->roce_info.p_key;
1097                 attr->retry_cnt = iwqp->udp_info.rexmit_thresh;
1098                 attr->rnr_retry = iwqp->udp_info.rnr_nak_thresh;
1099                 attr->max_rd_atomic = iwqp->roce_info.ord_size;
1100                 attr->max_dest_rd_atomic = iwqp->roce_info.ird_size;
1101         }
1102
1103         init_attr->event_handler = iwqp->ibqp.event_handler;
1104         init_attr->qp_context = iwqp->ibqp.qp_context;
1105         init_attr->send_cq = iwqp->ibqp.send_cq;
1106         init_attr->recv_cq = iwqp->ibqp.recv_cq;
1107         init_attr->cap = attr->cap;
1108
1109         return 0;
1110 }
1111
1112 /**
1113  * irdma_query_pkey - Query partition key
1114  * @ibdev: device pointer from stack
1115  * @port: port number
1116  * @index: index of pkey
1117  * @pkey: pointer to store the pkey
1118  */
1119 static int irdma_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
1120                             u16 *pkey)
1121 {
1122         if (index >= IRDMA_PKEY_TBL_SZ)
1123                 return -EINVAL;
1124
1125         *pkey = IRDMA_DEFAULT_PKEY;
1126         return 0;
1127 }
1128
1129 /**
1130  * irdma_modify_qp_roce - modify qp request
1131  * @ibqp: qp's pointer for modify
1132  * @attr: access attributes
1133  * @attr_mask: state mask
1134  * @udata: user data
1135  */
1136 int irdma_modify_qp_roce(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1137                          int attr_mask, struct ib_udata *udata)
1138 {
1139 #define IRDMA_MODIFY_QP_MIN_REQ_LEN offsetofend(struct irdma_modify_qp_req, rq_flush)
1140 #define IRDMA_MODIFY_QP_MIN_RESP_LEN offsetofend(struct irdma_modify_qp_resp, push_valid)
1141         struct irdma_pd *iwpd = to_iwpd(ibqp->pd);
1142         struct irdma_qp *iwqp = to_iwqp(ibqp);
1143         struct irdma_device *iwdev = iwqp->iwdev;
1144         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
1145         struct irdma_qp_host_ctx_info *ctx_info;
1146         struct irdma_roce_offload_info *roce_info;
1147         struct irdma_udp_offload_info *udp_info;
1148         struct irdma_modify_qp_info info = {};
1149         struct irdma_modify_qp_resp uresp = {};
1150         struct irdma_modify_qp_req ureq = {};
1151         unsigned long flags;
1152         u8 issue_modify_qp = 0;
1153         int ret = 0;
1154
1155         ctx_info = &iwqp->ctx_info;
1156         roce_info = &iwqp->roce_info;
1157         udp_info = &iwqp->udp_info;
1158
1159         if (udata) {
1160                 /* udata inlen/outlen can be 0 when supporting legacy libi40iw */
1161                 if ((udata->inlen && udata->inlen < IRDMA_MODIFY_QP_MIN_REQ_LEN) ||
1162                     (udata->outlen && udata->outlen < IRDMA_MODIFY_QP_MIN_RESP_LEN))
1163                         return -EINVAL;
1164         }
1165
1166         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1167                 return -EOPNOTSUPP;
1168
1169         if (attr_mask & IB_QP_DEST_QPN)
1170                 roce_info->dest_qp = attr->dest_qp_num;
1171
1172         if (attr_mask & IB_QP_PKEY_INDEX) {
1173                 ret = irdma_query_pkey(ibqp->device, 0, attr->pkey_index,
1174                                        &roce_info->p_key);
1175                 if (ret)
1176                         return ret;
1177         }
1178
1179         if (attr_mask & IB_QP_QKEY)
1180                 roce_info->qkey = attr->qkey;
1181
1182         if (attr_mask & IB_QP_PATH_MTU)
1183                 udp_info->snd_mss = ib_mtu_enum_to_int(attr->path_mtu);
1184
1185         if (attr_mask & IB_QP_SQ_PSN) {
1186                 udp_info->psn_nxt = attr->sq_psn;
1187                 udp_info->lsn =  0xffff;
1188                 udp_info->psn_una = attr->sq_psn;
1189                 udp_info->psn_max = attr->sq_psn;
1190         }
1191
1192         if (attr_mask & IB_QP_RQ_PSN)
1193                 udp_info->epsn = attr->rq_psn;
1194
1195         if (attr_mask & IB_QP_RNR_RETRY)
1196                 udp_info->rnr_nak_thresh = attr->rnr_retry;
1197
1198         if (attr_mask & IB_QP_RETRY_CNT)
1199                 udp_info->rexmit_thresh = attr->retry_cnt;
1200
1201         ctx_info->roce_info->pd_id = iwpd->sc_pd.pd_id;
1202
1203         if (attr_mask & IB_QP_AV) {
1204                 struct irdma_av *av = &iwqp->roce_ah.av;
1205                 const struct ib_gid_attr *sgid_attr;
1206                 u16 vlan_id = VLAN_N_VID;
1207                 u32 local_ip[4];
1208
1209                 memset(&iwqp->roce_ah, 0, sizeof(iwqp->roce_ah));
1210                 if (attr->ah_attr.ah_flags & IB_AH_GRH) {
1211                         udp_info->ttl = attr->ah_attr.grh.hop_limit;
1212                         udp_info->flow_label = attr->ah_attr.grh.flow_label;
1213                         udp_info->tos = attr->ah_attr.grh.traffic_class;
1214                         udp_info->src_port =
1215                                 rdma_get_udp_sport(udp_info->flow_label,
1216                                                    ibqp->qp_num,
1217                                                    roce_info->dest_qp);
1218                         irdma_qp_rem_qos(&iwqp->sc_qp);
1219                         dev->ws_remove(iwqp->sc_qp.vsi, ctx_info->user_pri);
1220                         ctx_info->user_pri = rt_tos2priority(udp_info->tos);
1221                         iwqp->sc_qp.user_pri = ctx_info->user_pri;
1222                         if (dev->ws_add(iwqp->sc_qp.vsi, ctx_info->user_pri))
1223                                 return -ENOMEM;
1224                         irdma_qp_add_qos(&iwqp->sc_qp);
1225                 }
1226                 sgid_attr = attr->ah_attr.grh.sgid_attr;
1227                 ret = rdma_read_gid_l2_fields(sgid_attr, &vlan_id,
1228                                               ctx_info->roce_info->mac_addr);
1229                 if (ret)
1230                         return ret;
1231
1232                 if (vlan_id >= VLAN_N_VID && iwdev->dcb_vlan_mode)
1233                         vlan_id = 0;
1234                 if (vlan_id < VLAN_N_VID) {
1235                         udp_info->insert_vlan_tag = true;
1236                         udp_info->vlan_tag = vlan_id |
1237                                 ctx_info->user_pri << VLAN_PRIO_SHIFT;
1238                 } else {
1239                         udp_info->insert_vlan_tag = false;
1240                 }
1241
1242                 av->attrs = attr->ah_attr;
1243                 rdma_gid2ip((struct sockaddr *)&av->sgid_addr, &sgid_attr->gid);
1244                 rdma_gid2ip((struct sockaddr *)&av->dgid_addr, &attr->ah_attr.grh.dgid);
1245                 if (av->net_type == RDMA_NETWORK_IPV6) {
1246                         __be32 *daddr =
1247                                 av->dgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32;
1248                         __be32 *saddr =
1249                                 av->sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32;
1250
1251                         irdma_copy_ip_ntohl(&udp_info->dest_ip_addr[0], daddr);
1252                         irdma_copy_ip_ntohl(&udp_info->local_ipaddr[0], saddr);
1253
1254                         udp_info->ipv4 = false;
1255                         irdma_copy_ip_ntohl(local_ip, daddr);
1256
1257                         udp_info->arp_idx = irdma_arp_table(iwdev->rf,
1258                                                             &local_ip[0],
1259                                                             false, NULL,
1260                                                             IRDMA_ARP_RESOLVE);
1261                 } else if (av->net_type == RDMA_NETWORK_IPV4) {
1262                         __be32 saddr = av->sgid_addr.saddr_in.sin_addr.s_addr;
1263                         __be32 daddr = av->dgid_addr.saddr_in.sin_addr.s_addr;
1264
1265                         local_ip[0] = ntohl(daddr);
1266
1267                         udp_info->ipv4 = true;
1268                         udp_info->dest_ip_addr[0] = 0;
1269                         udp_info->dest_ip_addr[1] = 0;
1270                         udp_info->dest_ip_addr[2] = 0;
1271                         udp_info->dest_ip_addr[3] = local_ip[0];
1272
1273                         udp_info->local_ipaddr[0] = 0;
1274                         udp_info->local_ipaddr[1] = 0;
1275                         udp_info->local_ipaddr[2] = 0;
1276                         udp_info->local_ipaddr[3] = ntohl(saddr);
1277                 }
1278                 udp_info->arp_idx =
1279                         irdma_add_arp(iwdev->rf, local_ip, udp_info->ipv4,
1280                                       attr->ah_attr.roce.dmac);
1281         }
1282
1283         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1284                 if (attr->max_rd_atomic > dev->hw_attrs.max_hw_ord) {
1285                         ibdev_err(&iwdev->ibdev,
1286                                   "rd_atomic = %d, above max_hw_ord=%d\n",
1287                                   attr->max_rd_atomic,
1288                                   dev->hw_attrs.max_hw_ord);
1289                         return -EINVAL;
1290                 }
1291                 if (attr->max_rd_atomic)
1292                         roce_info->ord_size = attr->max_rd_atomic;
1293                 info.ord_valid = true;
1294         }
1295
1296         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1297                 if (attr->max_dest_rd_atomic > dev->hw_attrs.max_hw_ird) {
1298                         ibdev_err(&iwdev->ibdev,
1299                                   "rd_atomic = %d, above max_hw_ird=%d\n",
1300                                    attr->max_rd_atomic,
1301                                    dev->hw_attrs.max_hw_ird);
1302                         return -EINVAL;
1303                 }
1304                 if (attr->max_dest_rd_atomic)
1305                         roce_info->ird_size = attr->max_dest_rd_atomic;
1306         }
1307
1308         if (attr_mask & IB_QP_ACCESS_FLAGS) {
1309                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
1310                         roce_info->wr_rdresp_en = true;
1311                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
1312                         roce_info->wr_rdresp_en = true;
1313                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
1314                         roce_info->rd_en = true;
1315         }
1316
1317         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
1318
1319         ibdev_dbg(&iwdev->ibdev,
1320                   "VERBS: caller: %pS qp_id=%d to_ibqpstate=%d ibqpstate=%d irdma_qpstate=%d attr_mask=0x%x\n",
1321                   __builtin_return_address(0), ibqp->qp_num, attr->qp_state,
1322                   iwqp->ibqp_state, iwqp->iwarp_state, attr_mask);
1323
1324         spin_lock_irqsave(&iwqp->lock, flags);
1325         if (attr_mask & IB_QP_STATE) {
1326                 if (!ib_modify_qp_is_ok(iwqp->ibqp_state, attr->qp_state,
1327                                         iwqp->ibqp.qp_type, attr_mask)) {
1328                         ibdev_warn(&iwdev->ibdev, "modify_qp invalid for qp_id=%d, old_state=0x%x, new_state=0x%x\n",
1329                                    iwqp->ibqp.qp_num, iwqp->ibqp_state,
1330                                    attr->qp_state);
1331                         ret = -EINVAL;
1332                         goto exit;
1333                 }
1334                 info.curr_iwarp_state = iwqp->iwarp_state;
1335
1336                 switch (attr->qp_state) {
1337                 case IB_QPS_INIT:
1338                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1339                                 ret = -EINVAL;
1340                                 goto exit;
1341                         }
1342
1343                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
1344                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
1345                                 issue_modify_qp = 1;
1346                         }
1347                         break;
1348                 case IB_QPS_RTR:
1349                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1350                                 ret = -EINVAL;
1351                                 goto exit;
1352                         }
1353                         info.arp_cache_idx_valid = true;
1354                         info.cq_num_valid = true;
1355                         info.next_iwarp_state = IRDMA_QP_STATE_RTR;
1356                         issue_modify_qp = 1;
1357                         break;
1358                 case IB_QPS_RTS:
1359                         if (iwqp->ibqp_state < IB_QPS_RTR ||
1360                             iwqp->ibqp_state == IB_QPS_ERR) {
1361                                 ret = -EINVAL;
1362                                 goto exit;
1363                         }
1364
1365                         info.arp_cache_idx_valid = true;
1366                         info.cq_num_valid = true;
1367                         info.ord_valid = true;
1368                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
1369                         issue_modify_qp = 1;
1370                         if (iwdev->push_mode && udata &&
1371                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1372                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1373                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1374                                 irdma_alloc_push_page(iwqp);
1375                                 spin_lock_irqsave(&iwqp->lock, flags);
1376                         }
1377                         break;
1378                 case IB_QPS_SQD:
1379                         if (iwqp->iwarp_state == IRDMA_QP_STATE_SQD)
1380                                 goto exit;
1381
1382                         if (iwqp->iwarp_state != IRDMA_QP_STATE_RTS) {
1383                                 ret = -EINVAL;
1384                                 goto exit;
1385                         }
1386
1387                         info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1388                         issue_modify_qp = 1;
1389                         break;
1390                 case IB_QPS_SQE:
1391                 case IB_QPS_ERR:
1392                 case IB_QPS_RESET:
1393                         if (iwqp->iwarp_state == IRDMA_QP_STATE_RTS) {
1394                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1395                                 info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1396                                 irdma_hw_modify_qp(iwdev, iwqp, &info, true);
1397                                 spin_lock_irqsave(&iwqp->lock, flags);
1398                         }
1399
1400                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1401                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1402                                 if (udata && udata->inlen) {
1403                                         if (ib_copy_from_udata(&ureq, udata,
1404                                             min(sizeof(ureq), udata->inlen)))
1405                                                 return -EINVAL;
1406
1407                                         irdma_flush_wqes(iwqp,
1408                                             (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1409                                             (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1410                                             IRDMA_REFLUSH);
1411                                 }
1412                                 return 0;
1413                         }
1414
1415                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1416                         issue_modify_qp = 1;
1417                         break;
1418                 default:
1419                         ret = -EINVAL;
1420                         goto exit;
1421                 }
1422
1423                 iwqp->ibqp_state = attr->qp_state;
1424         }
1425
1426         ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1427         ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1428         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1429         spin_unlock_irqrestore(&iwqp->lock, flags);
1430
1431         if (attr_mask & IB_QP_STATE) {
1432                 if (issue_modify_qp) {
1433                         ctx_info->rem_endpoint_idx = udp_info->arp_idx;
1434                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1435                                 return -EINVAL;
1436                         spin_lock_irqsave(&iwqp->lock, flags);
1437                         if (iwqp->iwarp_state == info.curr_iwarp_state) {
1438                                 iwqp->iwarp_state = info.next_iwarp_state;
1439                                 iwqp->ibqp_state = attr->qp_state;
1440                         }
1441                         if (iwqp->ibqp_state > IB_QPS_RTS &&
1442                             !iwqp->flush_issued) {
1443                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1444                                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_SQ |
1445                                                        IRDMA_FLUSH_RQ |
1446                                                        IRDMA_FLUSH_WAIT);
1447                                 iwqp->flush_issued = 1;
1448                         } else {
1449                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1450                         }
1451                 } else {
1452                         iwqp->ibqp_state = attr->qp_state;
1453                 }
1454                 if (udata && udata->outlen && dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1455                         struct irdma_ucontext *ucontext;
1456
1457                         ucontext = rdma_udata_to_drv_context(udata,
1458                                         struct irdma_ucontext, ibucontext);
1459                         if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1460                             !iwqp->push_wqe_mmap_entry &&
1461                             !irdma_setup_push_mmap_entries(ucontext, iwqp,
1462                                 &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1463                                 uresp.push_valid = 1;
1464                                 uresp.push_offset = iwqp->sc_qp.push_offset;
1465                         }
1466                         ret = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1467                                                udata->outlen));
1468                         if (ret) {
1469                                 irdma_remove_push_mmap_entries(iwqp);
1470                                 ibdev_dbg(&iwdev->ibdev,
1471                                           "VERBS: copy_to_udata failed\n");
1472                                 return ret;
1473                         }
1474                 }
1475         }
1476
1477         return 0;
1478 exit:
1479         spin_unlock_irqrestore(&iwqp->lock, flags);
1480
1481         return ret;
1482 }
1483
1484 /**
1485  * irdma_modify_qp - modify qp request
1486  * @ibqp: qp's pointer for modify
1487  * @attr: access attributes
1488  * @attr_mask: state mask
1489  * @udata: user data
1490  */
1491 int irdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
1492                     struct ib_udata *udata)
1493 {
1494 #define IRDMA_MODIFY_QP_MIN_REQ_LEN offsetofend(struct irdma_modify_qp_req, rq_flush)
1495 #define IRDMA_MODIFY_QP_MIN_RESP_LEN offsetofend(struct irdma_modify_qp_resp, push_valid)
1496         struct irdma_qp *iwqp = to_iwqp(ibqp);
1497         struct irdma_device *iwdev = iwqp->iwdev;
1498         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
1499         struct irdma_qp_host_ctx_info *ctx_info;
1500         struct irdma_tcp_offload_info *tcp_info;
1501         struct irdma_iwarp_offload_info *offload_info;
1502         struct irdma_modify_qp_info info = {};
1503         struct irdma_modify_qp_resp uresp = {};
1504         struct irdma_modify_qp_req ureq = {};
1505         u8 issue_modify_qp = 0;
1506         u8 dont_wait = 0;
1507         int err;
1508         unsigned long flags;
1509
1510         if (udata) {
1511                 /* udata inlen/outlen can be 0 when supporting legacy libi40iw */
1512                 if ((udata->inlen && udata->inlen < IRDMA_MODIFY_QP_MIN_REQ_LEN) ||
1513                     (udata->outlen && udata->outlen < IRDMA_MODIFY_QP_MIN_RESP_LEN))
1514                         return -EINVAL;
1515         }
1516
1517         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1518                 return -EOPNOTSUPP;
1519
1520         ctx_info = &iwqp->ctx_info;
1521         offload_info = &iwqp->iwarp_info;
1522         tcp_info = &iwqp->tcp_info;
1523         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
1524         ibdev_dbg(&iwdev->ibdev,
1525                   "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",
1526                   __builtin_return_address(0), ibqp->qp_num, attr->qp_state,
1527                   iwqp->ibqp_state, iwqp->iwarp_state, iwqp->last_aeq,
1528                   iwqp->hw_tcp_state, iwqp->hw_iwarp_state, attr_mask);
1529
1530         spin_lock_irqsave(&iwqp->lock, flags);
1531         if (attr_mask & IB_QP_STATE) {
1532                 info.curr_iwarp_state = iwqp->iwarp_state;
1533                 switch (attr->qp_state) {
1534                 case IB_QPS_INIT:
1535                 case IB_QPS_RTR:
1536                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1537                                 err = -EINVAL;
1538                                 goto exit;
1539                         }
1540
1541                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
1542                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
1543                                 issue_modify_qp = 1;
1544                         }
1545                         if (iwdev->push_mode && udata &&
1546                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1547                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1548                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1549                                 irdma_alloc_push_page(iwqp);
1550                                 spin_lock_irqsave(&iwqp->lock, flags);
1551                         }
1552                         break;
1553                 case IB_QPS_RTS:
1554                         if (iwqp->iwarp_state > IRDMA_QP_STATE_RTS ||
1555                             !iwqp->cm_id) {
1556                                 err = -EINVAL;
1557                                 goto exit;
1558                         }
1559
1560                         issue_modify_qp = 1;
1561                         iwqp->hw_tcp_state = IRDMA_TCP_STATE_ESTABLISHED;
1562                         iwqp->hte_added = 1;
1563                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
1564                         info.tcp_ctx_valid = true;
1565                         info.ord_valid = true;
1566                         info.arp_cache_idx_valid = true;
1567                         info.cq_num_valid = true;
1568                         break;
1569                 case IB_QPS_SQD:
1570                         if (iwqp->hw_iwarp_state > IRDMA_QP_STATE_RTS) {
1571                                 err = 0;
1572                                 goto exit;
1573                         }
1574
1575                         if (iwqp->iwarp_state == IRDMA_QP_STATE_CLOSING ||
1576                             iwqp->iwarp_state < IRDMA_QP_STATE_RTS) {
1577                                 err = 0;
1578                                 goto exit;
1579                         }
1580
1581                         if (iwqp->iwarp_state > IRDMA_QP_STATE_CLOSING) {
1582                                 err = -EINVAL;
1583                                 goto exit;
1584                         }
1585
1586                         info.next_iwarp_state = IRDMA_QP_STATE_CLOSING;
1587                         issue_modify_qp = 1;
1588                         break;
1589                 case IB_QPS_SQE:
1590                         if (iwqp->iwarp_state >= IRDMA_QP_STATE_TERMINATE) {
1591                                 err = -EINVAL;
1592                                 goto exit;
1593                         }
1594
1595                         info.next_iwarp_state = IRDMA_QP_STATE_TERMINATE;
1596                         issue_modify_qp = 1;
1597                         break;
1598                 case IB_QPS_ERR:
1599                 case IB_QPS_RESET:
1600                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1601                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1602                                 if (udata && udata->inlen) {
1603                                         if (ib_copy_from_udata(&ureq, udata,
1604                                             min(sizeof(ureq), udata->inlen)))
1605                                                 return -EINVAL;
1606
1607                                         irdma_flush_wqes(iwqp,
1608                                             (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1609                                             (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1610                                             IRDMA_REFLUSH);
1611                                 }
1612                                 return 0;
1613                         }
1614
1615                         if (iwqp->sc_qp.term_flags) {
1616                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1617                                 irdma_terminate_del_timer(&iwqp->sc_qp);
1618                                 spin_lock_irqsave(&iwqp->lock, flags);
1619                         }
1620                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1621                         if (iwqp->hw_tcp_state > IRDMA_TCP_STATE_CLOSED &&
1622                             iwdev->iw_status &&
1623                             iwqp->hw_tcp_state != IRDMA_TCP_STATE_TIME_WAIT)
1624                                 info.reset_tcp_conn = true;
1625                         else
1626                                 dont_wait = 1;
1627
1628                         issue_modify_qp = 1;
1629                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1630                         break;
1631                 default:
1632                         err = -EINVAL;
1633                         goto exit;
1634                 }
1635
1636                 iwqp->ibqp_state = attr->qp_state;
1637         }
1638         if (attr_mask & IB_QP_ACCESS_FLAGS) {
1639                 ctx_info->iwarp_info_valid = true;
1640                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
1641                         offload_info->wr_rdresp_en = true;
1642                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
1643                         offload_info->wr_rdresp_en = true;
1644                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
1645                         offload_info->rd_en = true;
1646         }
1647
1648         if (ctx_info->iwarp_info_valid) {
1649                 ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1650                 ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1651                 irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1652         }
1653         spin_unlock_irqrestore(&iwqp->lock, flags);
1654
1655         if (attr_mask & IB_QP_STATE) {
1656                 if (issue_modify_qp) {
1657                         ctx_info->rem_endpoint_idx = tcp_info->arp_idx;
1658                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1659                                 return -EINVAL;
1660                 }
1661
1662                 spin_lock_irqsave(&iwqp->lock, flags);
1663                 if (iwqp->iwarp_state == info.curr_iwarp_state) {
1664                         iwqp->iwarp_state = info.next_iwarp_state;
1665                         iwqp->ibqp_state = attr->qp_state;
1666                 }
1667                 spin_unlock_irqrestore(&iwqp->lock, flags);
1668         }
1669
1670         if (issue_modify_qp && iwqp->ibqp_state > IB_QPS_RTS) {
1671                 if (dont_wait) {
1672                         if (iwqp->hw_tcp_state) {
1673                                 spin_lock_irqsave(&iwqp->lock, flags);
1674                                 iwqp->hw_tcp_state = IRDMA_TCP_STATE_CLOSED;
1675                                 iwqp->last_aeq = IRDMA_AE_RESET_SENT;
1676                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1677                         }
1678                         irdma_cm_disconn(iwqp);
1679                 } else {
1680                         int close_timer_started;
1681
1682                         spin_lock_irqsave(&iwdev->cm_core.ht_lock, flags);
1683
1684                         if (iwqp->cm_node) {
1685                                 refcount_inc(&iwqp->cm_node->refcnt);
1686                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1687                                 close_timer_started = atomic_inc_return(&iwqp->close_timer_started);
1688                                 if (iwqp->cm_id && close_timer_started == 1)
1689                                         irdma_schedule_cm_timer(iwqp->cm_node,
1690                                                 (struct irdma_puda_buf *)iwqp,
1691                                                 IRDMA_TIMER_TYPE_CLOSE, 1, 0);
1692
1693                                 irdma_rem_ref_cm_node(iwqp->cm_node);
1694                         } else {
1695                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1696                         }
1697                 }
1698         }
1699         if (attr_mask & IB_QP_STATE && udata && udata->outlen &&
1700             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1701                 struct irdma_ucontext *ucontext;
1702
1703                 ucontext = rdma_udata_to_drv_context(udata,
1704                                         struct irdma_ucontext, ibucontext);
1705                 if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1706                     !iwqp->push_wqe_mmap_entry &&
1707                     !irdma_setup_push_mmap_entries(ucontext, iwqp,
1708                         &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1709                         uresp.push_valid = 1;
1710                         uresp.push_offset = iwqp->sc_qp.push_offset;
1711                 }
1712
1713                 err = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1714                                        udata->outlen));
1715                 if (err) {
1716                         irdma_remove_push_mmap_entries(iwqp);
1717                         ibdev_dbg(&iwdev->ibdev,
1718                                   "VERBS: copy_to_udata failed\n");
1719                         return err;
1720                 }
1721         }
1722
1723         return 0;
1724 exit:
1725         spin_unlock_irqrestore(&iwqp->lock, flags);
1726
1727         return err;
1728 }
1729
1730 /**
1731  * irdma_cq_free_rsrc - free up resources for cq
1732  * @rf: RDMA PCI function
1733  * @iwcq: cq ptr
1734  */
1735 static void irdma_cq_free_rsrc(struct irdma_pci_f *rf, struct irdma_cq *iwcq)
1736 {
1737         struct irdma_sc_cq *cq = &iwcq->sc_cq;
1738
1739         if (!iwcq->user_mode) {
1740                 dma_free_coherent(rf->sc_dev.hw->device, iwcq->kmem.size,
1741                                   iwcq->kmem.va, iwcq->kmem.pa);
1742                 iwcq->kmem.va = NULL;
1743                 dma_free_coherent(rf->sc_dev.hw->device,
1744                                   iwcq->kmem_shadow.size,
1745                                   iwcq->kmem_shadow.va, iwcq->kmem_shadow.pa);
1746                 iwcq->kmem_shadow.va = NULL;
1747         }
1748
1749         irdma_free_rsrc(rf, rf->allocated_cqs, cq->cq_uk.cq_id);
1750 }
1751
1752 /**
1753  * irdma_free_cqbuf - worker to free a cq buffer
1754  * @work: provides access to the cq buffer to free
1755  */
1756 static void irdma_free_cqbuf(struct work_struct *work)
1757 {
1758         struct irdma_cq_buf *cq_buf = container_of(work, struct irdma_cq_buf, work);
1759
1760         dma_free_coherent(cq_buf->hw->device, cq_buf->kmem_buf.size,
1761                           cq_buf->kmem_buf.va, cq_buf->kmem_buf.pa);
1762         cq_buf->kmem_buf.va = NULL;
1763         kfree(cq_buf);
1764 }
1765
1766 /**
1767  * irdma_process_resize_list - remove resized cq buffers from the resize_list
1768  * @iwcq: cq which owns the resize_list
1769  * @iwdev: irdma device
1770  * @lcqe_buf: the buffer where the last cqe is received
1771  */
1772 static int irdma_process_resize_list(struct irdma_cq *iwcq,
1773                                      struct irdma_device *iwdev,
1774                                      struct irdma_cq_buf *lcqe_buf)
1775 {
1776         struct list_head *tmp_node, *list_node;
1777         struct irdma_cq_buf *cq_buf;
1778         int cnt = 0;
1779
1780         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
1781                 cq_buf = list_entry(list_node, struct irdma_cq_buf, list);
1782                 if (cq_buf == lcqe_buf)
1783                         return cnt;
1784
1785                 list_del(&cq_buf->list);
1786                 queue_work(iwdev->cleanup_wq, &cq_buf->work);
1787                 cnt++;
1788         }
1789
1790         return cnt;
1791 }
1792
1793 /**
1794  * irdma_destroy_cq - destroy cq
1795  * @ib_cq: cq pointer
1796  * @udata: user data
1797  */
1798 static int irdma_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata)
1799 {
1800         struct irdma_device *iwdev = to_iwdev(ib_cq->device);
1801         struct irdma_cq *iwcq = to_iwcq(ib_cq);
1802         struct irdma_sc_cq *cq = &iwcq->sc_cq;
1803         struct irdma_sc_dev *dev = cq->dev;
1804         struct irdma_sc_ceq *ceq = dev->ceq[cq->ceq_id];
1805         struct irdma_ceq *iwceq = container_of(ceq, struct irdma_ceq, sc_ceq);
1806         unsigned long flags;
1807
1808         spin_lock_irqsave(&iwcq->lock, flags);
1809         if (!list_empty(&iwcq->cmpl_generated))
1810                 irdma_remove_cmpls_list(iwcq);
1811         if (!list_empty(&iwcq->resize_list))
1812                 irdma_process_resize_list(iwcq, iwdev, NULL);
1813         spin_unlock_irqrestore(&iwcq->lock, flags);
1814
1815         irdma_cq_wq_destroy(iwdev->rf, cq);
1816
1817         spin_lock_irqsave(&iwceq->ce_lock, flags);
1818         irdma_sc_cleanup_ceqes(cq, ceq);
1819         spin_unlock_irqrestore(&iwceq->ce_lock, flags);
1820         irdma_cq_free_rsrc(iwdev->rf, iwcq);
1821
1822         return 0;
1823 }
1824
1825 /**
1826  * irdma_resize_cq - resize cq
1827  * @ibcq: cq to be resized
1828  * @entries: desired cq size
1829  * @udata: user data
1830  */
1831 static int irdma_resize_cq(struct ib_cq *ibcq, int entries,
1832                            struct ib_udata *udata)
1833 {
1834 #define IRDMA_RESIZE_CQ_MIN_REQ_LEN offsetofend(struct irdma_resize_cq_req, user_cq_buffer)
1835         struct irdma_cq *iwcq = to_iwcq(ibcq);
1836         struct irdma_sc_dev *dev = iwcq->sc_cq.dev;
1837         struct irdma_cqp_request *cqp_request;
1838         struct cqp_cmds_info *cqp_info;
1839         struct irdma_modify_cq_info *m_info;
1840         struct irdma_modify_cq_info info = {};
1841         struct irdma_dma_mem kmem_buf;
1842         struct irdma_cq_mr *cqmr_buf;
1843         struct irdma_pbl *iwpbl_buf;
1844         struct irdma_device *iwdev;
1845         struct irdma_pci_f *rf;
1846         struct irdma_cq_buf *cq_buf = NULL;
1847         unsigned long flags;
1848         int ret;
1849
1850         iwdev = to_iwdev(ibcq->device);
1851         rf = iwdev->rf;
1852
1853         if (!(rf->sc_dev.hw_attrs.uk_attrs.feature_flags &
1854             IRDMA_FEATURE_CQ_RESIZE))
1855                 return -EOPNOTSUPP;
1856
1857         if (udata && udata->inlen < IRDMA_RESIZE_CQ_MIN_REQ_LEN)
1858                 return -EINVAL;
1859
1860         if (entries > rf->max_cqe)
1861                 return -EINVAL;
1862
1863         if (!iwcq->user_mode) {
1864                 entries++;
1865                 if (rf->sc_dev.hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
1866                         entries *= 2;
1867         }
1868
1869         info.cq_size = max(entries, 4);
1870
1871         if (info.cq_size == iwcq->sc_cq.cq_uk.cq_size - 1)
1872                 return 0;
1873
1874         if (udata) {
1875                 struct irdma_resize_cq_req req = {};
1876                 struct irdma_ucontext *ucontext =
1877                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
1878                                                   ibucontext);
1879
1880                 /* CQ resize not supported with legacy GEN_1 libi40iw */
1881                 if (ucontext->legacy_mode)
1882                         return -EOPNOTSUPP;
1883
1884                 if (ib_copy_from_udata(&req, udata,
1885                                        min(sizeof(req), udata->inlen)))
1886                         return -EINVAL;
1887
1888                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
1889                 iwpbl_buf = irdma_get_pbl((unsigned long)req.user_cq_buffer,
1890                                           &ucontext->cq_reg_mem_list);
1891                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
1892
1893                 if (!iwpbl_buf)
1894                         return -ENOMEM;
1895
1896                 cqmr_buf = &iwpbl_buf->cq_mr;
1897                 if (iwpbl_buf->pbl_allocated) {
1898                         info.virtual_map = true;
1899                         info.pbl_chunk_size = 1;
1900                         info.first_pm_pbl_idx = cqmr_buf->cq_pbl.idx;
1901                 } else {
1902                         info.cq_pa = cqmr_buf->cq_pbl.addr;
1903                 }
1904         } else {
1905                 /* Kmode CQ resize */
1906                 int rsize;
1907
1908                 rsize = info.cq_size * sizeof(struct irdma_cqe);
1909                 kmem_buf.size = ALIGN(round_up(rsize, 256), 256);
1910                 kmem_buf.va = dma_alloc_coherent(dev->hw->device,
1911                                                  kmem_buf.size, &kmem_buf.pa,
1912                                                  GFP_KERNEL);
1913                 if (!kmem_buf.va)
1914                         return -ENOMEM;
1915
1916                 info.cq_base = kmem_buf.va;
1917                 info.cq_pa = kmem_buf.pa;
1918                 cq_buf = kzalloc(sizeof(*cq_buf), GFP_KERNEL);
1919                 if (!cq_buf) {
1920                         ret = -ENOMEM;
1921                         goto error;
1922                 }
1923         }
1924
1925         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
1926         if (!cqp_request) {
1927                 ret = -ENOMEM;
1928                 goto error;
1929         }
1930
1931         info.shadow_read_threshold = iwcq->sc_cq.shadow_read_threshold;
1932         info.cq_resize = true;
1933
1934         cqp_info = &cqp_request->info;
1935         m_info = &cqp_info->in.u.cq_modify.info;
1936         memcpy(m_info, &info, sizeof(*m_info));
1937
1938         cqp_info->cqp_cmd = IRDMA_OP_CQ_MODIFY;
1939         cqp_info->in.u.cq_modify.cq = &iwcq->sc_cq;
1940         cqp_info->in.u.cq_modify.scratch = (uintptr_t)cqp_request;
1941         cqp_info->post_sq = 1;
1942         ret = irdma_handle_cqp_op(rf, cqp_request);
1943         irdma_put_cqp_request(&rf->cqp, cqp_request);
1944         if (ret)
1945                 goto error;
1946
1947         spin_lock_irqsave(&iwcq->lock, flags);
1948         if (cq_buf) {
1949                 cq_buf->kmem_buf = iwcq->kmem;
1950                 cq_buf->hw = dev->hw;
1951                 memcpy(&cq_buf->cq_uk, &iwcq->sc_cq.cq_uk, sizeof(cq_buf->cq_uk));
1952                 INIT_WORK(&cq_buf->work, irdma_free_cqbuf);
1953                 list_add_tail(&cq_buf->list, &iwcq->resize_list);
1954                 iwcq->kmem = kmem_buf;
1955         }
1956
1957         irdma_sc_cq_resize(&iwcq->sc_cq, &info);
1958         ibcq->cqe = info.cq_size - 1;
1959         spin_unlock_irqrestore(&iwcq->lock, flags);
1960
1961         return 0;
1962 error:
1963         if (!udata) {
1964                 dma_free_coherent(dev->hw->device, kmem_buf.size, kmem_buf.va,
1965                                   kmem_buf.pa);
1966                 kmem_buf.va = NULL;
1967         }
1968         kfree(cq_buf);
1969
1970         return ret;
1971 }
1972
1973 static inline int cq_validate_flags(u32 flags, u8 hw_rev)
1974 {
1975         /* GEN1 does not support CQ create flags */
1976         if (hw_rev == IRDMA_GEN_1)
1977                 return flags ? -EOPNOTSUPP : 0;
1978
1979         return flags & ~IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION ? -EOPNOTSUPP : 0;
1980 }
1981
1982 /**
1983  * irdma_create_cq - create cq
1984  * @ibcq: CQ allocated
1985  * @attr: attributes for cq
1986  * @udata: user data
1987  */
1988 static int irdma_create_cq(struct ib_cq *ibcq,
1989                            const struct ib_cq_init_attr *attr,
1990                            struct ib_udata *udata)
1991 {
1992 #define IRDMA_CREATE_CQ_MIN_REQ_LEN offsetofend(struct irdma_create_cq_req, user_cq_buf)
1993 #define IRDMA_CREATE_CQ_MIN_RESP_LEN offsetofend(struct irdma_create_cq_resp, cq_size)
1994         struct ib_device *ibdev = ibcq->device;
1995         struct irdma_device *iwdev = to_iwdev(ibdev);
1996         struct irdma_pci_f *rf = iwdev->rf;
1997         struct irdma_cq *iwcq = to_iwcq(ibcq);
1998         u32 cq_num = 0;
1999         struct irdma_sc_cq *cq;
2000         struct irdma_sc_dev *dev = &rf->sc_dev;
2001         struct irdma_cq_init_info info = {};
2002         struct irdma_cqp_request *cqp_request;
2003         struct cqp_cmds_info *cqp_info;
2004         struct irdma_cq_uk_init_info *ukinfo = &info.cq_uk_init_info;
2005         unsigned long flags;
2006         int err_code;
2007         int entries = attr->cqe;
2008
2009         err_code = cq_validate_flags(attr->flags, dev->hw_attrs.uk_attrs.hw_rev);
2010         if (err_code)
2011                 return err_code;
2012
2013         if (udata && (udata->inlen < IRDMA_CREATE_CQ_MIN_REQ_LEN ||
2014                       udata->outlen < IRDMA_CREATE_CQ_MIN_RESP_LEN))
2015                 return -EINVAL;
2016
2017         err_code = irdma_alloc_rsrc(rf, rf->allocated_cqs, rf->max_cq, &cq_num,
2018                                     &rf->next_cq);
2019         if (err_code)
2020                 return err_code;
2021
2022         cq = &iwcq->sc_cq;
2023         cq->back_cq = iwcq;
2024         spin_lock_init(&iwcq->lock);
2025         INIT_LIST_HEAD(&iwcq->resize_list);
2026         INIT_LIST_HEAD(&iwcq->cmpl_generated);
2027         info.dev = dev;
2028         ukinfo->cq_size = max(entries, 4);
2029         ukinfo->cq_id = cq_num;
2030         iwcq->ibcq.cqe = info.cq_uk_init_info.cq_size;
2031         if (attr->comp_vector < rf->ceqs_count)
2032                 info.ceq_id = attr->comp_vector;
2033         info.ceq_id_valid = true;
2034         info.ceqe_mask = 1;
2035         info.type = IRDMA_CQ_TYPE_IWARP;
2036         info.vsi = &iwdev->vsi;
2037
2038         if (udata) {
2039                 struct irdma_ucontext *ucontext;
2040                 struct irdma_create_cq_req req = {};
2041                 struct irdma_cq_mr *cqmr;
2042                 struct irdma_pbl *iwpbl;
2043                 struct irdma_pbl *iwpbl_shadow;
2044                 struct irdma_cq_mr *cqmr_shadow;
2045
2046                 iwcq->user_mode = true;
2047                 ucontext =
2048                         rdma_udata_to_drv_context(udata, struct irdma_ucontext,
2049                                                   ibucontext);
2050                 if (ib_copy_from_udata(&req, udata,
2051                                        min(sizeof(req), udata->inlen))) {
2052                         err_code = -EFAULT;
2053                         goto cq_free_rsrc;
2054                 }
2055
2056                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2057                 iwpbl = irdma_get_pbl((unsigned long)req.user_cq_buf,
2058                                       &ucontext->cq_reg_mem_list);
2059                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2060                 if (!iwpbl) {
2061                         err_code = -EPROTO;
2062                         goto cq_free_rsrc;
2063                 }
2064
2065                 iwcq->iwpbl = iwpbl;
2066                 iwcq->cq_mem_size = 0;
2067                 cqmr = &iwpbl->cq_mr;
2068
2069                 if (rf->sc_dev.hw_attrs.uk_attrs.feature_flags &
2070                     IRDMA_FEATURE_CQ_RESIZE && !ucontext->legacy_mode) {
2071                         spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2072                         iwpbl_shadow = irdma_get_pbl(
2073                                         (unsigned long)req.user_shadow_area,
2074                                         &ucontext->cq_reg_mem_list);
2075                         spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2076
2077                         if (!iwpbl_shadow) {
2078                                 err_code = -EPROTO;
2079                                 goto cq_free_rsrc;
2080                         }
2081                         iwcq->iwpbl_shadow = iwpbl_shadow;
2082                         cqmr_shadow = &iwpbl_shadow->cq_mr;
2083                         info.shadow_area_pa = cqmr_shadow->cq_pbl.addr;
2084                         cqmr->split = true;
2085                 } else {
2086                         info.shadow_area_pa = cqmr->shadow;
2087                 }
2088                 if (iwpbl->pbl_allocated) {
2089                         info.virtual_map = true;
2090                         info.pbl_chunk_size = 1;
2091                         info.first_pm_pbl_idx = cqmr->cq_pbl.idx;
2092                 } else {
2093                         info.cq_base_pa = cqmr->cq_pbl.addr;
2094                 }
2095         } else {
2096                 /* Kmode allocations */
2097                 int rsize;
2098
2099                 if (entries < 1 || entries > rf->max_cqe) {
2100                         err_code = -EINVAL;
2101                         goto cq_free_rsrc;
2102                 }
2103
2104                 entries++;
2105                 if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
2106                         entries *= 2;
2107                 ukinfo->cq_size = entries;
2108
2109                 rsize = info.cq_uk_init_info.cq_size * sizeof(struct irdma_cqe);
2110                 iwcq->kmem.size = ALIGN(round_up(rsize, 256), 256);
2111                 iwcq->kmem.va = dma_alloc_coherent(dev->hw->device,
2112                                                    iwcq->kmem.size,
2113                                                    &iwcq->kmem.pa, GFP_KERNEL);
2114                 if (!iwcq->kmem.va) {
2115                         err_code = -ENOMEM;
2116                         goto cq_free_rsrc;
2117                 }
2118
2119                 iwcq->kmem_shadow.size = ALIGN(IRDMA_SHADOW_AREA_SIZE << 3,
2120                                                64);
2121                 iwcq->kmem_shadow.va = dma_alloc_coherent(dev->hw->device,
2122                                                           iwcq->kmem_shadow.size,
2123                                                           &iwcq->kmem_shadow.pa,
2124                                                           GFP_KERNEL);
2125                 if (!iwcq->kmem_shadow.va) {
2126                         err_code = -ENOMEM;
2127                         goto cq_free_rsrc;
2128                 }
2129                 info.shadow_area_pa = iwcq->kmem_shadow.pa;
2130                 ukinfo->shadow_area = iwcq->kmem_shadow.va;
2131                 ukinfo->cq_base = iwcq->kmem.va;
2132                 info.cq_base_pa = iwcq->kmem.pa;
2133         }
2134
2135         if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
2136                 info.shadow_read_threshold = min(info.cq_uk_init_info.cq_size / 2,
2137                                                  (u32)IRDMA_MAX_CQ_READ_THRESH);
2138
2139         if (irdma_sc_cq_init(cq, &info)) {
2140                 ibdev_dbg(&iwdev->ibdev, "VERBS: init cq fail\n");
2141                 err_code = -EPROTO;
2142                 goto cq_free_rsrc;
2143         }
2144
2145         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
2146         if (!cqp_request) {
2147                 err_code = -ENOMEM;
2148                 goto cq_free_rsrc;
2149         }
2150
2151         cqp_info = &cqp_request->info;
2152         cqp_info->cqp_cmd = IRDMA_OP_CQ_CREATE;
2153         cqp_info->post_sq = 1;
2154         cqp_info->in.u.cq_create.cq = cq;
2155         cqp_info->in.u.cq_create.check_overflow = true;
2156         cqp_info->in.u.cq_create.scratch = (uintptr_t)cqp_request;
2157         err_code = irdma_handle_cqp_op(rf, cqp_request);
2158         irdma_put_cqp_request(&rf->cqp, cqp_request);
2159         if (err_code)
2160                 goto cq_free_rsrc;
2161
2162         if (udata) {
2163                 struct irdma_create_cq_resp resp = {};
2164
2165                 resp.cq_id = info.cq_uk_init_info.cq_id;
2166                 resp.cq_size = info.cq_uk_init_info.cq_size;
2167                 if (ib_copy_to_udata(udata, &resp,
2168                                      min(sizeof(resp), udata->outlen))) {
2169                         ibdev_dbg(&iwdev->ibdev,
2170                                   "VERBS: copy to user data\n");
2171                         err_code = -EPROTO;
2172                         goto cq_destroy;
2173                 }
2174         }
2175         return 0;
2176 cq_destroy:
2177         irdma_cq_wq_destroy(rf, cq);
2178 cq_free_rsrc:
2179         irdma_cq_free_rsrc(rf, iwcq);
2180
2181         return err_code;
2182 }
2183
2184 /**
2185  * irdma_get_mr_access - get hw MR access permissions from IB access flags
2186  * @access: IB access flags
2187  */
2188 static inline u16 irdma_get_mr_access(int access)
2189 {
2190         u16 hw_access = 0;
2191
2192         hw_access |= (access & IB_ACCESS_LOCAL_WRITE) ?
2193                      IRDMA_ACCESS_FLAGS_LOCALWRITE : 0;
2194         hw_access |= (access & IB_ACCESS_REMOTE_WRITE) ?
2195                      IRDMA_ACCESS_FLAGS_REMOTEWRITE : 0;
2196         hw_access |= (access & IB_ACCESS_REMOTE_READ) ?
2197                      IRDMA_ACCESS_FLAGS_REMOTEREAD : 0;
2198         hw_access |= (access & IB_ACCESS_MW_BIND) ?
2199                      IRDMA_ACCESS_FLAGS_BIND_WINDOW : 0;
2200         hw_access |= (access & IB_ZERO_BASED) ?
2201                      IRDMA_ACCESS_FLAGS_ZERO_BASED : 0;
2202         hw_access |= IRDMA_ACCESS_FLAGS_LOCALREAD;
2203
2204         return hw_access;
2205 }
2206
2207 /**
2208  * irdma_free_stag - free stag resource
2209  * @iwdev: irdma device
2210  * @stag: stag to free
2211  */
2212 static void irdma_free_stag(struct irdma_device *iwdev, u32 stag)
2213 {
2214         u32 stag_idx;
2215
2216         stag_idx = (stag & iwdev->rf->mr_stagmask) >> IRDMA_CQPSQ_STAG_IDX_S;
2217         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_mrs, stag_idx);
2218 }
2219
2220 /**
2221  * irdma_create_stag - create random stag
2222  * @iwdev: irdma device
2223  */
2224 static u32 irdma_create_stag(struct irdma_device *iwdev)
2225 {
2226         u32 stag = 0;
2227         u32 stag_index = 0;
2228         u32 next_stag_index;
2229         u32 driver_key;
2230         u32 random;
2231         u8 consumer_key;
2232         int ret;
2233
2234         get_random_bytes(&random, sizeof(random));
2235         consumer_key = (u8)random;
2236
2237         driver_key = random & ~iwdev->rf->mr_stagmask;
2238         next_stag_index = (random & iwdev->rf->mr_stagmask) >> 8;
2239         next_stag_index %= iwdev->rf->max_mr;
2240
2241         ret = irdma_alloc_rsrc(iwdev->rf, iwdev->rf->allocated_mrs,
2242                                iwdev->rf->max_mr, &stag_index,
2243                                &next_stag_index);
2244         if (ret)
2245                 return stag;
2246         stag = stag_index << IRDMA_CQPSQ_STAG_IDX_S;
2247         stag |= driver_key;
2248         stag += (u32)consumer_key;
2249
2250         return stag;
2251 }
2252
2253 /**
2254  * irdma_next_pbl_addr - Get next pbl address
2255  * @pbl: pointer to a pble
2256  * @pinfo: info pointer
2257  * @idx: index
2258  */
2259 static inline u64 *irdma_next_pbl_addr(u64 *pbl, struct irdma_pble_info **pinfo,
2260                                        u32 *idx)
2261 {
2262         *idx += 1;
2263         if (!(*pinfo) || *idx != (*pinfo)->cnt)
2264                 return ++pbl;
2265         *idx = 0;
2266         (*pinfo)++;
2267
2268         return (*pinfo)->addr;
2269 }
2270
2271 /**
2272  * irdma_copy_user_pgaddrs - copy user page address to pble's os locally
2273  * @iwmr: iwmr for IB's user page addresses
2274  * @pbl: ple pointer to save 1 level or 0 level pble
2275  * @level: indicated level 0, 1 or 2
2276  */
2277 static void irdma_copy_user_pgaddrs(struct irdma_mr *iwmr, u64 *pbl,
2278                                     enum irdma_pble_level level)
2279 {
2280         struct ib_umem *region = iwmr->region;
2281         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2282         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2283         struct irdma_pble_info *pinfo;
2284         struct ib_block_iter biter;
2285         u32 idx = 0;
2286         u32 pbl_cnt = 0;
2287
2288         pinfo = (level == PBLE_LEVEL_1) ? NULL : palloc->level2.leaf;
2289
2290         if (iwmr->type == IRDMA_MEMREG_TYPE_QP)
2291                 iwpbl->qp_mr.sq_page = sg_page(region->sgt_append.sgt.sgl);
2292
2293         rdma_umem_for_each_dma_block(region, &biter, iwmr->page_size) {
2294                 *pbl = rdma_block_iter_dma_address(&biter);
2295                 if (++pbl_cnt == palloc->total_cnt)
2296                         break;
2297                 pbl = irdma_next_pbl_addr(pbl, &pinfo, &idx);
2298         }
2299 }
2300
2301 /**
2302  * irdma_check_mem_contiguous - check if pbls stored in arr are contiguous
2303  * @arr: lvl1 pbl array
2304  * @npages: page count
2305  * @pg_size: page size
2306  *
2307  */
2308 static bool irdma_check_mem_contiguous(u64 *arr, u32 npages, u32 pg_size)
2309 {
2310         u32 pg_idx;
2311
2312         for (pg_idx = 0; pg_idx < npages; pg_idx++) {
2313                 if ((*arr + (pg_size * pg_idx)) != arr[pg_idx])
2314                         return false;
2315         }
2316
2317         return true;
2318 }
2319
2320 /**
2321  * irdma_check_mr_contiguous - check if MR is physically contiguous
2322  * @palloc: pbl allocation struct
2323  * @pg_size: page size
2324  */
2325 static bool irdma_check_mr_contiguous(struct irdma_pble_alloc *palloc,
2326                                       u32 pg_size)
2327 {
2328         struct irdma_pble_level2 *lvl2 = &palloc->level2;
2329         struct irdma_pble_info *leaf = lvl2->leaf;
2330         u64 *arr = NULL;
2331         u64 *start_addr = NULL;
2332         int i;
2333         bool ret;
2334
2335         if (palloc->level == PBLE_LEVEL_1) {
2336                 arr = palloc->level1.addr;
2337                 ret = irdma_check_mem_contiguous(arr, palloc->total_cnt,
2338                                                  pg_size);
2339                 return ret;
2340         }
2341
2342         start_addr = leaf->addr;
2343
2344         for (i = 0; i < lvl2->leaf_cnt; i++, leaf++) {
2345                 arr = leaf->addr;
2346                 if ((*start_addr + (i * pg_size * PBLE_PER_PAGE)) != *arr)
2347                         return false;
2348                 ret = irdma_check_mem_contiguous(arr, leaf->cnt, pg_size);
2349                 if (!ret)
2350                         return false;
2351         }
2352
2353         return true;
2354 }
2355
2356 /**
2357  * irdma_setup_pbles - copy user pg address to pble's
2358  * @rf: RDMA PCI function
2359  * @iwmr: mr pointer for this memory registration
2360  * @use_pbles: flag if to use pble's
2361  */
2362 static int irdma_setup_pbles(struct irdma_pci_f *rf, struct irdma_mr *iwmr,
2363                              bool use_pbles)
2364 {
2365         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2366         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2367         struct irdma_pble_info *pinfo;
2368         u64 *pbl;
2369         int status;
2370         enum irdma_pble_level level = PBLE_LEVEL_1;
2371
2372         if (use_pbles) {
2373                 status = irdma_get_pble(rf->pble_rsrc, palloc, iwmr->page_cnt,
2374                                         false);
2375                 if (status)
2376                         return status;
2377
2378                 iwpbl->pbl_allocated = true;
2379                 level = palloc->level;
2380                 pinfo = (level == PBLE_LEVEL_1) ? &palloc->level1 :
2381                                                   palloc->level2.leaf;
2382                 pbl = pinfo->addr;
2383         } else {
2384                 pbl = iwmr->pgaddrmem;
2385         }
2386
2387         irdma_copy_user_pgaddrs(iwmr, pbl, level);
2388
2389         if (use_pbles)
2390                 iwmr->pgaddrmem[0] = *pbl;
2391
2392         return 0;
2393 }
2394
2395 /**
2396  * irdma_handle_q_mem - handle memory for qp and cq
2397  * @iwdev: irdma device
2398  * @req: information for q memory management
2399  * @iwpbl: pble struct
2400  * @use_pbles: flag to use pble
2401  */
2402 static int irdma_handle_q_mem(struct irdma_device *iwdev,
2403                               struct irdma_mem_reg_req *req,
2404                               struct irdma_pbl *iwpbl, bool use_pbles)
2405 {
2406         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2407         struct irdma_mr *iwmr = iwpbl->iwmr;
2408         struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;
2409         struct irdma_cq_mr *cqmr = &iwpbl->cq_mr;
2410         struct irdma_hmc_pble *hmc_p;
2411         u64 *arr = iwmr->pgaddrmem;
2412         u32 pg_size, total;
2413         int err = 0;
2414         bool ret = true;
2415
2416         pg_size = iwmr->page_size;
2417         err = irdma_setup_pbles(iwdev->rf, iwmr, use_pbles);
2418         if (err)
2419                 return err;
2420
2421         if (use_pbles && palloc->level != PBLE_LEVEL_1) {
2422                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2423                 iwpbl->pbl_allocated = false;
2424                 return -ENOMEM;
2425         }
2426
2427         if (use_pbles)
2428                 arr = palloc->level1.addr;
2429
2430         switch (iwmr->type) {
2431         case IRDMA_MEMREG_TYPE_QP:
2432                 total = req->sq_pages + req->rq_pages;
2433                 hmc_p = &qpmr->sq_pbl;
2434                 qpmr->shadow = (dma_addr_t)arr[total];
2435
2436                 if (use_pbles) {
2437                         ret = irdma_check_mem_contiguous(arr, req->sq_pages,
2438                                                          pg_size);
2439                         if (ret)
2440                                 ret = irdma_check_mem_contiguous(&arr[req->sq_pages],
2441                                                                  req->rq_pages,
2442                                                                  pg_size);
2443                 }
2444
2445                 if (!ret) {
2446                         hmc_p->idx = palloc->level1.idx;
2447                         hmc_p = &qpmr->rq_pbl;
2448                         hmc_p->idx = palloc->level1.idx + req->sq_pages;
2449                 } else {
2450                         hmc_p->addr = arr[0];
2451                         hmc_p = &qpmr->rq_pbl;
2452                         hmc_p->addr = arr[req->sq_pages];
2453                 }
2454                 break;
2455         case IRDMA_MEMREG_TYPE_CQ:
2456                 hmc_p = &cqmr->cq_pbl;
2457
2458                 if (!cqmr->split)
2459                         cqmr->shadow = (dma_addr_t)arr[req->cq_pages];
2460
2461                 if (use_pbles)
2462                         ret = irdma_check_mem_contiguous(arr, req->cq_pages,
2463                                                          pg_size);
2464
2465                 if (!ret)
2466                         hmc_p->idx = palloc->level1.idx;
2467                 else
2468                         hmc_p->addr = arr[0];
2469         break;
2470         default:
2471                 ibdev_dbg(&iwdev->ibdev, "VERBS: MR type error\n");
2472                 err = -EINVAL;
2473         }
2474
2475         if (use_pbles && ret) {
2476                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2477                 iwpbl->pbl_allocated = false;
2478         }
2479
2480         return err;
2481 }
2482
2483 /**
2484  * irdma_hw_alloc_mw - create the hw memory window
2485  * @iwdev: irdma device
2486  * @iwmr: pointer to memory window info
2487  */
2488 static int irdma_hw_alloc_mw(struct irdma_device *iwdev, struct irdma_mr *iwmr)
2489 {
2490         struct irdma_mw_alloc_info *info;
2491         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2492         struct irdma_cqp_request *cqp_request;
2493         struct cqp_cmds_info *cqp_info;
2494         int status;
2495
2496         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2497         if (!cqp_request)
2498                 return -ENOMEM;
2499
2500         cqp_info = &cqp_request->info;
2501         info = &cqp_info->in.u.mw_alloc.info;
2502         memset(info, 0, sizeof(*info));
2503         if (iwmr->ibmw.type == IB_MW_TYPE_1)
2504                 info->mw_wide = true;
2505
2506         info->page_size = PAGE_SIZE;
2507         info->mw_stag_index = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2508         info->pd_id = iwpd->sc_pd.pd_id;
2509         info->remote_access = true;
2510         cqp_info->cqp_cmd = IRDMA_OP_MW_ALLOC;
2511         cqp_info->post_sq = 1;
2512         cqp_info->in.u.mw_alloc.dev = &iwdev->rf->sc_dev;
2513         cqp_info->in.u.mw_alloc.scratch = (uintptr_t)cqp_request;
2514         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2515         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2516
2517         return status;
2518 }
2519
2520 /**
2521  * irdma_alloc_mw - Allocate memory window
2522  * @ibmw: Memory Window
2523  * @udata: user data pointer
2524  */
2525 static int irdma_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata)
2526 {
2527         struct irdma_device *iwdev = to_iwdev(ibmw->device);
2528         struct irdma_mr *iwmr = to_iwmw(ibmw);
2529         int err_code;
2530         u32 stag;
2531
2532         stag = irdma_create_stag(iwdev);
2533         if (!stag)
2534                 return -ENOMEM;
2535
2536         iwmr->stag = stag;
2537         ibmw->rkey = stag;
2538
2539         err_code = irdma_hw_alloc_mw(iwdev, iwmr);
2540         if (err_code) {
2541                 irdma_free_stag(iwdev, stag);
2542                 return err_code;
2543         }
2544
2545         return 0;
2546 }
2547
2548 /**
2549  * irdma_dealloc_mw - Dealloc memory window
2550  * @ibmw: memory window structure.
2551  */
2552 static int irdma_dealloc_mw(struct ib_mw *ibmw)
2553 {
2554         struct ib_pd *ibpd = ibmw->pd;
2555         struct irdma_pd *iwpd = to_iwpd(ibpd);
2556         struct irdma_mr *iwmr = to_iwmr((struct ib_mr *)ibmw);
2557         struct irdma_device *iwdev = to_iwdev(ibmw->device);
2558         struct irdma_cqp_request *cqp_request;
2559         struct cqp_cmds_info *cqp_info;
2560         struct irdma_dealloc_stag_info *info;
2561
2562         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2563         if (!cqp_request)
2564                 return -ENOMEM;
2565
2566         cqp_info = &cqp_request->info;
2567         info = &cqp_info->in.u.dealloc_stag.info;
2568         memset(info, 0, sizeof(*info));
2569         info->pd_id = iwpd->sc_pd.pd_id;
2570         info->stag_idx = ibmw->rkey >> IRDMA_CQPSQ_STAG_IDX_S;
2571         info->mr = false;
2572         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
2573         cqp_info->post_sq = 1;
2574         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
2575         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
2576         irdma_handle_cqp_op(iwdev->rf, cqp_request);
2577         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2578         irdma_free_stag(iwdev, iwmr->stag);
2579
2580         return 0;
2581 }
2582
2583 /**
2584  * irdma_hw_alloc_stag - cqp command to allocate stag
2585  * @iwdev: irdma device
2586  * @iwmr: irdma mr pointer
2587  */
2588 static int irdma_hw_alloc_stag(struct irdma_device *iwdev,
2589                                struct irdma_mr *iwmr)
2590 {
2591         struct irdma_allocate_stag_info *info;
2592         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2593         int status;
2594         struct irdma_cqp_request *cqp_request;
2595         struct cqp_cmds_info *cqp_info;
2596
2597         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2598         if (!cqp_request)
2599                 return -ENOMEM;
2600
2601         cqp_info = &cqp_request->info;
2602         info = &cqp_info->in.u.alloc_stag.info;
2603         memset(info, 0, sizeof(*info));
2604         info->page_size = PAGE_SIZE;
2605         info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2606         info->pd_id = iwpd->sc_pd.pd_id;
2607         info->total_len = iwmr->len;
2608         info->remote_access = true;
2609         cqp_info->cqp_cmd = IRDMA_OP_ALLOC_STAG;
2610         cqp_info->post_sq = 1;
2611         cqp_info->in.u.alloc_stag.dev = &iwdev->rf->sc_dev;
2612         cqp_info->in.u.alloc_stag.scratch = (uintptr_t)cqp_request;
2613         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2614         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2615
2616         return status;
2617 }
2618
2619 /**
2620  * irdma_alloc_mr - register stag for fast memory registration
2621  * @pd: ibpd pointer
2622  * @mr_type: memory for stag registrion
2623  * @max_num_sg: man number of pages
2624  */
2625 static struct ib_mr *irdma_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
2626                                     u32 max_num_sg)
2627 {
2628         struct irdma_device *iwdev = to_iwdev(pd->device);
2629         struct irdma_pble_alloc *palloc;
2630         struct irdma_pbl *iwpbl;
2631         struct irdma_mr *iwmr;
2632         u32 stag;
2633         int err_code;
2634
2635         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2636         if (!iwmr)
2637                 return ERR_PTR(-ENOMEM);
2638
2639         stag = irdma_create_stag(iwdev);
2640         if (!stag) {
2641                 err_code = -ENOMEM;
2642                 goto err;
2643         }
2644
2645         iwmr->stag = stag;
2646         iwmr->ibmr.rkey = stag;
2647         iwmr->ibmr.lkey = stag;
2648         iwmr->ibmr.pd = pd;
2649         iwmr->ibmr.device = pd->device;
2650         iwpbl = &iwmr->iwpbl;
2651         iwpbl->iwmr = iwmr;
2652         iwmr->type = IRDMA_MEMREG_TYPE_MEM;
2653         palloc = &iwpbl->pble_alloc;
2654         iwmr->page_cnt = max_num_sg;
2655         err_code = irdma_get_pble(iwdev->rf->pble_rsrc, palloc, iwmr->page_cnt,
2656                                   false);
2657         if (err_code)
2658                 goto err_get_pble;
2659
2660         err_code = irdma_hw_alloc_stag(iwdev, iwmr);
2661         if (err_code)
2662                 goto err_alloc_stag;
2663
2664         iwpbl->pbl_allocated = true;
2665
2666         return &iwmr->ibmr;
2667 err_alloc_stag:
2668         irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2669 err_get_pble:
2670         irdma_free_stag(iwdev, stag);
2671 err:
2672         kfree(iwmr);
2673
2674         return ERR_PTR(err_code);
2675 }
2676
2677 /**
2678  * irdma_set_page - populate pbl list for fmr
2679  * @ibmr: ib mem to access iwarp mr pointer
2680  * @addr: page dma address fro pbl list
2681  */
2682 static int irdma_set_page(struct ib_mr *ibmr, u64 addr)
2683 {
2684         struct irdma_mr *iwmr = to_iwmr(ibmr);
2685         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2686         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2687         u64 *pbl;
2688
2689         if (unlikely(iwmr->npages == iwmr->page_cnt))
2690                 return -ENOMEM;
2691
2692         if (palloc->level == PBLE_LEVEL_2) {
2693                 struct irdma_pble_info *palloc_info =
2694                         palloc->level2.leaf + (iwmr->npages >> PBLE_512_SHIFT);
2695
2696                 palloc_info->addr[iwmr->npages & (PBLE_PER_PAGE - 1)] = addr;
2697         } else {
2698                 pbl = palloc->level1.addr;
2699                 pbl[iwmr->npages] = addr;
2700         }
2701         iwmr->npages++;
2702
2703         return 0;
2704 }
2705
2706 /**
2707  * irdma_map_mr_sg - map of sg list for fmr
2708  * @ibmr: ib mem to access iwarp mr pointer
2709  * @sg: scatter gather list
2710  * @sg_nents: number of sg pages
2711  * @sg_offset: scatter gather list for fmr
2712  */
2713 static int irdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
2714                            int sg_nents, unsigned int *sg_offset)
2715 {
2716         struct irdma_mr *iwmr = to_iwmr(ibmr);
2717
2718         iwmr->npages = 0;
2719
2720         return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, irdma_set_page);
2721 }
2722
2723 /**
2724  * irdma_hwreg_mr - send cqp command for memory registration
2725  * @iwdev: irdma device
2726  * @iwmr: irdma mr pointer
2727  * @access: access for MR
2728  */
2729 static int irdma_hwreg_mr(struct irdma_device *iwdev, struct irdma_mr *iwmr,
2730                           u16 access)
2731 {
2732         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2733         struct irdma_reg_ns_stag_info *stag_info;
2734         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
2735         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2736         struct irdma_cqp_request *cqp_request;
2737         struct cqp_cmds_info *cqp_info;
2738         int ret;
2739
2740         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2741         if (!cqp_request)
2742                 return -ENOMEM;
2743
2744         cqp_info = &cqp_request->info;
2745         stag_info = &cqp_info->in.u.mr_reg_non_shared.info;
2746         memset(stag_info, 0, sizeof(*stag_info));
2747         stag_info->va = iwpbl->user_base;
2748         stag_info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2749         stag_info->stag_key = (u8)iwmr->stag;
2750         stag_info->total_len = iwmr->len;
2751         stag_info->access_rights = irdma_get_mr_access(access);
2752         stag_info->pd_id = iwpd->sc_pd.pd_id;
2753         if (stag_info->access_rights & IRDMA_ACCESS_FLAGS_ZERO_BASED)
2754                 stag_info->addr_type = IRDMA_ADDR_TYPE_ZERO_BASED;
2755         else
2756                 stag_info->addr_type = IRDMA_ADDR_TYPE_VA_BASED;
2757         stag_info->page_size = iwmr->page_size;
2758
2759         if (iwpbl->pbl_allocated) {
2760                 if (palloc->level == PBLE_LEVEL_1) {
2761                         stag_info->first_pm_pbl_index = palloc->level1.idx;
2762                         stag_info->chunk_size = 1;
2763                 } else {
2764                         stag_info->first_pm_pbl_index = palloc->level2.root.idx;
2765                         stag_info->chunk_size = 3;
2766                 }
2767         } else {
2768                 stag_info->reg_addr_pa = iwmr->pgaddrmem[0];
2769         }
2770
2771         cqp_info->cqp_cmd = IRDMA_OP_MR_REG_NON_SHARED;
2772         cqp_info->post_sq = 1;
2773         cqp_info->in.u.mr_reg_non_shared.dev = &iwdev->rf->sc_dev;
2774         cqp_info->in.u.mr_reg_non_shared.scratch = (uintptr_t)cqp_request;
2775         ret = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2776         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2777
2778         return ret;
2779 }
2780
2781 /**
2782  * irdma_reg_user_mr - Register a user memory region
2783  * @pd: ptr of pd
2784  * @start: virtual start address
2785  * @len: length of mr
2786  * @virt: virtual address
2787  * @access: access of mr
2788  * @udata: user data
2789  */
2790 static struct ib_mr *irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,
2791                                        u64 virt, int access,
2792                                        struct ib_udata *udata)
2793 {
2794 #define IRDMA_MEM_REG_MIN_REQ_LEN offsetofend(struct irdma_mem_reg_req, sq_pages)
2795         struct irdma_device *iwdev = to_iwdev(pd->device);
2796         struct irdma_ucontext *ucontext;
2797         struct irdma_pble_alloc *palloc;
2798         struct irdma_pbl *iwpbl;
2799         struct irdma_mr *iwmr;
2800         struct ib_umem *region;
2801         struct irdma_mem_reg_req req;
2802         u32 total, stag = 0;
2803         u8 shadow_pgcnt = 1;
2804         bool use_pbles = false;
2805         unsigned long flags;
2806         int err = -EINVAL;
2807         int ret;
2808
2809         if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size)
2810                 return ERR_PTR(-EINVAL);
2811
2812         if (udata->inlen < IRDMA_MEM_REG_MIN_REQ_LEN)
2813                 return ERR_PTR(-EINVAL);
2814
2815         region = ib_umem_get(pd->device, start, len, access);
2816
2817         if (IS_ERR(region)) {
2818                 ibdev_dbg(&iwdev->ibdev,
2819                           "VERBS: Failed to create ib_umem region\n");
2820                 return (struct ib_mr *)region;
2821         }
2822
2823         if (ib_copy_from_udata(&req, udata, min(sizeof(req), udata->inlen))) {
2824                 ib_umem_release(region);
2825                 return ERR_PTR(-EFAULT);
2826         }
2827
2828         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2829         if (!iwmr) {
2830                 ib_umem_release(region);
2831                 return ERR_PTR(-ENOMEM);
2832         }
2833
2834         iwpbl = &iwmr->iwpbl;
2835         iwpbl->iwmr = iwmr;
2836         iwmr->region = region;
2837         iwmr->ibmr.pd = pd;
2838         iwmr->ibmr.device = pd->device;
2839         iwmr->ibmr.iova = virt;
2840         iwmr->page_size = PAGE_SIZE;
2841
2842         if (req.reg_type == IRDMA_MEMREG_TYPE_MEM) {
2843                 iwmr->page_size = ib_umem_find_best_pgsz(region,
2844                                                          iwdev->rf->sc_dev.hw_attrs.page_size_cap,
2845                                                          virt);
2846                 if (unlikely(!iwmr->page_size)) {
2847                         kfree(iwmr);
2848                         ib_umem_release(region);
2849                         return ERR_PTR(-EOPNOTSUPP);
2850                 }
2851         }
2852         iwmr->len = region->length;
2853         iwpbl->user_base = virt;
2854         palloc = &iwpbl->pble_alloc;
2855         iwmr->type = req.reg_type;
2856         iwmr->page_cnt = ib_umem_num_dma_blocks(region, iwmr->page_size);
2857
2858         switch (req.reg_type) {
2859         case IRDMA_MEMREG_TYPE_QP:
2860                 total = req.sq_pages + req.rq_pages + shadow_pgcnt;
2861                 if (total > iwmr->page_cnt) {
2862                         err = -EINVAL;
2863                         goto error;
2864                 }
2865                 total = req.sq_pages + req.rq_pages;
2866                 use_pbles = (total > 2);
2867                 err = irdma_handle_q_mem(iwdev, &req, iwpbl, use_pbles);
2868                 if (err)
2869                         goto error;
2870
2871                 ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext,
2872                                                      ibucontext);
2873                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
2874                 list_add_tail(&iwpbl->list, &ucontext->qp_reg_mem_list);
2875                 iwpbl->on_list = true;
2876                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
2877                 break;
2878         case IRDMA_MEMREG_TYPE_CQ:
2879                 if (iwdev->rf->sc_dev.hw_attrs.uk_attrs.feature_flags & IRDMA_FEATURE_CQ_RESIZE)
2880                         shadow_pgcnt = 0;
2881                 total = req.cq_pages + shadow_pgcnt;
2882                 if (total > iwmr->page_cnt) {
2883                         err = -EINVAL;
2884                         goto error;
2885                 }
2886
2887                 use_pbles = (req.cq_pages > 1);
2888                 err = irdma_handle_q_mem(iwdev, &req, iwpbl, use_pbles);
2889                 if (err)
2890                         goto error;
2891
2892                 ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext,
2893                                                      ibucontext);
2894                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2895                 list_add_tail(&iwpbl->list, &ucontext->cq_reg_mem_list);
2896                 iwpbl->on_list = true;
2897                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2898                 break;
2899         case IRDMA_MEMREG_TYPE_MEM:
2900                 use_pbles = (iwmr->page_cnt != 1);
2901
2902                 err = irdma_setup_pbles(iwdev->rf, iwmr, use_pbles);
2903                 if (err)
2904                         goto error;
2905
2906                 if (use_pbles) {
2907                         ret = irdma_check_mr_contiguous(palloc,
2908                                                         iwmr->page_size);
2909                         if (ret) {
2910                                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2911                                 iwpbl->pbl_allocated = false;
2912                         }
2913                 }
2914
2915                 stag = irdma_create_stag(iwdev);
2916                 if (!stag) {
2917                         err = -ENOMEM;
2918                         goto error;
2919                 }
2920
2921                 iwmr->stag = stag;
2922                 iwmr->ibmr.rkey = stag;
2923                 iwmr->ibmr.lkey = stag;
2924                 err = irdma_hwreg_mr(iwdev, iwmr, access);
2925                 if (err) {
2926                         irdma_free_stag(iwdev, stag);
2927                         goto error;
2928                 }
2929
2930                 break;
2931         default:
2932                 goto error;
2933         }
2934
2935         iwmr->type = req.reg_type;
2936
2937         return &iwmr->ibmr;
2938
2939 error:
2940         if (palloc->level != PBLE_LEVEL_0 && iwpbl->pbl_allocated)
2941                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2942         ib_umem_release(region);
2943         kfree(iwmr);
2944
2945         return ERR_PTR(err);
2946 }
2947
2948 /**
2949  * irdma_reg_phys_mr - register kernel physical memory
2950  * @pd: ibpd pointer
2951  * @addr: physical address of memory to register
2952  * @size: size of memory to register
2953  * @access: Access rights
2954  * @iova_start: start of virtual address for physical buffers
2955  */
2956 struct ib_mr *irdma_reg_phys_mr(struct ib_pd *pd, u64 addr, u64 size, int access,
2957                                 u64 *iova_start)
2958 {
2959         struct irdma_device *iwdev = to_iwdev(pd->device);
2960         struct irdma_pbl *iwpbl;
2961         struct irdma_mr *iwmr;
2962         u32 stag;
2963         int ret;
2964
2965         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2966         if (!iwmr)
2967                 return ERR_PTR(-ENOMEM);
2968
2969         iwmr->ibmr.pd = pd;
2970         iwmr->ibmr.device = pd->device;
2971         iwpbl = &iwmr->iwpbl;
2972         iwpbl->iwmr = iwmr;
2973         iwmr->type = IRDMA_MEMREG_TYPE_MEM;
2974         iwpbl->user_base = *iova_start;
2975         stag = irdma_create_stag(iwdev);
2976         if (!stag) {
2977                 ret = -ENOMEM;
2978                 goto err;
2979         }
2980
2981         iwmr->stag = stag;
2982         iwmr->ibmr.iova = *iova_start;
2983         iwmr->ibmr.rkey = stag;
2984         iwmr->ibmr.lkey = stag;
2985         iwmr->page_cnt = 1;
2986         iwmr->pgaddrmem[0] = addr;
2987         iwmr->len = size;
2988         iwmr->page_size = SZ_4K;
2989         ret = irdma_hwreg_mr(iwdev, iwmr, access);
2990         if (ret) {
2991                 irdma_free_stag(iwdev, stag);
2992                 goto err;
2993         }
2994
2995         return &iwmr->ibmr;
2996
2997 err:
2998         kfree(iwmr);
2999
3000         return ERR_PTR(ret);
3001 }
3002
3003 /**
3004  * irdma_get_dma_mr - register physical mem
3005  * @pd: ptr of pd
3006  * @acc: access for memory
3007  */
3008 static struct ib_mr *irdma_get_dma_mr(struct ib_pd *pd, int acc)
3009 {
3010         u64 kva = 0;
3011
3012         return irdma_reg_phys_mr(pd, 0, 0, acc, &kva);
3013 }
3014
3015 /**
3016  * irdma_del_memlist - Deleting pbl list entries for CQ/QP
3017  * @iwmr: iwmr for IB's user page addresses
3018  * @ucontext: ptr to user context
3019  */
3020 static void irdma_del_memlist(struct irdma_mr *iwmr,
3021                               struct irdma_ucontext *ucontext)
3022 {
3023         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
3024         unsigned long flags;
3025
3026         switch (iwmr->type) {
3027         case IRDMA_MEMREG_TYPE_CQ:
3028                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
3029                 if (iwpbl->on_list) {
3030                         iwpbl->on_list = false;
3031                         list_del(&iwpbl->list);
3032                 }
3033                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
3034                 break;
3035         case IRDMA_MEMREG_TYPE_QP:
3036                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
3037                 if (iwpbl->on_list) {
3038                         iwpbl->on_list = false;
3039                         list_del(&iwpbl->list);
3040                 }
3041                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
3042                 break;
3043         default:
3044                 break;
3045         }
3046 }
3047
3048 /**
3049  * irdma_dereg_mr - deregister mr
3050  * @ib_mr: mr ptr for dereg
3051  * @udata: user data
3052  */
3053 static int irdma_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
3054 {
3055         struct ib_pd *ibpd = ib_mr->pd;
3056         struct irdma_pd *iwpd = to_iwpd(ibpd);
3057         struct irdma_mr *iwmr = to_iwmr(ib_mr);
3058         struct irdma_device *iwdev = to_iwdev(ib_mr->device);
3059         struct irdma_dealloc_stag_info *info;
3060         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
3061         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
3062         struct irdma_cqp_request *cqp_request;
3063         struct cqp_cmds_info *cqp_info;
3064         int status;
3065
3066         if (iwmr->type != IRDMA_MEMREG_TYPE_MEM) {
3067                 if (iwmr->region) {
3068                         struct irdma_ucontext *ucontext;
3069
3070                         ucontext = rdma_udata_to_drv_context(udata,
3071                                                 struct irdma_ucontext,
3072                                                 ibucontext);
3073                         irdma_del_memlist(iwmr, ucontext);
3074                 }
3075                 goto done;
3076         }
3077
3078         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
3079         if (!cqp_request)
3080                 return -ENOMEM;
3081
3082         cqp_info = &cqp_request->info;
3083         info = &cqp_info->in.u.dealloc_stag.info;
3084         memset(info, 0, sizeof(*info));
3085         info->pd_id = iwpd->sc_pd.pd_id;
3086         info->stag_idx = ib_mr->rkey >> IRDMA_CQPSQ_STAG_IDX_S;
3087         info->mr = true;
3088         if (iwpbl->pbl_allocated)
3089                 info->dealloc_pbl = true;
3090
3091         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
3092         cqp_info->post_sq = 1;
3093         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
3094         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
3095         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
3096         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
3097         if (status)
3098                 return status;
3099
3100         irdma_free_stag(iwdev, iwmr->stag);
3101 done:
3102         if (iwpbl->pbl_allocated)
3103                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
3104         ib_umem_release(iwmr->region);
3105         kfree(iwmr);
3106
3107         return 0;
3108 }
3109
3110 /**
3111  * irdma_post_send -  kernel application wr
3112  * @ibqp: qp ptr for wr
3113  * @ib_wr: work request ptr
3114  * @bad_wr: return of bad wr if err
3115  */
3116 static int irdma_post_send(struct ib_qp *ibqp,
3117                            const struct ib_send_wr *ib_wr,
3118                            const struct ib_send_wr **bad_wr)
3119 {
3120         struct irdma_qp *iwqp;
3121         struct irdma_qp_uk *ukqp;
3122         struct irdma_sc_dev *dev;
3123         struct irdma_post_sq_info info;
3124         int err = 0;
3125         unsigned long flags;
3126         bool inv_stag;
3127         struct irdma_ah *ah;
3128
3129         iwqp = to_iwqp(ibqp);
3130         ukqp = &iwqp->sc_qp.qp_uk;
3131         dev = &iwqp->iwdev->rf->sc_dev;
3132
3133         spin_lock_irqsave(&iwqp->lock, flags);
3134         while (ib_wr) {
3135                 memset(&info, 0, sizeof(info));
3136                 inv_stag = false;
3137                 info.wr_id = (ib_wr->wr_id);
3138                 if ((ib_wr->send_flags & IB_SEND_SIGNALED) || iwqp->sig_all)
3139                         info.signaled = true;
3140                 if (ib_wr->send_flags & IB_SEND_FENCE)
3141                         info.read_fence = true;
3142                 switch (ib_wr->opcode) {
3143                 case IB_WR_SEND_WITH_IMM:
3144                         if (ukqp->qp_caps & IRDMA_SEND_WITH_IMM) {
3145                                 info.imm_data_valid = true;
3146                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
3147                         } else {
3148                                 err = -EINVAL;
3149                                 break;
3150                         }
3151                         fallthrough;
3152                 case IB_WR_SEND:
3153                 case IB_WR_SEND_WITH_INV:
3154                         if (ib_wr->opcode == IB_WR_SEND ||
3155                             ib_wr->opcode == IB_WR_SEND_WITH_IMM) {
3156                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
3157                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL;
3158                                 else
3159                                         info.op_type = IRDMA_OP_TYPE_SEND;
3160                         } else {
3161                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
3162                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL_INV;
3163                                 else
3164                                         info.op_type = IRDMA_OP_TYPE_SEND_INV;
3165                                 info.stag_to_inv = ib_wr->ex.invalidate_rkey;
3166                         }
3167
3168                         if (ib_wr->send_flags & IB_SEND_INLINE) {
3169                                 info.op.inline_send.data = (void *)(unsigned long)
3170                                                            ib_wr->sg_list[0].addr;
3171                                 info.op.inline_send.len = ib_wr->sg_list[0].length;
3172                                 if (iwqp->ibqp.qp_type == IB_QPT_UD ||
3173                                     iwqp->ibqp.qp_type == IB_QPT_GSI) {
3174                                         ah = to_iwah(ud_wr(ib_wr)->ah);
3175                                         info.op.inline_send.ah_id = ah->sc_ah.ah_info.ah_idx;
3176                                         info.op.inline_send.qkey = ud_wr(ib_wr)->remote_qkey;
3177                                         info.op.inline_send.dest_qp = ud_wr(ib_wr)->remote_qpn;
3178                                 }
3179                                 err = irdma_uk_inline_send(ukqp, &info, false);
3180                         } else {
3181                                 info.op.send.num_sges = ib_wr->num_sge;
3182                                 info.op.send.sg_list = ib_wr->sg_list;
3183                                 if (iwqp->ibqp.qp_type == IB_QPT_UD ||
3184                                     iwqp->ibqp.qp_type == IB_QPT_GSI) {
3185                                         ah = to_iwah(ud_wr(ib_wr)->ah);
3186                                         info.op.send.ah_id = ah->sc_ah.ah_info.ah_idx;
3187                                         info.op.send.qkey = ud_wr(ib_wr)->remote_qkey;
3188                                         info.op.send.dest_qp = ud_wr(ib_wr)->remote_qpn;
3189                                 }
3190                                 err = irdma_uk_send(ukqp, &info, false);
3191                         }
3192                         break;
3193                 case IB_WR_RDMA_WRITE_WITH_IMM:
3194                         if (ukqp->qp_caps & IRDMA_WRITE_WITH_IMM) {
3195                                 info.imm_data_valid = true;
3196                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
3197                         } else {
3198                                 err = -EINVAL;
3199                                 break;
3200                         }
3201                         fallthrough;
3202                 case IB_WR_RDMA_WRITE:
3203                         if (ib_wr->send_flags & IB_SEND_SOLICITED)
3204                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE_SOL;
3205                         else
3206                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE;
3207
3208                         if (ib_wr->send_flags & IB_SEND_INLINE) {
3209                                 info.op.inline_rdma_write.data = (void *)(uintptr_t)ib_wr->sg_list[0].addr;
3210                                 info.op.inline_rdma_write.len =
3211                                                 ib_wr->sg_list[0].length;
3212                                 info.op.inline_rdma_write.rem_addr.addr =
3213                                                 rdma_wr(ib_wr)->remote_addr;
3214                                 info.op.inline_rdma_write.rem_addr.lkey =
3215                                                 rdma_wr(ib_wr)->rkey;
3216                                 err = irdma_uk_inline_rdma_write(ukqp, &info, false);
3217                         } else {
3218                                 info.op.rdma_write.lo_sg_list = (void *)ib_wr->sg_list;
3219                                 info.op.rdma_write.num_lo_sges = ib_wr->num_sge;
3220                                 info.op.rdma_write.rem_addr.addr = rdma_wr(ib_wr)->remote_addr;
3221                                 info.op.rdma_write.rem_addr.lkey = rdma_wr(ib_wr)->rkey;
3222                                 err = irdma_uk_rdma_write(ukqp, &info, false);
3223                         }
3224                         break;
3225                 case IB_WR_RDMA_READ_WITH_INV:
3226                         inv_stag = true;
3227                         fallthrough;
3228                 case IB_WR_RDMA_READ:
3229                         if (ib_wr->num_sge >
3230                             dev->hw_attrs.uk_attrs.max_hw_read_sges) {
3231                                 err = -EINVAL;
3232                                 break;
3233                         }
3234                         info.op_type = IRDMA_OP_TYPE_RDMA_READ;
3235                         info.op.rdma_read.rem_addr.addr = rdma_wr(ib_wr)->remote_addr;
3236                         info.op.rdma_read.rem_addr.lkey = rdma_wr(ib_wr)->rkey;
3237                         info.op.rdma_read.lo_sg_list = (void *)ib_wr->sg_list;
3238                         info.op.rdma_read.num_lo_sges = ib_wr->num_sge;
3239                         err = irdma_uk_rdma_read(ukqp, &info, inv_stag, false);
3240                         break;
3241                 case IB_WR_LOCAL_INV:
3242                         info.op_type = IRDMA_OP_TYPE_INV_STAG;
3243                         info.op.inv_local_stag.target_stag = ib_wr->ex.invalidate_rkey;
3244                         err = irdma_uk_stag_local_invalidate(ukqp, &info, true);
3245                         break;
3246                 case IB_WR_REG_MR: {
3247                         struct irdma_mr *iwmr = to_iwmr(reg_wr(ib_wr)->mr);
3248                         struct irdma_pble_alloc *palloc = &iwmr->iwpbl.pble_alloc;
3249                         struct irdma_fast_reg_stag_info stag_info = {};
3250
3251                         stag_info.signaled = info.signaled;
3252                         stag_info.read_fence = info.read_fence;
3253                         stag_info.access_rights = irdma_get_mr_access(reg_wr(ib_wr)->access);
3254                         stag_info.stag_key = reg_wr(ib_wr)->key & 0xff;
3255                         stag_info.stag_idx = reg_wr(ib_wr)->key >> 8;
3256                         stag_info.page_size = reg_wr(ib_wr)->mr->page_size;
3257                         stag_info.wr_id = ib_wr->wr_id;
3258                         stag_info.addr_type = IRDMA_ADDR_TYPE_VA_BASED;
3259                         stag_info.va = (void *)(uintptr_t)iwmr->ibmr.iova;
3260                         stag_info.total_len = iwmr->ibmr.length;
3261                         stag_info.reg_addr_pa = *palloc->level1.addr;
3262                         stag_info.first_pm_pbl_index = palloc->level1.idx;
3263                         stag_info.local_fence = ib_wr->send_flags & IB_SEND_FENCE;
3264                         if (iwmr->npages > IRDMA_MIN_PAGES_PER_FMR)
3265                                 stag_info.chunk_size = 1;
3266                         err = irdma_sc_mr_fast_register(&iwqp->sc_qp, &stag_info,
3267                                                         true);
3268                         break;
3269                 }
3270                 default:
3271                         err = -EINVAL;
3272                         ibdev_dbg(&iwqp->iwdev->ibdev,
3273                                   "VERBS: upost_send bad opcode = 0x%x\n",
3274                                   ib_wr->opcode);
3275                         break;
3276                 }
3277
3278                 if (err)
3279                         break;
3280                 ib_wr = ib_wr->next;
3281         }
3282
3283         if (!iwqp->flush_issued) {
3284                 if (iwqp->hw_iwarp_state <= IRDMA_QP_STATE_RTS)
3285                         irdma_uk_qp_post_wr(ukqp);
3286                 spin_unlock_irqrestore(&iwqp->lock, flags);
3287         } else {
3288                 spin_unlock_irqrestore(&iwqp->lock, flags);
3289                 mod_delayed_work(iwqp->iwdev->cleanup_wq, &iwqp->dwork_flush,
3290                                  msecs_to_jiffies(IRDMA_FLUSH_DELAY_MS));
3291         }
3292         if (err)
3293                 *bad_wr = ib_wr;
3294
3295         return err;
3296 }
3297
3298 /**
3299  * irdma_post_recv - post receive wr for kernel application
3300  * @ibqp: ib qp pointer
3301  * @ib_wr: work request for receive
3302  * @bad_wr: bad wr caused an error
3303  */
3304 static int irdma_post_recv(struct ib_qp *ibqp,
3305                            const struct ib_recv_wr *ib_wr,
3306                            const struct ib_recv_wr **bad_wr)
3307 {
3308         struct irdma_qp *iwqp;
3309         struct irdma_qp_uk *ukqp;
3310         struct irdma_post_rq_info post_recv = {};
3311         unsigned long flags;
3312         int err = 0;
3313
3314         iwqp = to_iwqp(ibqp);
3315         ukqp = &iwqp->sc_qp.qp_uk;
3316
3317         spin_lock_irqsave(&iwqp->lock, flags);
3318         while (ib_wr) {
3319                 post_recv.num_sges = ib_wr->num_sge;
3320                 post_recv.wr_id = ib_wr->wr_id;
3321                 post_recv.sg_list = ib_wr->sg_list;
3322                 err = irdma_uk_post_receive(ukqp, &post_recv);
3323                 if (err) {
3324                         ibdev_dbg(&iwqp->iwdev->ibdev,
3325                                   "VERBS: post_recv err %d\n", err);
3326                         goto out;
3327                 }
3328
3329                 ib_wr = ib_wr->next;
3330         }
3331
3332 out:
3333         spin_unlock_irqrestore(&iwqp->lock, flags);
3334         if (iwqp->flush_issued)
3335                 mod_delayed_work(iwqp->iwdev->cleanup_wq, &iwqp->dwork_flush,
3336                                  msecs_to_jiffies(IRDMA_FLUSH_DELAY_MS));
3337
3338         if (err)
3339                 *bad_wr = ib_wr;
3340
3341         return err;
3342 }
3343
3344 /**
3345  * irdma_flush_err_to_ib_wc_status - return change flush error code to IB status
3346  * @opcode: iwarp flush code
3347  */
3348 static enum ib_wc_status irdma_flush_err_to_ib_wc_status(enum irdma_flush_opcode opcode)
3349 {
3350         switch (opcode) {
3351         case FLUSH_PROT_ERR:
3352                 return IB_WC_LOC_PROT_ERR;
3353         case FLUSH_REM_ACCESS_ERR:
3354                 return IB_WC_REM_ACCESS_ERR;
3355         case FLUSH_LOC_QP_OP_ERR:
3356                 return IB_WC_LOC_QP_OP_ERR;
3357         case FLUSH_REM_OP_ERR:
3358                 return IB_WC_REM_OP_ERR;
3359         case FLUSH_LOC_LEN_ERR:
3360                 return IB_WC_LOC_LEN_ERR;
3361         case FLUSH_GENERAL_ERR:
3362                 return IB_WC_WR_FLUSH_ERR;
3363         case FLUSH_RETRY_EXC_ERR:
3364                 return IB_WC_RETRY_EXC_ERR;
3365         case FLUSH_MW_BIND_ERR:
3366                 return IB_WC_MW_BIND_ERR;
3367         case FLUSH_REM_INV_REQ_ERR:
3368                 return IB_WC_REM_INV_REQ_ERR;
3369         case FLUSH_FATAL_ERR:
3370         default:
3371                 return IB_WC_FATAL_ERR;
3372         }
3373 }
3374
3375 /**
3376  * irdma_process_cqe - process cqe info
3377  * @entry: processed cqe
3378  * @cq_poll_info: cqe info
3379  */
3380 static void irdma_process_cqe(struct ib_wc *entry,
3381                               struct irdma_cq_poll_info *cq_poll_info)
3382 {
3383         struct irdma_qp *iwqp;
3384         struct irdma_sc_qp *qp;
3385
3386         entry->wc_flags = 0;
3387         entry->pkey_index = 0;
3388         entry->wr_id = cq_poll_info->wr_id;
3389
3390         qp = cq_poll_info->qp_handle;
3391         iwqp = qp->qp_uk.back_qp;
3392         entry->qp = qp->qp_uk.back_qp;
3393
3394         if (cq_poll_info->error) {
3395                 entry->status = (cq_poll_info->comp_status == IRDMA_COMPL_STATUS_FLUSHED) ?
3396                                 irdma_flush_err_to_ib_wc_status(cq_poll_info->minor_err) : IB_WC_GENERAL_ERR;
3397
3398                 entry->vendor_err = cq_poll_info->major_err << 16 |
3399                                     cq_poll_info->minor_err;
3400         } else {
3401                 entry->status = IB_WC_SUCCESS;
3402                 if (cq_poll_info->imm_valid) {
3403                         entry->ex.imm_data = htonl(cq_poll_info->imm_data);
3404                         entry->wc_flags |= IB_WC_WITH_IMM;
3405                 }
3406                 if (cq_poll_info->ud_smac_valid) {
3407                         ether_addr_copy(entry->smac, cq_poll_info->ud_smac);
3408                         entry->wc_flags |= IB_WC_WITH_SMAC;
3409                 }
3410
3411                 if (cq_poll_info->ud_vlan_valid) {
3412                         u16 vlan = cq_poll_info->ud_vlan & VLAN_VID_MASK;
3413
3414                         entry->sl = cq_poll_info->ud_vlan >> VLAN_PRIO_SHIFT;
3415                         if (vlan) {
3416                                 entry->vlan_id = vlan;
3417                                 entry->wc_flags |= IB_WC_WITH_VLAN;
3418                         }
3419                 } else {
3420                         entry->sl = 0;
3421                 }
3422         }
3423
3424         switch (cq_poll_info->op_type) {
3425         case IRDMA_OP_TYPE_RDMA_WRITE:
3426         case IRDMA_OP_TYPE_RDMA_WRITE_SOL:
3427                 entry->opcode = IB_WC_RDMA_WRITE;
3428                 break;
3429         case IRDMA_OP_TYPE_RDMA_READ_INV_STAG:
3430         case IRDMA_OP_TYPE_RDMA_READ:
3431                 entry->opcode = IB_WC_RDMA_READ;
3432                 break;
3433         case IRDMA_OP_TYPE_SEND_INV:
3434         case IRDMA_OP_TYPE_SEND_SOL:
3435         case IRDMA_OP_TYPE_SEND_SOL_INV:
3436         case IRDMA_OP_TYPE_SEND:
3437                 entry->opcode = IB_WC_SEND;
3438                 break;
3439         case IRDMA_OP_TYPE_FAST_REG_NSMR:
3440                 entry->opcode = IB_WC_REG_MR;
3441                 break;
3442         case IRDMA_OP_TYPE_INV_STAG:
3443                 entry->opcode = IB_WC_LOCAL_INV;
3444                 break;
3445         case IRDMA_OP_TYPE_REC_IMM:
3446         case IRDMA_OP_TYPE_REC:
3447                 entry->opcode = cq_poll_info->op_type == IRDMA_OP_TYPE_REC_IMM ?
3448                         IB_WC_RECV_RDMA_WITH_IMM : IB_WC_RECV;
3449                 if (qp->qp_uk.qp_type != IRDMA_QP_TYPE_ROCE_UD &&
3450                     cq_poll_info->stag_invalid_set) {
3451                         entry->ex.invalidate_rkey = cq_poll_info->inv_stag;
3452                         entry->wc_flags |= IB_WC_WITH_INVALIDATE;
3453                 }
3454                 break;
3455         default:
3456                 ibdev_err(&iwqp->iwdev->ibdev,
3457                           "Invalid opcode = %d in CQE\n", cq_poll_info->op_type);
3458                 entry->status = IB_WC_GENERAL_ERR;
3459                 return;
3460         }
3461
3462         if (qp->qp_uk.qp_type == IRDMA_QP_TYPE_ROCE_UD) {
3463                 entry->src_qp = cq_poll_info->ud_src_qpn;
3464                 entry->slid = 0;
3465                 entry->wc_flags |=
3466                         (IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE);
3467                 entry->network_hdr_type = cq_poll_info->ipv4 ?
3468                                                   RDMA_NETWORK_IPV4 :
3469                                                   RDMA_NETWORK_IPV6;
3470         } else {
3471                 entry->src_qp = cq_poll_info->qp_id;
3472         }
3473
3474         entry->byte_len = cq_poll_info->bytes_xfered;
3475 }
3476
3477 /**
3478  * irdma_poll_one - poll one entry of the CQ
3479  * @ukcq: ukcq to poll
3480  * @cur_cqe: current CQE info to be filled in
3481  * @entry: ibv_wc object to be filled for non-extended CQ or NULL for extended CQ
3482  *
3483  * Returns the internal irdma device error code or 0 on success
3484  */
3485 static inline int irdma_poll_one(struct irdma_cq_uk *ukcq,
3486                                  struct irdma_cq_poll_info *cur_cqe,
3487                                  struct ib_wc *entry)
3488 {
3489         int ret = irdma_uk_cq_poll_cmpl(ukcq, cur_cqe);
3490
3491         if (ret)
3492                 return ret;
3493
3494         irdma_process_cqe(entry, cur_cqe);
3495
3496         return 0;
3497 }
3498
3499 /**
3500  * __irdma_poll_cq - poll cq for completion (kernel apps)
3501  * @iwcq: cq to poll
3502  * @num_entries: number of entries to poll
3503  * @entry: wr of a completed entry
3504  */
3505 static int __irdma_poll_cq(struct irdma_cq *iwcq, int num_entries, struct ib_wc *entry)
3506 {
3507         struct list_head *tmp_node, *list_node;
3508         struct irdma_cq_buf *last_buf = NULL;
3509         struct irdma_cq_poll_info *cur_cqe = &iwcq->cur_cqe;
3510         struct irdma_cq_buf *cq_buf;
3511         int ret;
3512         struct irdma_device *iwdev;
3513         struct irdma_cq_uk *ukcq;
3514         bool cq_new_cqe = false;
3515         int resized_bufs = 0;
3516         int npolled = 0;
3517
3518         iwdev = to_iwdev(iwcq->ibcq.device);
3519         ukcq = &iwcq->sc_cq.cq_uk;
3520
3521         /* go through the list of previously resized CQ buffers */
3522         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
3523                 cq_buf = container_of(list_node, struct irdma_cq_buf, list);
3524                 while (npolled < num_entries) {
3525                         ret = irdma_poll_one(&cq_buf->cq_uk, cur_cqe, entry + npolled);
3526                         if (!ret) {
3527                                 ++npolled;
3528                                 cq_new_cqe = true;
3529                                 continue;
3530                         }
3531                         if (ret == -ENOENT)
3532                                 break;
3533                          /* QP using the CQ is destroyed. Skip reporting this CQE */
3534                         if (ret == -EFAULT) {
3535                                 cq_new_cqe = true;
3536                                 continue;
3537                         }
3538                         goto error;
3539                 }
3540
3541                 /* save the resized CQ buffer which received the last cqe */
3542                 if (cq_new_cqe)
3543                         last_buf = cq_buf;
3544                 cq_new_cqe = false;
3545         }
3546
3547         /* check the current CQ for new cqes */
3548         while (npolled < num_entries) {
3549                 ret = irdma_poll_one(ukcq, cur_cqe, entry + npolled);
3550                 if (ret == -ENOENT) {
3551                         ret = irdma_generated_cmpls(iwcq, cur_cqe);
3552                         if (!ret)
3553                                 irdma_process_cqe(entry + npolled, cur_cqe);
3554                 }
3555                 if (!ret) {
3556                         ++npolled;
3557                         cq_new_cqe = true;
3558                         continue;
3559                 }
3560
3561                 if (ret == -ENOENT)
3562                         break;
3563                 /* QP using the CQ is destroyed. Skip reporting this CQE */
3564                 if (ret == -EFAULT) {
3565                         cq_new_cqe = true;
3566                         continue;
3567                 }
3568                 goto error;
3569         }
3570
3571         if (cq_new_cqe)
3572                 /* all previous CQ resizes are complete */
3573                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, NULL);
3574         else if (last_buf)
3575                 /* only CQ resizes up to the last_buf are complete */
3576                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, last_buf);
3577         if (resized_bufs)
3578                 /* report to the HW the number of complete CQ resizes */
3579                 irdma_uk_cq_set_resized_cnt(ukcq, resized_bufs);
3580
3581         return npolled;
3582 error:
3583         ibdev_dbg(&iwdev->ibdev, "%s: Error polling CQ, irdma_err: %d\n",
3584                   __func__, ret);
3585
3586         return ret;
3587 }
3588
3589 /**
3590  * irdma_poll_cq - poll cq for completion (kernel apps)
3591  * @ibcq: cq to poll
3592  * @num_entries: number of entries to poll
3593  * @entry: wr of a completed entry
3594  */
3595 static int irdma_poll_cq(struct ib_cq *ibcq, int num_entries,
3596                          struct ib_wc *entry)
3597 {
3598         struct irdma_cq *iwcq;
3599         unsigned long flags;
3600         int ret;
3601
3602         iwcq = to_iwcq(ibcq);
3603
3604         spin_lock_irqsave(&iwcq->lock, flags);
3605         ret = __irdma_poll_cq(iwcq, num_entries, entry);
3606         spin_unlock_irqrestore(&iwcq->lock, flags);
3607
3608         return ret;
3609 }
3610
3611 /**
3612  * irdma_req_notify_cq - arm cq kernel application
3613  * @ibcq: cq to arm
3614  * @notify_flags: notofication flags
3615  */
3616 static int irdma_req_notify_cq(struct ib_cq *ibcq,
3617                                enum ib_cq_notify_flags notify_flags)
3618 {
3619         struct irdma_cq *iwcq;
3620         struct irdma_cq_uk *ukcq;
3621         unsigned long flags;
3622         enum irdma_cmpl_notify cq_notify;
3623         bool promo_event = false;
3624         int ret = 0;
3625
3626         cq_notify = notify_flags == IB_CQ_SOLICITED ?
3627                     IRDMA_CQ_COMPL_SOLICITED : IRDMA_CQ_COMPL_EVENT;
3628         iwcq = to_iwcq(ibcq);
3629         ukcq = &iwcq->sc_cq.cq_uk;
3630
3631         spin_lock_irqsave(&iwcq->lock, flags);
3632         /* Only promote to arm the CQ for any event if the last arm event was solicited. */
3633         if (iwcq->last_notify == IRDMA_CQ_COMPL_SOLICITED && notify_flags != IB_CQ_SOLICITED)
3634                 promo_event = true;
3635
3636         if (!atomic_cmpxchg(&iwcq->armed, 0, 1) || promo_event) {
3637                 iwcq->last_notify = cq_notify;
3638                 irdma_uk_cq_request_notification(ukcq, cq_notify);
3639         }
3640
3641         if ((notify_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
3642             (!irdma_cq_empty(iwcq) || !list_empty(&iwcq->cmpl_generated)))
3643                 ret = 1;
3644         spin_unlock_irqrestore(&iwcq->lock, flags);
3645
3646         return ret;
3647 }
3648
3649 static int irdma_roce_port_immutable(struct ib_device *ibdev, u32 port_num,
3650                                      struct ib_port_immutable *immutable)
3651 {
3652         struct ib_port_attr attr;
3653         int err;
3654
3655         immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
3656         err = ib_query_port(ibdev, port_num, &attr);
3657         if (err)
3658                 return err;
3659
3660         immutable->max_mad_size = IB_MGMT_MAD_SIZE;
3661         immutable->pkey_tbl_len = attr.pkey_tbl_len;
3662         immutable->gid_tbl_len = attr.gid_tbl_len;
3663
3664         return 0;
3665 }
3666
3667 static int irdma_iw_port_immutable(struct ib_device *ibdev, u32 port_num,
3668                                    struct ib_port_immutable *immutable)
3669 {
3670         struct ib_port_attr attr;
3671         int err;
3672
3673         immutable->core_cap_flags = RDMA_CORE_PORT_IWARP;
3674         err = ib_query_port(ibdev, port_num, &attr);
3675         if (err)
3676                 return err;
3677         immutable->gid_tbl_len = attr.gid_tbl_len;
3678
3679         return 0;
3680 }
3681
3682 static const struct rdma_stat_desc irdma_hw_stat_descs[] = {
3683         /* 32bit names */
3684         [IRDMA_HW_STAT_INDEX_RXVLANERR].name = "rxVlanErrors",
3685         [IRDMA_HW_STAT_INDEX_IP4RXDISCARD].name = "ip4InDiscards",
3686         [IRDMA_HW_STAT_INDEX_IP4RXTRUNC].name = "ip4InTruncatedPkts",
3687         [IRDMA_HW_STAT_INDEX_IP4TXNOROUTE].name = "ip4OutNoRoutes",
3688         [IRDMA_HW_STAT_INDEX_IP6RXDISCARD].name = "ip6InDiscards",
3689         [IRDMA_HW_STAT_INDEX_IP6RXTRUNC].name = "ip6InTruncatedPkts",
3690         [IRDMA_HW_STAT_INDEX_IP6TXNOROUTE].name = "ip6OutNoRoutes",
3691         [IRDMA_HW_STAT_INDEX_TCPRTXSEG].name = "tcpRetransSegs",
3692         [IRDMA_HW_STAT_INDEX_TCPRXOPTERR].name = "tcpInOptErrors",
3693         [IRDMA_HW_STAT_INDEX_TCPRXPROTOERR].name = "tcpInProtoErrors",
3694         [IRDMA_HW_STAT_INDEX_RXRPCNPHANDLED].name = "cnpHandled",
3695         [IRDMA_HW_STAT_INDEX_RXRPCNPIGNORED].name = "cnpIgnored",
3696         [IRDMA_HW_STAT_INDEX_TXNPCNPSENT].name = "cnpSent",
3697
3698         /* 64bit names */
3699         [IRDMA_HW_STAT_INDEX_IP4RXOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3700                 "ip4InOctets",
3701         [IRDMA_HW_STAT_INDEX_IP4RXPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3702                 "ip4InPkts",
3703         [IRDMA_HW_STAT_INDEX_IP4RXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3704                 "ip4InReasmRqd",
3705         [IRDMA_HW_STAT_INDEX_IP4RXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3706                 "ip4InMcastOctets",
3707         [IRDMA_HW_STAT_INDEX_IP4RXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3708                 "ip4InMcastPkts",
3709         [IRDMA_HW_STAT_INDEX_IP4TXOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3710                 "ip4OutOctets",
3711         [IRDMA_HW_STAT_INDEX_IP4TXPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3712                 "ip4OutPkts",
3713         [IRDMA_HW_STAT_INDEX_IP4TXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3714                 "ip4OutSegRqd",
3715         [IRDMA_HW_STAT_INDEX_IP4TXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3716                 "ip4OutMcastOctets",
3717         [IRDMA_HW_STAT_INDEX_IP4TXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3718                 "ip4OutMcastPkts",
3719         [IRDMA_HW_STAT_INDEX_IP6RXOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3720                 "ip6InOctets",
3721         [IRDMA_HW_STAT_INDEX_IP6RXPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3722                 "ip6InPkts",
3723         [IRDMA_HW_STAT_INDEX_IP6RXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3724                 "ip6InReasmRqd",
3725         [IRDMA_HW_STAT_INDEX_IP6RXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3726                 "ip6InMcastOctets",
3727         [IRDMA_HW_STAT_INDEX_IP6RXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3728                 "ip6InMcastPkts",
3729         [IRDMA_HW_STAT_INDEX_IP6TXOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3730                 "ip6OutOctets",
3731         [IRDMA_HW_STAT_INDEX_IP6TXPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3732                 "ip6OutPkts",
3733         [IRDMA_HW_STAT_INDEX_IP6TXFRAGS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3734                 "ip6OutSegRqd",
3735         [IRDMA_HW_STAT_INDEX_IP6TXMCOCTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3736                 "ip6OutMcastOctets",
3737         [IRDMA_HW_STAT_INDEX_IP6TXMCPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3738                 "ip6OutMcastPkts",
3739         [IRDMA_HW_STAT_INDEX_TCPRXSEGS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3740                 "tcpInSegs",
3741         [IRDMA_HW_STAT_INDEX_TCPTXSEG + IRDMA_HW_STAT_INDEX_MAX_32].name =
3742                 "tcpOutSegs",
3743         [IRDMA_HW_STAT_INDEX_RDMARXRDS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3744                 "iwInRdmaReads",
3745         [IRDMA_HW_STAT_INDEX_RDMARXSNDS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3746                 "iwInRdmaSends",
3747         [IRDMA_HW_STAT_INDEX_RDMARXWRS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3748                 "iwInRdmaWrites",
3749         [IRDMA_HW_STAT_INDEX_RDMATXRDS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3750                 "iwOutRdmaReads",
3751         [IRDMA_HW_STAT_INDEX_RDMATXSNDS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3752                 "iwOutRdmaSends",
3753         [IRDMA_HW_STAT_INDEX_RDMATXWRS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3754                 "iwOutRdmaWrites",
3755         [IRDMA_HW_STAT_INDEX_RDMAVBND + IRDMA_HW_STAT_INDEX_MAX_32].name =
3756                 "iwRdmaBnd",
3757         [IRDMA_HW_STAT_INDEX_RDMAVINV + IRDMA_HW_STAT_INDEX_MAX_32].name =
3758                 "iwRdmaInv",
3759         [IRDMA_HW_STAT_INDEX_UDPRXPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3760                 "RxUDP",
3761         [IRDMA_HW_STAT_INDEX_UDPTXPKTS + IRDMA_HW_STAT_INDEX_MAX_32].name =
3762                 "TxUDP",
3763         [IRDMA_HW_STAT_INDEX_RXNPECNMARKEDPKTS + IRDMA_HW_STAT_INDEX_MAX_32]
3764                 .name = "RxECNMrkd",
3765 };
3766
3767 static void irdma_get_dev_fw_str(struct ib_device *dev, char *str)
3768 {
3769         struct irdma_device *iwdev = to_iwdev(dev);
3770
3771         snprintf(str, IB_FW_VERSION_NAME_MAX, "%u.%u",
3772                  irdma_fw_major_ver(&iwdev->rf->sc_dev),
3773                  irdma_fw_minor_ver(&iwdev->rf->sc_dev));
3774 }
3775
3776 /**
3777  * irdma_alloc_hw_port_stats - Allocate a hw stats structure
3778  * @ibdev: device pointer from stack
3779  * @port_num: port number
3780  */
3781 static struct rdma_hw_stats *irdma_alloc_hw_port_stats(struct ib_device *ibdev,
3782                                                        u32 port_num)
3783 {
3784         int num_counters = IRDMA_HW_STAT_INDEX_MAX_32 +
3785                            IRDMA_HW_STAT_INDEX_MAX_64;
3786         unsigned long lifespan = RDMA_HW_STATS_DEFAULT_LIFESPAN;
3787
3788         BUILD_BUG_ON(ARRAY_SIZE(irdma_hw_stat_descs) !=
3789                      (IRDMA_HW_STAT_INDEX_MAX_32 + IRDMA_HW_STAT_INDEX_MAX_64));
3790
3791         return rdma_alloc_hw_stats_struct(irdma_hw_stat_descs, num_counters,
3792                                           lifespan);
3793 }
3794
3795 /**
3796  * irdma_get_hw_stats - Populates the rdma_hw_stats structure
3797  * @ibdev: device pointer from stack
3798  * @stats: stats pointer from stack
3799  * @port_num: port number
3800  * @index: which hw counter the stack is requesting we update
3801  */
3802 static int irdma_get_hw_stats(struct ib_device *ibdev,
3803                               struct rdma_hw_stats *stats, u32 port_num,
3804                               int index)
3805 {
3806         struct irdma_device *iwdev = to_iwdev(ibdev);
3807         struct irdma_dev_hw_stats *hw_stats = &iwdev->vsi.pestat->hw_stats;
3808
3809         if (iwdev->rf->rdma_ver >= IRDMA_GEN_2)
3810                 irdma_cqp_gather_stats_cmd(&iwdev->rf->sc_dev, iwdev->vsi.pestat, true);
3811         else
3812                 irdma_cqp_gather_stats_gen1(&iwdev->rf->sc_dev, iwdev->vsi.pestat);
3813
3814         memcpy(&stats->value[0], hw_stats, sizeof(*hw_stats));
3815
3816         return stats->num_counters;
3817 }
3818
3819 /**
3820  * irdma_query_gid - Query port GID
3821  * @ibdev: device pointer from stack
3822  * @port: port number
3823  * @index: Entry index
3824  * @gid: Global ID
3825  */
3826 static int irdma_query_gid(struct ib_device *ibdev, u32 port, int index,
3827                            union ib_gid *gid)
3828 {
3829         struct irdma_device *iwdev = to_iwdev(ibdev);
3830
3831         memset(gid->raw, 0, sizeof(gid->raw));
3832         ether_addr_copy(gid->raw, iwdev->netdev->dev_addr);
3833
3834         return 0;
3835 }
3836
3837 /**
3838  * mcast_list_add -  Add a new mcast item to list
3839  * @rf: RDMA PCI function
3840  * @new_elem: pointer to element to add
3841  */
3842 static void mcast_list_add(struct irdma_pci_f *rf,
3843                            struct mc_table_list *new_elem)
3844 {
3845         list_add(&new_elem->list, &rf->mc_qht_list.list);
3846 }
3847
3848 /**
3849  * mcast_list_del - Remove an mcast item from list
3850  * @mc_qht_elem: pointer to mcast table list element
3851  */
3852 static void mcast_list_del(struct mc_table_list *mc_qht_elem)
3853 {
3854         if (mc_qht_elem)
3855                 list_del(&mc_qht_elem->list);
3856 }
3857
3858 /**
3859  * mcast_list_lookup_ip - Search mcast list for address
3860  * @rf: RDMA PCI function
3861  * @ip_mcast: pointer to mcast IP address
3862  */
3863 static struct mc_table_list *mcast_list_lookup_ip(struct irdma_pci_f *rf,
3864                                                   u32 *ip_mcast)
3865 {
3866         struct mc_table_list *mc_qht_el;
3867         struct list_head *pos, *q;
3868
3869         list_for_each_safe (pos, q, &rf->mc_qht_list.list) {
3870                 mc_qht_el = list_entry(pos, struct mc_table_list, list);
3871                 if (!memcmp(mc_qht_el->mc_info.dest_ip, ip_mcast,
3872                             sizeof(mc_qht_el->mc_info.dest_ip)))
3873                         return mc_qht_el;
3874         }
3875
3876         return NULL;
3877 }
3878
3879 /**
3880  * irdma_mcast_cqp_op - perform a mcast cqp operation
3881  * @iwdev: irdma device
3882  * @mc_grp_ctx: mcast group info
3883  * @op: operation
3884  *
3885  * returns error status
3886  */
3887 static int irdma_mcast_cqp_op(struct irdma_device *iwdev,
3888                               struct irdma_mcast_grp_info *mc_grp_ctx, u8 op)
3889 {
3890         struct cqp_cmds_info *cqp_info;
3891         struct irdma_cqp_request *cqp_request;
3892         int status;
3893
3894         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
3895         if (!cqp_request)
3896                 return -ENOMEM;
3897
3898         cqp_request->info.in.u.mc_create.info = *mc_grp_ctx;
3899         cqp_info = &cqp_request->info;
3900         cqp_info->cqp_cmd = op;
3901         cqp_info->post_sq = 1;
3902         cqp_info->in.u.mc_create.scratch = (uintptr_t)cqp_request;
3903         cqp_info->in.u.mc_create.cqp = &iwdev->rf->cqp.sc_cqp;
3904         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
3905         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
3906
3907         return status;
3908 }
3909
3910 /**
3911  * irdma_mcast_mac - Get the multicast MAC for an IP address
3912  * @ip_addr: IPv4 or IPv6 address
3913  * @mac: pointer to result MAC address
3914  * @ipv4: flag indicating IPv4 or IPv6
3915  *
3916  */
3917 void irdma_mcast_mac(u32 *ip_addr, u8 *mac, bool ipv4)
3918 {
3919         u8 *ip = (u8 *)ip_addr;
3920
3921         if (ipv4) {
3922                 unsigned char mac4[ETH_ALEN] = {0x01, 0x00, 0x5E, 0x00,
3923                                                 0x00, 0x00};
3924
3925                 mac4[3] = ip[2] & 0x7F;
3926                 mac4[4] = ip[1];
3927                 mac4[5] = ip[0];
3928                 ether_addr_copy(mac, mac4);
3929         } else {
3930                 unsigned char mac6[ETH_ALEN] = {0x33, 0x33, 0x00, 0x00,
3931                                                 0x00, 0x00};
3932
3933                 mac6[2] = ip[3];
3934                 mac6[3] = ip[2];
3935                 mac6[4] = ip[1];
3936                 mac6[5] = ip[0];
3937                 ether_addr_copy(mac, mac6);
3938         }
3939 }
3940
3941 /**
3942  * irdma_attach_mcast - attach a qp to a multicast group
3943  * @ibqp: ptr to qp
3944  * @ibgid: pointer to global ID
3945  * @lid: local ID
3946  *
3947  * returns error status
3948  */
3949 static int irdma_attach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
3950 {
3951         struct irdma_qp *iwqp = to_iwqp(ibqp);
3952         struct irdma_device *iwdev = iwqp->iwdev;
3953         struct irdma_pci_f *rf = iwdev->rf;
3954         struct mc_table_list *mc_qht_elem;
3955         struct irdma_mcast_grp_ctx_entry_info mcg_info = {};
3956         unsigned long flags;
3957         u32 ip_addr[4] = {};
3958         u32 mgn;
3959         u32 no_mgs;
3960         int ret = 0;
3961         bool ipv4;
3962         u16 vlan_id;
3963         union irdma_sockaddr sgid_addr;
3964         unsigned char dmac[ETH_ALEN];
3965
3966         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
3967
3968         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid)) {
3969                 irdma_copy_ip_ntohl(ip_addr,
3970                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
3971                 irdma_netdev_vlan_ipv6(ip_addr, &vlan_id, NULL);
3972                 ipv4 = false;
3973                 ibdev_dbg(&iwdev->ibdev,
3974                           "VERBS: qp_id=%d, IP6address=%pI6\n", ibqp->qp_num,
3975                           ip_addr);
3976                 irdma_mcast_mac(ip_addr, dmac, false);
3977         } else {
3978                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
3979                 ipv4 = true;
3980                 vlan_id = irdma_get_vlan_ipv4(ip_addr);
3981                 irdma_mcast_mac(ip_addr, dmac, true);
3982                 ibdev_dbg(&iwdev->ibdev,
3983                           "VERBS: qp_id=%d, IP4address=%pI4, MAC=%pM\n",
3984                           ibqp->qp_num, ip_addr, dmac);
3985         }
3986
3987         spin_lock_irqsave(&rf->qh_list_lock, flags);
3988         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
3989         if (!mc_qht_elem) {
3990                 struct irdma_dma_mem *dma_mem_mc;
3991
3992                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3993                 mc_qht_elem = kzalloc(sizeof(*mc_qht_elem), GFP_KERNEL);
3994                 if (!mc_qht_elem)
3995                         return -ENOMEM;
3996
3997                 mc_qht_elem->mc_info.ipv4_valid = ipv4;
3998                 memcpy(mc_qht_elem->mc_info.dest_ip, ip_addr,
3999                        sizeof(mc_qht_elem->mc_info.dest_ip));
4000                 ret = irdma_alloc_rsrc(rf, rf->allocated_mcgs, rf->max_mcg,
4001                                        &mgn, &rf->next_mcg);
4002                 if (ret) {
4003                         kfree(mc_qht_elem);
4004                         return -ENOMEM;
4005                 }
4006
4007                 mc_qht_elem->mc_info.mgn = mgn;
4008                 dma_mem_mc = &mc_qht_elem->mc_grp_ctx.dma_mem_mc;
4009                 dma_mem_mc->size = ALIGN(sizeof(u64) * IRDMA_MAX_MGS_PER_CTX,
4010                                          IRDMA_HW_PAGE_SIZE);
4011                 dma_mem_mc->va = dma_alloc_coherent(rf->hw.device,
4012                                                     dma_mem_mc->size,
4013                                                     &dma_mem_mc->pa,
4014                                                     GFP_KERNEL);
4015                 if (!dma_mem_mc->va) {
4016                         irdma_free_rsrc(rf, rf->allocated_mcgs, mgn);
4017                         kfree(mc_qht_elem);
4018                         return -ENOMEM;
4019                 }
4020
4021                 mc_qht_elem->mc_grp_ctx.mg_id = (u16)mgn;
4022                 memcpy(mc_qht_elem->mc_grp_ctx.dest_ip_addr, ip_addr,
4023                        sizeof(mc_qht_elem->mc_grp_ctx.dest_ip_addr));
4024                 mc_qht_elem->mc_grp_ctx.ipv4_valid = ipv4;
4025                 mc_qht_elem->mc_grp_ctx.vlan_id = vlan_id;
4026                 if (vlan_id < VLAN_N_VID)
4027                         mc_qht_elem->mc_grp_ctx.vlan_valid = true;
4028                 mc_qht_elem->mc_grp_ctx.hmc_fcn_id = iwdev->vsi.fcn_id;
4029                 mc_qht_elem->mc_grp_ctx.qs_handle =
4030                         iwqp->sc_qp.vsi->qos[iwqp->sc_qp.user_pri].qs_handle;
4031                 ether_addr_copy(mc_qht_elem->mc_grp_ctx.dest_mac_addr, dmac);
4032
4033                 spin_lock_irqsave(&rf->qh_list_lock, flags);
4034                 mcast_list_add(rf, mc_qht_elem);
4035         } else {
4036                 if (mc_qht_elem->mc_grp_ctx.no_of_mgs ==
4037                     IRDMA_MAX_MGS_PER_CTX) {
4038                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4039                         return -ENOMEM;
4040                 }
4041         }
4042
4043         mcg_info.qp_id = iwqp->ibqp.qp_num;
4044         no_mgs = mc_qht_elem->mc_grp_ctx.no_of_mgs;
4045         irdma_sc_add_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
4046         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4047
4048         /* Only if there is a change do we need to modify or create */
4049         if (!no_mgs) {
4050                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4051                                          IRDMA_OP_MC_CREATE);
4052         } else if (no_mgs != mc_qht_elem->mc_grp_ctx.no_of_mgs) {
4053                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4054                                          IRDMA_OP_MC_MODIFY);
4055         } else {
4056                 return 0;
4057         }
4058
4059         if (ret)
4060                 goto error;
4061
4062         return 0;
4063
4064 error:
4065         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
4066         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
4067                 mcast_list_del(mc_qht_elem);
4068                 dma_free_coherent(rf->hw.device,
4069                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.size,
4070                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.va,
4071                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.pa);
4072                 mc_qht_elem->mc_grp_ctx.dma_mem_mc.va = NULL;
4073                 irdma_free_rsrc(rf, rf->allocated_mcgs,
4074                                 mc_qht_elem->mc_grp_ctx.mg_id);
4075                 kfree(mc_qht_elem);
4076         }
4077
4078         return ret;
4079 }
4080
4081 /**
4082  * irdma_detach_mcast - detach a qp from a multicast group
4083  * @ibqp: ptr to qp
4084  * @ibgid: pointer to global ID
4085  * @lid: local ID
4086  *
4087  * returns error status
4088  */
4089 static int irdma_detach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
4090 {
4091         struct irdma_qp *iwqp = to_iwqp(ibqp);
4092         struct irdma_device *iwdev = iwqp->iwdev;
4093         struct irdma_pci_f *rf = iwdev->rf;
4094         u32 ip_addr[4] = {};
4095         struct mc_table_list *mc_qht_elem;
4096         struct irdma_mcast_grp_ctx_entry_info mcg_info = {};
4097         int ret;
4098         unsigned long flags;
4099         union irdma_sockaddr sgid_addr;
4100
4101         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
4102         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid))
4103                 irdma_copy_ip_ntohl(ip_addr,
4104                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4105         else
4106                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
4107
4108         spin_lock_irqsave(&rf->qh_list_lock, flags);
4109         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
4110         if (!mc_qht_elem) {
4111                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4112                 ibdev_dbg(&iwdev->ibdev,
4113                           "VERBS: address not found MCG\n");
4114                 return 0;
4115         }
4116
4117         mcg_info.qp_id = iwqp->ibqp.qp_num;
4118         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
4119         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
4120                 mcast_list_del(mc_qht_elem);
4121                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4122                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4123                                          IRDMA_OP_MC_DESTROY);
4124                 if (ret) {
4125                         ibdev_dbg(&iwdev->ibdev,
4126                                   "VERBS: failed MC_DESTROY MCG\n");
4127                         spin_lock_irqsave(&rf->qh_list_lock, flags);
4128                         mcast_list_add(rf, mc_qht_elem);
4129                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4130                         return -EAGAIN;
4131                 }
4132
4133                 dma_free_coherent(rf->hw.device,
4134                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.size,
4135                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.va,
4136                                   mc_qht_elem->mc_grp_ctx.dma_mem_mc.pa);
4137                 mc_qht_elem->mc_grp_ctx.dma_mem_mc.va = NULL;
4138                 irdma_free_rsrc(rf, rf->allocated_mcgs,
4139                                 mc_qht_elem->mc_grp_ctx.mg_id);
4140                 kfree(mc_qht_elem);
4141         } else {
4142                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
4143                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
4144                                          IRDMA_OP_MC_MODIFY);
4145                 if (ret) {
4146                         ibdev_dbg(&iwdev->ibdev,
4147                                   "VERBS: failed Modify MCG\n");
4148                         return ret;
4149                 }
4150         }
4151
4152         return 0;
4153 }
4154
4155 static int irdma_create_hw_ah(struct irdma_device *iwdev, struct irdma_ah *ah, bool sleep)
4156 {
4157         struct irdma_pci_f *rf = iwdev->rf;
4158         int err;
4159
4160         err = irdma_alloc_rsrc(rf, rf->allocated_ahs, rf->max_ah, &ah->sc_ah.ah_info.ah_idx,
4161                                &rf->next_ah);
4162         if (err)
4163                 return err;
4164
4165         err = irdma_ah_cqp_op(rf, &ah->sc_ah, IRDMA_OP_AH_CREATE, sleep,
4166                               irdma_gsi_ud_qp_ah_cb, &ah->sc_ah);
4167
4168         if (err) {
4169                 ibdev_dbg(&iwdev->ibdev, "VERBS: CQP-OP Create AH fail");
4170                 goto err_ah_create;
4171         }
4172
4173         if (!sleep) {
4174                 int cnt = CQP_COMPL_WAIT_TIME_MS * CQP_TIMEOUT_THRESHOLD;
4175
4176                 do {
4177                         irdma_cqp_ce_handler(rf, &rf->ccq.sc_cq);
4178                         mdelay(1);
4179                 } while (!ah->sc_ah.ah_info.ah_valid && --cnt);
4180
4181                 if (!cnt) {
4182                         ibdev_dbg(&iwdev->ibdev, "VERBS: CQP create AH timed out");
4183                         err = -ETIMEDOUT;
4184                         goto err_ah_create;
4185                 }
4186         }
4187         return 0;
4188
4189 err_ah_create:
4190         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_ahs, ah->sc_ah.ah_info.ah_idx);
4191
4192         return err;
4193 }
4194
4195 static int irdma_setup_ah(struct ib_ah *ibah, struct rdma_ah_init_attr *attr)
4196 {
4197         struct irdma_pd *pd = to_iwpd(ibah->pd);
4198         struct irdma_ah *ah = container_of(ibah, struct irdma_ah, ibah);
4199         struct rdma_ah_attr *ah_attr = attr->ah_attr;
4200         const struct ib_gid_attr *sgid_attr;
4201         struct irdma_device *iwdev = to_iwdev(ibah->pd->device);
4202         struct irdma_pci_f *rf = iwdev->rf;
4203         struct irdma_sc_ah *sc_ah;
4204         struct irdma_ah_info *ah_info;
4205         union irdma_sockaddr sgid_addr, dgid_addr;
4206         int err;
4207         u8 dmac[ETH_ALEN];
4208
4209         ah->pd = pd;
4210         sc_ah = &ah->sc_ah;
4211         sc_ah->ah_info.vsi = &iwdev->vsi;
4212         irdma_sc_init_ah(&rf->sc_dev, sc_ah);
4213         ah->sgid_index = ah_attr->grh.sgid_index;
4214         sgid_attr = ah_attr->grh.sgid_attr;
4215         memcpy(&ah->dgid, &ah_attr->grh.dgid, sizeof(ah->dgid));
4216         rdma_gid2ip((struct sockaddr *)&sgid_addr, &sgid_attr->gid);
4217         rdma_gid2ip((struct sockaddr *)&dgid_addr, &ah_attr->grh.dgid);
4218         ah->av.attrs = *ah_attr;
4219         ah->av.net_type = rdma_gid_attr_network_type(sgid_attr);
4220         ah_info = &sc_ah->ah_info;
4221         ah_info->pd_idx = pd->sc_pd.pd_id;
4222         if (ah_attr->ah_flags & IB_AH_GRH) {
4223                 ah_info->flow_label = ah_attr->grh.flow_label;
4224                 ah_info->hop_ttl = ah_attr->grh.hop_limit;
4225                 ah_info->tc_tos = ah_attr->grh.traffic_class;
4226         }
4227
4228         ether_addr_copy(dmac, ah_attr->roce.dmac);
4229         if (ah->av.net_type == RDMA_NETWORK_IPV4) {
4230                 ah_info->ipv4_valid = true;
4231                 ah_info->dest_ip_addr[0] =
4232                         ntohl(dgid_addr.saddr_in.sin_addr.s_addr);
4233                 ah_info->src_ip_addr[0] =
4234                         ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
4235                 ah_info->do_lpbk = irdma_ipv4_is_lpb(ah_info->src_ip_addr[0],
4236                                                      ah_info->dest_ip_addr[0]);
4237                 if (ipv4_is_multicast(dgid_addr.saddr_in.sin_addr.s_addr)) {
4238                         ah_info->do_lpbk = true;
4239                         irdma_mcast_mac(ah_info->dest_ip_addr, dmac, true);
4240                 }
4241         } else {
4242                 irdma_copy_ip_ntohl(ah_info->dest_ip_addr,
4243                                     dgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4244                 irdma_copy_ip_ntohl(ah_info->src_ip_addr,
4245                                     sgid_addr.saddr_in6.sin6_addr.in6_u.u6_addr32);
4246                 ah_info->do_lpbk = irdma_ipv6_is_lpb(ah_info->src_ip_addr,
4247                                                      ah_info->dest_ip_addr);
4248                 if (rdma_is_multicast_addr(&dgid_addr.saddr_in6.sin6_addr)) {
4249                         ah_info->do_lpbk = true;
4250                         irdma_mcast_mac(ah_info->dest_ip_addr, dmac, false);
4251                 }
4252         }
4253
4254         err = rdma_read_gid_l2_fields(sgid_attr, &ah_info->vlan_tag,
4255                                       ah_info->mac_addr);
4256         if (err)
4257                 return err;
4258
4259         ah_info->dst_arpindex = irdma_add_arp(iwdev->rf, ah_info->dest_ip_addr,
4260                                               ah_info->ipv4_valid, dmac);
4261
4262         if (ah_info->dst_arpindex == -1)
4263                 return -EINVAL;
4264
4265         if (ah_info->vlan_tag >= VLAN_N_VID && iwdev->dcb_vlan_mode)
4266                 ah_info->vlan_tag = 0;
4267
4268         if (ah_info->vlan_tag < VLAN_N_VID) {
4269                 ah_info->insert_vlan_tag = true;
4270                 ah_info->vlan_tag |=
4271                         rt_tos2priority(ah_info->tc_tos) << VLAN_PRIO_SHIFT;
4272         }
4273
4274         return 0;
4275 }
4276
4277 /**
4278  * irdma_ah_exists - Check for existing identical AH
4279  * @iwdev: irdma device
4280  * @new_ah: AH to check for
4281  *
4282  * returns true if AH is found, false if not found.
4283  */
4284 static bool irdma_ah_exists(struct irdma_device *iwdev,
4285                             struct irdma_ah *new_ah)
4286 {
4287         struct irdma_ah *ah;
4288         u32 key = new_ah->sc_ah.ah_info.dest_ip_addr[0] ^
4289                   new_ah->sc_ah.ah_info.dest_ip_addr[1] ^
4290                   new_ah->sc_ah.ah_info.dest_ip_addr[2] ^
4291                   new_ah->sc_ah.ah_info.dest_ip_addr[3];
4292
4293         hash_for_each_possible(iwdev->ah_hash_tbl, ah, list, key) {
4294                 /* Set ah_valid and ah_id the same so memcmp can work */
4295                 new_ah->sc_ah.ah_info.ah_idx = ah->sc_ah.ah_info.ah_idx;
4296                 new_ah->sc_ah.ah_info.ah_valid = ah->sc_ah.ah_info.ah_valid;
4297                 if (!memcmp(&ah->sc_ah.ah_info, &new_ah->sc_ah.ah_info,
4298                             sizeof(ah->sc_ah.ah_info))) {
4299                         refcount_inc(&ah->refcnt);
4300                         new_ah->parent_ah = ah;
4301                         return true;
4302                 }
4303         }
4304
4305         return false;
4306 }
4307
4308 /**
4309  * irdma_destroy_ah - Destroy address handle
4310  * @ibah: pointer to address handle
4311  * @ah_flags: flags for sleepable
4312  */
4313 static int irdma_destroy_ah(struct ib_ah *ibah, u32 ah_flags)
4314 {
4315         struct irdma_device *iwdev = to_iwdev(ibah->device);
4316         struct irdma_ah *ah = to_iwah(ibah);
4317
4318         if ((ah_flags & RDMA_DESTROY_AH_SLEEPABLE) && ah->parent_ah) {
4319                 mutex_lock(&iwdev->ah_tbl_lock);
4320                 if (!refcount_dec_and_test(&ah->parent_ah->refcnt)) {
4321                         mutex_unlock(&iwdev->ah_tbl_lock);
4322                         return 0;
4323                 }
4324                 hash_del(&ah->parent_ah->list);
4325                 kfree(ah->parent_ah);
4326                 mutex_unlock(&iwdev->ah_tbl_lock);
4327         }
4328
4329         irdma_ah_cqp_op(iwdev->rf, &ah->sc_ah, IRDMA_OP_AH_DESTROY,
4330                         false, NULL, ah);
4331
4332         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_ahs,
4333                         ah->sc_ah.ah_info.ah_idx);
4334
4335         return 0;
4336 }
4337
4338 /**
4339  * irdma_create_user_ah - create user address handle
4340  * @ibah: address handle
4341  * @attr: address handle attributes
4342  * @udata: User data
4343  *
4344  * returns 0 on success, error otherwise
4345  */
4346 static int irdma_create_user_ah(struct ib_ah *ibah,
4347                                 struct rdma_ah_init_attr *attr,
4348                                 struct ib_udata *udata)
4349 {
4350 #define IRDMA_CREATE_AH_MIN_RESP_LEN offsetofend(struct irdma_create_ah_resp, rsvd)
4351         struct irdma_ah *ah = container_of(ibah, struct irdma_ah, ibah);
4352         struct irdma_device *iwdev = to_iwdev(ibah->pd->device);
4353         struct irdma_create_ah_resp uresp;
4354         struct irdma_ah *parent_ah;
4355         int err;
4356
4357         if (udata && udata->outlen < IRDMA_CREATE_AH_MIN_RESP_LEN)
4358                 return -EINVAL;
4359
4360         err = irdma_setup_ah(ibah, attr);
4361         if (err)
4362                 return err;
4363         mutex_lock(&iwdev->ah_tbl_lock);
4364         if (!irdma_ah_exists(iwdev, ah)) {
4365                 err = irdma_create_hw_ah(iwdev, ah, true);
4366                 if (err) {
4367                         mutex_unlock(&iwdev->ah_tbl_lock);
4368                         return err;
4369                 }
4370                 /* Add new AH to list */
4371                 parent_ah = kmemdup(ah, sizeof(*ah), GFP_KERNEL);
4372                 if (parent_ah) {
4373                         u32 key = parent_ah->sc_ah.ah_info.dest_ip_addr[0] ^
4374                                   parent_ah->sc_ah.ah_info.dest_ip_addr[1] ^
4375                                   parent_ah->sc_ah.ah_info.dest_ip_addr[2] ^
4376                                   parent_ah->sc_ah.ah_info.dest_ip_addr[3];
4377
4378                         ah->parent_ah = parent_ah;
4379                         hash_add(iwdev->ah_hash_tbl, &parent_ah->list, key);
4380                         refcount_set(&parent_ah->refcnt, 1);
4381                 }
4382         }
4383         mutex_unlock(&iwdev->ah_tbl_lock);
4384
4385         uresp.ah_id = ah->sc_ah.ah_info.ah_idx;
4386         err = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp), udata->outlen));
4387         if (err)
4388                 irdma_destroy_ah(ibah, attr->flags);
4389
4390         return err;
4391 }
4392
4393 /**
4394  * irdma_create_ah - create address handle
4395  * @ibah: address handle
4396  * @attr: address handle attributes
4397  * @udata: NULL
4398  *
4399  * returns 0 on success, error otherwise
4400  */
4401 static int irdma_create_ah(struct ib_ah *ibah, struct rdma_ah_init_attr *attr,
4402                            struct ib_udata *udata)
4403 {
4404         struct irdma_ah *ah = container_of(ibah, struct irdma_ah, ibah);
4405         struct irdma_device *iwdev = to_iwdev(ibah->pd->device);
4406         int err;
4407
4408         err = irdma_setup_ah(ibah, attr);
4409         if (err)
4410                 return err;
4411         err = irdma_create_hw_ah(iwdev, ah, attr->flags & RDMA_CREATE_AH_SLEEPABLE);
4412
4413         return err;
4414 }
4415
4416 /**
4417  * irdma_query_ah - Query address handle
4418  * @ibah: pointer to address handle
4419  * @ah_attr: address handle attributes
4420  */
4421 static int irdma_query_ah(struct ib_ah *ibah, struct rdma_ah_attr *ah_attr)
4422 {
4423         struct irdma_ah *ah = to_iwah(ibah);
4424
4425         memset(ah_attr, 0, sizeof(*ah_attr));
4426         if (ah->av.attrs.ah_flags & IB_AH_GRH) {
4427                 ah_attr->ah_flags = IB_AH_GRH;
4428                 ah_attr->grh.flow_label = ah->sc_ah.ah_info.flow_label;
4429                 ah_attr->grh.traffic_class = ah->sc_ah.ah_info.tc_tos;
4430                 ah_attr->grh.hop_limit = ah->sc_ah.ah_info.hop_ttl;
4431                 ah_attr->grh.sgid_index = ah->sgid_index;
4432                 ah_attr->grh.sgid_index = ah->sgid_index;
4433                 memcpy(&ah_attr->grh.dgid, &ah->dgid,
4434                        sizeof(ah_attr->grh.dgid));
4435         }
4436
4437         return 0;
4438 }
4439
4440 static enum rdma_link_layer irdma_get_link_layer(struct ib_device *ibdev,
4441                                                  u32 port_num)
4442 {
4443         return IB_LINK_LAYER_ETHERNET;
4444 }
4445
4446 static const struct ib_device_ops irdma_roce_dev_ops = {
4447         .attach_mcast = irdma_attach_mcast,
4448         .create_ah = irdma_create_ah,
4449         .create_user_ah = irdma_create_user_ah,
4450         .destroy_ah = irdma_destroy_ah,
4451         .detach_mcast = irdma_detach_mcast,
4452         .get_link_layer = irdma_get_link_layer,
4453         .get_port_immutable = irdma_roce_port_immutable,
4454         .modify_qp = irdma_modify_qp_roce,
4455         .query_ah = irdma_query_ah,
4456         .query_pkey = irdma_query_pkey,
4457 };
4458
4459 static const struct ib_device_ops irdma_iw_dev_ops = {
4460         .modify_qp = irdma_modify_qp,
4461         .get_port_immutable = irdma_iw_port_immutable,
4462         .query_gid = irdma_query_gid,
4463 };
4464
4465 static const struct ib_device_ops irdma_dev_ops = {
4466         .owner = THIS_MODULE,
4467         .driver_id = RDMA_DRIVER_IRDMA,
4468         .uverbs_abi_ver = IRDMA_ABI_VER,
4469
4470         .alloc_hw_port_stats = irdma_alloc_hw_port_stats,
4471         .alloc_mr = irdma_alloc_mr,
4472         .alloc_mw = irdma_alloc_mw,
4473         .alloc_pd = irdma_alloc_pd,
4474         .alloc_ucontext = irdma_alloc_ucontext,
4475         .create_cq = irdma_create_cq,
4476         .create_qp = irdma_create_qp,
4477         .dealloc_driver = irdma_ib_dealloc_device,
4478         .dealloc_mw = irdma_dealloc_mw,
4479         .dealloc_pd = irdma_dealloc_pd,
4480         .dealloc_ucontext = irdma_dealloc_ucontext,
4481         .dereg_mr = irdma_dereg_mr,
4482         .destroy_cq = irdma_destroy_cq,
4483         .destroy_qp = irdma_destroy_qp,
4484         .disassociate_ucontext = irdma_disassociate_ucontext,
4485         .get_dev_fw_str = irdma_get_dev_fw_str,
4486         .get_dma_mr = irdma_get_dma_mr,
4487         .get_hw_stats = irdma_get_hw_stats,
4488         .map_mr_sg = irdma_map_mr_sg,
4489         .mmap = irdma_mmap,
4490         .mmap_free = irdma_mmap_free,
4491         .poll_cq = irdma_poll_cq,
4492         .post_recv = irdma_post_recv,
4493         .post_send = irdma_post_send,
4494         .query_device = irdma_query_device,
4495         .query_port = irdma_query_port,
4496         .query_qp = irdma_query_qp,
4497         .reg_user_mr = irdma_reg_user_mr,
4498         .req_notify_cq = irdma_req_notify_cq,
4499         .resize_cq = irdma_resize_cq,
4500         INIT_RDMA_OBJ_SIZE(ib_pd, irdma_pd, ibpd),
4501         INIT_RDMA_OBJ_SIZE(ib_ucontext, irdma_ucontext, ibucontext),
4502         INIT_RDMA_OBJ_SIZE(ib_ah, irdma_ah, ibah),
4503         INIT_RDMA_OBJ_SIZE(ib_cq, irdma_cq, ibcq),
4504         INIT_RDMA_OBJ_SIZE(ib_mw, irdma_mr, ibmw),
4505         INIT_RDMA_OBJ_SIZE(ib_qp, irdma_qp, ibqp),
4506 };
4507
4508 /**
4509  * irdma_init_roce_device - initialization of roce rdma device
4510  * @iwdev: irdma device
4511  */
4512 static void irdma_init_roce_device(struct irdma_device *iwdev)
4513 {
4514         iwdev->ibdev.node_type = RDMA_NODE_IB_CA;
4515         addrconf_addr_eui48((u8 *)&iwdev->ibdev.node_guid,
4516                             iwdev->netdev->dev_addr);
4517         ib_set_device_ops(&iwdev->ibdev, &irdma_roce_dev_ops);
4518 }
4519
4520 /**
4521  * irdma_init_iw_device - initialization of iwarp rdma device
4522  * @iwdev: irdma device
4523  */
4524 static int irdma_init_iw_device(struct irdma_device *iwdev)
4525 {
4526         struct net_device *netdev = iwdev->netdev;
4527
4528         iwdev->ibdev.node_type = RDMA_NODE_RNIC;
4529         addrconf_addr_eui48((u8 *)&iwdev->ibdev.node_guid,
4530                             netdev->dev_addr);
4531         iwdev->ibdev.ops.iw_add_ref = irdma_qp_add_ref;
4532         iwdev->ibdev.ops.iw_rem_ref = irdma_qp_rem_ref;
4533         iwdev->ibdev.ops.iw_get_qp = irdma_get_qp;
4534         iwdev->ibdev.ops.iw_connect = irdma_connect;
4535         iwdev->ibdev.ops.iw_accept = irdma_accept;
4536         iwdev->ibdev.ops.iw_reject = irdma_reject;
4537         iwdev->ibdev.ops.iw_create_listen = irdma_create_listen;
4538         iwdev->ibdev.ops.iw_destroy_listen = irdma_destroy_listen;
4539         memcpy(iwdev->ibdev.iw_ifname, netdev->name,
4540                sizeof(iwdev->ibdev.iw_ifname));
4541         ib_set_device_ops(&iwdev->ibdev, &irdma_iw_dev_ops);
4542
4543         return 0;
4544 }
4545
4546 /**
4547  * irdma_init_rdma_device - initialization of rdma device
4548  * @iwdev: irdma device
4549  */
4550 static int irdma_init_rdma_device(struct irdma_device *iwdev)
4551 {
4552         struct pci_dev *pcidev = iwdev->rf->pcidev;
4553         int ret;
4554
4555         if (iwdev->roce_mode) {
4556                 irdma_init_roce_device(iwdev);
4557         } else {
4558                 ret = irdma_init_iw_device(iwdev);
4559                 if (ret)
4560                         return ret;
4561         }
4562         iwdev->ibdev.phys_port_cnt = 1;
4563         iwdev->ibdev.num_comp_vectors = iwdev->rf->ceqs_count;
4564         iwdev->ibdev.dev.parent = &pcidev->dev;
4565         ib_set_device_ops(&iwdev->ibdev, &irdma_dev_ops);
4566
4567         return 0;
4568 }
4569
4570 /**
4571  * irdma_port_ibevent - indicate port event
4572  * @iwdev: irdma device
4573  */
4574 void irdma_port_ibevent(struct irdma_device *iwdev)
4575 {
4576         struct ib_event event;
4577
4578         event.device = &iwdev->ibdev;
4579         event.element.port_num = 1;
4580         event.event =
4581                 iwdev->iw_status ? IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
4582         ib_dispatch_event(&event);
4583 }
4584
4585 /**
4586  * irdma_ib_unregister_device - unregister rdma device from IB
4587  * core
4588  * @iwdev: irdma device
4589  */
4590 void irdma_ib_unregister_device(struct irdma_device *iwdev)
4591 {
4592         iwdev->iw_status = 0;
4593         irdma_port_ibevent(iwdev);
4594         ib_unregister_device(&iwdev->ibdev);
4595 }
4596
4597 /**
4598  * irdma_ib_register_device - register irdma device to IB core
4599  * @iwdev: irdma device
4600  */
4601 int irdma_ib_register_device(struct irdma_device *iwdev)
4602 {
4603         int ret;
4604
4605         ret = irdma_init_rdma_device(iwdev);
4606         if (ret)
4607                 return ret;
4608
4609         ret = ib_device_set_netdev(&iwdev->ibdev, iwdev->netdev, 1);
4610         if (ret)
4611                 goto error;
4612         dma_set_max_seg_size(iwdev->rf->hw.device, UINT_MAX);
4613         ret = ib_register_device(&iwdev->ibdev, "irdma%d", iwdev->rf->hw.device);
4614         if (ret)
4615                 goto error;
4616
4617         iwdev->iw_status = 1;
4618         irdma_port_ibevent(iwdev);
4619
4620         return 0;
4621
4622 error:
4623         if (ret)
4624                 ibdev_dbg(&iwdev->ibdev, "VERBS: Register RDMA device fail\n");
4625
4626         return ret;
4627 }
4628
4629 /**
4630  * irdma_ib_dealloc_device
4631  * @ibdev: ib device
4632  *
4633  * callback from ibdev dealloc_driver to deallocate resources
4634  * unber irdma device
4635  */
4636 void irdma_ib_dealloc_device(struct ib_device *ibdev)
4637 {
4638         struct irdma_device *iwdev = to_iwdev(ibdev);
4639
4640         irdma_rt_deinit_hw(iwdev);
4641         irdma_ctrl_deinit_hw(iwdev->rf);
4642         kfree(iwdev->rf);
4643 }