sys/dev/irdma/irdma_verbs.c

   1 /*-
   2  * SPDX-License-Identifier: GPL-2.0 or Linux-OpenIB
   3  *
   4  * Copyright (c) 2015 - 2022 Intel Corporation
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenFabrics.org BSD license below:
  11  *
  12  *   Redistribution and use in source and binary forms, with or
  13  *   without modification, are permitted provided that the following
  14  *   conditions are met:
  15  *
  16  *    - Redistributions of source code must retain the above
  17  *      copyright notice, this list of conditions and the following
  18  *      disclaimer.
  19  *
  20  *    - Redistributions in binary form must reproduce the above
  21  *      copyright notice, this list of conditions and the following
  22  *      disclaimer in the documentation and/or other materials
  23  *      provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34 /*$FreeBSD$*/
  35
  36 #include "irdma_main.h"
  37
  38 /**
  39  * irdma_query_device - get device attributes
  40  * @ibdev: device pointer from stack
  41  * @props: returning device attributes
  42  * @udata: user data
  43  */
  44 static int
  45 irdma_query_device(struct ib_device *ibdev,
  46                    struct ib_device_attr *props,
  47                    struct ib_udata *udata)
  48 {
  49         struct irdma_device *iwdev = to_iwdev(ibdev);
  50         struct irdma_pci_f *rf = iwdev->rf;
  51         struct pci_dev *pcidev = iwdev->rf->pcidev;
  52         struct irdma_hw_attrs *hw_attrs = &rf->sc_dev.hw_attrs;
  53
  54         if (udata->inlen || udata->outlen)
  55                 return -EINVAL;
  56
  57         memset(props, 0, sizeof(*props));
  58         ether_addr_copy((u8 *)&props->sys_image_guid, IF_LLADDR(iwdev->netdev));
  59         props->fw_ver = (u64)irdma_fw_major_ver(&rf->sc_dev) << 32 |
  60             irdma_fw_minor_ver(&rf->sc_dev);
  61         props->device_cap_flags = IB_DEVICE_MEM_WINDOW |
  62             IB_DEVICE_MEM_MGT_EXTENSIONS;
  63         props->device_cap_flags |= IB_DEVICE_LOCAL_DMA_LKEY;
  64         props->vendor_id = pcidev->vendor;
  65         props->vendor_part_id = pcidev->device;
  66         props->hw_ver = pcidev->revision;
  67         props->page_size_cap = hw_attrs->page_size_cap;
  68         props->max_mr_size = hw_attrs->max_mr_size;
  69         props->max_qp = rf->max_qp - rf->used_qps;
  70         props->max_qp_wr = hw_attrs->max_qp_wr;
  71         set_max_sge(props, rf);
  72         props->max_cq = rf->max_cq - rf->used_cqs;
  73         props->max_cqe = rf->max_cqe - 1;
  74         props->max_mr = rf->max_mr - rf->used_mrs;
  75         props->max_mw = props->max_mr;
  76         props->max_pd = rf->max_pd - rf->used_pds;
  77         props->max_sge_rd = hw_attrs->uk_attrs.max_hw_read_sges;
  78         props->max_qp_rd_atom = hw_attrs->max_hw_ird;
  79         props->max_qp_init_rd_atom = hw_attrs->max_hw_ord;
  80         if (rdma_protocol_roce(ibdev, 1)) {
  81                 props->device_cap_flags |= IB_DEVICE_RC_RNR_NAK_GEN;
  82                 props->max_pkeys = IRDMA_PKEY_TBL_SZ;
  83                 props->max_ah = rf->max_ah;
  84                 if (hw_attrs->uk_attrs.hw_rev == IRDMA_GEN_2) {
  85                         props->max_mcast_grp = rf->max_mcg;
  86                         props->max_mcast_qp_attach = IRDMA_MAX_MGS_PER_CTX;
  87                         props->max_total_mcast_qp_attach = rf->max_qp * IRDMA_MAX_MGS_PER_CTX;
  88                 }
  89         }
  90         props->max_fast_reg_page_list_len = IRDMA_MAX_PAGES_PER_FMR;
  91         if (hw_attrs->uk_attrs.hw_rev >= IRDMA_GEN_2)
  92                 props->device_cap_flags |= IB_DEVICE_MEM_WINDOW_TYPE_2B;
  93
  94         return 0;
  95 }
  96
  97 static int
  98 irdma_mmap_legacy(struct irdma_ucontext *ucontext,
  99                   struct vm_area_struct *vma)
 100 {
 101         u64 pfn;
 102
 103         if (vma->vm_pgoff || vma->vm_end - vma->vm_start != PAGE_SIZE)
 104                 return -EINVAL;
 105
 106         vma->vm_private_data = ucontext;
 107         pfn = ((uintptr_t)ucontext->iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET] +
 108                pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;
 109
 110         return rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn, PAGE_SIZE,
 111                                  pgprot_noncached(vma->vm_page_prot), NULL);
 112 }
 113
 114 static void
 115 irdma_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
 116 {
 117         struct irdma_user_mmap_entry *entry = to_irdma_mmap_entry(rdma_entry);
 118
 119         kfree(entry);
 120 }
 121
 122 struct rdma_user_mmap_entry *
 123 irdma_user_mmap_entry_insert(struct irdma_ucontext *ucontext, u64 bar_offset,
 124                              enum irdma_mmap_flag mmap_flag, u64 *mmap_offset)
 125 {
 126         struct irdma_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 127         int ret;
 128
 129         if (!entry)
 130                 return NULL;
 131
 132         entry->bar_offset = bar_offset;
 133         entry->mmap_flag = mmap_flag;
 134
 135         ret = rdma_user_mmap_entry_insert(&ucontext->ibucontext,
 136                                           &entry->rdma_entry, PAGE_SIZE);
 137         if (ret) {
 138                 kfree(entry);
 139                 return NULL;
 140         }
 141         *mmap_offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
 142
 143         return &entry->rdma_entry;
 144 }
 145
 146 /**
 147  * irdma_mmap - user memory map
 148  * @context: context created during alloc
 149  * @vma: kernel info for user memory map
 150  */
 151 static int
 152 irdma_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
 153 {
 154         struct rdma_user_mmap_entry *rdma_entry;
 155         struct irdma_user_mmap_entry *entry;
 156         struct irdma_ucontext *ucontext;
 157         u64 pfn;
 158         int ret;
 159
 160         ucontext = to_ucontext(context);
 161
 162         /* Legacy support for libi40iw with hard-coded mmap key */
 163         if (ucontext->legacy_mode)
 164                 return irdma_mmap_legacy(ucontext, vma);
 165
 166         rdma_entry = rdma_user_mmap_entry_get(&ucontext->ibucontext, vma);
 167         if (!rdma_entry) {
 168                 irdma_debug(iwdev_to_idev(ucontext->iwdev), IRDMA_DEBUG_VERBS,
 169                             "pgoff[0x%lx] does not have valid entry\n",
 170                             vma->vm_pgoff);
 171                 return -EINVAL;
 172         }
 173
 174         entry = to_irdma_mmap_entry(rdma_entry);
 175         irdma_debug(iwdev_to_idev(ucontext->iwdev), IRDMA_DEBUG_VERBS,
 176                     "bar_offset [0x%lx] mmap_flag [%d]\n", entry->bar_offset,
 177                     entry->mmap_flag);
 178
 179         pfn = (entry->bar_offset +
 180                pci_resource_start(ucontext->iwdev->rf->pcidev, 0)) >> PAGE_SHIFT;
 181
 182         switch (entry->mmap_flag) {
 183         case IRDMA_MMAP_IO_NC:
 184                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
 185                                         pgprot_noncached(vma->vm_page_prot),
 186                                         rdma_entry);
 187                 break;
 188         case IRDMA_MMAP_IO_WC:
 189                 ret = rdma_user_mmap_io(context, vma, pfn, PAGE_SIZE,
 190                                         pgprot_writecombine(vma->vm_page_prot),
 191                                         rdma_entry);
 192                 break;
 193         default:
 194                 ret = -EINVAL;
 195         }
 196
 197         if (ret)
 198                 irdma_debug(iwdev_to_idev(ucontext->iwdev), IRDMA_DEBUG_VERBS,
 199                             "bar_offset [0x%lx] mmap_flag[%d] err[%d]\n",
 200                             entry->bar_offset, entry->mmap_flag, ret);
 201         rdma_user_mmap_entry_put(rdma_entry);
 202
 203         return ret;
 204 }
 205
 206 /**
 207  * irdma_alloc_push_page - allocate a push page for qp
 208  * @iwqp: qp pointer
 209  */
 210 static void
 211 irdma_alloc_push_page(struct irdma_qp *iwqp)
 212 {
 213         struct irdma_cqp_request *cqp_request;
 214         struct cqp_cmds_info *cqp_info;
 215         struct irdma_device *iwdev = iwqp->iwdev;
 216         struct irdma_sc_qp *qp = &iwqp->sc_qp;
 217         int status;
 218
 219         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
 220         if (!cqp_request)
 221                 return;
 222
 223         cqp_info = &cqp_request->info;
 224         cqp_info->cqp_cmd = IRDMA_OP_MANAGE_PUSH_PAGE;
 225         cqp_info->post_sq = 1;
 226         cqp_info->in.u.manage_push_page.info.push_idx = 0;
 227         cqp_info->in.u.manage_push_page.info.qs_handle =
 228             qp->vsi->qos[qp->user_pri].qs_handle;
 229         cqp_info->in.u.manage_push_page.info.free_page = 0;
 230         cqp_info->in.u.manage_push_page.info.push_page_type = 0;
 231         cqp_info->in.u.manage_push_page.cqp = &iwdev->rf->cqp.sc_cqp;
 232         cqp_info->in.u.manage_push_page.scratch = (uintptr_t)cqp_request;
 233
 234         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
 235         if (!status && cqp_request->compl_info.op_ret_val <
 236             iwdev->rf->sc_dev.hw_attrs.max_hw_device_pages) {
 237                 qp->push_idx = cqp_request->compl_info.op_ret_val;
 238                 qp->push_offset = 0;
 239         }
 240
 241         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
 242 }
 243
 244 /**
 245  * irdma_get_pbl - Retrieve pbl from a list given a virtual
 246  * address
 247  * @va: user virtual address
 248  * @pbl_list: pbl list to search in (QP's or CQ's)
 249  */
 250 struct irdma_pbl *
 251 irdma_get_pbl(unsigned long va,
 252               struct list_head *pbl_list)
 253 {
 254         struct irdma_pbl *iwpbl;
 255
 256         list_for_each_entry(iwpbl, pbl_list, list) {
 257                 if (iwpbl->user_base == va) {
 258                         list_del(&iwpbl->list);
 259                         iwpbl->on_list = false;
 260                         return iwpbl;
 261                 }
 262         }
 263
 264         return NULL;
 265 }
 266
 267 /**
 268  * irdma_clean_cqes - clean cq entries for qp
 269  * @iwqp: qp ptr (user or kernel)
 270  * @iwcq: cq ptr
 271  */
 272 void
 273 irdma_clean_cqes(struct irdma_qp *iwqp, struct irdma_cq *iwcq)
 274 {
 275         struct irdma_cq_uk *ukcq = &iwcq->sc_cq.cq_uk;
 276         unsigned long flags;
 277
 278         spin_lock_irqsave(&iwcq->lock, flags);
 279         irdma_uk_clean_cq(&iwqp->sc_qp.qp_uk, ukcq);
 280         spin_unlock_irqrestore(&iwcq->lock, flags);
 281 }
 282
 283 static u64 irdma_compute_push_wqe_offset(struct irdma_device *iwdev, u32 page_idx){
 284         u64 bar_off = (uintptr_t)iwdev->rf->sc_dev.hw_regs[IRDMA_DB_ADDR_OFFSET];
 285
 286         if (iwdev->rf->sc_dev.hw_attrs.uk_attrs.hw_rev == IRDMA_GEN_2) {
 287                 /* skip over db page */
 288                 bar_off += IRDMA_HW_PAGE_SIZE;
 289                 /* skip over reserved space */
 290                 bar_off += IRDMA_PF_BAR_RSVD;
 291         }
 292
 293         /* push wqe page */
 294         bar_off += (u64)page_idx * IRDMA_HW_PAGE_SIZE;
 295
 296         return bar_off;
 297 }
 298
 299 void
 300 irdma_remove_push_mmap_entries(struct irdma_qp *iwqp)
 301 {
 302         if (iwqp->push_db_mmap_entry) {
 303                 rdma_user_mmap_entry_remove(iwqp->push_db_mmap_entry);
 304                 iwqp->push_db_mmap_entry = NULL;
 305         }
 306         if (iwqp->push_wqe_mmap_entry) {
 307                 rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
 308                 iwqp->push_wqe_mmap_entry = NULL;
 309         }
 310 }
 311
 312 static int
 313 irdma_setup_push_mmap_entries(struct irdma_ucontext *ucontext,
 314                               struct irdma_qp *iwqp,
 315                               u64 *push_wqe_mmap_key,
 316                               u64 *push_db_mmap_key)
 317 {
 318         struct irdma_device *iwdev = ucontext->iwdev;
 319         u64 bar_off;
 320
 321         WARN_ON_ONCE(iwdev->rf->sc_dev.hw_attrs.uk_attrs.hw_rev < IRDMA_GEN_2);
 322
 323         bar_off = irdma_compute_push_wqe_offset(iwdev, iwqp->sc_qp.push_idx);
 324
 325         iwqp->push_wqe_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
 326                                                                  bar_off, IRDMA_MMAP_IO_WC,
 327                                                                  push_wqe_mmap_key);
 328         if (!iwqp->push_wqe_mmap_entry)
 329                 return -ENOMEM;
 330
 331         /* push doorbell page */
 332         bar_off += IRDMA_HW_PAGE_SIZE;
 333         iwqp->push_db_mmap_entry = irdma_user_mmap_entry_insert(ucontext,
 334                                                                 bar_off, IRDMA_MMAP_IO_NC,
 335                                                                 push_db_mmap_key);
 336         if (!iwqp->push_db_mmap_entry) {
 337                 rdma_user_mmap_entry_remove(iwqp->push_wqe_mmap_entry);
 338                 return -ENOMEM;
 339         }
 340
 341         return 0;
 342 }
 343
 344 /**
 345  * irdma_setup_virt_qp - setup for allocation of virtual qp
 346  * @iwdev: irdma device
 347  * @iwqp: qp ptr
 348  * @init_info: initialize info to return
 349  */
 350 void
 351 irdma_setup_virt_qp(struct irdma_device *iwdev,
 352                     struct irdma_qp *iwqp,
 353                     struct irdma_qp_init_info *init_info)
 354 {
 355         struct irdma_pbl *iwpbl = iwqp->iwpbl;
 356         struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;
 357
 358         iwqp->page = qpmr->sq_page;
 359         init_info->shadow_area_pa = qpmr->shadow;
 360         if (iwpbl->pbl_allocated) {
 361                 init_info->virtual_map = true;
 362                 init_info->sq_pa = qpmr->sq_pbl.idx;
 363                 init_info->rq_pa = qpmr->rq_pbl.idx;
 364         } else {
 365                 init_info->sq_pa = qpmr->sq_pbl.addr;
 366                 init_info->rq_pa = qpmr->rq_pbl.addr;
 367         }
 368 }
 369
 370 /**
 371  * irdma_setup_umode_qp - setup sq and rq size in user mode qp
 372  * @udata: user data
 373  * @iwdev: iwarp device
 374  * @iwqp: qp ptr (user or kernel)
 375  * @info: initialize info to return
 376  * @init_attr: Initial QP create attributes
 377  */
 378 int
 379 irdma_setup_umode_qp(struct ib_udata *udata,
 380                      struct irdma_device *iwdev,
 381                      struct irdma_qp *iwqp,
 382                      struct irdma_qp_init_info *info,
 383                      struct ib_qp_init_attr *init_attr)
 384 {
 385         struct irdma_qp_uk_init_info *ukinfo = &info->qp_uk_init_info;
 386         struct irdma_create_qp_req req = {0};
 387         unsigned long flags;
 388         int ret;
 389
 390         ret = ib_copy_from_udata(&req, udata,
 391                                  min(sizeof(req), udata->inlen));
 392         if (ret) {
 393                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
 394                             "ib_copy_from_data fail\n");
 395                 return ret;
 396         }
 397
 398         iwqp->ctx_info.qp_compl_ctx = req.user_compl_ctx;
 399         iwqp->user_mode = 1;
 400         if (req.user_wqe_bufs) {
 401                 struct irdma_ucontext *ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext, ibucontext);
 402
 403                 info->qp_uk_init_info.legacy_mode = ucontext->legacy_mode;
 404                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
 405                 iwqp->iwpbl = irdma_get_pbl((unsigned long)req.user_wqe_bufs,
 406                                             &ucontext->qp_reg_mem_list);
 407                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
 408
 409                 if (!iwqp->iwpbl) {
 410                         ret = -ENODATA;
 411                         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
 412                                     "no pbl info\n");
 413                         return ret;
 414                 }
 415         }
 416
 417         if (ukinfo->abi_ver <= 5) {
 418                 /**
 419                  * For ABI version less than 6 passes raw sq and rq
 420                  * quanta in cap.max_send_wr and cap.max_recv_wr.
 421                  */
 422                 iwqp->max_send_wr = init_attr->cap.max_send_wr;
 423                 iwqp->max_recv_wr = init_attr->cap.max_recv_wr;
 424                 ukinfo->sq_size = init_attr->cap.max_send_wr;
 425                 ukinfo->rq_size = init_attr->cap.max_recv_wr;
 426                 irdma_uk_calc_shift_wq(ukinfo, &ukinfo->sq_shift, &ukinfo->rq_shift);
 427         } else {
 428                 ret = irdma_uk_calc_depth_shift_sq(ukinfo, &ukinfo->sq_depth,
 429                                                    &ukinfo->sq_shift);
 430                 if (ret)
 431                         return ret;
 432
 433                 ret = irdma_uk_calc_depth_shift_rq(ukinfo, &ukinfo->rq_depth,
 434                                                    &ukinfo->rq_shift);
 435                 if (ret)
 436                         return ret;
 437
 438                 iwqp->max_send_wr = (ukinfo->sq_depth - IRDMA_SQ_RSVD) >> ukinfo->sq_shift;
 439                 iwqp->max_recv_wr = (ukinfo->rq_depth - IRDMA_RQ_RSVD) >> ukinfo->rq_shift;
 440                 ukinfo->sq_size = ukinfo->sq_depth >> ukinfo->sq_shift;
 441                 ukinfo->rq_size = ukinfo->rq_depth >> ukinfo->rq_shift;
 442         }
 443         irdma_setup_virt_qp(iwdev, iwqp, info);
 444
 445         return 0;
 446 }
 447
 448 /**
 449  * irdma_setup_kmode_qp - setup initialization for kernel mode qp
 450  * @iwdev: iwarp device
 451  * @iwqp: qp ptr (user or kernel)
 452  * @info: initialize info to return
 453  * @init_attr: Initial QP create attributes
 454  */
 455 int
 456 irdma_setup_kmode_qp(struct irdma_device *iwdev,
 457                      struct irdma_qp *iwqp,
 458                      struct irdma_qp_init_info *info,
 459                      struct ib_qp_init_attr *init_attr)
 460 {
 461         struct irdma_dma_mem *mem = &iwqp->kqp.dma_mem;
 462         u32 size;
 463         int status;
 464         struct irdma_qp_uk_init_info *ukinfo = &info->qp_uk_init_info;
 465
 466         status = irdma_uk_calc_depth_shift_sq(ukinfo, &ukinfo->sq_depth,
 467                                               &ukinfo->sq_shift);
 468         if (status)
 469                 return status;
 470
 471         status = irdma_uk_calc_depth_shift_rq(ukinfo, &ukinfo->rq_depth,
 472                                               &ukinfo->rq_shift);
 473         if (status)
 474                 return status;
 475
 476         iwqp->kqp.sq_wrid_mem =
 477             kcalloc(ukinfo->sq_depth, sizeof(*iwqp->kqp.sq_wrid_mem), GFP_KERNEL);
 478         if (!iwqp->kqp.sq_wrid_mem)
 479                 return -ENOMEM;
 480
 481         iwqp->kqp.rq_wrid_mem =
 482             kcalloc(ukinfo->rq_depth, sizeof(*iwqp->kqp.rq_wrid_mem), GFP_KERNEL);
 483         if (!iwqp->kqp.rq_wrid_mem) {
 484                 kfree(iwqp->kqp.sq_wrid_mem);
 485                 iwqp->kqp.sq_wrid_mem = NULL;
 486                 return -ENOMEM;
 487         }
 488
 489         iwqp->kqp.sig_trk_mem = kcalloc(ukinfo->sq_depth, sizeof(u32), GFP_KERNEL);
 490         memset(iwqp->kqp.sig_trk_mem, 0, ukinfo->sq_depth * sizeof(u32));
 491         if (!iwqp->kqp.sig_trk_mem) {
 492                 kfree(iwqp->kqp.sq_wrid_mem);
 493                 iwqp->kqp.sq_wrid_mem = NULL;
 494                 kfree(iwqp->kqp.rq_wrid_mem);
 495                 iwqp->kqp.rq_wrid_mem = NULL;
 496                 return -ENOMEM;
 497         }
 498         ukinfo->sq_sigwrtrk_array = (void *)iwqp->kqp.sig_trk_mem;
 499         ukinfo->sq_wrtrk_array = iwqp->kqp.sq_wrid_mem;
 500         ukinfo->rq_wrid_array = iwqp->kqp.rq_wrid_mem;
 501
 502         size = (ukinfo->sq_depth + ukinfo->rq_depth) * IRDMA_QP_WQE_MIN_SIZE;
 503         size += (IRDMA_SHADOW_AREA_SIZE << 3);
 504
 505         mem->size = size;
 506         mem->va = irdma_allocate_dma_mem(&iwdev->rf->hw, mem, mem->size,
 507                                          256);
 508         if (!mem->va) {
 509                 kfree(iwqp->kqp.sq_wrid_mem);
 510                 iwqp->kqp.sq_wrid_mem = NULL;
 511                 kfree(iwqp->kqp.rq_wrid_mem);
 512                 iwqp->kqp.rq_wrid_mem = NULL;
 513                 return -ENOMEM;
 514         }
 515
 516         ukinfo->sq = mem->va;
 517         info->sq_pa = mem->pa;
 518         ukinfo->rq = &ukinfo->sq[ukinfo->sq_depth];
 519         info->rq_pa = info->sq_pa + (ukinfo->sq_depth * IRDMA_QP_WQE_MIN_SIZE);
 520         ukinfo->shadow_area = ukinfo->rq[ukinfo->rq_depth].elem;
 521         info->shadow_area_pa = info->rq_pa + (ukinfo->rq_depth * IRDMA_QP_WQE_MIN_SIZE);
 522         ukinfo->sq_size = ukinfo->sq_depth >> ukinfo->sq_shift;
 523         ukinfo->rq_size = ukinfo->rq_depth >> ukinfo->rq_shift;
 524         ukinfo->qp_id = iwqp->ibqp.qp_num;
 525
 526         iwqp->max_send_wr = (ukinfo->sq_depth - IRDMA_SQ_RSVD) >> ukinfo->sq_shift;
 527         iwqp->max_recv_wr = (ukinfo->rq_depth - IRDMA_RQ_RSVD) >> ukinfo->rq_shift;
 528         init_attr->cap.max_send_wr = iwqp->max_send_wr;
 529         init_attr->cap.max_recv_wr = iwqp->max_recv_wr;
 530
 531         return 0;
 532 }
 533
 534 int
 535 irdma_cqp_create_qp_cmd(struct irdma_qp *iwqp)
 536 {
 537         struct irdma_pci_f *rf = iwqp->iwdev->rf;
 538         struct irdma_cqp_request *cqp_request;
 539         struct cqp_cmds_info *cqp_info;
 540         struct irdma_create_qp_info *qp_info;
 541         int status;
 542
 543         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
 544         if (!cqp_request)
 545                 return -ENOMEM;
 546
 547         cqp_info = &cqp_request->info;
 548         qp_info = &cqp_request->info.in.u.qp_create.info;
 549         memset(qp_info, 0, sizeof(*qp_info));
 550         qp_info->mac_valid = true;
 551         qp_info->cq_num_valid = true;
 552         qp_info->next_iwarp_state = IRDMA_QP_STATE_IDLE;
 553
 554         cqp_info->cqp_cmd = IRDMA_OP_QP_CREATE;
 555         cqp_info->post_sq = 1;
 556         cqp_info->in.u.qp_create.qp = &iwqp->sc_qp;
 557         cqp_info->in.u.qp_create.scratch = (uintptr_t)cqp_request;
 558         status = irdma_handle_cqp_op(rf, cqp_request);
 559         irdma_put_cqp_request(&rf->cqp, cqp_request);
 560
 561         return status;
 562 }
 563
 564 void
 565 irdma_roce_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
 566                                    struct irdma_qp_host_ctx_info *ctx_info)
 567 {
 568         struct irdma_device *iwdev = iwqp->iwdev;
 569         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 570         struct irdma_roce_offload_info *roce_info;
 571         struct irdma_udp_offload_info *udp_info;
 572
 573         udp_info = &iwqp->udp_info;
 574         udp_info->snd_mss = ib_mtu_enum_to_int(ib_mtu_int_to_enum(iwdev->vsi.mtu));
 575         udp_info->cwnd = iwdev->roce_cwnd;
 576         udp_info->rexmit_thresh = 2;
 577         udp_info->rnr_nak_thresh = 2;
 578         udp_info->src_port = 0xc000;
 579         udp_info->dst_port = ROCE_V2_UDP_DPORT;
 580         roce_info = &iwqp->roce_info;
 581         ether_addr_copy(roce_info->mac_addr, IF_LLADDR(iwdev->netdev));
 582
 583         roce_info->rd_en = true;
 584         roce_info->wr_rdresp_en = true;
 585         roce_info->bind_en = true;
 586         roce_info->dcqcn_en = false;
 587         roce_info->rtomin = 5;
 588
 589         roce_info->ack_credits = iwdev->roce_ackcreds;
 590         roce_info->ird_size = dev->hw_attrs.max_hw_ird;
 591         roce_info->ord_size = dev->hw_attrs.max_hw_ord;
 592
 593         if (!iwqp->user_mode) {
 594                 roce_info->priv_mode_en = true;
 595                 roce_info->fast_reg_en = true;
 596                 roce_info->udprivcq_en = true;
 597         }
 598         roce_info->roce_tver = 0;
 599
 600         ctx_info->roce_info = &iwqp->roce_info;
 601         ctx_info->udp_info = &iwqp->udp_info;
 602         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
 603 }
 604
 605 void
 606 irdma_iw_fill_and_set_qpctx_info(struct irdma_qp *iwqp,
 607                                  struct irdma_qp_host_ctx_info *ctx_info)
 608 {
 609         struct irdma_device *iwdev = iwqp->iwdev;
 610         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 611         struct irdma_iwarp_offload_info *iwarp_info;
 612
 613         iwarp_info = &iwqp->iwarp_info;
 614         ether_addr_copy(iwarp_info->mac_addr, IF_LLADDR(iwdev->netdev));
 615         iwarp_info->rd_en = true;
 616         iwarp_info->wr_rdresp_en = true;
 617         iwarp_info->bind_en = true;
 618         iwarp_info->ecn_en = true;
 619         iwarp_info->rtomin = 5;
 620
 621         if (dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
 622                 iwarp_info->ib_rd_en = true;
 623         if (!iwqp->user_mode) {
 624                 iwarp_info->priv_mode_en = true;
 625                 iwarp_info->fast_reg_en = true;
 626         }
 627         iwarp_info->ddp_ver = 1;
 628         iwarp_info->rdmap_ver = 1;
 629
 630         ctx_info->iwarp_info = &iwqp->iwarp_info;
 631         ctx_info->iwarp_info_valid = true;
 632         irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
 633         ctx_info->iwarp_info_valid = false;
 634 }
 635
 636 int
 637 irdma_validate_qp_attrs(struct ib_qp_init_attr *init_attr,
 638                         struct irdma_device *iwdev)
 639 {
 640         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 641         struct irdma_uk_attrs *uk_attrs = &dev->hw_attrs.uk_attrs;
 642
 643         if (init_attr->create_flags)
 644                 return -EOPNOTSUPP;
 645
 646         if (init_attr->cap.max_inline_data > uk_attrs->max_hw_inline ||
 647             init_attr->cap.max_send_sge > uk_attrs->max_hw_wq_frags ||
 648             init_attr->cap.max_recv_sge > uk_attrs->max_hw_wq_frags)
 649                 return -EINVAL;
 650
 651         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
 652                 if (init_attr->qp_type != IB_QPT_RC &&
 653                     init_attr->qp_type != IB_QPT_UD &&
 654                     init_attr->qp_type != IB_QPT_GSI)
 655                         return -EOPNOTSUPP;
 656         } else {
 657                 if (init_attr->qp_type != IB_QPT_RC)
 658                         return -EOPNOTSUPP;
 659         }
 660
 661         return 0;
 662 }
 663
 664 void
 665 irdma_sched_qp_flush_work(struct irdma_qp *iwqp)
 666 {
 667         irdma_qp_add_ref(&iwqp->ibqp);
 668         if (mod_delayed_work(iwqp->iwdev->cleanup_wq, &iwqp->dwork_flush,
 669                              msecs_to_jiffies(IRDMA_FLUSH_DELAY_MS)))
 670                 irdma_qp_rem_ref(&iwqp->ibqp);
 671 }
 672
 673 void
 674 irdma_flush_worker(struct work_struct *work)
 675 {
 676         struct delayed_work *dwork = to_delayed_work(work);
 677         struct irdma_qp *iwqp = container_of(dwork, struct irdma_qp, dwork_flush);
 678
 679         irdma_generate_flush_completions(iwqp);
 680         /* For the add in irdma_sched_qp_flush_work */
 681         irdma_qp_rem_ref(&iwqp->ibqp);
 682 }
 683
 684 static int
 685 irdma_get_ib_acc_flags(struct irdma_qp *iwqp)
 686 {
 687         int acc_flags = 0;
 688
 689         if (rdma_protocol_roce(iwqp->ibqp.device, 1)) {
 690                 if (iwqp->roce_info.wr_rdresp_en) {
 691                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
 692                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
 693                 }
 694                 if (iwqp->roce_info.rd_en)
 695                         acc_flags |= IB_ACCESS_REMOTE_READ;
 696                 if (iwqp->roce_info.bind_en)
 697                         acc_flags |= IB_ACCESS_MW_BIND;
 698         } else {
 699                 if (iwqp->iwarp_info.wr_rdresp_en) {
 700                         acc_flags |= IB_ACCESS_LOCAL_WRITE;
 701                         acc_flags |= IB_ACCESS_REMOTE_WRITE;
 702                 }
 703                 if (iwqp->iwarp_info.rd_en)
 704                         acc_flags |= IB_ACCESS_REMOTE_READ;
 705                 if (iwqp->iwarp_info.bind_en)
 706                         acc_flags |= IB_ACCESS_MW_BIND;
 707         }
 708         return acc_flags;
 709 }
 710
 711 /**
 712  * irdma_query_qp - query qp attributes
 713  * @ibqp: qp pointer
 714  * @attr: attributes pointer
 715  * @attr_mask: Not used
 716  * @init_attr: qp attributes to return
 717  */
 718 static int
 719 irdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 720                int attr_mask, struct ib_qp_init_attr *init_attr)
 721 {
 722         struct irdma_qp *iwqp = to_iwqp(ibqp);
 723         struct irdma_sc_qp *qp = &iwqp->sc_qp;
 724
 725         memset(attr, 0, sizeof(*attr));
 726         memset(init_attr, 0, sizeof(*init_attr));
 727
 728         attr->qp_state = iwqp->ibqp_state;
 729         attr->cur_qp_state = iwqp->ibqp_state;
 730         attr->cap.max_send_wr = iwqp->max_send_wr;
 731         attr->cap.max_recv_wr = iwqp->max_recv_wr;
 732         attr->cap.max_inline_data = qp->qp_uk.max_inline_data;
 733         attr->cap.max_send_sge = qp->qp_uk.max_sq_frag_cnt;
 734         attr->cap.max_recv_sge = qp->qp_uk.max_rq_frag_cnt;
 735         attr->qp_access_flags = irdma_get_ib_acc_flags(iwqp);
 736         attr->port_num = 1;
 737         if (rdma_protocol_roce(ibqp->device, 1)) {
 738                 attr->path_mtu = ib_mtu_int_to_enum(iwqp->udp_info.snd_mss);
 739                 attr->qkey = iwqp->roce_info.qkey;
 740                 attr->rq_psn = iwqp->udp_info.epsn;
 741                 attr->sq_psn = iwqp->udp_info.psn_nxt;
 742                 attr->dest_qp_num = iwqp->roce_info.dest_qp;
 743                 attr->pkey_index = iwqp->roce_info.p_key;
 744                 attr->retry_cnt = iwqp->udp_info.rexmit_thresh;
 745                 attr->rnr_retry = iwqp->udp_info.rnr_nak_thresh;
 746                 attr->max_rd_atomic = iwqp->roce_info.ord_size;
 747                 attr->max_dest_rd_atomic = iwqp->roce_info.ird_size;
 748         }
 749
 750         init_attr->event_handler = iwqp->ibqp.event_handler;
 751         init_attr->qp_context = iwqp->ibqp.qp_context;
 752         init_attr->send_cq = iwqp->ibqp.send_cq;
 753         init_attr->recv_cq = iwqp->ibqp.recv_cq;
 754         init_attr->cap = attr->cap;
 755
 756         return 0;
 757 }
 758
 759 /**
 760  * irdma_modify_qp_roce - modify qp request
 761  * @ibqp: qp's pointer for modify
 762  * @attr: access attributes
 763  * @attr_mask: state mask
 764  * @udata: user data
 765  */
 766 int
 767 irdma_modify_qp_roce(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 768                      int attr_mask, struct ib_udata *udata)
 769 {
 770 #define IRDMA_MODIFY_QP_MIN_REQ_LEN offsetofend(struct irdma_modify_qp_req, rq_flush)
 771 #define IRDMA_MODIFY_QP_MIN_RESP_LEN offsetofend(struct irdma_modify_qp_resp, push_valid)
 772         struct irdma_pd *iwpd = to_iwpd(ibqp->pd);
 773         struct irdma_qp *iwqp = to_iwqp(ibqp);
 774         struct irdma_device *iwdev = iwqp->iwdev;
 775         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
 776         struct irdma_qp_host_ctx_info *ctx_info;
 777         struct irdma_roce_offload_info *roce_info;
 778         struct irdma_udp_offload_info *udp_info;
 779         struct irdma_modify_qp_info info = {0};
 780         struct irdma_modify_qp_resp uresp = {};
 781         struct irdma_modify_qp_req ureq;
 782         unsigned long flags;
 783         u8 issue_modify_qp = 0;
 784         int ret = 0;
 785
 786         ctx_info = &iwqp->ctx_info;
 787         roce_info = &iwqp->roce_info;
 788         udp_info = &iwqp->udp_info;
 789
 790         if (udata) {
 791                 if ((udata->inlen && udata->inlen < IRDMA_MODIFY_QP_MIN_REQ_LEN) ||
 792                     (udata->outlen && udata->outlen < IRDMA_MODIFY_QP_MIN_RESP_LEN))
 793                         return -EINVAL;
 794         }
 795
 796         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
 797                 return -EOPNOTSUPP;
 798
 799         if (attr_mask & IB_QP_DEST_QPN)
 800                 roce_info->dest_qp = attr->dest_qp_num;
 801
 802         if (attr_mask & IB_QP_PKEY_INDEX) {
 803                 ret = irdma_query_pkey(ibqp->device, 0, attr->pkey_index,
 804                                        &roce_info->p_key);
 805                 if (ret)
 806                         return ret;
 807         }
 808
 809         if (attr_mask & IB_QP_QKEY)
 810                 roce_info->qkey = attr->qkey;
 811
 812         if (attr_mask & IB_QP_PATH_MTU)
 813                 udp_info->snd_mss = ib_mtu_enum_to_int(attr->path_mtu);
 814
 815         if (attr_mask & IB_QP_SQ_PSN) {
 816                 udp_info->psn_nxt = attr->sq_psn;
 817                 udp_info->lsn = 0xffff;
 818                 udp_info->psn_una = attr->sq_psn;
 819                 udp_info->psn_max = attr->sq_psn;
 820         }
 821
 822         if (attr_mask & IB_QP_RQ_PSN)
 823                 udp_info->epsn = attr->rq_psn;
 824
 825         if (attr_mask & IB_QP_RNR_RETRY)
 826                 udp_info->rnr_nak_thresh = attr->rnr_retry;
 827
 828         if (attr_mask & IB_QP_RETRY_CNT)
 829                 udp_info->rexmit_thresh = attr->retry_cnt;
 830
 831         ctx_info->roce_info->pd_id = iwpd->sc_pd.pd_id;
 832
 833         if (attr_mask & IB_QP_AV) {
 834                 struct irdma_av *av = &iwqp->roce_ah.av;
 835                 u16 vlan_id = VLAN_N_VID;
 836                 u32 local_ip[4] = {};
 837
 838                 memset(&iwqp->roce_ah, 0, sizeof(iwqp->roce_ah));
 839                 if (attr->ah_attr.ah_flags & IB_AH_GRH) {
 840                         udp_info->ttl = attr->ah_attr.grh.hop_limit;
 841                         udp_info->flow_label = attr->ah_attr.grh.flow_label;
 842                         udp_info->tos = attr->ah_attr.grh.traffic_class;
 843
 844                         udp_info->src_port = kc_rdma_get_udp_sport(udp_info->flow_label,
 845                                                                    ibqp->qp_num,
 846                                                                    roce_info->dest_qp);
 847
 848                         irdma_qp_rem_qos(&iwqp->sc_qp);
 849                         dev->ws_remove(iwqp->sc_qp.vsi, ctx_info->user_pri);
 850                         if (iwqp->sc_qp.vsi->dscp_mode)
 851                                 ctx_info->user_pri =
 852                                     iwqp->sc_qp.vsi->dscp_map[irdma_tos2dscp(udp_info->tos)];
 853                         else
 854                                 ctx_info->user_pri = rt_tos2priority(udp_info->tos);
 855                 }
 856                 ret = kc_irdma_set_roce_cm_info(iwqp, attr, &vlan_id);
 857                 if (ret)
 858                         return ret;
 859                 if (dev->ws_add(iwqp->sc_qp.vsi, ctx_info->user_pri))
 860                         return -ENOMEM;
 861                 iwqp->sc_qp.user_pri = ctx_info->user_pri;
 862                 irdma_qp_add_qos(&iwqp->sc_qp);
 863
 864                 if (vlan_id >= VLAN_N_VID && iwdev->dcb_vlan_mode)
 865                         vlan_id = 0;
 866                 if (vlan_id < VLAN_N_VID) {
 867                         udp_info->insert_vlan_tag = true;
 868                         udp_info->vlan_tag = vlan_id |
 869                             ctx_info->user_pri << VLAN_PRIO_SHIFT;
 870                 } else {
 871                         udp_info->insert_vlan_tag = false;
 872                 }
 873
 874                 av->attrs = attr->ah_attr;
 875                 rdma_gid2ip((struct sockaddr *)&av->dgid_addr, &attr->ah_attr.grh.dgid);
 876                 if (av->sgid_addr.saddr.sa_family == AF_INET6) {
 877                         __be32 *daddr =
 878                         av->dgid_addr.saddr_in6.sin6_addr.__u6_addr.__u6_addr32;
 879                         __be32 *saddr =
 880                         av->sgid_addr.saddr_in6.sin6_addr.__u6_addr.__u6_addr32;
 881
 882                         irdma_copy_ip_ntohl(&udp_info->dest_ip_addr[0], daddr);
 883                         irdma_copy_ip_ntohl(&udp_info->local_ipaddr[0], saddr);
 884
 885                         udp_info->ipv4 = false;
 886                         irdma_copy_ip_ntohl(local_ip, daddr);
 887
 888                         udp_info->arp_idx = irdma_arp_table(iwdev->rf, local_ip,
 889                                                             NULL, IRDMA_ARP_RESOLVE);
 890                 } else {
 891                         __be32 saddr = av->sgid_addr.saddr_in.sin_addr.s_addr;
 892                         __be32 daddr = av->dgid_addr.saddr_in.sin_addr.s_addr;
 893
 894                         local_ip[0] = ntohl(daddr);
 895
 896                         udp_info->ipv4 = true;
 897                         udp_info->dest_ip_addr[0] = 0;
 898                         udp_info->dest_ip_addr[1] = 0;
 899                         udp_info->dest_ip_addr[2] = 0;
 900                         udp_info->dest_ip_addr[3] = local_ip[0];
 901
 902                         udp_info->local_ipaddr[0] = 0;
 903                         udp_info->local_ipaddr[1] = 0;
 904                         udp_info->local_ipaddr[2] = 0;
 905                         udp_info->local_ipaddr[3] = ntohl(saddr);
 906                 }
 907                 udp_info->arp_idx =
 908                     irdma_add_arp(iwdev->rf, local_ip,
 909                                   ah_attr_to_dmac(attr->ah_attr));
 910         }
 911
 912         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
 913                 if (attr->max_rd_atomic > dev->hw_attrs.max_hw_ord) {
 914                         ibdev_err(&iwdev->ibdev,
 915                                   "rd_atomic = %d, above max_hw_ord=%d\n",
 916                                   attr->max_rd_atomic,
 917                                   dev->hw_attrs.max_hw_ord);
 918                         return -EINVAL;
 919                 }
 920                 if (attr->max_rd_atomic)
 921                         roce_info->ord_size = attr->max_rd_atomic;
 922                 info.ord_valid = true;
 923         }
 924
 925         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
 926                 if (attr->max_dest_rd_atomic > dev->hw_attrs.max_hw_ird) {
 927                         ibdev_err(&iwdev->ibdev,
 928                                   "rd_atomic = %d, above max_hw_ird=%d\n",
 929                                   attr->max_rd_atomic,
 930                                   dev->hw_attrs.max_hw_ird);
 931                         return -EINVAL;
 932                 }
 933                 if (attr->max_dest_rd_atomic)
 934                         roce_info->ird_size = attr->max_dest_rd_atomic;
 935         }
 936
 937         if (attr_mask & IB_QP_ACCESS_FLAGS) {
 938                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
 939                         roce_info->wr_rdresp_en = true;
 940                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
 941                         roce_info->wr_rdresp_en = true;
 942                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
 943                         roce_info->rd_en = true;
 944         }
 945
 946         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
 947
 948         irdma_debug(dev, IRDMA_DEBUG_VERBS,
 949                     "caller: %pS qp_id=%d to_ibqpstate=%d ibqpstate=%d irdma_qpstate=%d attr_mask=0x%x\n",
 950                     __builtin_return_address(0), ibqp->qp_num, attr->qp_state, iwqp->ibqp_state,
 951                     iwqp->iwarp_state, attr_mask);
 952
 953         spin_lock_irqsave(&iwqp->lock, flags);
 954         if (attr_mask & IB_QP_STATE) {
 955                 if (!kc_ib_modify_qp_is_ok(iwqp->ibqp_state, attr->qp_state,
 956                                            iwqp->ibqp.qp_type, attr_mask,
 957                                            IB_LINK_LAYER_ETHERNET)) {
 958                         irdma_print("modify_qp invalid for qp_id=%d, old_state=0x%x, new_state=0x%x\n",
 959                                     iwqp->ibqp.qp_num, iwqp->ibqp_state,
 960                                     attr->qp_state);
 961                         ret = -EINVAL;
 962                         goto exit;
 963                 }
 964                 info.curr_iwarp_state = iwqp->iwarp_state;
 965
 966                 switch (attr->qp_state) {
 967                 case IB_QPS_INIT:
 968                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
 969                                 ret = -EINVAL;
 970                                 goto exit;
 971                         }
 972
 973                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
 974                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
 975                                 issue_modify_qp = 1;
 976                         }
 977                         break;
 978                 case IB_QPS_RTR:
 979                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
 980                                 ret = -EINVAL;
 981                                 goto exit;
 982                         }
 983                         info.arp_cache_idx_valid = true;
 984                         info.cq_num_valid = true;
 985                         info.next_iwarp_state = IRDMA_QP_STATE_RTR;
 986                         issue_modify_qp = 1;
 987                         break;
 988                 case IB_QPS_RTS:
 989                         if (iwqp->ibqp_state < IB_QPS_RTR ||
 990                             iwqp->ibqp_state == IB_QPS_ERR) {
 991                                 ret = -EINVAL;
 992                                 goto exit;
 993                         }
 994
 995                         info.arp_cache_idx_valid = true;
 996                         info.cq_num_valid = true;
 997                         info.ord_valid = true;
 998                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
 999                         issue_modify_qp = 1;
1000                         if (dev->hw_attrs.uk_attrs.hw_rev == IRDMA_GEN_2)
1001                                 iwdev->rf->check_fc(&iwdev->vsi, &iwqp->sc_qp);
1002                         udp_info->cwnd = iwdev->roce_cwnd;
1003                         roce_info->ack_credits = iwdev->roce_ackcreds;
1004                         if (iwdev->push_mode && udata &&
1005                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1006                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1007                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1008                                 irdma_alloc_push_page(iwqp);
1009                                 spin_lock_irqsave(&iwqp->lock, flags);
1010                         }
1011                         break;
1012                 case IB_QPS_SQD:
1013                         if (iwqp->iwarp_state == IRDMA_QP_STATE_SQD)
1014                                 goto exit;
1015
1016                         if (iwqp->iwarp_state != IRDMA_QP_STATE_RTS) {
1017                                 ret = -EINVAL;
1018                                 goto exit;
1019                         }
1020
1021                         info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1022                         issue_modify_qp = 1;
1023                         break;
1024                 case IB_QPS_SQE:
1025                 case IB_QPS_ERR:
1026                 case IB_QPS_RESET:
1027                         if (iwqp->iwarp_state == IRDMA_QP_STATE_RTS) {
1028                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1029                                 info.next_iwarp_state = IRDMA_QP_STATE_SQD;
1030                                 irdma_hw_modify_qp(iwdev, iwqp, &info, true);
1031                                 spin_lock_irqsave(&iwqp->lock, flags);
1032                         }
1033
1034                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1035                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1036                                 if (udata && udata->inlen) {
1037                                         if (ib_copy_from_udata(&ureq, udata,
1038                                                                min(sizeof(ureq), udata->inlen)))
1039                                                 return -EINVAL;
1040
1041                                         irdma_flush_wqes(iwqp,
1042                                                          (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1043                                                          (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1044                                                          IRDMA_REFLUSH);
1045                                 }
1046                                 return 0;
1047                         }
1048
1049                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1050                         issue_modify_qp = 1;
1051                         break;
1052                 default:
1053                         ret = -EINVAL;
1054                         goto exit;
1055                 }
1056
1057                 iwqp->ibqp_state = attr->qp_state;
1058         }
1059
1060         ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1061         ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1062         irdma_sc_qp_setctx_roce(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1063         spin_unlock_irqrestore(&iwqp->lock, flags);
1064
1065         if (attr_mask & IB_QP_STATE) {
1066                 if (issue_modify_qp) {
1067                         ctx_info->rem_endpoint_idx = udp_info->arp_idx;
1068                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1069                                 return -EINVAL;
1070                         spin_lock_irqsave(&iwqp->lock, flags);
1071                         if (iwqp->iwarp_state == info.curr_iwarp_state) {
1072                                 iwqp->iwarp_state = info.next_iwarp_state;
1073                                 iwqp->ibqp_state = attr->qp_state;
1074                         }
1075                         if (iwqp->ibqp_state > IB_QPS_RTS &&
1076                             !iwqp->flush_issued) {
1077                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1078                                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_SQ |
1079                                                  IRDMA_FLUSH_RQ |
1080                                                  IRDMA_FLUSH_WAIT);
1081                                 iwqp->flush_issued = 1;
1082
1083                         } else {
1084                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1085                         }
1086                 } else {
1087                         iwqp->ibqp_state = attr->qp_state;
1088                 }
1089                 if (udata && udata->outlen && dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1090                         struct irdma_ucontext *ucontext;
1091
1092                         ucontext = rdma_udata_to_drv_context(udata,
1093                                                              struct irdma_ucontext, ibucontext);
1094                         if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1095                             !iwqp->push_wqe_mmap_entry &&
1096                             !irdma_setup_push_mmap_entries(ucontext, iwqp,
1097                                                            &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1098                                 uresp.push_valid = 1;
1099                                 uresp.push_offset = iwqp->sc_qp.push_offset;
1100                         }
1101                         uresp.rd_fence_rate = iwdev->rd_fence_rate;
1102                         ret = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1103                                                                   udata->outlen));
1104                         if (ret) {
1105                                 irdma_remove_push_mmap_entries(iwqp);
1106                                 irdma_debug(iwdev_to_idev(iwdev),
1107                                             IRDMA_DEBUG_VERBS,
1108                                             "copy_to_udata failed\n");
1109                                 return ret;
1110                         }
1111                 }
1112         }
1113
1114         return 0;
1115 exit:
1116         spin_unlock_irqrestore(&iwqp->lock, flags);
1117
1118         return ret;
1119 }
1120
1121 /**
1122  * irdma_modify_qp - modify qp request
1123  * @ibqp: qp's pointer for modify
1124  * @attr: access attributes
1125  * @attr_mask: state mask
1126  * @udata: user data
1127  */
1128 int
1129 irdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
1130                 struct ib_udata *udata)
1131 {
1132 #define IRDMA_MODIFY_QP_MIN_REQ_LEN offsetofend(struct irdma_modify_qp_req, rq_flush)
1133 #define IRDMA_MODIFY_QP_MIN_RESP_LEN offsetofend(struct irdma_modify_qp_resp, push_valid)
1134         struct irdma_qp *iwqp = to_iwqp(ibqp);
1135         struct irdma_device *iwdev = iwqp->iwdev;
1136         struct irdma_sc_dev *dev = &iwdev->rf->sc_dev;
1137         struct irdma_qp_host_ctx_info *ctx_info;
1138         struct irdma_tcp_offload_info *tcp_info;
1139         struct irdma_iwarp_offload_info *offload_info;
1140         struct irdma_modify_qp_info info = {0};
1141         struct irdma_modify_qp_resp uresp = {};
1142         struct irdma_modify_qp_req ureq = {};
1143         u8 issue_modify_qp = 0;
1144         u8 dont_wait = 0;
1145         int err;
1146         unsigned long flags;
1147
1148         if (udata) {
1149                 if ((udata->inlen && udata->inlen < IRDMA_MODIFY_QP_MIN_REQ_LEN) ||
1150                     (udata->outlen && udata->outlen < IRDMA_MODIFY_QP_MIN_RESP_LEN))
1151                         return -EINVAL;
1152         }
1153
1154         if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1155                 return -EOPNOTSUPP;
1156
1157         ctx_info = &iwqp->ctx_info;
1158         offload_info = &iwqp->iwarp_info;
1159         tcp_info = &iwqp->tcp_info;
1160         wait_event(iwqp->mod_qp_waitq, !atomic_read(&iwqp->hw_mod_qp_pend));
1161         irdma_debug(dev, IRDMA_DEBUG_VERBS,
1162                     "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",
1163                     __builtin_return_address(0), ibqp->qp_num, attr->qp_state, iwqp->ibqp_state, iwqp->iwarp_state,
1164                     iwqp->last_aeq, iwqp->hw_tcp_state, iwqp->hw_iwarp_state, attr_mask);
1165
1166         spin_lock_irqsave(&iwqp->lock, flags);
1167         if (attr_mask & IB_QP_STATE) {
1168                 info.curr_iwarp_state = iwqp->iwarp_state;
1169                 switch (attr->qp_state) {
1170                 case IB_QPS_INIT:
1171                 case IB_QPS_RTR:
1172                         if (iwqp->iwarp_state > IRDMA_QP_STATE_IDLE) {
1173                                 err = -EINVAL;
1174                                 goto exit;
1175                         }
1176
1177                         if (iwqp->iwarp_state == IRDMA_QP_STATE_INVALID) {
1178                                 info.next_iwarp_state = IRDMA_QP_STATE_IDLE;
1179                                 issue_modify_qp = 1;
1180                         }
1181                         if (iwdev->push_mode && udata &&
1182                             iwqp->sc_qp.push_idx == IRDMA_INVALID_PUSH_PAGE_INDEX &&
1183                             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1184                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1185                                 irdma_alloc_push_page(iwqp);
1186                                 spin_lock_irqsave(&iwqp->lock, flags);
1187                         }
1188                         break;
1189                 case IB_QPS_RTS:
1190                         if (iwqp->iwarp_state > IRDMA_QP_STATE_RTS ||
1191                             !iwqp->cm_id) {
1192                                 err = -EINVAL;
1193                                 goto exit;
1194                         }
1195
1196                         issue_modify_qp = 1;
1197                         iwqp->hw_tcp_state = IRDMA_TCP_STATE_ESTABLISHED;
1198                         iwqp->hte_added = 1;
1199                         info.next_iwarp_state = IRDMA_QP_STATE_RTS;
1200                         info.tcp_ctx_valid = true;
1201                         info.ord_valid = true;
1202                         info.arp_cache_idx_valid = true;
1203                         info.cq_num_valid = true;
1204                         break;
1205                 case IB_QPS_SQD:
1206                         if (iwqp->hw_iwarp_state > IRDMA_QP_STATE_RTS) {
1207                                 err = 0;
1208                                 goto exit;
1209                         }
1210
1211                         if (iwqp->iwarp_state == IRDMA_QP_STATE_CLOSING ||
1212                             iwqp->iwarp_state < IRDMA_QP_STATE_RTS) {
1213                                 err = 0;
1214                                 goto exit;
1215                         }
1216
1217                         if (iwqp->iwarp_state > IRDMA_QP_STATE_CLOSING) {
1218                                 err = -EINVAL;
1219                                 goto exit;
1220                         }
1221
1222                         info.next_iwarp_state = IRDMA_QP_STATE_CLOSING;
1223                         issue_modify_qp = 1;
1224                         break;
1225                 case IB_QPS_SQE:
1226                         if (iwqp->iwarp_state >= IRDMA_QP_STATE_TERMINATE) {
1227                                 err = -EINVAL;
1228                                 goto exit;
1229                         }
1230
1231                         info.next_iwarp_state = IRDMA_QP_STATE_TERMINATE;
1232                         issue_modify_qp = 1;
1233                         break;
1234                 case IB_QPS_ERR:
1235                 case IB_QPS_RESET:
1236                         if (iwqp->iwarp_state == IRDMA_QP_STATE_ERROR) {
1237                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1238                                 if (udata && udata->inlen) {
1239                                         if (ib_copy_from_udata(&ureq, udata,
1240                                                                min(sizeof(ureq), udata->inlen)))
1241                                                 return -EINVAL;
1242
1243                                         irdma_flush_wqes(iwqp,
1244                                                          (ureq.sq_flush ? IRDMA_FLUSH_SQ : 0) |
1245                                                          (ureq.rq_flush ? IRDMA_FLUSH_RQ : 0) |
1246                                                          IRDMA_REFLUSH);
1247                                 }
1248                                 return 0;
1249                         }
1250
1251                         if (iwqp->sc_qp.term_flags) {
1252                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1253                                 irdma_terminate_del_timer(&iwqp->sc_qp);
1254                                 spin_lock_irqsave(&iwqp->lock, flags);
1255                         }
1256                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1257                         if (iwqp->hw_tcp_state > IRDMA_TCP_STATE_CLOSED &&
1258                             iwdev->iw_status &&
1259                             iwqp->hw_tcp_state != IRDMA_TCP_STATE_TIME_WAIT)
1260                                 info.reset_tcp_conn = true;
1261                         else
1262                                 dont_wait = 1;
1263
1264                         issue_modify_qp = 1;
1265                         info.next_iwarp_state = IRDMA_QP_STATE_ERROR;
1266                         break;
1267                 default:
1268                         err = -EINVAL;
1269                         goto exit;
1270                 }
1271
1272                 iwqp->ibqp_state = attr->qp_state;
1273         }
1274         if (attr_mask & IB_QP_ACCESS_FLAGS) {
1275                 ctx_info->iwarp_info_valid = true;
1276                 if (attr->qp_access_flags & IB_ACCESS_LOCAL_WRITE)
1277                         offload_info->wr_rdresp_en = true;
1278                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
1279                         offload_info->wr_rdresp_en = true;
1280                 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
1281                         offload_info->rd_en = true;
1282         }
1283
1284         if (ctx_info->iwarp_info_valid) {
1285                 ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
1286                 ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
1287                 irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
1288         }
1289         spin_unlock_irqrestore(&iwqp->lock, flags);
1290
1291         if (attr_mask & IB_QP_STATE) {
1292                 if (issue_modify_qp) {
1293                         ctx_info->rem_endpoint_idx = tcp_info->arp_idx;
1294                         if (irdma_hw_modify_qp(iwdev, iwqp, &info, true))
1295                                 return -EINVAL;
1296                 }
1297
1298                 spin_lock_irqsave(&iwqp->lock, flags);
1299                 if (iwqp->iwarp_state == info.curr_iwarp_state) {
1300                         iwqp->iwarp_state = info.next_iwarp_state;
1301                         iwqp->ibqp_state = attr->qp_state;
1302                 }
1303                 spin_unlock_irqrestore(&iwqp->lock, flags);
1304         }
1305
1306         if (issue_modify_qp && iwqp->ibqp_state > IB_QPS_RTS) {
1307                 if (dont_wait) {
1308                         if (iwqp->hw_tcp_state) {
1309                                 spin_lock_irqsave(&iwqp->lock, flags);
1310                                 iwqp->hw_tcp_state = IRDMA_TCP_STATE_CLOSED;
1311                                 iwqp->last_aeq = IRDMA_AE_RESET_SENT;
1312                                 spin_unlock_irqrestore(&iwqp->lock, flags);
1313                         }
1314                         irdma_cm_disconn(iwqp);
1315                 } else {
1316                         int close_timer_started;
1317
1318                         spin_lock_irqsave(&iwdev->cm_core.ht_lock, flags);
1319
1320                         if (iwqp->cm_node) {
1321                                 atomic_inc(&iwqp->cm_node->refcnt);
1322                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1323                                 close_timer_started = atomic_inc_return(&iwqp->close_timer_started);
1324                                 if (iwqp->cm_id && close_timer_started == 1)
1325                                         irdma_schedule_cm_timer(iwqp->cm_node,
1326                                                                 (struct irdma_puda_buf *)iwqp,
1327                                                                 IRDMA_TIMER_TYPE_CLOSE, 1, 0);
1328
1329                                 irdma_rem_ref_cm_node(iwqp->cm_node);
1330                         } else {
1331                                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
1332                         }
1333                 }
1334         }
1335         if (attr_mask & IB_QP_STATE && udata && udata->outlen &&
1336             dev->hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
1337                 struct irdma_ucontext *ucontext;
1338
1339                 ucontext = rdma_udata_to_drv_context(udata,
1340                                                      struct irdma_ucontext, ibucontext);
1341                 if (iwqp->sc_qp.push_idx != IRDMA_INVALID_PUSH_PAGE_INDEX &&
1342                     !iwqp->push_wqe_mmap_entry &&
1343                     !irdma_setup_push_mmap_entries(ucontext, iwqp,
1344                                                    &uresp.push_wqe_mmap_key, &uresp.push_db_mmap_key)) {
1345                         uresp.push_valid = 1;
1346                         uresp.push_offset = iwqp->sc_qp.push_offset;
1347                 }
1348                 uresp.rd_fence_rate = iwdev->rd_fence_rate;
1349
1350                 err = ib_copy_to_udata(udata, &uresp, min(sizeof(uresp),
1351                                                           udata->outlen));
1352                 if (err) {
1353                         irdma_remove_push_mmap_entries(iwqp);
1354                         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
1355                                     "copy_to_udata failed\n");
1356                         return err;
1357                 }
1358         }
1359
1360         return 0;
1361 exit:
1362         spin_unlock_irqrestore(&iwqp->lock, flags);
1363
1364         return err;
1365 }
1366
1367 /**
1368  * irdma_cq_free_rsrc - free up resources for cq
1369  * @rf: RDMA PCI function
1370  * @iwcq: cq ptr
1371  */
1372 void
1373 irdma_cq_free_rsrc(struct irdma_pci_f *rf, struct irdma_cq *iwcq)
1374 {
1375         struct irdma_sc_cq *cq = &iwcq->sc_cq;
1376
1377         if (!iwcq->user_mode) {
1378                 irdma_free_dma_mem(rf->sc_dev.hw, &iwcq->kmem);
1379                 irdma_free_dma_mem(rf->sc_dev.hw, &iwcq->kmem_shadow);
1380         }
1381
1382         irdma_free_rsrc(rf, rf->allocated_cqs, cq->cq_uk.cq_id);
1383 }
1384
1385 /**
1386  * irdma_free_cqbuf - worker to free a cq buffer
1387  * @work: provides access to the cq buffer to free
1388  */
1389 static void
1390 irdma_free_cqbuf(struct work_struct *work)
1391 {
1392         struct irdma_cq_buf *cq_buf = container_of(work, struct irdma_cq_buf, work);
1393
1394         irdma_free_dma_mem(cq_buf->hw, &cq_buf->kmem_buf);
1395         kfree(cq_buf);
1396 }
1397
1398 /**
1399  * irdma_process_resize_list - remove resized cq buffers from the resize_list
1400  * @iwcq: cq which owns the resize_list
1401  * @iwdev: irdma device
1402  * @lcqe_buf: the buffer where the last cqe is received
1403  */
1404 int
1405 irdma_process_resize_list(struct irdma_cq *iwcq,
1406                           struct irdma_device *iwdev,
1407                           struct irdma_cq_buf *lcqe_buf)
1408 {
1409         struct list_head *tmp_node, *list_node;
1410         struct irdma_cq_buf *cq_buf;
1411         int cnt = 0;
1412
1413         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
1414                 cq_buf = list_entry(list_node, struct irdma_cq_buf, list);
1415                 if (cq_buf == lcqe_buf)
1416                         return cnt;
1417
1418                 list_del(&cq_buf->list);
1419                 queue_work(iwdev->cleanup_wq, &cq_buf->work);
1420                 cnt++;
1421         }
1422
1423         return cnt;
1424 }
1425
1426 /**
1427  * irdma_resize_cq - resize cq
1428  * @ibcq: cq to be resized
1429  * @entries: desired cq size
1430  * @udata: user data
1431  */
1432 static int
1433 irdma_resize_cq(struct ib_cq *ibcq, int entries,
1434                 struct ib_udata *udata)
1435 {
1436 #define IRDMA_RESIZE_CQ_MIN_REQ_LEN offsetofend(struct irdma_resize_cq_req, user_cq_buffer)
1437         struct irdma_cq *iwcq = to_iwcq(ibcq);
1438         struct irdma_sc_dev *dev = iwcq->sc_cq.dev;
1439         struct irdma_cqp_request *cqp_request;
1440         struct cqp_cmds_info *cqp_info;
1441         struct irdma_modify_cq_info *m_info;
1442         struct irdma_modify_cq_info info = {0};
1443         struct irdma_dma_mem kmem_buf;
1444         struct irdma_cq_mr *cqmr_buf;
1445         struct irdma_pbl *iwpbl_buf;
1446         struct irdma_device *iwdev;
1447         struct irdma_pci_f *rf;
1448         struct irdma_cq_buf *cq_buf = NULL;
1449         unsigned long flags;
1450         int ret;
1451
1452         iwdev = to_iwdev(ibcq->device);
1453         rf = iwdev->rf;
1454
1455         if (!(rf->sc_dev.hw_attrs.uk_attrs.feature_flags &
1456               IRDMA_FEATURE_CQ_RESIZE))
1457                 return -EOPNOTSUPP;
1458
1459         if (udata && udata->inlen < IRDMA_RESIZE_CQ_MIN_REQ_LEN)
1460                 return -EINVAL;
1461
1462         if (entries > rf->max_cqe)
1463                 return -EINVAL;
1464
1465         if (!iwcq->user_mode) {
1466                 entries++;
1467                 if (rf->sc_dev.hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2)
1468                         entries *= 2;
1469         }
1470
1471         info.cq_size = max(entries, 4);
1472
1473         if (info.cq_size == iwcq->sc_cq.cq_uk.cq_size - 1)
1474                 return 0;
1475
1476         if (udata) {
1477                 struct irdma_resize_cq_req req = {};
1478                 struct irdma_ucontext *ucontext =
1479                 rdma_udata_to_drv_context(udata, struct irdma_ucontext,
1480                                           ibucontext);
1481
1482                 /* CQ resize not supported with legacy GEN_1 libi40iw */
1483                 if (ucontext->legacy_mode)
1484                         return -EOPNOTSUPP;
1485
1486                 if (ib_copy_from_udata(&req, udata,
1487                                        min(sizeof(req), udata->inlen)))
1488                         return -EINVAL;
1489
1490                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
1491                 iwpbl_buf = irdma_get_pbl((unsigned long)req.user_cq_buffer,
1492                                           &ucontext->cq_reg_mem_list);
1493                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
1494
1495                 if (!iwpbl_buf)
1496                         return -ENOMEM;
1497
1498                 cqmr_buf = &iwpbl_buf->cq_mr;
1499                 if (iwpbl_buf->pbl_allocated) {
1500                         info.virtual_map = true;
1501                         info.pbl_chunk_size = 1;
1502                         info.first_pm_pbl_idx = cqmr_buf->cq_pbl.idx;
1503                 } else {
1504                         info.cq_pa = cqmr_buf->cq_pbl.addr;
1505                 }
1506         } else {
1507                 /* Kmode CQ resize */
1508                 int rsize;
1509
1510                 rsize = info.cq_size * sizeof(struct irdma_cqe);
1511                 kmem_buf.size = round_up(rsize, 256);
1512                 kmem_buf.va = irdma_allocate_dma_mem(dev->hw, &kmem_buf,
1513                                                      kmem_buf.size, 256);
1514                 if (!kmem_buf.va)
1515                         return -ENOMEM;
1516
1517                 info.cq_base = kmem_buf.va;
1518                 info.cq_pa = kmem_buf.pa;
1519                 cq_buf = kzalloc(sizeof(*cq_buf), GFP_KERNEL);
1520                 if (!cq_buf) {
1521                         ret = -ENOMEM;
1522                         goto error;
1523                 }
1524         }
1525
1526         cqp_request = irdma_alloc_and_get_cqp_request(&rf->cqp, true);
1527         if (!cqp_request) {
1528                 ret = -ENOMEM;
1529                 goto error;
1530         }
1531
1532         info.shadow_read_threshold = iwcq->sc_cq.shadow_read_threshold;
1533         info.cq_resize = true;
1534
1535         cqp_info = &cqp_request->info;
1536         m_info = &cqp_info->in.u.cq_modify.info;
1537         memcpy(m_info, &info, sizeof(*m_info));
1538
1539         cqp_info->cqp_cmd = IRDMA_OP_CQ_MODIFY;
1540         cqp_info->in.u.cq_modify.cq = &iwcq->sc_cq;
1541         cqp_info->in.u.cq_modify.scratch = (uintptr_t)cqp_request;
1542         cqp_info->post_sq = 1;
1543         ret = irdma_handle_cqp_op(rf, cqp_request);
1544         irdma_put_cqp_request(&rf->cqp, cqp_request);
1545         if (ret)
1546                 goto error;
1547
1548         spin_lock_irqsave(&iwcq->lock, flags);
1549         if (cq_buf) {
1550                 cq_buf->kmem_buf = iwcq->kmem;
1551                 cq_buf->hw = dev->hw;
1552                 memcpy(&cq_buf->cq_uk, &iwcq->sc_cq.cq_uk, sizeof(cq_buf->cq_uk));
1553                 INIT_WORK(&cq_buf->work, irdma_free_cqbuf);
1554                 list_add_tail(&cq_buf->list, &iwcq->resize_list);
1555                 iwcq->kmem = kmem_buf;
1556         }
1557
1558         irdma_sc_cq_resize(&iwcq->sc_cq, &info);
1559         ibcq->cqe = info.cq_size - 1;
1560         spin_unlock_irqrestore(&iwcq->lock, flags);
1561
1562         return 0;
1563 error:
1564         if (!udata)
1565                 irdma_free_dma_mem(dev->hw, &kmem_buf);
1566         kfree(cq_buf);
1567
1568         return ret;
1569 }
1570
1571 /**
1572  * irdma_get_mr_access - get hw MR access permissions from IB access flags
1573  * @access: IB access flags
1574  */
1575 static inline u16 irdma_get_mr_access(int access){
1576         u16 hw_access = 0;
1577
1578         hw_access |= (access & IB_ACCESS_LOCAL_WRITE) ?
1579             IRDMA_ACCESS_FLAGS_LOCALWRITE : 0;
1580         hw_access |= (access & IB_ACCESS_REMOTE_WRITE) ?
1581             IRDMA_ACCESS_FLAGS_REMOTEWRITE : 0;
1582         hw_access |= (access & IB_ACCESS_REMOTE_READ) ?
1583             IRDMA_ACCESS_FLAGS_REMOTEREAD : 0;
1584         hw_access |= (access & IB_ACCESS_MW_BIND) ?
1585             IRDMA_ACCESS_FLAGS_BIND_WINDOW : 0;
1586         hw_access |= (access & IB_ZERO_BASED) ?
1587             IRDMA_ACCESS_FLAGS_ZERO_BASED : 0;
1588         hw_access |= IRDMA_ACCESS_FLAGS_LOCALREAD;
1589
1590         return hw_access;
1591 }
1592
1593 /**
1594  * irdma_free_stag - free stag resource
1595  * @iwdev: irdma device
1596  * @stag: stag to free
1597  */
1598 void
1599 irdma_free_stag(struct irdma_device *iwdev, u32 stag)
1600 {
1601         u32 stag_idx;
1602
1603         stag_idx = (stag & iwdev->rf->mr_stagmask) >> IRDMA_CQPSQ_STAG_IDX_S;
1604         irdma_free_rsrc(iwdev->rf, iwdev->rf->allocated_mrs, stag_idx);
1605 }
1606
1607 /**
1608  * irdma_create_stag - create random stag
1609  * @iwdev: irdma device
1610  */
1611 u32
1612 irdma_create_stag(struct irdma_device *iwdev)
1613 {
1614         u32 stag;
1615         u32 stag_index = 0;
1616         u32 next_stag_index;
1617         u32 driver_key;
1618         u32 random;
1619         u8 consumer_key;
1620         int ret;
1621
1622         get_random_bytes(&random, sizeof(random));
1623         consumer_key = (u8)random;
1624
1625         driver_key = random & ~iwdev->rf->mr_stagmask;
1626         next_stag_index = (random & iwdev->rf->mr_stagmask) >> 8;
1627         next_stag_index %= iwdev->rf->max_mr;
1628
1629         ret = irdma_alloc_rsrc(iwdev->rf, iwdev->rf->allocated_mrs,
1630                                iwdev->rf->max_mr, &stag_index,
1631                                &next_stag_index);
1632         if (ret)
1633                 return 0;
1634         stag = stag_index << IRDMA_CQPSQ_STAG_IDX_S;
1635         stag |= driver_key;
1636         stag += (u32)consumer_key;
1637
1638         return stag;
1639 }
1640
1641 /**
1642  * irdma_check_mem_contiguous - check if pbls stored in arr are contiguous
1643  * @arr: lvl1 pbl array
1644  * @npages: page count
1645  * @pg_size: page size
1646  *
1647  */
1648 static bool
1649 irdma_check_mem_contiguous(u64 *arr, u32 npages, u32 pg_size)
1650 {
1651         u32 pg_idx;
1652
1653         for (pg_idx = 0; pg_idx < npages; pg_idx++) {
1654                 if ((*arr + (pg_size * pg_idx)) != arr[pg_idx])
1655                         return false;
1656         }
1657
1658         return true;
1659 }
1660
1661 /**
1662  * irdma_check_mr_contiguous - check if MR is physically contiguous
1663  * @palloc: pbl allocation struct
1664  * @pg_size: page size
1665  */
1666 static bool
1667 irdma_check_mr_contiguous(struct irdma_pble_alloc *palloc,
1668                           u32 pg_size)
1669 {
1670         struct irdma_pble_level2 *lvl2 = &palloc->level2;
1671         struct irdma_pble_info *leaf = lvl2->leaf;
1672         u64 *arr = NULL;
1673         u64 *start_addr = NULL;
1674         int i;
1675         bool ret;
1676
1677         if (palloc->level == PBLE_LEVEL_1) {
1678                 arr = palloc->level1.addr;
1679                 ret = irdma_check_mem_contiguous(arr, palloc->total_cnt,
1680                                                  pg_size);
1681                 return ret;
1682         }
1683
1684         start_addr = leaf->addr;
1685
1686         for (i = 0; i < lvl2->leaf_cnt; i++, leaf++) {
1687                 arr = leaf->addr;
1688                 if ((*start_addr + (i * pg_size * PBLE_PER_PAGE)) != *arr)
1689                         return false;
1690                 ret = irdma_check_mem_contiguous(arr, leaf->cnt, pg_size);
1691                 if (!ret)
1692                         return false;
1693         }
1694
1695         return true;
1696 }
1697
1698 /**
1699  * irdma_setup_pbles - copy user pg address to pble's
1700  * @rf: RDMA PCI function
1701  * @iwmr: mr pointer for this memory registration
1702  * @use_pbles: flag if to use pble's
1703  * @lvl_1_only: request only level 1 pble if true
1704  */
1705 static int
1706 irdma_setup_pbles(struct irdma_pci_f *rf, struct irdma_mr *iwmr,
1707                   bool use_pbles, bool lvl_1_only)
1708 {
1709         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
1710         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
1711         struct irdma_pble_info *pinfo;
1712         u64 *pbl;
1713         int status;
1714         enum irdma_pble_level level = PBLE_LEVEL_1;
1715
1716         if (use_pbles) {
1717                 status = irdma_get_pble(rf->pble_rsrc, palloc, iwmr->page_cnt,
1718                                         lvl_1_only);
1719                 if (status)
1720                         return status;
1721
1722                 iwpbl->pbl_allocated = true;
1723                 level = palloc->level;
1724                 pinfo = (level == PBLE_LEVEL_1) ? &palloc->level1 :
1725                     palloc->level2.leaf;
1726                 pbl = pinfo->addr;
1727         } else {
1728                 pbl = iwmr->pgaddrmem;
1729         }
1730
1731         irdma_copy_user_pgaddrs(iwmr, pbl, level);
1732
1733         if (use_pbles)
1734                 iwmr->pgaddrmem[0] = *pbl;
1735
1736         return 0;
1737 }
1738
1739 /**
1740  * irdma_handle_q_mem - handle memory for qp and cq
1741  * @iwdev: irdma device
1742  * @req: information for q memory management
1743  * @iwpbl: pble struct
1744  * @use_pbles: flag to use pble
1745  */
1746 static int
1747 irdma_handle_q_mem(struct irdma_device *iwdev,
1748                    struct irdma_mem_reg_req *req,
1749                    struct irdma_pbl *iwpbl, bool use_pbles)
1750 {
1751         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
1752         struct irdma_mr *iwmr = iwpbl->iwmr;
1753         struct irdma_qp_mr *qpmr = &iwpbl->qp_mr;
1754         struct irdma_cq_mr *cqmr = &iwpbl->cq_mr;
1755         struct irdma_hmc_pble *hmc_p;
1756         u64 *arr = iwmr->pgaddrmem;
1757         u32 pg_size, total;
1758         int err = 0;
1759         bool ret = true;
1760
1761         pg_size = iwmr->page_size;
1762         err = irdma_setup_pbles(iwdev->rf, iwmr, use_pbles, true);
1763         if (err)
1764                 return err;
1765
1766         if (use_pbles)
1767                 arr = palloc->level1.addr;
1768
1769         switch (iwmr->type) {
1770         case IRDMA_MEMREG_TYPE_QP:
1771                 total = req->sq_pages + req->rq_pages;
1772                 hmc_p = &qpmr->sq_pbl;
1773                 qpmr->shadow = (dma_addr_t) arr[total];
1774                 if (use_pbles) {
1775                         ret = irdma_check_mem_contiguous(arr, req->sq_pages,
1776                                                          pg_size);
1777                         if (ret)
1778                                 ret = irdma_check_mem_contiguous(&arr[req->sq_pages],
1779                                                                  req->rq_pages,
1780                                                                  pg_size);
1781                 }
1782
1783                 if (!ret) {
1784                         hmc_p->idx = palloc->level1.idx;
1785                         hmc_p = &qpmr->rq_pbl;
1786                         hmc_p->idx = palloc->level1.idx + req->sq_pages;
1787                 } else {
1788                         hmc_p->addr = arr[0];
1789                         hmc_p = &qpmr->rq_pbl;
1790                         hmc_p->addr = arr[req->sq_pages];
1791                 }
1792                 break;
1793         case IRDMA_MEMREG_TYPE_CQ:
1794                 hmc_p = &cqmr->cq_pbl;
1795
1796                 if (!cqmr->split)
1797                         cqmr->shadow = (dma_addr_t) arr[req->cq_pages];
1798
1799                 if (use_pbles)
1800                         ret = irdma_check_mem_contiguous(arr, req->cq_pages,
1801                                                          pg_size);
1802
1803                 if (!ret)
1804                         hmc_p->idx = palloc->level1.idx;
1805                 else
1806                         hmc_p->addr = arr[0];
1807                 break;
1808         default:
1809                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
1810                             "MR type error\n");
1811                 err = -EINVAL;
1812         }
1813
1814         if (use_pbles && ret) {
1815                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
1816                 iwpbl->pbl_allocated = false;
1817         }
1818
1819         return err;
1820 }
1821
1822 /**
1823  * irdma_hw_alloc_mw - create the hw memory window
1824  * @iwdev: irdma device
1825  * @iwmr: pointer to memory window info
1826  */
1827 int
1828 irdma_hw_alloc_mw(struct irdma_device *iwdev, struct irdma_mr *iwmr)
1829 {
1830         struct irdma_mw_alloc_info *info;
1831         struct irdma_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
1832         struct irdma_cqp_request *cqp_request;
1833         struct cqp_cmds_info *cqp_info;
1834         int status;
1835
1836         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
1837         if (!cqp_request)
1838                 return -ENOMEM;
1839
1840         cqp_info = &cqp_request->info;
1841         info = &cqp_info->in.u.mw_alloc.info;
1842         memset(info, 0, sizeof(*info));
1843         if (iwmr->ibmw.type == IB_MW_TYPE_1)
1844                 info->mw_wide = true;
1845
1846         info->page_size = PAGE_SIZE;
1847         info->mw_stag_index = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
1848         info->pd_id = iwpd->sc_pd.pd_id;
1849         info->remote_access = true;
1850         cqp_info->cqp_cmd = IRDMA_OP_MW_ALLOC;
1851         cqp_info->post_sq = 1;
1852         cqp_info->in.u.mw_alloc.dev = &iwdev->rf->sc_dev;
1853         cqp_info->in.u.mw_alloc.scratch = (uintptr_t)cqp_request;
1854         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
1855         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
1856
1857         return status;
1858 }
1859
1860 /**
1861  * irdma_dealloc_mw - Dealloc memory window
1862  * @ibmw: memory window structure.
1863  */
1864 static int
1865 irdma_dealloc_mw(struct ib_mw *ibmw)
1866 {
1867         struct ib_pd *ibpd = ibmw->pd;
1868         struct irdma_pd *iwpd = to_iwpd(ibpd);
1869         struct irdma_mr *iwmr = to_iwmr((struct ib_mr *)ibmw);
1870         struct irdma_device *iwdev = to_iwdev(ibmw->device);
1871         struct irdma_cqp_request *cqp_request;
1872         struct cqp_cmds_info *cqp_info;
1873         struct irdma_dealloc_stag_info *info;
1874
1875         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
1876         if (!cqp_request)
1877                 return -ENOMEM;
1878
1879         cqp_info = &cqp_request->info;
1880         info = &cqp_info->in.u.dealloc_stag.info;
1881         memset(info, 0, sizeof(*info));
1882         info->pd_id = iwpd->sc_pd.pd_id;
1883         info->stag_idx = RS_64_1(ibmw->rkey, IRDMA_CQPSQ_STAG_IDX_S);
1884         info->mr = false;
1885         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
1886         cqp_info->post_sq = 1;
1887         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
1888         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
1889         irdma_handle_cqp_op(iwdev->rf, cqp_request);
1890         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
1891         irdma_free_stag(iwdev, iwmr->stag);
1892         kfree(iwmr);
1893
1894         return 0;
1895 }
1896
1897 /**
1898  * irdma_hw_alloc_stag - cqp command to allocate stag
1899  * @iwdev: irdma device
1900  * @iwmr: irdma mr pointer
1901  */
1902 int
1903 irdma_hw_alloc_stag(struct irdma_device *iwdev,
1904                     struct irdma_mr *iwmr)
1905 {
1906         struct irdma_allocate_stag_info *info;
1907         struct ib_pd *pd = iwmr->ibmr.pd;
1908         struct irdma_pd *iwpd = to_iwpd(pd);
1909         struct irdma_cqp_request *cqp_request;
1910         struct cqp_cmds_info *cqp_info;
1911         int status;
1912
1913         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
1914         if (!cqp_request)
1915                 return -ENOMEM;
1916
1917         cqp_info = &cqp_request->info;
1918         info = &cqp_info->in.u.alloc_stag.info;
1919         memset(info, 0, sizeof(*info));
1920         info->page_size = PAGE_SIZE;
1921         info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
1922         info->pd_id = iwpd->sc_pd.pd_id;
1923         info->total_len = iwmr->len;
1924         info->all_memory = (pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) ? true : false;
1925         info->remote_access = true;
1926         cqp_info->cqp_cmd = IRDMA_OP_ALLOC_STAG;
1927         cqp_info->post_sq = 1;
1928         cqp_info->in.u.alloc_stag.dev = &iwdev->rf->sc_dev;
1929         cqp_info->in.u.alloc_stag.scratch = (uintptr_t)cqp_request;
1930         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
1931         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
1932         if (!status)
1933                 iwmr->is_hwreg = 1;
1934
1935         return status;
1936 }
1937
1938 /**
1939  * irdma_set_page - populate pbl list for fmr
1940  * @ibmr: ib mem to access iwarp mr pointer
1941  * @addr: page dma address fro pbl list
1942  */
1943 static int
1944 irdma_set_page(struct ib_mr *ibmr, u64 addr)
1945 {
1946         struct irdma_mr *iwmr = to_iwmr(ibmr);
1947         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
1948         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
1949         u64 *pbl;
1950
1951         if (unlikely(iwmr->npages == iwmr->page_cnt))
1952                 return -ENOMEM;
1953
1954         if (palloc->level == PBLE_LEVEL_2) {
1955                 struct irdma_pble_info *palloc_info =
1956                 palloc->level2.leaf + (iwmr->npages >> PBLE_512_SHIFT);
1957
1958                 palloc_info->addr[iwmr->npages & (PBLE_PER_PAGE - 1)] = addr;
1959         } else {
1960                 pbl = palloc->level1.addr;
1961                 pbl[iwmr->npages] = addr;
1962         }
1963
1964         iwmr->npages++;
1965         return 0;
1966 }
1967
1968 /**
1969  * irdma_map_mr_sg - map of sg list for fmr
1970  * @ibmr: ib mem to access iwarp mr pointer
1971  * @sg: scatter gather list
1972  * @sg_nents: number of sg pages
1973  * @sg_offset: scatter gather list for fmr
1974  */
1975 static int
1976 irdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
1977                 int sg_nents, unsigned int *sg_offset)
1978 {
1979         struct irdma_mr *iwmr = to_iwmr(ibmr);
1980
1981         iwmr->npages = 0;
1982
1983         return ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, irdma_set_page);
1984 }
1985
1986 /**
1987  * irdma_hwreg_mr - send cqp command for memory registration
1988  * @iwdev: irdma device
1989  * @iwmr: irdma mr pointer
1990  * @access: access for MR
1991  */
1992 int
1993 irdma_hwreg_mr(struct irdma_device *iwdev, struct irdma_mr *iwmr,
1994                u16 access)
1995 {
1996         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
1997         struct irdma_reg_ns_stag_info *stag_info;
1998         struct ib_pd *pd = iwmr->ibmr.pd;
1999         struct irdma_pd *iwpd = to_iwpd(pd);
2000         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2001         struct irdma_cqp_request *cqp_request;
2002         struct cqp_cmds_info *cqp_info;
2003         int ret;
2004
2005         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2006         if (!cqp_request)
2007                 return -ENOMEM;
2008
2009         cqp_info = &cqp_request->info;
2010         stag_info = &cqp_info->in.u.mr_reg_non_shared.info;
2011         memset(stag_info, 0, sizeof(*stag_info));
2012         stag_info->va = iwpbl->user_base;
2013         stag_info->stag_idx = iwmr->stag >> IRDMA_CQPSQ_STAG_IDX_S;
2014         stag_info->stag_key = (u8)iwmr->stag;
2015         stag_info->total_len = iwmr->len;
2016         stag_info->all_memory = (pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) ? true : false;
2017         stag_info->access_rights = irdma_get_mr_access(access);
2018         stag_info->pd_id = iwpd->sc_pd.pd_id;
2019         if (stag_info->access_rights & IRDMA_ACCESS_FLAGS_ZERO_BASED)
2020                 stag_info->addr_type = IRDMA_ADDR_TYPE_ZERO_BASED;
2021         else
2022                 stag_info->addr_type = IRDMA_ADDR_TYPE_VA_BASED;
2023         stag_info->page_size = iwmr->page_size;
2024
2025         if (iwpbl->pbl_allocated) {
2026                 if (palloc->level == PBLE_LEVEL_1) {
2027                         stag_info->first_pm_pbl_index = palloc->level1.idx;
2028                         stag_info->chunk_size = 1;
2029                 } else {
2030                         stag_info->first_pm_pbl_index = palloc->level2.root.idx;
2031                         stag_info->chunk_size = 3;
2032                 }
2033         } else {
2034                 stag_info->reg_addr_pa = iwmr->pgaddrmem[0];
2035         }
2036
2037         cqp_info->cqp_cmd = IRDMA_OP_MR_REG_NON_SHARED;
2038         cqp_info->post_sq = 1;
2039         cqp_info->in.u.mr_reg_non_shared.dev = &iwdev->rf->sc_dev;
2040         cqp_info->in.u.mr_reg_non_shared.scratch = (uintptr_t)cqp_request;
2041         ret = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2042         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2043
2044         if (!ret)
2045                 iwmr->is_hwreg = 1;
2046
2047         return ret;
2048 }
2049
2050 /**
2051  * irdma_reg_user_mr - Register a user memory region
2052  * @pd: ptr of pd
2053  * @start: virtual start address
2054  * @len: length of mr
2055  * @virt: virtual address
2056  * @access: access of mr
2057  * @udata: user data
2058  */
2059 static struct ib_mr *
2060 irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,
2061                   u64 virt, int access,
2062                   struct ib_udata *udata)
2063 {
2064 #define IRDMA_MEM_REG_MIN_REQ_LEN offsetofend(struct irdma_mem_reg_req, sq_pages)
2065         struct irdma_device *iwdev = to_iwdev(pd->device);
2066         struct irdma_ucontext *ucontext;
2067         struct irdma_pble_alloc *palloc;
2068         struct irdma_pbl *iwpbl;
2069         struct irdma_mr *iwmr;
2070         struct ib_umem *region;
2071         struct irdma_mem_reg_req req = {};
2072         u32 total, stag = 0;
2073         u8 shadow_pgcnt = 1;
2074         bool use_pbles = false;
2075         unsigned long flags;
2076         int err = -EINVAL;
2077         int ret;
2078
2079         if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size)
2080                 return ERR_PTR(-EINVAL);
2081
2082         if (udata->inlen < IRDMA_MEM_REG_MIN_REQ_LEN)
2083                 return ERR_PTR(-EINVAL);
2084
2085         region = ib_umem_get(pd->uobject->context, start, len, access, 0);
2086
2087         if (IS_ERR(region)) {
2088                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
2089                             "Failed to create ib_umem region\n");
2090                 return (struct ib_mr *)region;
2091         }
2092
2093         if (ib_copy_from_udata(&req, udata, min(sizeof(req), udata->inlen))) {
2094                 ib_umem_release(region);
2095                 return ERR_PTR(-EFAULT);
2096         }
2097
2098         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2099         if (!iwmr) {
2100                 ib_umem_release(region);
2101                 return ERR_PTR(-ENOMEM);
2102         }
2103
2104         iwpbl = &iwmr->iwpbl;
2105         iwpbl->iwmr = iwmr;
2106         iwmr->region = region;
2107         iwmr->ibmr.pd = pd;
2108         iwmr->ibmr.device = pd->device;
2109         iwmr->ibmr.iova = virt;
2110         iwmr->page_size = IRDMA_HW_PAGE_SIZE;
2111         iwmr->page_msk = ~(IRDMA_HW_PAGE_SIZE - 1);
2112
2113         iwmr->len = region->length;
2114         iwpbl->user_base = virt;
2115         palloc = &iwpbl->pble_alloc;
2116         iwmr->type = req.reg_type;
2117         iwmr->page_cnt = irdma_ib_umem_num_dma_blocks(region, iwmr->page_size, virt);
2118
2119         switch (req.reg_type) {
2120         case IRDMA_MEMREG_TYPE_QP:
2121                 total = req.sq_pages + req.rq_pages + shadow_pgcnt;
2122                 if (total > iwmr->page_cnt) {
2123                         err = -EINVAL;
2124                         goto error;
2125                 }
2126                 total = req.sq_pages + req.rq_pages;
2127                 use_pbles = (total > 2);
2128                 err = irdma_handle_q_mem(iwdev, &req, iwpbl, use_pbles);
2129                 if (err)
2130                         goto error;
2131
2132                 ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext,
2133                                                      ibucontext);
2134                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
2135                 list_add_tail(&iwpbl->list, &ucontext->qp_reg_mem_list);
2136                 iwpbl->on_list = true;
2137                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
2138                 break;
2139         case IRDMA_MEMREG_TYPE_CQ:
2140                 if (iwdev->rf->sc_dev.hw_attrs.uk_attrs.feature_flags & IRDMA_FEATURE_CQ_RESIZE)
2141                         shadow_pgcnt = 0;
2142                 total = req.cq_pages + shadow_pgcnt;
2143                 if (total > iwmr->page_cnt) {
2144                         err = -EINVAL;
2145                         goto error;
2146                 }
2147
2148                 use_pbles = (req.cq_pages > 1);
2149                 err = irdma_handle_q_mem(iwdev, &req, iwpbl, use_pbles);
2150                 if (err)
2151                         goto error;
2152
2153                 ucontext = rdma_udata_to_drv_context(udata, struct irdma_ucontext,
2154                                                      ibucontext);
2155                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2156                 list_add_tail(&iwpbl->list, &ucontext->cq_reg_mem_list);
2157                 iwpbl->on_list = true;
2158                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2159                 break;
2160         case IRDMA_MEMREG_TYPE_MEM:
2161                 use_pbles = (iwmr->page_cnt != 1);
2162
2163                 err = irdma_setup_pbles(iwdev->rf, iwmr, use_pbles, false);
2164                 if (err)
2165                         goto error;
2166
2167                 if (use_pbles) {
2168                         ret = irdma_check_mr_contiguous(palloc,
2169                                                         iwmr->page_size);
2170                         if (ret) {
2171                                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2172                                 iwpbl->pbl_allocated = false;
2173                         }
2174                 }
2175
2176                 stag = irdma_create_stag(iwdev);
2177                 if (!stag) {
2178                         err = -ENOMEM;
2179                         goto error;
2180                 }
2181
2182                 iwmr->stag = stag;
2183                 iwmr->ibmr.rkey = stag;
2184                 iwmr->ibmr.lkey = stag;
2185                 iwmr->access = access;
2186                 err = irdma_hwreg_mr(iwdev, iwmr, access);
2187                 if (err) {
2188                         irdma_free_stag(iwdev, stag);
2189                         goto error;
2190                 }
2191
2192                 break;
2193         default:
2194                 goto error;
2195         }
2196
2197         iwmr->type = req.reg_type;
2198
2199         return &iwmr->ibmr;
2200
2201 error:
2202         if (palloc->level != PBLE_LEVEL_0 && iwpbl->pbl_allocated)
2203                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2204         ib_umem_release(region);
2205         kfree(iwmr);
2206
2207         return ERR_PTR(err);
2208 }
2209
2210 int
2211 irdma_hwdereg_mr(struct ib_mr *ib_mr)
2212 {
2213         struct irdma_device *iwdev = to_iwdev(ib_mr->device);
2214         struct irdma_mr *iwmr = to_iwmr(ib_mr);
2215         struct irdma_pd *iwpd = to_iwpd(ib_mr->pd);
2216         struct irdma_dealloc_stag_info *info;
2217         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2218         struct irdma_cqp_request *cqp_request;
2219         struct cqp_cmds_info *cqp_info;
2220         int status;
2221
2222         /*
2223          * Skip HW MR de-register when it is already de-registered during an MR re-reregister and the re-registration
2224          * fails
2225          */
2226         if (!iwmr->is_hwreg)
2227                 return 0;
2228
2229         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
2230         if (!cqp_request)
2231                 return -ENOMEM;
2232
2233         cqp_info = &cqp_request->info;
2234         info = &cqp_info->in.u.dealloc_stag.info;
2235         memset(info, 0, sizeof(*info));
2236         info->pd_id = iwpd->sc_pd.pd_id;
2237         info->stag_idx = RS_64_1(ib_mr->rkey, IRDMA_CQPSQ_STAG_IDX_S);
2238         info->mr = true;
2239         if (iwpbl->pbl_allocated)
2240                 info->dealloc_pbl = true;
2241
2242         cqp_info->cqp_cmd = IRDMA_OP_DEALLOC_STAG;
2243         cqp_info->post_sq = 1;
2244         cqp_info->in.u.dealloc_stag.dev = &iwdev->rf->sc_dev;
2245         cqp_info->in.u.dealloc_stag.scratch = (uintptr_t)cqp_request;
2246         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
2247         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
2248
2249         if (!status)
2250                 iwmr->is_hwreg = 0;
2251
2252         return status;
2253 }
2254
2255 /*
2256  * irdma_rereg_mr_trans - Re-register a user MR for a change translation. @iwmr: ptr of iwmr @start: virtual start
2257  * address @len: length of mr @virt: virtual address
2258  *
2259  * Re-register a user memory region when a change translation is requested. Re-register a new region while reusing the
2260  * stag from the original registration.
2261  */
2262 struct ib_mr *
2263 irdma_rereg_mr_trans(struct irdma_mr *iwmr, u64 start, u64 len,
2264                      u64 virt, struct ib_udata *udata)
2265 {
2266         struct irdma_device *iwdev = to_iwdev(iwmr->ibmr.device);
2267         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2268         struct irdma_pble_alloc *palloc = &iwpbl->pble_alloc;
2269         struct ib_pd *pd = iwmr->ibmr.pd;
2270         struct ib_umem *region;
2271         bool use_pbles;
2272         int err;
2273
2274         region = ib_umem_get(pd->uobject->context, start, len, iwmr->access, 0);
2275
2276         if (IS_ERR(region)) {
2277                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
2278                             "Failed to create ib_umem region\n");
2279                 return (struct ib_mr *)region;
2280         }
2281
2282         iwmr->region = region;
2283         iwmr->ibmr.iova = virt;
2284         iwmr->ibmr.pd = pd;
2285         iwmr->page_size = PAGE_SIZE;
2286
2287         iwmr->len = region->length;
2288         iwpbl->user_base = virt;
2289         iwmr->page_cnt = irdma_ib_umem_num_dma_blocks(region, iwmr->page_size,
2290                                                       virt);
2291
2292         use_pbles = (iwmr->page_cnt != 1);
2293
2294         err = irdma_setup_pbles(iwdev->rf, iwmr, use_pbles, false);
2295         if (err)
2296                 goto error;
2297
2298         if (use_pbles) {
2299                 err = irdma_check_mr_contiguous(palloc,
2300                                                 iwmr->page_size);
2301                 if (err) {
2302                         irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2303                         iwpbl->pbl_allocated = false;
2304                 }
2305         }
2306
2307         err = irdma_hwreg_mr(iwdev, iwmr, iwmr->access);
2308         if (err)
2309                 goto error;
2310
2311         return &iwmr->ibmr;
2312
2313 error:
2314         if (palloc->level != PBLE_LEVEL_0 && iwpbl->pbl_allocated) {
2315                 irdma_free_pble(iwdev->rf->pble_rsrc, palloc);
2316                 iwpbl->pbl_allocated = false;
2317         }
2318         ib_umem_release(region);
2319         iwmr->region = NULL;
2320
2321         return ERR_PTR(err);
2322 }
2323
2324 /**
2325  * irdma_reg_phys_mr - register kernel physical memory
2326  * @pd: ibpd pointer
2327  * @addr: physical address of memory to register
2328  * @size: size of memory to register
2329  * @access: Access rights
2330  * @iova_start: start of virtual address for physical buffers
2331  */
2332 struct ib_mr *
2333 irdma_reg_phys_mr(struct ib_pd *pd, u64 addr, u64 size, int access,
2334                   u64 *iova_start)
2335 {
2336         struct irdma_device *iwdev = to_iwdev(pd->device);
2337         struct irdma_pbl *iwpbl;
2338         struct irdma_mr *iwmr;
2339         u32 stag;
2340         int ret;
2341
2342         iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
2343         if (!iwmr)
2344                 return ERR_PTR(-ENOMEM);
2345
2346         iwmr->ibmr.pd = pd;
2347         iwmr->ibmr.device = pd->device;
2348         iwpbl = &iwmr->iwpbl;
2349         iwpbl->iwmr = iwmr;
2350         iwmr->type = IRDMA_MEMREG_TYPE_MEM;
2351         iwpbl->user_base = *iova_start;
2352         stag = irdma_create_stag(iwdev);
2353         if (!stag) {
2354                 ret = -ENOMEM;
2355                 goto err;
2356         }
2357
2358         iwmr->stag = stag;
2359         iwmr->ibmr.iova = *iova_start;
2360         iwmr->ibmr.rkey = stag;
2361         iwmr->ibmr.lkey = stag;
2362         iwmr->page_cnt = 1;
2363         iwmr->pgaddrmem[0] = addr;
2364         iwmr->len = size;
2365         iwmr->page_size = SZ_4K;
2366         ret = irdma_hwreg_mr(iwdev, iwmr, access);
2367         if (ret) {
2368                 irdma_free_stag(iwdev, stag);
2369                 goto err;
2370         }
2371
2372         return &iwmr->ibmr;
2373
2374 err:
2375         kfree(iwmr);
2376
2377         return ERR_PTR(ret);
2378 }
2379
2380 /**
2381  * irdma_get_dma_mr - register physical mem
2382  * @pd: ptr of pd
2383  * @acc: access for memory
2384  */
2385 static struct ib_mr *
2386 irdma_get_dma_mr(struct ib_pd *pd, int acc)
2387 {
2388         u64 kva = 0;
2389
2390         return irdma_reg_phys_mr(pd, 0, 0, acc, &kva);
2391 }
2392
2393 /**
2394  * irdma_del_memlist - Deleting pbl list entries for CQ/QP
2395  * @iwmr: iwmr for IB's user page addresses
2396  * @ucontext: ptr to user context
2397  */
2398 void
2399 irdma_del_memlist(struct irdma_mr *iwmr,
2400                   struct irdma_ucontext *ucontext)
2401 {
2402         struct irdma_pbl *iwpbl = &iwmr->iwpbl;
2403         unsigned long flags;
2404
2405         switch (iwmr->type) {
2406         case IRDMA_MEMREG_TYPE_CQ:
2407                 spin_lock_irqsave(&ucontext->cq_reg_mem_list_lock, flags);
2408                 if (iwpbl->on_list) {
2409                         iwpbl->on_list = false;
2410                         list_del(&iwpbl->list);
2411                 }
2412                 spin_unlock_irqrestore(&ucontext->cq_reg_mem_list_lock, flags);
2413                 break;
2414         case IRDMA_MEMREG_TYPE_QP:
2415                 spin_lock_irqsave(&ucontext->qp_reg_mem_list_lock, flags);
2416                 if (iwpbl->on_list) {
2417                         iwpbl->on_list = false;
2418                         list_del(&iwpbl->list);
2419                 }
2420                 spin_unlock_irqrestore(&ucontext->qp_reg_mem_list_lock, flags);
2421                 break;
2422         default:
2423                 break;
2424         }
2425 }
2426
2427 /**
2428  * irdma_copy_sg_list - copy sg list for qp
2429  * @sg_list: copied into sg_list
2430  * @sgl: copy from sgl
2431  * @num_sges: count of sg entries
2432  */
2433 static void
2434 irdma_copy_sg_list(struct irdma_sge *sg_list, struct ib_sge *sgl,
2435                    int num_sges)
2436 {
2437         unsigned int i;
2438
2439         for (i = 0; i < num_sges; i++) {
2440                 sg_list[i].tag_off = sgl[i].addr;
2441                 sg_list[i].len = sgl[i].length;
2442                 sg_list[i].stag = sgl[i].lkey;
2443         }
2444 }
2445
2446 /**
2447  * irdma_post_send -  kernel application wr
2448  * @ibqp: qp ptr for wr
2449  * @ib_wr: work request ptr
2450  * @bad_wr: return of bad wr if err
2451  */
2452 static int
2453 irdma_post_send(struct ib_qp *ibqp,
2454                 const struct ib_send_wr *ib_wr,
2455                 const struct ib_send_wr **bad_wr)
2456 {
2457         struct irdma_qp *iwqp;
2458         struct irdma_qp_uk *ukqp;
2459         struct irdma_sc_dev *dev;
2460         struct irdma_post_sq_info info;
2461         int err = 0;
2462         unsigned long flags;
2463         bool inv_stag;
2464         struct irdma_ah *ah;
2465
2466         iwqp = to_iwqp(ibqp);
2467         ukqp = &iwqp->sc_qp.qp_uk;
2468         dev = &iwqp->iwdev->rf->sc_dev;
2469
2470         spin_lock_irqsave(&iwqp->lock, flags);
2471         while (ib_wr) {
2472                 memset(&info, 0, sizeof(info));
2473                 inv_stag = false;
2474                 info.wr_id = (ib_wr->wr_id);
2475                 if ((ib_wr->send_flags & IB_SEND_SIGNALED) || iwqp->sig_all)
2476                         info.signaled = true;
2477                 if (ib_wr->send_flags & IB_SEND_FENCE)
2478                         info.read_fence = true;
2479                 switch (ib_wr->opcode) {
2480                 case IB_WR_SEND_WITH_IMM:
2481                         if (ukqp->qp_caps & IRDMA_SEND_WITH_IMM) {
2482                                 info.imm_data_valid = true;
2483                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
2484                         } else {
2485                                 err = -EINVAL;
2486                                 break;
2487                         }
2488                         /* fallthrough */
2489                 case IB_WR_SEND:
2490                 case IB_WR_SEND_WITH_INV:
2491                         if (ib_wr->opcode == IB_WR_SEND ||
2492                             ib_wr->opcode == IB_WR_SEND_WITH_IMM) {
2493                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
2494                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL;
2495                                 else
2496                                         info.op_type = IRDMA_OP_TYPE_SEND;
2497                         } else {
2498                                 if (ib_wr->send_flags & IB_SEND_SOLICITED)
2499                                         info.op_type = IRDMA_OP_TYPE_SEND_SOL_INV;
2500                                 else
2501                                         info.op_type = IRDMA_OP_TYPE_SEND_INV;
2502                                 info.stag_to_inv = ib_wr->ex.invalidate_rkey;
2503                         }
2504
2505                         info.op.send.num_sges = ib_wr->num_sge;
2506                         info.op.send.sg_list = (struct irdma_sge *)ib_wr->sg_list;
2507                         if (iwqp->ibqp.qp_type == IB_QPT_UD ||
2508                             iwqp->ibqp.qp_type == IB_QPT_GSI) {
2509                                 ah = to_iwah(ud_wr(ib_wr)->ah);
2510                                 info.op.send.ah_id = ah->sc_ah.ah_info.ah_idx;
2511                                 info.op.send.qkey = ud_wr(ib_wr)->remote_qkey;
2512                                 info.op.send.dest_qp = ud_wr(ib_wr)->remote_qpn;
2513                         }
2514
2515                         if (ib_wr->send_flags & IB_SEND_INLINE)
2516                                 err = irdma_uk_inline_send(ukqp, &info, false);
2517                         else
2518                                 err = irdma_uk_send(ukqp, &info, false);
2519                         break;
2520                 case IB_WR_RDMA_WRITE_WITH_IMM:
2521                         if (ukqp->qp_caps & IRDMA_WRITE_WITH_IMM) {
2522                                 info.imm_data_valid = true;
2523                                 info.imm_data = ntohl(ib_wr->ex.imm_data);
2524                         } else {
2525                                 err = -EINVAL;
2526                                 break;
2527                         }
2528                         /* fallthrough */
2529                 case IB_WR_RDMA_WRITE:
2530                         if (ib_wr->send_flags & IB_SEND_SOLICITED)
2531                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE_SOL;
2532                         else
2533                                 info.op_type = IRDMA_OP_TYPE_RDMA_WRITE;
2534
2535                         info.op.rdma_write.num_lo_sges = ib_wr->num_sge;
2536                         info.op.rdma_write.lo_sg_list = (void *)ib_wr->sg_list;
2537                         info.op.rdma_write.rem_addr.tag_off = rdma_wr(ib_wr)->remote_addr;
2538                         info.op.rdma_write.rem_addr.stag = rdma_wr(ib_wr)->rkey;
2539                         if (ib_wr->send_flags & IB_SEND_INLINE)
2540                                 err = irdma_uk_inline_rdma_write(ukqp, &info, false);
2541                         else
2542                                 err = irdma_uk_rdma_write(ukqp, &info, false);
2543                         break;
2544                 case IB_WR_RDMA_READ_WITH_INV:
2545                         inv_stag = true;
2546                         /* fallthrough */
2547                 case IB_WR_RDMA_READ:
2548                         if (ib_wr->num_sge >
2549                             dev->hw_attrs.uk_attrs.max_hw_read_sges) {
2550                                 err = -EINVAL;
2551                                 break;
2552                         }
2553                         info.op_type = IRDMA_OP_TYPE_RDMA_READ;
2554                         info.op.rdma_read.rem_addr.tag_off = rdma_wr(ib_wr)->remote_addr;
2555                         info.op.rdma_read.rem_addr.stag = rdma_wr(ib_wr)->rkey;
2556                         info.op.rdma_read.lo_sg_list = (void *)ib_wr->sg_list;
2557                         info.op.rdma_read.num_lo_sges = ib_wr->num_sge;
2558                         err = irdma_uk_rdma_read(ukqp, &info, inv_stag, false);
2559                         break;
2560                 case IB_WR_LOCAL_INV:
2561                         info.op_type = IRDMA_OP_TYPE_INV_STAG;
2562                         info.op.inv_local_stag.target_stag = ib_wr->ex.invalidate_rkey;
2563                         err = irdma_uk_stag_local_invalidate(ukqp, &info, true);
2564                         break;
2565                 case IB_WR_REG_MR:{
2566                                 struct irdma_mr *iwmr = to_iwmr(reg_wr(ib_wr)->mr);
2567                                 struct irdma_pble_alloc *palloc = &iwmr->iwpbl.pble_alloc;
2568                                 struct irdma_fast_reg_stag_info stag_info = {0};
2569
2570                                 stag_info.signaled = info.signaled;
2571                                 stag_info.read_fence = info.read_fence;
2572                                 stag_info.access_rights = irdma_get_mr_access(reg_wr(ib_wr)->access);
2573                                 stag_info.stag_key = reg_wr(ib_wr)->key & 0xff;
2574                                 stag_info.stag_idx = reg_wr(ib_wr)->key >> 8;
2575                                 stag_info.page_size = reg_wr(ib_wr)->mr->page_size;
2576                                 stag_info.wr_id = ib_wr->wr_id;
2577                                 stag_info.addr_type = IRDMA_ADDR_TYPE_VA_BASED;
2578                                 stag_info.va = (void *)(uintptr_t)iwmr->ibmr.iova;
2579                                 stag_info.total_len = iwmr->ibmr.length;
2580                                 if (palloc->level == PBLE_LEVEL_2) {
2581                                         stag_info.chunk_size = 3;
2582                                         stag_info.first_pm_pbl_index = palloc->level2.root.idx;
2583                                 } else {
2584                                         stag_info.chunk_size = 1;
2585                                         stag_info.first_pm_pbl_index = palloc->level1.idx;
2586                                 }
2587                                 stag_info.local_fence = ib_wr->send_flags & IB_SEND_FENCE;
2588                                 err = irdma_sc_mr_fast_register(&iwqp->sc_qp, &stag_info,
2589                                                                 true);
2590                                 break;
2591                         }
2592                 default:
2593                         err = -EINVAL;
2594                         irdma_debug(iwdev_to_idev(iwqp->iwdev),
2595                                     IRDMA_DEBUG_VERBS,
2596                                     "upost_send bad opcode = 0x%x\n",
2597                                     ib_wr->opcode);
2598                         break;
2599                 }
2600
2601                 if (err)
2602                         break;
2603                 ib_wr = ib_wr->next;
2604         }
2605
2606         if (!iwqp->flush_issued) {
2607                 if (iwqp->hw_iwarp_state <= IRDMA_QP_STATE_RTS)
2608                         irdma_uk_qp_post_wr(ukqp);
2609                 spin_unlock_irqrestore(&iwqp->lock, flags);
2610         } else {
2611                 spin_unlock_irqrestore(&iwqp->lock, flags);
2612                 irdma_sched_qp_flush_work(iwqp);
2613         }
2614         if (err)
2615                 *bad_wr = ib_wr;
2616
2617         return err;
2618 }
2619
2620 /**
2621  * irdma_post_recv - post receive wr for kernel application
2622  * @ibqp: ib qp pointer
2623  * @ib_wr: work request for receive
2624  * @bad_wr: bad wr caused an error
2625  */
2626 static int
2627 irdma_post_recv(struct ib_qp *ibqp,
2628                 const struct ib_recv_wr *ib_wr,
2629                 const struct ib_recv_wr **bad_wr)
2630 {
2631         struct irdma_qp *iwqp = to_iwqp(ibqp);
2632         struct irdma_qp_uk *ukqp = &iwqp->sc_qp.qp_uk;
2633         struct irdma_post_rq_info post_recv = {0};
2634         struct irdma_sge *sg_list = iwqp->sg_list;
2635         unsigned long flags;
2636         int err = 0;
2637
2638         spin_lock_irqsave(&iwqp->lock, flags);
2639
2640         while (ib_wr) {
2641                 if (ib_wr->num_sge > ukqp->max_rq_frag_cnt) {
2642                         err = -EINVAL;
2643                         goto out;
2644                 }
2645                 post_recv.num_sges = ib_wr->num_sge;
2646                 post_recv.wr_id = ib_wr->wr_id;
2647                 irdma_copy_sg_list(sg_list, ib_wr->sg_list, ib_wr->num_sge);
2648                 post_recv.sg_list = sg_list;
2649                 err = irdma_uk_post_receive(ukqp, &post_recv);
2650                 if (err) {
2651                         irdma_debug(iwdev_to_idev(iwqp->iwdev),
2652                                     IRDMA_DEBUG_VERBS, "post_recv err %d\n",
2653                                     err);
2654                         goto out;
2655                 }
2656
2657                 ib_wr = ib_wr->next;
2658         }
2659
2660 out:
2661         spin_unlock_irqrestore(&iwqp->lock, flags);
2662         if (iwqp->flush_issued)
2663                 irdma_sched_qp_flush_work(iwqp);
2664         if (err)
2665                 *bad_wr = ib_wr;
2666
2667         return err;
2668 }
2669
2670 /**
2671  * irdma_flush_err_to_ib_wc_status - return change flush error code to IB status
2672  * @opcode: iwarp flush code
2673  */
2674 static enum ib_wc_status
2675 irdma_flush_err_to_ib_wc_status(enum irdma_flush_opcode opcode)
2676 {
2677         switch (opcode) {
2678         case FLUSH_PROT_ERR:
2679                 return IB_WC_LOC_PROT_ERR;
2680         case FLUSH_REM_ACCESS_ERR:
2681                 return IB_WC_REM_ACCESS_ERR;
2682         case FLUSH_LOC_QP_OP_ERR:
2683                 return IB_WC_LOC_QP_OP_ERR;
2684         case FLUSH_REM_OP_ERR:
2685                 return IB_WC_REM_OP_ERR;
2686         case FLUSH_LOC_LEN_ERR:
2687                 return IB_WC_LOC_LEN_ERR;
2688         case FLUSH_GENERAL_ERR:
2689                 return IB_WC_WR_FLUSH_ERR;
2690         case FLUSH_MW_BIND_ERR:
2691                 return IB_WC_MW_BIND_ERR;
2692         case FLUSH_REM_INV_REQ_ERR:
2693                 return IB_WC_REM_INV_REQ_ERR;
2694         case FLUSH_RETRY_EXC_ERR:
2695                 return IB_WC_RETRY_EXC_ERR;
2696         case FLUSH_FATAL_ERR:
2697         default:
2698                 return IB_WC_FATAL_ERR;
2699         }
2700 }
2701
2702 static inline void
2703 set_ib_wc_op_sq(struct irdma_cq_poll_info *cq_poll_info,
2704                 struct ib_wc *entry)
2705 {
2706         struct irdma_sc_qp *qp;
2707
2708         switch (cq_poll_info->op_type) {
2709         case IRDMA_OP_TYPE_RDMA_WRITE:
2710         case IRDMA_OP_TYPE_RDMA_WRITE_SOL:
2711                 entry->opcode = IB_WC_RDMA_WRITE;
2712                 break;
2713         case IRDMA_OP_TYPE_RDMA_READ_INV_STAG:
2714         case IRDMA_OP_TYPE_RDMA_READ:
2715                 entry->opcode = IB_WC_RDMA_READ;
2716                 break;
2717         case IRDMA_OP_TYPE_SEND_SOL:
2718         case IRDMA_OP_TYPE_SEND_SOL_INV:
2719         case IRDMA_OP_TYPE_SEND_INV:
2720         case IRDMA_OP_TYPE_SEND:
2721                 entry->opcode = IB_WC_SEND;
2722                 break;
2723         case IRDMA_OP_TYPE_FAST_REG_NSMR:
2724                 entry->opcode = IB_WC_REG_MR;
2725                 break;
2726         case IRDMA_OP_TYPE_INV_STAG:
2727                 entry->opcode = IB_WC_LOCAL_INV;
2728                 break;
2729         default:
2730                 qp = cq_poll_info->qp_handle;
2731                 ibdev_err(irdma_get_ibdev(qp->dev), "Invalid opcode = %d in CQE\n",
2732                           cq_poll_info->op_type);
2733                 entry->status = IB_WC_GENERAL_ERR;
2734         }
2735 }
2736
2737 static inline void
2738 set_ib_wc_op_rq(struct irdma_cq_poll_info *cq_poll_info,
2739                 struct ib_wc *entry, bool send_imm_support)
2740 {
2741         /**
2742          * iWARP does not support sendImm, so the presence of Imm data
2743          * must be WriteImm.
2744          */
2745         if (!send_imm_support) {
2746                 entry->opcode = cq_poll_info->imm_valid ?
2747                     IB_WC_RECV_RDMA_WITH_IMM : IB_WC_RECV;
2748                 return;
2749         }
2750         switch (cq_poll_info->op_type) {
2751         case IB_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE:
2752         case IB_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE:
2753                 entry->opcode = IB_WC_RECV_RDMA_WITH_IMM;
2754                 break;
2755         default:
2756                 entry->opcode = IB_WC_RECV;
2757         }
2758 }
2759
2760 /**
2761  * irdma_process_cqe - process cqe info
2762  * @entry: processed cqe
2763  * @cq_poll_info: cqe info
2764  */
2765 static void
2766 irdma_process_cqe(struct ib_wc *entry,
2767                   struct irdma_cq_poll_info *cq_poll_info)
2768 {
2769         struct irdma_sc_qp *qp;
2770
2771         entry->wc_flags = 0;
2772         entry->pkey_index = 0;
2773         entry->wr_id = cq_poll_info->wr_id;
2774
2775         qp = cq_poll_info->qp_handle;
2776         entry->qp = qp->qp_uk.back_qp;
2777
2778         if (cq_poll_info->error) {
2779                 entry->status = (cq_poll_info->comp_status == IRDMA_COMPL_STATUS_FLUSHED) ?
2780                     irdma_flush_err_to_ib_wc_status(cq_poll_info->minor_err) : IB_WC_GENERAL_ERR;
2781
2782                 entry->vendor_err = cq_poll_info->major_err << 16 |
2783                     cq_poll_info->minor_err;
2784         } else {
2785                 entry->status = IB_WC_SUCCESS;
2786                 if (cq_poll_info->imm_valid) {
2787                         entry->ex.imm_data = htonl(cq_poll_info->imm_data);
2788                         entry->wc_flags |= IB_WC_WITH_IMM;
2789                 }
2790                 if (cq_poll_info->ud_smac_valid) {
2791                         ether_addr_copy(entry->smac, cq_poll_info->ud_smac);
2792                         entry->wc_flags |= IB_WC_WITH_SMAC;
2793                 }
2794
2795                 if (cq_poll_info->ud_vlan_valid) {
2796                         u16 vlan = cq_poll_info->ud_vlan & EVL_VLID_MASK;
2797
2798                         entry->sl = cq_poll_info->ud_vlan >> VLAN_PRIO_SHIFT;
2799                         if (vlan) {
2800                                 entry->vlan_id = vlan;
2801                                 entry->wc_flags |= IB_WC_WITH_VLAN;
2802                         }
2803                 } else {
2804                         entry->sl = 0;
2805                 }
2806         }
2807
2808         if (cq_poll_info->q_type == IRDMA_CQE_QTYPE_SQ) {
2809                 set_ib_wc_op_sq(cq_poll_info, entry);
2810         } else {
2811                 set_ib_wc_op_rq(cq_poll_info, entry,
2812                                 qp->qp_uk.qp_caps & IRDMA_SEND_WITH_IMM ?
2813                                 true : false);
2814                 if (qp->qp_uk.qp_type != IRDMA_QP_TYPE_ROCE_UD &&
2815                     cq_poll_info->stag_invalid_set) {
2816                         entry->ex.invalidate_rkey = cq_poll_info->inv_stag;
2817                         entry->wc_flags |= IB_WC_WITH_INVALIDATE;
2818                 }
2819         }
2820
2821         if (qp->qp_uk.qp_type == IRDMA_QP_TYPE_ROCE_UD) {
2822                 entry->src_qp = cq_poll_info->ud_src_qpn;
2823                 entry->slid = 0;
2824                 entry->wc_flags |=
2825                     (IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE);
2826                 entry->network_hdr_type = cq_poll_info->ipv4 ?
2827                     RDMA_NETWORK_IPV4 :
2828                     RDMA_NETWORK_IPV6;
2829         } else {
2830                 entry->src_qp = cq_poll_info->qp_id;
2831         }
2832
2833         entry->byte_len = cq_poll_info->bytes_xfered;
2834 }
2835
2836 /**
2837  * irdma_poll_one - poll one entry of the CQ
2838  * @ukcq: ukcq to poll
2839  * @cur_cqe: current CQE info to be filled in
2840  * @entry: ibv_wc object to be filled for non-extended CQ or NULL for extended CQ
2841  *
2842  * Returns the internal irdma device error code or 0 on success
2843  */
2844 static inline int
2845 irdma_poll_one(struct irdma_cq_uk *ukcq,
2846                struct irdma_cq_poll_info *cur_cqe,
2847                struct ib_wc *entry)
2848 {
2849         int ret = irdma_uk_cq_poll_cmpl(ukcq, cur_cqe);
2850
2851         if (ret)
2852                 return ret;
2853
2854         irdma_process_cqe(entry, cur_cqe);
2855
2856         return 0;
2857 }
2858
2859 /**
2860  * __irdma_poll_cq - poll cq for completion (kernel apps)
2861  * @iwcq: cq to poll
2862  * @num_entries: number of entries to poll
2863  * @entry: wr of a completed entry
2864  */
2865 static int
2866 __irdma_poll_cq(struct irdma_cq *iwcq, int num_entries, struct ib_wc *entry)
2867 {
2868         struct list_head *tmp_node, *list_node;
2869         struct irdma_cq_buf *last_buf = NULL;
2870         struct irdma_cq_poll_info *cur_cqe = &iwcq->cur_cqe;
2871         struct irdma_cq_buf *cq_buf;
2872         int ret;
2873         struct irdma_device *iwdev;
2874         struct irdma_cq_uk *ukcq;
2875         bool cq_new_cqe = false;
2876         int resized_bufs = 0;
2877         int npolled = 0;
2878
2879         iwdev = to_iwdev(iwcq->ibcq.device);
2880         ukcq = &iwcq->sc_cq.cq_uk;
2881
2882         /* go through the list of previously resized CQ buffers */
2883         list_for_each_safe(list_node, tmp_node, &iwcq->resize_list) {
2884                 cq_buf = container_of(list_node, struct irdma_cq_buf, list);
2885                 while (npolled < num_entries) {
2886                         ret = irdma_poll_one(&cq_buf->cq_uk, cur_cqe, entry + npolled);
2887                         if (!ret) {
2888                                 ++npolled;
2889                                 cq_new_cqe = true;
2890                                 continue;
2891                         }
2892                         if (ret == -ENOENT)
2893                                 break;
2894                         /* QP using the CQ is destroyed. Skip reporting this CQE */
2895                         if (ret == -EFAULT) {
2896                                 cq_new_cqe = true;
2897                                 continue;
2898                         }
2899                         goto error;
2900                 }
2901
2902                 /* save the resized CQ buffer which received the last cqe */
2903                 if (cq_new_cqe)
2904                         last_buf = cq_buf;
2905                 cq_new_cqe = false;
2906         }
2907
2908         /* check the current CQ for new cqes */
2909         while (npolled < num_entries) {
2910                 ret = irdma_poll_one(ukcq, cur_cqe, entry + npolled);
2911                 if (ret == -ENOENT) {
2912                         ret = irdma_generated_cmpls(iwcq, cur_cqe);
2913                         if (!ret)
2914                                 irdma_process_cqe(entry + npolled, cur_cqe);
2915                 }
2916                 if (!ret) {
2917                         ++npolled;
2918                         cq_new_cqe = true;
2919                         continue;
2920                 }
2921
2922                 if (ret == -ENOENT)
2923                         break;
2924                 /* QP using the CQ is destroyed. Skip reporting this CQE */
2925                 if (ret == -EFAULT) {
2926                         cq_new_cqe = true;
2927                         continue;
2928                 }
2929                 goto error;
2930         }
2931
2932         if (cq_new_cqe)
2933                 /* all previous CQ resizes are complete */
2934                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, NULL);
2935         else if (last_buf)
2936                 /* only CQ resizes up to the last_buf are complete */
2937                 resized_bufs = irdma_process_resize_list(iwcq, iwdev, last_buf);
2938         if (resized_bufs)
2939                 /* report to the HW the number of complete CQ resizes */
2940                 irdma_uk_cq_set_resized_cnt(ukcq, resized_bufs);
2941
2942         return npolled;
2943 error:
2944         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
2945                     "%s: Error polling CQ, irdma_err: %d\n", __func__, ret);
2946
2947         return ret;
2948 }
2949
2950 /**
2951  * irdma_poll_cq - poll cq for completion (kernel apps)
2952  * @ibcq: cq to poll
2953  * @num_entries: number of entries to poll
2954  * @entry: wr of a completed entry
2955  */
2956 static int
2957 irdma_poll_cq(struct ib_cq *ibcq, int num_entries,
2958               struct ib_wc *entry)
2959 {
2960         struct irdma_cq *iwcq;
2961         unsigned long flags;
2962         int ret;
2963
2964         iwcq = to_iwcq(ibcq);
2965
2966         spin_lock_irqsave(&iwcq->lock, flags);
2967         ret = __irdma_poll_cq(iwcq, num_entries, entry);
2968         spin_unlock_irqrestore(&iwcq->lock, flags);
2969
2970         return ret;
2971 }
2972
2973 /**
2974  * irdma_req_notify_cq - arm cq kernel application
2975  * @ibcq: cq to arm
2976  * @notify_flags: notofication flags
2977  */
2978 static int
2979 irdma_req_notify_cq(struct ib_cq *ibcq,
2980                     enum ib_cq_notify_flags notify_flags)
2981 {
2982         struct irdma_cq *iwcq;
2983         struct irdma_cq_uk *ukcq;
2984         unsigned long flags;
2985         enum irdma_cmpl_notify cq_notify = IRDMA_CQ_COMPL_EVENT;
2986         bool promo_event = false;
2987         int ret = 0;
2988
2989         iwcq = to_iwcq(ibcq);
2990         ukcq = &iwcq->sc_cq.cq_uk;
2991
2992         spin_lock_irqsave(&iwcq->lock, flags);
2993         if (notify_flags == IB_CQ_SOLICITED) {
2994                 cq_notify = IRDMA_CQ_COMPL_SOLICITED;
2995         } else {
2996                 if (iwcq->last_notify == IRDMA_CQ_COMPL_SOLICITED)
2997                         promo_event = true;
2998         }
2999
3000         if (!atomic_cmpxchg(&iwcq->armed, 0, 1) || promo_event) {
3001                 iwcq->last_notify = cq_notify;
3002                 irdma_uk_cq_request_notification(ukcq, cq_notify);
3003         }
3004
3005         if ((notify_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
3006             (!irdma_cq_empty(iwcq) || !list_empty(&iwcq->cmpl_generated)))
3007                 ret = 1;
3008         spin_unlock_irqrestore(&iwcq->lock, flags);
3009
3010         return ret;
3011 }
3012
3013 /**
3014  * mcast_list_add -  Add a new mcast item to list
3015  * @rf: RDMA PCI function
3016  * @new_elem: pointer to element to add
3017  */
3018 static void
3019 mcast_list_add(struct irdma_pci_f *rf,
3020                struct mc_table_list *new_elem)
3021 {
3022         list_add(&new_elem->list, &rf->mc_qht_list.list);
3023 }
3024
3025 /**
3026  * mcast_list_del - Remove an mcast item from list
3027  * @mc_qht_elem: pointer to mcast table list element
3028  */
3029 static void
3030 mcast_list_del(struct mc_table_list *mc_qht_elem)
3031 {
3032         if (mc_qht_elem)
3033                 list_del(&mc_qht_elem->list);
3034 }
3035
3036 /**
3037  * mcast_list_lookup_ip - Search mcast list for address
3038  * @rf: RDMA PCI function
3039  * @ip_mcast: pointer to mcast IP address
3040  */
3041 static struct mc_table_list *
3042 mcast_list_lookup_ip(struct irdma_pci_f *rf,
3043                      u32 *ip_mcast)
3044 {
3045         struct mc_table_list *mc_qht_el;
3046         struct list_head *pos, *q;
3047
3048         list_for_each_safe(pos, q, &rf->mc_qht_list.list) {
3049                 mc_qht_el = list_entry(pos, struct mc_table_list, list);
3050                 if (!memcmp(mc_qht_el->mc_info.dest_ip, ip_mcast,
3051                             sizeof(mc_qht_el->mc_info.dest_ip)))
3052                         return mc_qht_el;
3053         }
3054
3055         return NULL;
3056 }
3057
3058 /**
3059  * irdma_mcast_cqp_op - perform a mcast cqp operation
3060  * @iwdev: irdma device
3061  * @mc_grp_ctx: mcast group info
3062  * @op: operation
3063  *
3064  * returns error status
3065  */
3066 static int
3067 irdma_mcast_cqp_op(struct irdma_device *iwdev,
3068                    struct irdma_mcast_grp_info *mc_grp_ctx, u8 op)
3069 {
3070         struct cqp_cmds_info *cqp_info;
3071         struct irdma_cqp_request *cqp_request;
3072         int status;
3073
3074         cqp_request = irdma_alloc_and_get_cqp_request(&iwdev->rf->cqp, true);
3075         if (!cqp_request)
3076                 return -ENOMEM;
3077
3078         cqp_request->info.in.u.mc_create.info = *mc_grp_ctx;
3079         cqp_info = &cqp_request->info;
3080         cqp_info->cqp_cmd = op;
3081         cqp_info->post_sq = 1;
3082         cqp_info->in.u.mc_create.scratch = (uintptr_t)cqp_request;
3083         cqp_info->in.u.mc_create.cqp = &iwdev->rf->cqp.sc_cqp;
3084         status = irdma_handle_cqp_op(iwdev->rf, cqp_request);
3085         irdma_put_cqp_request(&iwdev->rf->cqp, cqp_request);
3086
3087         return status;
3088 }
3089
3090 /**
3091  * irdma_attach_mcast - attach a qp to a multicast group
3092  * @ibqp: ptr to qp
3093  * @ibgid: pointer to global ID
3094  * @lid: local ID
3095  *
3096  * returns error status
3097  */
3098 static int
3099 irdma_attach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
3100 {
3101         struct irdma_qp *iwqp = to_iwqp(ibqp);
3102         struct irdma_device *iwdev = iwqp->iwdev;
3103         struct irdma_pci_f *rf = iwdev->rf;
3104         struct mc_table_list *mc_qht_elem;
3105         struct irdma_mcast_grp_ctx_entry_info mcg_info = {0};
3106         unsigned long flags;
3107         u32 ip_addr[4] = {0};
3108         u32 mgn;
3109         u32 no_mgs;
3110         int ret = 0;
3111         bool ipv4;
3112         u16 vlan_id;
3113         union {
3114                 struct sockaddr saddr;
3115                 struct sockaddr_in saddr_in;
3116                 struct sockaddr_in6 saddr_in6;
3117         } sgid_addr;
3118         unsigned char dmac[ETH_ALEN];
3119
3120         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
3121
3122         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid)) {
3123                 irdma_copy_ip_ntohl(ip_addr,
3124                                     sgid_addr.saddr_in6.sin6_addr.__u6_addr.__u6_addr32);
3125                 irdma_netdev_vlan_ipv6(ip_addr, &vlan_id, NULL);
3126                 ipv4 = false;
3127                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
3128                             "qp_id=%d, IP6address=%pI6\n", ibqp->qp_num,
3129                             ip_addr);
3130                 irdma_mcast_mac_v6(ip_addr, dmac);
3131         } else {
3132                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
3133                 ipv4 = true;
3134                 vlan_id = irdma_get_vlan_ipv4(ip_addr);
3135                 irdma_mcast_mac_v4(ip_addr, dmac);
3136                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
3137                             "qp_id=%d, IP4address=%pI4, MAC=%pM\n",
3138                             ibqp->qp_num, ip_addr, dmac);
3139         }
3140
3141         spin_lock_irqsave(&rf->qh_list_lock, flags);
3142         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
3143         if (!mc_qht_elem) {
3144                 struct irdma_dma_mem *dma_mem_mc;
3145
3146                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3147                 mc_qht_elem = kzalloc(sizeof(*mc_qht_elem), GFP_KERNEL);
3148                 if (!mc_qht_elem)
3149                         return -ENOMEM;
3150
3151                 mc_qht_elem->mc_info.ipv4_valid = ipv4;
3152                 memcpy(mc_qht_elem->mc_info.dest_ip, ip_addr,
3153                        sizeof(mc_qht_elem->mc_info.dest_ip));
3154                 ret = irdma_alloc_rsrc(rf, rf->allocated_mcgs, rf->max_mcg,
3155                                        &mgn, &rf->next_mcg);
3156                 if (ret) {
3157                         kfree(mc_qht_elem);
3158                         return -ENOMEM;
3159                 }
3160
3161                 mc_qht_elem->mc_info.mgn = mgn;
3162                 dma_mem_mc = &mc_qht_elem->mc_grp_ctx.dma_mem_mc;
3163                 dma_mem_mc->size = sizeof(u64)* IRDMA_MAX_MGS_PER_CTX;
3164                 dma_mem_mc->va = irdma_allocate_dma_mem(&rf->hw, dma_mem_mc,
3165                                                         dma_mem_mc->size,
3166                                                         IRDMA_HW_PAGE_SIZE);
3167                 if (!dma_mem_mc->va) {
3168                         irdma_free_rsrc(rf, rf->allocated_mcgs, mgn);
3169                         kfree(mc_qht_elem);
3170                         return -ENOMEM;
3171                 }
3172
3173                 mc_qht_elem->mc_grp_ctx.mg_id = (u16)mgn;
3174                 memcpy(mc_qht_elem->mc_grp_ctx.dest_ip_addr, ip_addr,
3175                        sizeof(mc_qht_elem->mc_grp_ctx.dest_ip_addr));
3176                 mc_qht_elem->mc_grp_ctx.ipv4_valid = ipv4;
3177                 mc_qht_elem->mc_grp_ctx.vlan_id = vlan_id;
3178                 if (vlan_id < VLAN_N_VID)
3179                         mc_qht_elem->mc_grp_ctx.vlan_valid = true;
3180                 mc_qht_elem->mc_grp_ctx.hmc_fcn_id = iwdev->rf->sc_dev.hmc_fn_id;
3181                 mc_qht_elem->mc_grp_ctx.qs_handle =
3182                     iwqp->sc_qp.vsi->qos[iwqp->sc_qp.user_pri].qs_handle;
3183                 ether_addr_copy(mc_qht_elem->mc_grp_ctx.dest_mac_addr, dmac);
3184
3185                 spin_lock_irqsave(&rf->qh_list_lock, flags);
3186                 mcast_list_add(rf, mc_qht_elem);
3187         } else {
3188                 if (mc_qht_elem->mc_grp_ctx.no_of_mgs ==
3189                     IRDMA_MAX_MGS_PER_CTX) {
3190                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3191                         return -ENOMEM;
3192                 }
3193         }
3194
3195         mcg_info.qp_id = iwqp->ibqp.qp_num;
3196         no_mgs = mc_qht_elem->mc_grp_ctx.no_of_mgs;
3197         irdma_sc_add_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
3198         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3199
3200         /* Only if there is a change do we need to modify or create */
3201         if (!no_mgs) {
3202                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
3203                                          IRDMA_OP_MC_CREATE);
3204         } else if (no_mgs != mc_qht_elem->mc_grp_ctx.no_of_mgs) {
3205                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
3206                                          IRDMA_OP_MC_MODIFY);
3207         } else {
3208                 return 0;
3209         }
3210
3211         if (ret)
3212                 goto error;
3213
3214         return 0;
3215
3216 error:
3217         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
3218         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
3219                 mcast_list_del(mc_qht_elem);
3220                 irdma_free_dma_mem(&rf->hw,
3221                                    &mc_qht_elem->mc_grp_ctx.dma_mem_mc);
3222                 irdma_free_rsrc(rf, rf->allocated_mcgs,
3223                                 mc_qht_elem->mc_grp_ctx.mg_id);
3224                 kfree(mc_qht_elem);
3225         }
3226
3227         return ret;
3228 }
3229
3230 /**
3231  * irdma_detach_mcast - detach a qp from a multicast group
3232  * @ibqp: ptr to qp
3233  * @ibgid: pointer to global ID
3234  * @lid: local ID
3235  *
3236  * returns error status
3237  */
3238 static int
3239 irdma_detach_mcast(struct ib_qp *ibqp, union ib_gid *ibgid, u16 lid)
3240 {
3241         struct irdma_qp *iwqp = to_iwqp(ibqp);
3242         struct irdma_device *iwdev = iwqp->iwdev;
3243         struct irdma_pci_f *rf = iwdev->rf;
3244         u32 ip_addr[4] = {0};
3245         struct mc_table_list *mc_qht_elem;
3246         struct irdma_mcast_grp_ctx_entry_info mcg_info = {0};
3247         int ret;
3248         unsigned long flags;
3249         union {
3250                 struct sockaddr saddr;
3251                 struct sockaddr_in saddr_in;
3252                 struct sockaddr_in6 saddr_in6;
3253         } sgid_addr;
3254
3255         rdma_gid2ip((struct sockaddr *)&sgid_addr, ibgid);
3256         if (!ipv6_addr_v4mapped((struct in6_addr *)ibgid))
3257                 irdma_copy_ip_ntohl(ip_addr,
3258                                     sgid_addr.saddr_in6.sin6_addr.__u6_addr.__u6_addr32);
3259         else
3260                 ip_addr[0] = ntohl(sgid_addr.saddr_in.sin_addr.s_addr);
3261
3262         spin_lock_irqsave(&rf->qh_list_lock, flags);
3263         mc_qht_elem = mcast_list_lookup_ip(rf, ip_addr);
3264         if (!mc_qht_elem) {
3265                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3266                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
3267                             "address not found MCG\n");
3268                 return 0;
3269         }
3270
3271         mcg_info.qp_id = iwqp->ibqp.qp_num;
3272         irdma_sc_del_mcast_grp(&mc_qht_elem->mc_grp_ctx, &mcg_info);
3273         if (!mc_qht_elem->mc_grp_ctx.no_of_mgs) {
3274                 mcast_list_del(mc_qht_elem);
3275                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3276                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
3277                                          IRDMA_OP_MC_DESTROY);
3278                 if (ret) {
3279                         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
3280                                     "failed MC_DESTROY MCG\n");
3281                         spin_lock_irqsave(&rf->qh_list_lock, flags);
3282                         mcast_list_add(rf, mc_qht_elem);
3283                         spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3284                         return -EAGAIN;
3285                 }
3286
3287                 irdma_free_dma_mem(&rf->hw,
3288                                    &mc_qht_elem->mc_grp_ctx.dma_mem_mc);
3289                 irdma_free_rsrc(rf, rf->allocated_mcgs,
3290                                 mc_qht_elem->mc_grp_ctx.mg_id);
3291                 kfree(mc_qht_elem);
3292         } else {
3293                 spin_unlock_irqrestore(&rf->qh_list_lock, flags);
3294                 ret = irdma_mcast_cqp_op(iwdev, &mc_qht_elem->mc_grp_ctx,
3295                                          IRDMA_OP_MC_MODIFY);
3296                 if (ret) {
3297                         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_VERBS,
3298                                     "failed Modify MCG\n");
3299                         return ret;
3300                 }
3301         }
3302
3303         return 0;
3304 }
3305
3306 /**
3307  * irdma_query_ah - Query address handle
3308  * @ibah: pointer to address handle
3309  * @ah_attr: address handle attributes
3310  */
3311 static int
3312 irdma_query_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
3313 {
3314         struct irdma_ah *ah = to_iwah(ibah);
3315
3316         memset(ah_attr, 0, sizeof(*ah_attr));
3317         if (ah->av.attrs.ah_flags & IB_AH_GRH) {
3318                 ah_attr->ah_flags = IB_AH_GRH;
3319                 ah_attr->grh.flow_label = ah->sc_ah.ah_info.flow_label;
3320                 ah_attr->grh.traffic_class = ah->sc_ah.ah_info.tc_tos;
3321                 ah_attr->grh.hop_limit = ah->sc_ah.ah_info.hop_ttl;
3322                 ah_attr->grh.sgid_index = ah->sgid_index;
3323                 ah_attr->grh.sgid_index = ah->sgid_index;
3324                 memcpy(&ah_attr->grh.dgid, &ah->dgid,
3325                        sizeof(ah_attr->grh.dgid));
3326         }
3327
3328         return 0;
3329 }
3330
3331 static __be64 irdma_mac_to_guid(struct ifnet *ndev){
3332         const unsigned char *mac = IF_LLADDR(ndev);
3333         __be64 guid;
3334         unsigned char *dst = (unsigned char *)&guid;
3335
3336         dst[0] = mac[0] ^ 2;
3337         dst[1] = mac[1];
3338         dst[2] = mac[2];
3339         dst[3] = 0xff;
3340         dst[4] = 0xfe;
3341         dst[5] = mac[3];
3342         dst[6] = mac[4];
3343         dst[7] = mac[5];
3344
3345         return guid;
3346 }
3347
3348 static struct ifnet *
3349 irdma_get_netdev(struct ib_device *ibdev, u8 port_num)
3350 {
3351         struct irdma_device *iwdev = to_iwdev(ibdev);
3352
3353         if (iwdev->netdev) {
3354                 dev_hold(iwdev->netdev);
3355                 return iwdev->netdev;
3356         }
3357
3358         return NULL;
3359 }
3360
3361 static void
3362 irdma_set_device_ops(struct ib_device *ibdev)
3363 {
3364         struct ib_device *dev_ops = ibdev;
3365
3366         dev_ops->ops.driver_id = RDMA_DRIVER_I40IW;
3367         dev_ops->ops.size_ib_ah = IRDMA_SET_RDMA_OBJ_SIZE(ib_ah, irdma_ah, ibah);
3368         dev_ops->ops.size_ib_cq = IRDMA_SET_RDMA_OBJ_SIZE(ib_cq, irdma_cq, ibcq);
3369         dev_ops->ops.size_ib_pd = IRDMA_SET_RDMA_OBJ_SIZE(ib_pd, irdma_pd, ibpd);
3370         dev_ops->ops.size_ib_ucontext = IRDMA_SET_RDMA_OBJ_SIZE(ib_ucontext,
3371                                                                 irdma_ucontext,
3372                                                                 ibucontext);
3373
3374         dev_ops->alloc_hw_stats = irdma_alloc_hw_stats;
3375         dev_ops->alloc_mr = irdma_alloc_mr;
3376         dev_ops->alloc_mw = irdma_alloc_mw;
3377         dev_ops->alloc_pd = irdma_alloc_pd;
3378         dev_ops->alloc_ucontext = irdma_alloc_ucontext;
3379         dev_ops->create_cq = irdma_create_cq;
3380         dev_ops->create_qp = irdma_create_qp;
3381         dev_ops->dealloc_mw = irdma_dealloc_mw;
3382         dev_ops->dealloc_pd = irdma_dealloc_pd;
3383         dev_ops->dealloc_ucontext = irdma_dealloc_ucontext;
3384         dev_ops->dereg_mr = irdma_dereg_mr;
3385         dev_ops->destroy_cq = irdma_destroy_cq;
3386         dev_ops->destroy_qp = irdma_destroy_qp;
3387         dev_ops->disassociate_ucontext = irdma_disassociate_ucontext;
3388         dev_ops->get_dev_fw_str = irdma_get_dev_fw_str;
3389         dev_ops->get_dma_mr = irdma_get_dma_mr;
3390         dev_ops->get_hw_stats = irdma_get_hw_stats;
3391         dev_ops->get_netdev = irdma_get_netdev;
3392         dev_ops->map_mr_sg = irdma_map_mr_sg;
3393         dev_ops->mmap = irdma_mmap;
3394         dev_ops->mmap_free = irdma_mmap_free;
3395         dev_ops->poll_cq = irdma_poll_cq;
3396         dev_ops->post_recv = irdma_post_recv;
3397         dev_ops->post_send = irdma_post_send;
3398         dev_ops->query_device = irdma_query_device;
3399         dev_ops->query_port = irdma_query_port;
3400         dev_ops->modify_port = irdma_modify_port;
3401         dev_ops->query_qp = irdma_query_qp;
3402         dev_ops->reg_user_mr = irdma_reg_user_mr;
3403         dev_ops->rereg_user_mr = irdma_rereg_user_mr;
3404         dev_ops->req_notify_cq = irdma_req_notify_cq;
3405         dev_ops->resize_cq = irdma_resize_cq;
3406 }
3407
3408 static void
3409 irdma_set_device_mcast_ops(struct ib_device *ibdev)
3410 {
3411         struct ib_device *dev_ops = ibdev;
3412         dev_ops->attach_mcast = irdma_attach_mcast;
3413         dev_ops->detach_mcast = irdma_detach_mcast;
3414 }
3415
3416 static void
3417 irdma_set_device_roce_ops(struct ib_device *ibdev)
3418 {
3419         struct ib_device *dev_ops = ibdev;
3420         dev_ops->create_ah = irdma_create_ah;
3421         dev_ops->destroy_ah = irdma_destroy_ah;
3422         dev_ops->get_link_layer = irdma_get_link_layer;
3423         dev_ops->get_port_immutable = irdma_roce_port_immutable;
3424         dev_ops->modify_qp = irdma_modify_qp_roce;
3425         dev_ops->query_ah = irdma_query_ah;
3426         dev_ops->query_gid = irdma_query_gid_roce;
3427         dev_ops->query_pkey = irdma_query_pkey;
3428         ibdev->add_gid = irdma_add_gid;
3429         ibdev->del_gid = irdma_del_gid;
3430 }
3431
3432 static void
3433 irdma_set_device_iw_ops(struct ib_device *ibdev)
3434 {
3435         struct ib_device *dev_ops = ibdev;
3436
3437         ibdev->uverbs_cmd_mask |=
3438             (1ull << IB_USER_VERBS_CMD_CREATE_AH) |
3439             (1ull << IB_USER_VERBS_CMD_DESTROY_AH);
3440
3441         dev_ops->create_ah = irdma_create_ah_stub;
3442         dev_ops->destroy_ah = irdma_destroy_ah_stub;
3443         dev_ops->get_port_immutable = irdma_iw_port_immutable;
3444         dev_ops->modify_qp = irdma_modify_qp;
3445         dev_ops->query_gid = irdma_query_gid;
3446         dev_ops->query_pkey = irdma_iw_query_pkey;
3447 }
3448
3449 static inline void
3450 irdma_set_device_gen1_ops(struct ib_device *ibdev)
3451 {
3452 }
3453
3454 /**
3455  * irdma_init_roce_device - initialization of roce rdma device
3456  * @iwdev: irdma device
3457  */
3458 static void
3459 irdma_init_roce_device(struct irdma_device *iwdev)
3460 {
3461         kc_set_roce_uverbs_cmd_mask(iwdev);
3462         iwdev->ibdev.node_type = RDMA_NODE_IB_CA;
3463         iwdev->ibdev.node_guid = irdma_mac_to_guid(iwdev->netdev);
3464         irdma_set_device_roce_ops(&iwdev->ibdev);
3465         if (iwdev->rf->rdma_ver == IRDMA_GEN_2)
3466                 irdma_set_device_mcast_ops(&iwdev->ibdev);
3467 }
3468
3469 /**
3470  * irdma_init_iw_device - initialization of iwarp rdma device
3471  * @iwdev: irdma device
3472  */
3473 static int
3474 irdma_init_iw_device(struct irdma_device *iwdev)
3475 {
3476         struct ifnet *netdev = iwdev->netdev;
3477
3478         iwdev->ibdev.node_type = RDMA_NODE_RNIC;
3479         ether_addr_copy((u8 *)&iwdev->ibdev.node_guid, IF_LLADDR(netdev));
3480         iwdev->ibdev.iwcm = kzalloc(sizeof(*iwdev->ibdev.iwcm), GFP_KERNEL);
3481         if (!iwdev->ibdev.iwcm)
3482                 return -ENOMEM;
3483
3484         iwdev->ibdev.iwcm->add_ref = irdma_qp_add_ref;
3485         iwdev->ibdev.iwcm->rem_ref = irdma_qp_rem_ref;
3486         iwdev->ibdev.iwcm->get_qp = irdma_get_qp;
3487         iwdev->ibdev.iwcm->connect = irdma_connect;
3488         iwdev->ibdev.iwcm->accept = irdma_accept;
3489         iwdev->ibdev.iwcm->reject = irdma_reject;
3490         iwdev->ibdev.iwcm->create_listen = irdma_create_listen;
3491         iwdev->ibdev.iwcm->destroy_listen = irdma_destroy_listen;
3492         memcpy(iwdev->ibdev.iwcm->ifname, if_name(netdev),
3493                sizeof(iwdev->ibdev.iwcm->ifname));
3494         irdma_set_device_iw_ops(&iwdev->ibdev);
3495
3496         return 0;
3497 }
3498
3499 /**
3500  * irdma_init_rdma_device - initialization of rdma device
3501  * @iwdev: irdma device
3502  */
3503 static int
3504 irdma_init_rdma_device(struct irdma_device *iwdev)
3505 {
3506         struct pci_dev *pcidev = iwdev->rf->pcidev;
3507         int ret;
3508
3509         iwdev->ibdev.owner = THIS_MODULE;
3510         iwdev->ibdev.uverbs_abi_ver = IRDMA_ABI_VER;
3511         kc_set_rdma_uverbs_cmd_mask(iwdev);
3512
3513         if (iwdev->roce_mode) {
3514                 irdma_init_roce_device(iwdev);
3515         } else {
3516                 ret = irdma_init_iw_device(iwdev);
3517                 if (ret)
3518                         return ret;
3519         }
3520
3521         iwdev->ibdev.phys_port_cnt = 1;
3522         iwdev->ibdev.num_comp_vectors = iwdev->rf->ceqs_count;
3523         iwdev->ibdev.dev.parent = iwdev->rf->dev_ctx.dev;
3524         set_ibdev_dma_device(iwdev->ibdev, &pcidev->dev);
3525         irdma_set_device_ops(&iwdev->ibdev);
3526         if (iwdev->rf->rdma_ver == IRDMA_GEN_1)
3527                 irdma_set_device_gen1_ops(&iwdev->ibdev);
3528
3529         return 0;
3530 }
3531
3532 /**
3533  * irdma_port_ibevent - indicate port event
3534  * @iwdev: irdma device
3535  */
3536 void
3537 irdma_port_ibevent(struct irdma_device *iwdev)
3538 {
3539         struct ib_event event;
3540
3541         event.device = &iwdev->ibdev;
3542         event.element.port_num = 1;
3543         event.event =
3544             iwdev->iw_status ? IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
3545         ib_dispatch_event(&event);
3546 }
3547
3548 /**
3549  * irdma_ib_unregister_device - unregister rdma device from IB
3550  * core
3551  * @iwdev: irdma device
3552  */
3553 void
3554 irdma_ib_unregister_device(struct irdma_device *iwdev)
3555 {
3556         iwdev->iw_status = 0;
3557         irdma_port_ibevent(iwdev);
3558         ib_unregister_device(&iwdev->ibdev);
3559         dev_put(iwdev->netdev);
3560         kfree(iwdev->ibdev.iwcm);
3561         iwdev->ibdev.iwcm = NULL;
3562 }
3563
3564 /**
3565  * irdma_ib_register_device - register irdma device to IB core
3566  * @iwdev: irdma device
3567  */
3568 int
3569 irdma_ib_register_device(struct irdma_device *iwdev)
3570 {
3571         int ret;
3572
3573         ret = irdma_init_rdma_device(iwdev);
3574         if (ret)
3575                 return ret;
3576
3577         dev_hold(iwdev->netdev);
3578         sprintf(iwdev->ibdev.name, "irdma-%s", if_name(iwdev->netdev));
3579         ret = ib_register_device(&iwdev->ibdev, NULL);
3580         if (ret)
3581                 goto error;
3582
3583         iwdev->iw_status = 1;
3584         irdma_port_ibevent(iwdev);
3585
3586         return 0;
3587
3588 error:
3589         kfree(iwdev->ibdev.iwcm);
3590         iwdev->ibdev.iwcm = NULL;
3591         irdma_debug(&iwdev->rf->sc_dev, IRDMA_DEBUG_VERBS, "Register RDMA device fail\n");
3592
3593         return ret;
3594 }