drivers/infiniband/hw/efa/efa_verbs.c

   1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
   2 /*
   3  * Copyright 2018-2019 Amazon.com, Inc. or its affiliates. All rights reserved.
   4  */
   5
   6 #include <linux/vmalloc.h>
   7
   8 #include <rdma/ib_addr.h>
   9 #include <rdma/ib_umem.h>
  10 #include <rdma/ib_user_verbs.h>
  11 #include <rdma/ib_verbs.h>
  12 #include <rdma/uverbs_ioctl.h>
  13
  14 #include "efa.h"
  15
  16 #define EFA_MMAP_FLAG_SHIFT 56
  17 #define EFA_MMAP_PAGE_MASK GENMASK(EFA_MMAP_FLAG_SHIFT - 1, 0)
  18 #define EFA_MMAP_INVALID U64_MAX
  19
  20 enum {
  21         EFA_MMAP_DMA_PAGE = 0,
  22         EFA_MMAP_IO_WC,
  23         EFA_MMAP_IO_NC,
  24 };
  25
  26 #define EFA_AENQ_ENABLED_GROUPS \
  27         (BIT(EFA_ADMIN_FATAL_ERROR) | BIT(EFA_ADMIN_WARNING) | \
  28          BIT(EFA_ADMIN_NOTIFICATION) | BIT(EFA_ADMIN_KEEP_ALIVE))
  29
  30 struct efa_mmap_entry {
  31         void  *obj;
  32         u64 address;
  33         u64 length;
  34         u32 mmap_page;
  35         u8 mmap_flag;
  36 };
  37
  38 static inline u64 get_mmap_key(const struct efa_mmap_entry *efa)
  39 {
  40         return ((u64)efa->mmap_flag << EFA_MMAP_FLAG_SHIFT) |
  41                ((u64)efa->mmap_page << PAGE_SHIFT);
  42 }
  43
  44 #define EFA_CHUNK_PAYLOAD_SHIFT       12
  45 #define EFA_CHUNK_PAYLOAD_SIZE        BIT(EFA_CHUNK_PAYLOAD_SHIFT)
  46 #define EFA_CHUNK_PAYLOAD_PTR_SIZE    8
  47
  48 #define EFA_CHUNK_SHIFT               12
  49 #define EFA_CHUNK_SIZE                BIT(EFA_CHUNK_SHIFT)
  50 #define EFA_CHUNK_PTR_SIZE            sizeof(struct efa_com_ctrl_buff_info)
  51
  52 #define EFA_PTRS_PER_CHUNK \
  53         ((EFA_CHUNK_SIZE - EFA_CHUNK_PTR_SIZE) / EFA_CHUNK_PAYLOAD_PTR_SIZE)
  54
  55 #define EFA_CHUNK_USED_SIZE \
  56         ((EFA_PTRS_PER_CHUNK * EFA_CHUNK_PAYLOAD_PTR_SIZE) + EFA_CHUNK_PTR_SIZE)
  57
  58 #define EFA_SUPPORTED_ACCESS_FLAGS IB_ACCESS_LOCAL_WRITE
  59
  60 struct pbl_chunk {
  61         dma_addr_t dma_addr;
  62         u64 *buf;
  63         u32 length;
  64 };
  65
  66 struct pbl_chunk_list {
  67         struct pbl_chunk *chunks;
  68         unsigned int size;
  69 };
  70
  71 struct pbl_context {
  72         union {
  73                 struct {
  74                         dma_addr_t dma_addr;
  75                 } continuous;
  76                 struct {
  77                         u32 pbl_buf_size_in_pages;
  78                         struct scatterlist *sgl;
  79                         int sg_dma_cnt;
  80                         struct pbl_chunk_list chunk_list;
  81                 } indirect;
  82         } phys;
  83         u64 *pbl_buf;
  84         u32 pbl_buf_size_in_bytes;
  85         u8 physically_continuous;
  86 };
  87
  88 static inline struct efa_dev *to_edev(struct ib_device *ibdev)
  89 {
  90         return container_of(ibdev, struct efa_dev, ibdev);
  91 }
  92
  93 static inline struct efa_ucontext *to_eucontext(struct ib_ucontext *ibucontext)
  94 {
  95         return container_of(ibucontext, struct efa_ucontext, ibucontext);
  96 }
  97
  98 static inline struct efa_pd *to_epd(struct ib_pd *ibpd)
  99 {
 100         return container_of(ibpd, struct efa_pd, ibpd);
 101 }
 102
 103 static inline struct efa_mr *to_emr(struct ib_mr *ibmr)
 104 {
 105         return container_of(ibmr, struct efa_mr, ibmr);
 106 }
 107
 108 static inline struct efa_qp *to_eqp(struct ib_qp *ibqp)
 109 {
 110         return container_of(ibqp, struct efa_qp, ibqp);
 111 }
 112
 113 static inline struct efa_cq *to_ecq(struct ib_cq *ibcq)
 114 {
 115         return container_of(ibcq, struct efa_cq, ibcq);
 116 }
 117
 118 static inline struct efa_ah *to_eah(struct ib_ah *ibah)
 119 {
 120         return container_of(ibah, struct efa_ah, ibah);
 121 }
 122
 123 #define field_avail(x, fld, sz) (offsetof(typeof(x), fld) + \
 124                                  sizeof(((typeof(x) *)0)->fld) <= (sz))
 125
 126 #define is_reserved_cleared(reserved) \
 127         !memchr_inv(reserved, 0, sizeof(reserved))
 128
 129 static void *efa_zalloc_mapped(struct efa_dev *dev, dma_addr_t *dma_addr,
 130                                size_t size, enum dma_data_direction dir)
 131 {
 132         void *addr;
 133
 134         addr = alloc_pages_exact(size, GFP_KERNEL | __GFP_ZERO);
 135         if (!addr)
 136                 return NULL;
 137
 138         *dma_addr = dma_map_single(&dev->pdev->dev, addr, size, dir);
 139         if (dma_mapping_error(&dev->pdev->dev, *dma_addr)) {
 140                 ibdev_err(&dev->ibdev, "Failed to map DMA address\n");
 141                 free_pages_exact(addr, size);
 142                 return NULL;
 143         }
 144
 145         return addr;
 146 }
 147
 148 /*
 149  * This is only called when the ucontext is destroyed and there can be no
 150  * concurrent query via mmap or allocate on the xarray, thus we can be sure no
 151  * other thread is using the entry pointer. We also know that all the BAR
 152  * pages have either been zap'd or munmaped at this point.  Normal pages are
 153  * refcounted and will be freed at the proper time.
 154  */
 155 static void mmap_entries_remove_free(struct efa_dev *dev,
 156                                      struct efa_ucontext *ucontext)
 157 {
 158         struct efa_mmap_entry *entry;
 159         unsigned long mmap_page;
 160
 161         xa_for_each(&ucontext->mmap_xa, mmap_page, entry) {
 162                 xa_erase(&ucontext->mmap_xa, mmap_page);
 163
 164                 ibdev_dbg(
 165                         &dev->ibdev,
 166                         "mmap: obj[0x%p] key[%#llx] addr[%#llx] len[%#llx] removed\n",
 167                         entry->obj, get_mmap_key(entry), entry->address,
 168                         entry->length);
 169                 if (entry->mmap_flag == EFA_MMAP_DMA_PAGE)
 170                         /* DMA mapping is already gone, now free the pages */
 171                         free_pages_exact(phys_to_virt(entry->address),
 172                                          entry->length);
 173                 kfree(entry);
 174         }
 175 }
 176
 177 static struct efa_mmap_entry *mmap_entry_get(struct efa_dev *dev,
 178                                              struct efa_ucontext *ucontext,
 179                                              u64 key, u64 len)
 180 {
 181         struct efa_mmap_entry *entry;
 182         u64 mmap_page;
 183
 184         mmap_page = (key & EFA_MMAP_PAGE_MASK) >> PAGE_SHIFT;
 185         if (mmap_page > U32_MAX)
 186                 return NULL;
 187
 188         entry = xa_load(&ucontext->mmap_xa, mmap_page);
 189         if (!entry || get_mmap_key(entry) != key || entry->length != len)
 190                 return NULL;
 191
 192         ibdev_dbg(&dev->ibdev,
 193                   "mmap: obj[0x%p] key[%#llx] addr[%#llx] len[%#llx] removed\n",
 194                   entry->obj, key, entry->address, entry->length);
 195
 196         return entry;
 197 }
 198
 199 /*
 200  * Note this locking scheme cannot support removal of entries, except during
 201  * ucontext destruction when the core code guarentees no concurrency.
 202  */
 203 static u64 mmap_entry_insert(struct efa_dev *dev, struct efa_ucontext *ucontext,
 204                              void *obj, u64 address, u64 length, u8 mmap_flag)
 205 {
 206         struct efa_mmap_entry *entry;
 207         int err;
 208
 209         entry = kmalloc(sizeof(*entry), GFP_KERNEL);
 210         if (!entry)
 211                 return EFA_MMAP_INVALID;
 212
 213         entry->obj = obj;
 214         entry->address = address;
 215         entry->length = length;
 216         entry->mmap_flag = mmap_flag;
 217
 218         xa_lock(&ucontext->mmap_xa);
 219         entry->mmap_page = ucontext->mmap_xa_page;
 220         ucontext->mmap_xa_page += DIV_ROUND_UP(length, PAGE_SIZE);
 221         err = __xa_insert(&ucontext->mmap_xa, entry->mmap_page, entry,
 222                           GFP_KERNEL);
 223         xa_unlock(&ucontext->mmap_xa);
 224         if (err){
 225                 kfree(entry);
 226                 return EFA_MMAP_INVALID;
 227         }
 228
 229         ibdev_dbg(
 230                 &dev->ibdev,
 231                 "mmap: obj[0x%p] addr[%#llx], len[%#llx], key[%#llx] inserted\n",
 232                 entry->obj, entry->address, entry->length, get_mmap_key(entry));
 233
 234         return get_mmap_key(entry);
 235 }
 236
 237 int efa_query_device(struct ib_device *ibdev,
 238                      struct ib_device_attr *props,
 239                      struct ib_udata *udata)
 240 {
 241         struct efa_com_get_device_attr_result *dev_attr;
 242         struct efa_ibv_ex_query_device_resp resp = {};
 243         struct efa_dev *dev = to_edev(ibdev);
 244         int err;
 245
 246         if (udata && udata->inlen &&
 247             !ib_is_udata_cleared(udata, 0, udata->inlen)) {
 248                 ibdev_dbg(ibdev,
 249                           "Incompatible ABI params, udata not cleared\n");
 250                 return -EINVAL;
 251         }
 252
 253         dev_attr = &dev->dev_attr;
 254
 255         memset(props, 0, sizeof(*props));
 256         props->max_mr_size = dev_attr->max_mr_pages * PAGE_SIZE;
 257         props->page_size_cap = dev_attr->page_size_cap;
 258         props->vendor_id = dev->pdev->vendor;
 259         props->vendor_part_id = dev->pdev->device;
 260         props->hw_ver = dev->pdev->subsystem_device;
 261         props->max_qp = dev_attr->max_qp;
 262         props->max_cq = dev_attr->max_cq;
 263         props->max_pd = dev_attr->max_pd;
 264         props->max_mr = dev_attr->max_mr;
 265         props->max_ah = dev_attr->max_ah;
 266         props->max_cqe = dev_attr->max_cq_depth;
 267         props->max_qp_wr = min_t(u32, dev_attr->max_sq_depth,
 268                                  dev_attr->max_rq_depth);
 269         props->max_send_sge = dev_attr->max_sq_sge;
 270         props->max_recv_sge = dev_attr->max_rq_sge;
 271
 272         if (udata && udata->outlen) {
 273                 resp.max_sq_sge = dev_attr->max_sq_sge;
 274                 resp.max_rq_sge = dev_attr->max_rq_sge;
 275                 resp.max_sq_wr = dev_attr->max_sq_depth;
 276                 resp.max_rq_wr = dev_attr->max_rq_depth;
 277
 278                 err = ib_copy_to_udata(udata, &resp,
 279                                        min(sizeof(resp), udata->outlen));
 280                 if (err) {
 281                         ibdev_dbg(ibdev,
 282                                   "Failed to copy udata for query_device\n");
 283                         return err;
 284                 }
 285         }
 286
 287         return 0;
 288 }
 289
 290 int efa_query_port(struct ib_device *ibdev, u8 port,
 291                    struct ib_port_attr *props)
 292 {
 293         struct efa_dev *dev = to_edev(ibdev);
 294
 295         props->lmc = 1;
 296
 297         props->state = IB_PORT_ACTIVE;
 298         props->phys_state = 5;
 299         props->gid_tbl_len = 1;
 300         props->pkey_tbl_len = 1;
 301         props->active_speed = IB_SPEED_EDR;
 302         props->active_width = IB_WIDTH_4X;
 303         props->max_mtu = ib_mtu_int_to_enum(dev->mtu);
 304         props->active_mtu = ib_mtu_int_to_enum(dev->mtu);
 305         props->max_msg_sz = dev->mtu;
 306         props->max_vl_num = 1;
 307
 308         return 0;
 309 }
 310
 311 int efa_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
 312                  int qp_attr_mask,
 313                  struct ib_qp_init_attr *qp_init_attr)
 314 {
 315         struct efa_dev *dev = to_edev(ibqp->device);
 316         struct efa_com_query_qp_params params = {};
 317         struct efa_com_query_qp_result result;
 318         struct efa_qp *qp = to_eqp(ibqp);
 319         int err;
 320
 321 #define EFA_QUERY_QP_SUPP_MASK \
 322         (IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT | \
 323          IB_QP_QKEY | IB_QP_SQ_PSN | IB_QP_CAP)
 324
 325         if (qp_attr_mask & ~EFA_QUERY_QP_SUPP_MASK) {
 326                 ibdev_dbg(&dev->ibdev,
 327                           "Unsupported qp_attr_mask[%#x] supported[%#x]\n",
 328                           qp_attr_mask, EFA_QUERY_QP_SUPP_MASK);
 329                 return -EOPNOTSUPP;
 330         }
 331
 332         memset(qp_attr, 0, sizeof(*qp_attr));
 333         memset(qp_init_attr, 0, sizeof(*qp_init_attr));
 334
 335         params.qp_handle = qp->qp_handle;
 336         err = efa_com_query_qp(&dev->edev, &params, &result);
 337         if (err)
 338                 return err;
 339
 340         qp_attr->qp_state = result.qp_state;
 341         qp_attr->qkey = result.qkey;
 342         qp_attr->sq_psn = result.sq_psn;
 343         qp_attr->sq_draining = result.sq_draining;
 344         qp_attr->port_num = 1;
 345
 346         qp_attr->cap.max_send_wr = qp->max_send_wr;
 347         qp_attr->cap.max_recv_wr = qp->max_recv_wr;
 348         qp_attr->cap.max_send_sge = qp->max_send_sge;
 349         qp_attr->cap.max_recv_sge = qp->max_recv_sge;
 350         qp_attr->cap.max_inline_data = qp->max_inline_data;
 351
 352         qp_init_attr->qp_type = ibqp->qp_type;
 353         qp_init_attr->recv_cq = ibqp->recv_cq;
 354         qp_init_attr->send_cq = ibqp->send_cq;
 355         qp_init_attr->qp_context = ibqp->qp_context;
 356         qp_init_attr->cap = qp_attr->cap;
 357
 358         return 0;
 359 }
 360
 361 int efa_query_gid(struct ib_device *ibdev, u8 port, int index,
 362                   union ib_gid *gid)
 363 {
 364         struct efa_dev *dev = to_edev(ibdev);
 365
 366         memcpy(gid->raw, dev->addr, sizeof(dev->addr));
 367
 368         return 0;
 369 }
 370
 371 int efa_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
 372                    u16 *pkey)
 373 {
 374         if (index > 0)
 375                 return -EINVAL;
 376
 377         *pkey = 0xffff;
 378         return 0;
 379 }
 380
 381 static int efa_pd_dealloc(struct efa_dev *dev, u16 pdn)
 382 {
 383         struct efa_com_dealloc_pd_params params = {
 384                 .pdn = pdn,
 385         };
 386
 387         return efa_com_dealloc_pd(&dev->edev, &params);
 388 }
 389
 390 int efa_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 391 {
 392         struct efa_dev *dev = to_edev(ibpd->device);
 393         struct efa_ibv_alloc_pd_resp resp = {};
 394         struct efa_com_alloc_pd_result result;
 395         struct efa_pd *pd = to_epd(ibpd);
 396         int err;
 397
 398         if (udata->inlen &&
 399             !ib_is_udata_cleared(udata, 0, udata->inlen)) {
 400                 ibdev_dbg(&dev->ibdev,
 401                           "Incompatible ABI params, udata not cleared\n");
 402                 err = -EINVAL;
 403                 goto err_out;
 404         }
 405
 406         err = efa_com_alloc_pd(&dev->edev, &result);
 407         if (err)
 408                 goto err_out;
 409
 410         pd->pdn = result.pdn;
 411         resp.pdn = result.pdn;
 412
 413         if (udata->outlen) {
 414                 err = ib_copy_to_udata(udata, &resp,
 415                                        min(sizeof(resp), udata->outlen));
 416                 if (err) {
 417                         ibdev_dbg(&dev->ibdev,
 418                                   "Failed to copy udata for alloc_pd\n");
 419                         goto err_dealloc_pd;
 420                 }
 421         }
 422
 423         ibdev_dbg(&dev->ibdev, "Allocated pd[%d]\n", pd->pdn);
 424
 425         return 0;
 426
 427 err_dealloc_pd:
 428         efa_pd_dealloc(dev, result.pdn);
 429 err_out:
 430         atomic64_inc(&dev->stats.sw_stats.alloc_pd_err);
 431         return err;
 432 }
 433
 434 void efa_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 435 {
 436         struct efa_dev *dev = to_edev(ibpd->device);
 437         struct efa_pd *pd = to_epd(ibpd);
 438
 439         ibdev_dbg(&dev->ibdev, "Dealloc pd[%d]\n", pd->pdn);
 440         efa_pd_dealloc(dev, pd->pdn);
 441 }
 442
 443 static int efa_destroy_qp_handle(struct efa_dev *dev, u32 qp_handle)
 444 {
 445         struct efa_com_destroy_qp_params params = { .qp_handle = qp_handle };
 446
 447         return efa_com_destroy_qp(&dev->edev, &params);
 448 }
 449
 450 int efa_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
 451 {
 452         struct efa_dev *dev = to_edev(ibqp->pd->device);
 453         struct efa_qp *qp = to_eqp(ibqp);
 454         int err;
 455
 456         ibdev_dbg(&dev->ibdev, "Destroy qp[%u]\n", ibqp->qp_num);
 457         err = efa_destroy_qp_handle(dev, qp->qp_handle);
 458         if (err)
 459                 return err;
 460
 461         if (qp->rq_cpu_addr) {
 462                 ibdev_dbg(&dev->ibdev,
 463                           "qp->cpu_addr[0x%p] freed: size[%lu], dma[%pad]\n",
 464                           qp->rq_cpu_addr, qp->rq_size,
 465                           &qp->rq_dma_addr);
 466                 dma_unmap_single(&dev->pdev->dev, qp->rq_dma_addr, qp->rq_size,
 467                                  DMA_TO_DEVICE);
 468         }
 469
 470         kfree(qp);
 471         return 0;
 472 }
 473
 474 static int qp_mmap_entries_setup(struct efa_qp *qp,
 475                                  struct efa_dev *dev,
 476                                  struct efa_ucontext *ucontext,
 477                                  struct efa_com_create_qp_params *params,
 478                                  struct efa_ibv_create_qp_resp *resp)
 479 {
 480         /*
 481          * Once an entry is inserted it might be mmapped, hence cannot be
 482          * cleaned up until dealloc_ucontext.
 483          */
 484         resp->sq_db_mmap_key =
 485                 mmap_entry_insert(dev, ucontext, qp,
 486                                   dev->db_bar_addr + resp->sq_db_offset,
 487                                   PAGE_SIZE, EFA_MMAP_IO_NC);
 488         if (resp->sq_db_mmap_key == EFA_MMAP_INVALID)
 489                 return -ENOMEM;
 490
 491         resp->sq_db_offset &= ~PAGE_MASK;
 492
 493         resp->llq_desc_mmap_key =
 494                 mmap_entry_insert(dev, ucontext, qp,
 495                                   dev->mem_bar_addr + resp->llq_desc_offset,
 496                                   PAGE_ALIGN(params->sq_ring_size_in_bytes +
 497                                              (resp->llq_desc_offset & ~PAGE_MASK)),
 498                                   EFA_MMAP_IO_WC);
 499         if (resp->llq_desc_mmap_key == EFA_MMAP_INVALID)
 500                 return -ENOMEM;
 501
 502         resp->llq_desc_offset &= ~PAGE_MASK;
 503
 504         if (qp->rq_size) {
 505                 resp->rq_db_mmap_key =
 506                         mmap_entry_insert(dev, ucontext, qp,
 507                                           dev->db_bar_addr + resp->rq_db_offset,
 508                                           PAGE_SIZE, EFA_MMAP_IO_NC);
 509                 if (resp->rq_db_mmap_key == EFA_MMAP_INVALID)
 510                         return -ENOMEM;
 511
 512                 resp->rq_db_offset &= ~PAGE_MASK;
 513
 514                 resp->rq_mmap_key =
 515                         mmap_entry_insert(dev, ucontext, qp,
 516                                           virt_to_phys(qp->rq_cpu_addr),
 517                                           qp->rq_size, EFA_MMAP_DMA_PAGE);
 518                 if (resp->rq_mmap_key == EFA_MMAP_INVALID)
 519                         return -ENOMEM;
 520
 521                 resp->rq_mmap_size = qp->rq_size;
 522         }
 523
 524         return 0;
 525 }
 526
 527 static int efa_qp_validate_cap(struct efa_dev *dev,
 528                                struct ib_qp_init_attr *init_attr)
 529 {
 530         if (init_attr->cap.max_send_wr > dev->dev_attr.max_sq_depth) {
 531                 ibdev_dbg(&dev->ibdev,
 532                           "qp: requested send wr[%u] exceeds the max[%u]\n",
 533                           init_attr->cap.max_send_wr,
 534                           dev->dev_attr.max_sq_depth);
 535                 return -EINVAL;
 536         }
 537         if (init_attr->cap.max_recv_wr > dev->dev_attr.max_rq_depth) {
 538                 ibdev_dbg(&dev->ibdev,
 539                           "qp: requested receive wr[%u] exceeds the max[%u]\n",
 540                           init_attr->cap.max_recv_wr,
 541                           dev->dev_attr.max_rq_depth);
 542                 return -EINVAL;
 543         }
 544         if (init_attr->cap.max_send_sge > dev->dev_attr.max_sq_sge) {
 545                 ibdev_dbg(&dev->ibdev,
 546                           "qp: requested sge send[%u] exceeds the max[%u]\n",
 547                           init_attr->cap.max_send_sge, dev->dev_attr.max_sq_sge);
 548                 return -EINVAL;
 549         }
 550         if (init_attr->cap.max_recv_sge > dev->dev_attr.max_rq_sge) {
 551                 ibdev_dbg(&dev->ibdev,
 552                           "qp: requested sge recv[%u] exceeds the max[%u]\n",
 553                           init_attr->cap.max_recv_sge, dev->dev_attr.max_rq_sge);
 554                 return -EINVAL;
 555         }
 556         if (init_attr->cap.max_inline_data > dev->dev_attr.inline_buf_size) {
 557                 ibdev_dbg(&dev->ibdev,
 558                           "qp: requested inline data[%u] exceeds the max[%u]\n",
 559                           init_attr->cap.max_inline_data,
 560                           dev->dev_attr.inline_buf_size);
 561                 return -EINVAL;
 562         }
 563
 564         return 0;
 565 }
 566
 567 static int efa_qp_validate_attr(struct efa_dev *dev,
 568                                 struct ib_qp_init_attr *init_attr)
 569 {
 570         if (init_attr->qp_type != IB_QPT_DRIVER &&
 571             init_attr->qp_type != IB_QPT_UD) {
 572                 ibdev_dbg(&dev->ibdev,
 573                           "Unsupported qp type %d\n", init_attr->qp_type);
 574                 return -EOPNOTSUPP;
 575         }
 576
 577         if (init_attr->srq) {
 578                 ibdev_dbg(&dev->ibdev, "SRQ is not supported\n");
 579                 return -EOPNOTSUPP;
 580         }
 581
 582         if (init_attr->create_flags) {
 583                 ibdev_dbg(&dev->ibdev, "Unsupported create flags\n");
 584                 return -EOPNOTSUPP;
 585         }
 586
 587         return 0;
 588 }
 589
 590 struct ib_qp *efa_create_qp(struct ib_pd *ibpd,
 591                             struct ib_qp_init_attr *init_attr,
 592                             struct ib_udata *udata)
 593 {
 594         struct efa_com_create_qp_params create_qp_params = {};
 595         struct efa_com_create_qp_result create_qp_resp;
 596         struct efa_dev *dev = to_edev(ibpd->device);
 597         struct efa_ibv_create_qp_resp resp = {};
 598         struct efa_ibv_create_qp cmd = {};
 599         bool rq_entry_inserted = false;
 600         struct efa_ucontext *ucontext;
 601         struct efa_qp *qp;
 602         int err;
 603
 604         ucontext = rdma_udata_to_drv_context(udata, struct efa_ucontext,
 605                                              ibucontext);
 606
 607         err = efa_qp_validate_cap(dev, init_attr);
 608         if (err)
 609                 goto err_out;
 610
 611         err = efa_qp_validate_attr(dev, init_attr);
 612         if (err)
 613                 goto err_out;
 614
 615         if (!field_avail(cmd, driver_qp_type, udata->inlen)) {
 616                 ibdev_dbg(&dev->ibdev,
 617                           "Incompatible ABI params, no input udata\n");
 618                 err = -EINVAL;
 619                 goto err_out;
 620         }
 621
 622         if (udata->inlen > sizeof(cmd) &&
 623             !ib_is_udata_cleared(udata, sizeof(cmd),
 624                                  udata->inlen - sizeof(cmd))) {
 625                 ibdev_dbg(&dev->ibdev,
 626                           "Incompatible ABI params, unknown fields in udata\n");
 627                 err = -EINVAL;
 628                 goto err_out;
 629         }
 630
 631         err = ib_copy_from_udata(&cmd, udata,
 632                                  min(sizeof(cmd), udata->inlen));
 633         if (err) {
 634                 ibdev_dbg(&dev->ibdev,
 635                           "Cannot copy udata for create_qp\n");
 636                 goto err_out;
 637         }
 638
 639         if (cmd.comp_mask) {
 640                 ibdev_dbg(&dev->ibdev,
 641                           "Incompatible ABI params, unknown fields in udata\n");
 642                 err = -EINVAL;
 643                 goto err_out;
 644         }
 645
 646         qp = kzalloc(sizeof(*qp), GFP_KERNEL);
 647         if (!qp) {
 648                 err = -ENOMEM;
 649                 goto err_out;
 650         }
 651
 652         create_qp_params.uarn = ucontext->uarn;
 653         create_qp_params.pd = to_epd(ibpd)->pdn;
 654
 655         if (init_attr->qp_type == IB_QPT_UD) {
 656                 create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_UD;
 657         } else if (cmd.driver_qp_type == EFA_QP_DRIVER_TYPE_SRD) {
 658                 create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_SRD;
 659         } else {
 660                 ibdev_dbg(&dev->ibdev,
 661                           "Unsupported qp type %d driver qp type %d\n",
 662                           init_attr->qp_type, cmd.driver_qp_type);
 663                 err = -EOPNOTSUPP;
 664                 goto err_free_qp;
 665         }
 666
 667         ibdev_dbg(&dev->ibdev, "Create QP: qp type %d driver qp type %#x\n",
 668                   init_attr->qp_type, cmd.driver_qp_type);
 669         create_qp_params.send_cq_idx = to_ecq(init_attr->send_cq)->cq_idx;
 670         create_qp_params.recv_cq_idx = to_ecq(init_attr->recv_cq)->cq_idx;
 671         create_qp_params.sq_depth = init_attr->cap.max_send_wr;
 672         create_qp_params.sq_ring_size_in_bytes = cmd.sq_ring_size;
 673
 674         create_qp_params.rq_depth = init_attr->cap.max_recv_wr;
 675         create_qp_params.rq_ring_size_in_bytes = cmd.rq_ring_size;
 676         qp->rq_size = PAGE_ALIGN(create_qp_params.rq_ring_size_in_bytes);
 677         if (qp->rq_size) {
 678                 qp->rq_cpu_addr = efa_zalloc_mapped(dev, &qp->rq_dma_addr,
 679                                                     qp->rq_size, DMA_TO_DEVICE);
 680                 if (!qp->rq_cpu_addr) {
 681                         err = -ENOMEM;
 682                         goto err_free_qp;
 683                 }
 684
 685                 ibdev_dbg(&dev->ibdev,
 686                           "qp->cpu_addr[0x%p] allocated: size[%lu], dma[%pad]\n",
 687                           qp->rq_cpu_addr, qp->rq_size, &qp->rq_dma_addr);
 688                 create_qp_params.rq_base_addr = qp->rq_dma_addr;
 689         }
 690
 691         err = efa_com_create_qp(&dev->edev, &create_qp_params,
 692                                 &create_qp_resp);
 693         if (err)
 694                 goto err_free_mapped;
 695
 696         resp.sq_db_offset = create_qp_resp.sq_db_offset;
 697         resp.rq_db_offset = create_qp_resp.rq_db_offset;
 698         resp.llq_desc_offset = create_qp_resp.llq_descriptors_offset;
 699         resp.send_sub_cq_idx = create_qp_resp.send_sub_cq_idx;
 700         resp.recv_sub_cq_idx = create_qp_resp.recv_sub_cq_idx;
 701
 702         err = qp_mmap_entries_setup(qp, dev, ucontext, &create_qp_params,
 703                                     &resp);
 704         if (err)
 705                 goto err_destroy_qp;
 706
 707         rq_entry_inserted = true;
 708         qp->qp_handle = create_qp_resp.qp_handle;
 709         qp->ibqp.qp_num = create_qp_resp.qp_num;
 710         qp->ibqp.qp_type = init_attr->qp_type;
 711         qp->max_send_wr = init_attr->cap.max_send_wr;
 712         qp->max_recv_wr = init_attr->cap.max_recv_wr;
 713         qp->max_send_sge = init_attr->cap.max_send_sge;
 714         qp->max_recv_sge = init_attr->cap.max_recv_sge;
 715         qp->max_inline_data = init_attr->cap.max_inline_data;
 716
 717         if (udata->outlen) {
 718                 err = ib_copy_to_udata(udata, &resp,
 719                                        min(sizeof(resp), udata->outlen));
 720                 if (err) {
 721                         ibdev_dbg(&dev->ibdev,
 722                                   "Failed to copy udata for qp[%u]\n",
 723                                   create_qp_resp.qp_num);
 724                         goto err_destroy_qp;
 725                 }
 726         }
 727
 728         ibdev_dbg(&dev->ibdev, "Created qp[%d]\n", qp->ibqp.qp_num);
 729
 730         return &qp->ibqp;
 731
 732 err_destroy_qp:
 733         efa_destroy_qp_handle(dev, create_qp_resp.qp_handle);
 734 err_free_mapped:
 735         if (qp->rq_size) {
 736                 dma_unmap_single(&dev->pdev->dev, qp->rq_dma_addr, qp->rq_size,
 737                                  DMA_TO_DEVICE);
 738                 if (!rq_entry_inserted)
 739                         free_pages_exact(qp->rq_cpu_addr, qp->rq_size);
 740         }
 741 err_free_qp:
 742         kfree(qp);
 743 err_out:
 744         atomic64_inc(&dev->stats.sw_stats.create_qp_err);
 745         return ERR_PTR(err);
 746 }
 747
 748 static int efa_modify_qp_validate(struct efa_dev *dev, struct efa_qp *qp,
 749                                   struct ib_qp_attr *qp_attr, int qp_attr_mask,
 750                                   enum ib_qp_state cur_state,
 751                                   enum ib_qp_state new_state)
 752 {
 753 #define EFA_MODIFY_QP_SUPP_MASK \
 754         (IB_QP_STATE | IB_QP_CUR_STATE | IB_QP_EN_SQD_ASYNC_NOTIFY | \
 755          IB_QP_PKEY_INDEX | IB_QP_PORT | IB_QP_QKEY | IB_QP_SQ_PSN)
 756
 757         if (qp_attr_mask & ~EFA_MODIFY_QP_SUPP_MASK) {
 758                 ibdev_dbg(&dev->ibdev,
 759                           "Unsupported qp_attr_mask[%#x] supported[%#x]\n",
 760                           qp_attr_mask, EFA_MODIFY_QP_SUPP_MASK);
 761                 return -EOPNOTSUPP;
 762         }
 763
 764         if (!ib_modify_qp_is_ok(cur_state, new_state, IB_QPT_UD,
 765                                 qp_attr_mask)) {
 766                 ibdev_dbg(&dev->ibdev, "Invalid modify QP parameters\n");
 767                 return -EINVAL;
 768         }
 769
 770         if ((qp_attr_mask & IB_QP_PORT) && qp_attr->port_num != 1) {
 771                 ibdev_dbg(&dev->ibdev, "Can't change port num\n");
 772                 return -EOPNOTSUPP;
 773         }
 774
 775         if ((qp_attr_mask & IB_QP_PKEY_INDEX) && qp_attr->pkey_index) {
 776                 ibdev_dbg(&dev->ibdev, "Can't change pkey index\n");
 777                 return -EOPNOTSUPP;
 778         }
 779
 780         return 0;
 781 }
 782
 783 int efa_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
 784                   int qp_attr_mask, struct ib_udata *udata)
 785 {
 786         struct efa_dev *dev = to_edev(ibqp->device);
 787         struct efa_com_modify_qp_params params = {};
 788         struct efa_qp *qp = to_eqp(ibqp);
 789         enum ib_qp_state cur_state;
 790         enum ib_qp_state new_state;
 791         int err;
 792
 793         if (udata->inlen &&
 794             !ib_is_udata_cleared(udata, 0, udata->inlen)) {
 795                 ibdev_dbg(&dev->ibdev,
 796                           "Incompatible ABI params, udata not cleared\n");
 797                 return -EINVAL;
 798         }
 799
 800         cur_state = qp_attr_mask & IB_QP_CUR_STATE ? qp_attr->cur_qp_state :
 801                                                      qp->state;
 802         new_state = qp_attr_mask & IB_QP_STATE ? qp_attr->qp_state : cur_state;
 803
 804         err = efa_modify_qp_validate(dev, qp, qp_attr, qp_attr_mask, cur_state,
 805                                      new_state);
 806         if (err)
 807                 return err;
 808
 809         params.qp_handle = qp->qp_handle;
 810
 811         if (qp_attr_mask & IB_QP_STATE) {
 812                 params.modify_mask |= BIT(EFA_ADMIN_QP_STATE_BIT) |
 813                                       BIT(EFA_ADMIN_CUR_QP_STATE_BIT);
 814                 params.cur_qp_state = qp_attr->cur_qp_state;
 815                 params.qp_state = qp_attr->qp_state;
 816         }
 817
 818         if (qp_attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) {
 819                 params.modify_mask |=
 820                         BIT(EFA_ADMIN_SQ_DRAINED_ASYNC_NOTIFY_BIT);
 821                 params.sq_drained_async_notify = qp_attr->en_sqd_async_notify;
 822         }
 823
 824         if (qp_attr_mask & IB_QP_QKEY) {
 825                 params.modify_mask |= BIT(EFA_ADMIN_QKEY_BIT);
 826                 params.qkey = qp_attr->qkey;
 827         }
 828
 829         if (qp_attr_mask & IB_QP_SQ_PSN) {
 830                 params.modify_mask |= BIT(EFA_ADMIN_SQ_PSN_BIT);
 831                 params.sq_psn = qp_attr->sq_psn;
 832         }
 833
 834         err = efa_com_modify_qp(&dev->edev, &params);
 835         if (err)
 836                 return err;
 837
 838         qp->state = new_state;
 839
 840         return 0;
 841 }
 842
 843 static int efa_destroy_cq_idx(struct efa_dev *dev, int cq_idx)
 844 {
 845         struct efa_com_destroy_cq_params params = { .cq_idx = cq_idx };
 846
 847         return efa_com_destroy_cq(&dev->edev, &params);
 848 }
 849
 850 void efa_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
 851 {
 852         struct efa_dev *dev = to_edev(ibcq->device);
 853         struct efa_cq *cq = to_ecq(ibcq);
 854
 855         ibdev_dbg(&dev->ibdev,
 856                   "Destroy cq[%d] virt[0x%p] freed: size[%lu], dma[%pad]\n",
 857                   cq->cq_idx, cq->cpu_addr, cq->size, &cq->dma_addr);
 858
 859         efa_destroy_cq_idx(dev, cq->cq_idx);
 860         dma_unmap_single(&dev->pdev->dev, cq->dma_addr, cq->size,
 861                          DMA_FROM_DEVICE);
 862 }
 863
 864 static int cq_mmap_entries_setup(struct efa_dev *dev, struct efa_cq *cq,
 865                                  struct efa_ibv_create_cq_resp *resp)
 866 {
 867         resp->q_mmap_size = cq->size;
 868         resp->q_mmap_key = mmap_entry_insert(dev, cq->ucontext, cq,
 869                                              virt_to_phys(cq->cpu_addr),
 870                                              cq->size, EFA_MMAP_DMA_PAGE);
 871         if (resp->q_mmap_key == EFA_MMAP_INVALID)
 872                 return -ENOMEM;
 873
 874         return 0;
 875 }
 876
 877 int efa_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 878                   struct ib_udata *udata)
 879 {
 880         struct efa_ucontext *ucontext = rdma_udata_to_drv_context(
 881                 udata, struct efa_ucontext, ibucontext);
 882         struct efa_ibv_create_cq_resp resp = {};
 883         struct efa_com_create_cq_params params;
 884         struct efa_com_create_cq_result result;
 885         struct ib_device *ibdev = ibcq->device;
 886         struct efa_dev *dev = to_edev(ibdev);
 887         struct efa_ibv_create_cq cmd = {};
 888         struct efa_cq *cq = to_ecq(ibcq);
 889         bool cq_entry_inserted = false;
 890         int entries = attr->cqe;
 891         int err;
 892
 893         ibdev_dbg(ibdev, "create_cq entries %d\n", entries);
 894
 895         if (entries < 1 || entries > dev->dev_attr.max_cq_depth) {
 896                 ibdev_dbg(ibdev,
 897                           "cq: requested entries[%u] non-positive or greater than max[%u]\n",
 898                           entries, dev->dev_attr.max_cq_depth);
 899                 err = -EINVAL;
 900                 goto err_out;
 901         }
 902
 903         if (!field_avail(cmd, num_sub_cqs, udata->inlen)) {
 904                 ibdev_dbg(ibdev,
 905                           "Incompatible ABI params, no input udata\n");
 906                 err = -EINVAL;
 907                 goto err_out;
 908         }
 909
 910         if (udata->inlen > sizeof(cmd) &&
 911             !ib_is_udata_cleared(udata, sizeof(cmd),
 912                                  udata->inlen - sizeof(cmd))) {
 913                 ibdev_dbg(ibdev,
 914                           "Incompatible ABI params, unknown fields in udata\n");
 915                 err = -EINVAL;
 916                 goto err_out;
 917         }
 918
 919         err = ib_copy_from_udata(&cmd, udata,
 920                                  min(sizeof(cmd), udata->inlen));
 921         if (err) {
 922                 ibdev_dbg(ibdev, "Cannot copy udata for create_cq\n");
 923                 goto err_out;
 924         }
 925
 926         if (cmd.comp_mask || !is_reserved_cleared(cmd.reserved_50)) {
 927                 ibdev_dbg(ibdev,
 928                           "Incompatible ABI params, unknown fields in udata\n");
 929                 err = -EINVAL;
 930                 goto err_out;
 931         }
 932
 933         if (!cmd.cq_entry_size) {
 934                 ibdev_dbg(ibdev,
 935                           "Invalid entry size [%u]\n", cmd.cq_entry_size);
 936                 err = -EINVAL;
 937                 goto err_out;
 938         }
 939
 940         if (cmd.num_sub_cqs != dev->dev_attr.sub_cqs_per_cq) {
 941                 ibdev_dbg(ibdev,
 942                           "Invalid number of sub cqs[%u] expected[%u]\n",
 943                           cmd.num_sub_cqs, dev->dev_attr.sub_cqs_per_cq);
 944                 err = -EINVAL;
 945                 goto err_out;
 946         }
 947
 948         cq->ucontext = ucontext;
 949         cq->size = PAGE_ALIGN(cmd.cq_entry_size * entries * cmd.num_sub_cqs);
 950         cq->cpu_addr = efa_zalloc_mapped(dev, &cq->dma_addr, cq->size,
 951                                          DMA_FROM_DEVICE);
 952         if (!cq->cpu_addr) {
 953                 err = -ENOMEM;
 954                 goto err_out;
 955         }
 956
 957         params.uarn = cq->ucontext->uarn;
 958         params.cq_depth = entries;
 959         params.dma_addr = cq->dma_addr;
 960         params.entry_size_in_bytes = cmd.cq_entry_size;
 961         params.num_sub_cqs = cmd.num_sub_cqs;
 962         err = efa_com_create_cq(&dev->edev, &params, &result);
 963         if (err)
 964                 goto err_free_mapped;
 965
 966         resp.cq_idx = result.cq_idx;
 967         cq->cq_idx = result.cq_idx;
 968         cq->ibcq.cqe = result.actual_depth;
 969         WARN_ON_ONCE(entries != result.actual_depth);
 970
 971         err = cq_mmap_entries_setup(dev, cq, &resp);
 972         if (err) {
 973                 ibdev_dbg(ibdev, "Could not setup cq[%u] mmap entries\n",
 974                           cq->cq_idx);
 975                 goto err_destroy_cq;
 976         }
 977
 978         cq_entry_inserted = true;
 979
 980         if (udata->outlen) {
 981                 err = ib_copy_to_udata(udata, &resp,
 982                                        min(sizeof(resp), udata->outlen));
 983                 if (err) {
 984                         ibdev_dbg(ibdev,
 985                                   "Failed to copy udata for create_cq\n");
 986                         goto err_destroy_cq;
 987                 }
 988         }
 989
 990         ibdev_dbg(ibdev, "Created cq[%d], cq depth[%u]. dma[%pad] virt[0x%p]\n",
 991                   cq->cq_idx, result.actual_depth, &cq->dma_addr, cq->cpu_addr);
 992
 993         return 0;
 994
 995 err_destroy_cq:
 996         efa_destroy_cq_idx(dev, cq->cq_idx);
 997 err_free_mapped:
 998         dma_unmap_single(&dev->pdev->dev, cq->dma_addr, cq->size,
 999                          DMA_FROM_DEVICE);
1000         if (!cq_entry_inserted)
1001                 free_pages_exact(cq->cpu_addr, cq->size);
1002 err_out:
1003         atomic64_inc(&dev->stats.sw_stats.create_cq_err);
1004         return err;
1005 }
1006
1007 static int umem_to_page_list(struct efa_dev *dev,
1008                              struct ib_umem *umem,
1009                              u64 *page_list,
1010                              u32 hp_cnt,
1011                              u8 hp_shift)
1012 {
1013         u32 pages_in_hp = BIT(hp_shift - PAGE_SHIFT);
1014         struct ib_block_iter biter;
1015         unsigned int hp_idx = 0;
1016
1017         ibdev_dbg(&dev->ibdev, "hp_cnt[%u], pages_in_hp[%u]\n",
1018                   hp_cnt, pages_in_hp);
1019
1020         rdma_for_each_block(umem->sg_head.sgl, &biter, umem->nmap,
1021                             BIT(hp_shift))
1022                 page_list[hp_idx++] = rdma_block_iter_dma_address(&biter);
1023
1024         return 0;
1025 }
1026
1027 static struct scatterlist *efa_vmalloc_buf_to_sg(u64 *buf, int page_cnt)
1028 {
1029         struct scatterlist *sglist;
1030         struct page *pg;
1031         int i;
1032
1033         sglist = kcalloc(page_cnt, sizeof(*sglist), GFP_KERNEL);
1034         if (!sglist)
1035                 return NULL;
1036         sg_init_table(sglist, page_cnt);
1037         for (i = 0; i < page_cnt; i++) {
1038                 pg = vmalloc_to_page(buf);
1039                 if (!pg)
1040                         goto err;
1041                 sg_set_page(&sglist[i], pg, PAGE_SIZE, 0);
1042                 buf += PAGE_SIZE / sizeof(*buf);
1043         }
1044         return sglist;
1045
1046 err:
1047         kfree(sglist);
1048         return NULL;
1049 }
1050
1051 /*
1052  * create a chunk list of physical pages dma addresses from the supplied
1053  * scatter gather list
1054  */
1055 static int pbl_chunk_list_create(struct efa_dev *dev, struct pbl_context *pbl)
1056 {
1057         struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list;
1058         int page_cnt = pbl->phys.indirect.pbl_buf_size_in_pages;
1059         struct scatterlist *pages_sgl = pbl->phys.indirect.sgl;
1060         unsigned int chunk_list_size, chunk_idx, payload_idx;
1061         int sg_dma_cnt = pbl->phys.indirect.sg_dma_cnt;
1062         struct efa_com_ctrl_buff_info *ctrl_buf;
1063         u64 *cur_chunk_buf, *prev_chunk_buf;
1064         struct ib_block_iter biter;
1065         dma_addr_t dma_addr;
1066         int i;
1067
1068         /* allocate a chunk list that consists of 4KB chunks */
1069         chunk_list_size = DIV_ROUND_UP(page_cnt, EFA_PTRS_PER_CHUNK);
1070
1071         chunk_list->size = chunk_list_size;
1072         chunk_list->chunks = kcalloc(chunk_list_size,
1073                                      sizeof(*chunk_list->chunks),
1074                                      GFP_KERNEL);
1075         if (!chunk_list->chunks)
1076                 return -ENOMEM;
1077
1078         ibdev_dbg(&dev->ibdev,
1079                   "chunk_list_size[%u] - pages[%u]\n", chunk_list_size,
1080                   page_cnt);
1081
1082         /* allocate chunk buffers: */
1083         for (i = 0; i < chunk_list_size; i++) {
1084                 chunk_list->chunks[i].buf = kzalloc(EFA_CHUNK_SIZE, GFP_KERNEL);
1085                 if (!chunk_list->chunks[i].buf)
1086                         goto chunk_list_dealloc;
1087
1088                 chunk_list->chunks[i].length = EFA_CHUNK_USED_SIZE;
1089         }
1090         chunk_list->chunks[chunk_list_size - 1].length =
1091                 ((page_cnt % EFA_PTRS_PER_CHUNK) * EFA_CHUNK_PAYLOAD_PTR_SIZE) +
1092                         EFA_CHUNK_PTR_SIZE;
1093
1094         /* fill the dma addresses of sg list pages to chunks: */
1095         chunk_idx = 0;
1096         payload_idx = 0;
1097         cur_chunk_buf = chunk_list->chunks[0].buf;
1098         rdma_for_each_block(pages_sgl, &biter, sg_dma_cnt,
1099                             EFA_CHUNK_PAYLOAD_SIZE) {
1100                 cur_chunk_buf[payload_idx++] =
1101                         rdma_block_iter_dma_address(&biter);
1102
1103                 if (payload_idx == EFA_PTRS_PER_CHUNK) {
1104                         chunk_idx++;
1105                         cur_chunk_buf = chunk_list->chunks[chunk_idx].buf;
1106                         payload_idx = 0;
1107                 }
1108         }
1109
1110         /* map chunks to dma and fill chunks next ptrs */
1111         for (i = chunk_list_size - 1; i >= 0; i--) {
1112                 dma_addr = dma_map_single(&dev->pdev->dev,
1113                                           chunk_list->chunks[i].buf,
1114                                           chunk_list->chunks[i].length,
1115                                           DMA_TO_DEVICE);
1116                 if (dma_mapping_error(&dev->pdev->dev, dma_addr)) {
1117                         ibdev_err(&dev->ibdev,
1118                                   "chunk[%u] dma_map_failed\n", i);
1119                         goto chunk_list_unmap;
1120                 }
1121
1122                 chunk_list->chunks[i].dma_addr = dma_addr;
1123                 ibdev_dbg(&dev->ibdev,
1124                           "chunk[%u] mapped at [%pad]\n", i, &dma_addr);
1125
1126                 if (!i)
1127                         break;
1128
1129                 prev_chunk_buf = chunk_list->chunks[i - 1].buf;
1130
1131                 ctrl_buf = (struct efa_com_ctrl_buff_info *)
1132                                 &prev_chunk_buf[EFA_PTRS_PER_CHUNK];
1133                 ctrl_buf->length = chunk_list->chunks[i].length;
1134
1135                 efa_com_set_dma_addr(dma_addr,
1136                                      &ctrl_buf->address.mem_addr_high,
1137                                      &ctrl_buf->address.mem_addr_low);
1138         }
1139
1140         return 0;
1141
1142 chunk_list_unmap:
1143         for (; i < chunk_list_size; i++) {
1144                 dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr,
1145                                  chunk_list->chunks[i].length, DMA_TO_DEVICE);
1146         }
1147 chunk_list_dealloc:
1148         for (i = 0; i < chunk_list_size; i++)
1149                 kfree(chunk_list->chunks[i].buf);
1150
1151         kfree(chunk_list->chunks);
1152         return -ENOMEM;
1153 }
1154
1155 static void pbl_chunk_list_destroy(struct efa_dev *dev, struct pbl_context *pbl)
1156 {
1157         struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list;
1158         int i;
1159
1160         for (i = 0; i < chunk_list->size; i++) {
1161                 dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr,
1162                                  chunk_list->chunks[i].length, DMA_TO_DEVICE);
1163                 kfree(chunk_list->chunks[i].buf);
1164         }
1165
1166         kfree(chunk_list->chunks);
1167 }
1168
1169 /* initialize pbl continuous mode: map pbl buffer to a dma address. */
1170 static int pbl_continuous_initialize(struct efa_dev *dev,
1171                                      struct pbl_context *pbl)
1172 {
1173         dma_addr_t dma_addr;
1174
1175         dma_addr = dma_map_single(&dev->pdev->dev, pbl->pbl_buf,
1176                                   pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE);
1177         if (dma_mapping_error(&dev->pdev->dev, dma_addr)) {
1178                 ibdev_err(&dev->ibdev, "Unable to map pbl to DMA address\n");
1179                 return -ENOMEM;
1180         }
1181
1182         pbl->phys.continuous.dma_addr = dma_addr;
1183         ibdev_dbg(&dev->ibdev,
1184                   "pbl continuous - dma_addr = %pad, size[%u]\n",
1185                   &dma_addr, pbl->pbl_buf_size_in_bytes);
1186
1187         return 0;
1188 }
1189
1190 /*
1191  * initialize pbl indirect mode:
1192  * create a chunk list out of the dma addresses of the physical pages of
1193  * pbl buffer.
1194  */
1195 static int pbl_indirect_initialize(struct efa_dev *dev, struct pbl_context *pbl)
1196 {
1197         u32 size_in_pages = DIV_ROUND_UP(pbl->pbl_buf_size_in_bytes, PAGE_SIZE);
1198         struct scatterlist *sgl;
1199         int sg_dma_cnt, err;
1200
1201         BUILD_BUG_ON(EFA_CHUNK_PAYLOAD_SIZE > PAGE_SIZE);
1202         sgl = efa_vmalloc_buf_to_sg(pbl->pbl_buf, size_in_pages);
1203         if (!sgl)
1204                 return -ENOMEM;
1205
1206         sg_dma_cnt = dma_map_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE);
1207         if (!sg_dma_cnt) {
1208                 err = -EINVAL;
1209                 goto err_map;
1210         }
1211
1212         pbl->phys.indirect.pbl_buf_size_in_pages = size_in_pages;
1213         pbl->phys.indirect.sgl = sgl;
1214         pbl->phys.indirect.sg_dma_cnt = sg_dma_cnt;
1215         err = pbl_chunk_list_create(dev, pbl);
1216         if (err) {
1217                 ibdev_dbg(&dev->ibdev,
1218                           "chunk_list creation failed[%d]\n", err);
1219                 goto err_chunk;
1220         }
1221
1222         ibdev_dbg(&dev->ibdev,
1223                   "pbl indirect - size[%u], chunks[%u]\n",
1224                   pbl->pbl_buf_size_in_bytes,
1225                   pbl->phys.indirect.chunk_list.size);
1226
1227         return 0;
1228
1229 err_chunk:
1230         dma_unmap_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE);
1231 err_map:
1232         kfree(sgl);
1233         return err;
1234 }
1235
1236 static void pbl_indirect_terminate(struct efa_dev *dev, struct pbl_context *pbl)
1237 {
1238         pbl_chunk_list_destroy(dev, pbl);
1239         dma_unmap_sg(&dev->pdev->dev, pbl->phys.indirect.sgl,
1240                      pbl->phys.indirect.pbl_buf_size_in_pages, DMA_TO_DEVICE);
1241         kfree(pbl->phys.indirect.sgl);
1242 }
1243
1244 /* create a page buffer list from a mapped user memory region */
1245 static int pbl_create(struct efa_dev *dev,
1246                       struct pbl_context *pbl,
1247                       struct ib_umem *umem,
1248                       int hp_cnt,
1249                       u8 hp_shift)
1250 {
1251         int err;
1252
1253         pbl->pbl_buf_size_in_bytes = hp_cnt * EFA_CHUNK_PAYLOAD_PTR_SIZE;
1254         pbl->pbl_buf = kvzalloc(pbl->pbl_buf_size_in_bytes, GFP_KERNEL);
1255         if (!pbl->pbl_buf)
1256                 return -ENOMEM;
1257
1258         if (is_vmalloc_addr(pbl->pbl_buf)) {
1259                 pbl->physically_continuous = 0;
1260                 err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt,
1261                                         hp_shift);
1262                 if (err)
1263                         goto err_free;
1264
1265                 err = pbl_indirect_initialize(dev, pbl);
1266                 if (err)
1267                         goto err_free;
1268         } else {
1269                 pbl->physically_continuous = 1;
1270                 err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt,
1271                                         hp_shift);
1272                 if (err)
1273                         goto err_free;
1274
1275                 err = pbl_continuous_initialize(dev, pbl);
1276                 if (err)
1277                         goto err_free;
1278         }
1279
1280         ibdev_dbg(&dev->ibdev,
1281                   "user_pbl_created: user_pages[%u], continuous[%u]\n",
1282                   hp_cnt, pbl->physically_continuous);
1283
1284         return 0;
1285
1286 err_free:
1287         kvfree(pbl->pbl_buf);
1288         return err;
1289 }
1290
1291 static void pbl_destroy(struct efa_dev *dev, struct pbl_context *pbl)
1292 {
1293         if (pbl->physically_continuous)
1294                 dma_unmap_single(&dev->pdev->dev, pbl->phys.continuous.dma_addr,
1295                                  pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE);
1296         else
1297                 pbl_indirect_terminate(dev, pbl);
1298
1299         kvfree(pbl->pbl_buf);
1300 }
1301
1302 static int efa_create_inline_pbl(struct efa_dev *dev, struct efa_mr *mr,
1303                                  struct efa_com_reg_mr_params *params)
1304 {
1305         int err;
1306
1307         params->inline_pbl = 1;
1308         err = umem_to_page_list(dev, mr->umem, params->pbl.inline_pbl_array,
1309                                 params->page_num, params->page_shift);
1310         if (err)
1311                 return err;
1312
1313         ibdev_dbg(&dev->ibdev,
1314                   "inline_pbl_array - pages[%u]\n", params->page_num);
1315
1316         return 0;
1317 }
1318
1319 static int efa_create_pbl(struct efa_dev *dev,
1320                           struct pbl_context *pbl,
1321                           struct efa_mr *mr,
1322                           struct efa_com_reg_mr_params *params)
1323 {
1324         int err;
1325
1326         err = pbl_create(dev, pbl, mr->umem, params->page_num,
1327                          params->page_shift);
1328         if (err) {
1329                 ibdev_dbg(&dev->ibdev, "Failed to create pbl[%d]\n", err);
1330                 return err;
1331         }
1332
1333         params->inline_pbl = 0;
1334         params->indirect = !pbl->physically_continuous;
1335         if (pbl->physically_continuous) {
1336                 params->pbl.pbl.length = pbl->pbl_buf_size_in_bytes;
1337
1338                 efa_com_set_dma_addr(pbl->phys.continuous.dma_addr,
1339                                      &params->pbl.pbl.address.mem_addr_high,
1340                                      &params->pbl.pbl.address.mem_addr_low);
1341         } else {
1342                 params->pbl.pbl.length =
1343                         pbl->phys.indirect.chunk_list.chunks[0].length;
1344
1345                 efa_com_set_dma_addr(pbl->phys.indirect.chunk_list.chunks[0].dma_addr,
1346                                      &params->pbl.pbl.address.mem_addr_high,
1347                                      &params->pbl.pbl.address.mem_addr_low);
1348         }
1349
1350         return 0;
1351 }
1352
1353 struct ib_mr *efa_reg_mr(struct ib_pd *ibpd, u64 start, u64 length,
1354                          u64 virt_addr, int access_flags,
1355                          struct ib_udata *udata)
1356 {
1357         struct efa_dev *dev = to_edev(ibpd->device);
1358         struct efa_com_reg_mr_params params = {};
1359         struct efa_com_reg_mr_result result = {};
1360         struct pbl_context pbl;
1361         unsigned int pg_sz;
1362         struct efa_mr *mr;
1363         int inline_size;
1364         int err;
1365
1366         if (udata->inlen &&
1367             !ib_is_udata_cleared(udata, 0, sizeof(udata->inlen))) {
1368                 ibdev_dbg(&dev->ibdev,
1369                           "Incompatible ABI params, udata not cleared\n");
1370                 err = -EINVAL;
1371                 goto err_out;
1372         }
1373
1374         if (access_flags & ~EFA_SUPPORTED_ACCESS_FLAGS) {
1375                 ibdev_dbg(&dev->ibdev,
1376                           "Unsupported access flags[%#x], supported[%#x]\n",
1377                           access_flags, EFA_SUPPORTED_ACCESS_FLAGS);
1378                 err = -EOPNOTSUPP;
1379                 goto err_out;
1380         }
1381
1382         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1383         if (!mr) {
1384                 err = -ENOMEM;
1385                 goto err_out;
1386         }
1387
1388         mr->umem = ib_umem_get(udata, start, length, access_flags, 0);
1389         if (IS_ERR(mr->umem)) {
1390                 err = PTR_ERR(mr->umem);
1391                 ibdev_dbg(&dev->ibdev,
1392                           "Failed to pin and map user space memory[%d]\n", err);
1393                 goto err_free;
1394         }
1395
1396         params.pd = to_epd(ibpd)->pdn;
1397         params.iova = virt_addr;
1398         params.mr_length_in_bytes = length;
1399         params.permissions = access_flags & 0x1;
1400
1401         pg_sz = ib_umem_find_best_pgsz(mr->umem,
1402                                        dev->dev_attr.page_size_cap,
1403                                        virt_addr);
1404         if (!pg_sz) {
1405                 err = -EOPNOTSUPP;
1406                 ibdev_dbg(&dev->ibdev, "Failed to find a suitable page size in page_size_cap %#llx\n",
1407                           dev->dev_attr.page_size_cap);
1408                 goto err_unmap;
1409         }
1410
1411         params.page_shift = __ffs(pg_sz);
1412         params.page_num = DIV_ROUND_UP(length + (start & (pg_sz - 1)),
1413                                        pg_sz);
1414
1415         ibdev_dbg(&dev->ibdev,
1416                   "start %#llx length %#llx params.page_shift %u params.page_num %u\n",
1417                   start, length, params.page_shift, params.page_num);
1418
1419         inline_size = ARRAY_SIZE(params.pbl.inline_pbl_array);
1420         if (params.page_num <= inline_size) {
1421                 err = efa_create_inline_pbl(dev, mr, &params);
1422                 if (err)
1423                         goto err_unmap;
1424
1425                 err = efa_com_register_mr(&dev->edev, &params, &result);
1426                 if (err)
1427                         goto err_unmap;
1428         } else {
1429                 err = efa_create_pbl(dev, &pbl, mr, &params);
1430                 if (err)
1431                         goto err_unmap;
1432
1433                 err = efa_com_register_mr(&dev->edev, &params, &result);
1434                 pbl_destroy(dev, &pbl);
1435
1436                 if (err)
1437                         goto err_unmap;
1438         }
1439
1440         mr->ibmr.lkey = result.l_key;
1441         mr->ibmr.rkey = result.r_key;
1442         mr->ibmr.length = length;
1443         ibdev_dbg(&dev->ibdev, "Registered mr[%d]\n", mr->ibmr.lkey);
1444
1445         return &mr->ibmr;
1446
1447 err_unmap:
1448         ib_umem_release(mr->umem);
1449 err_free:
1450         kfree(mr);
1451 err_out:
1452         atomic64_inc(&dev->stats.sw_stats.reg_mr_err);
1453         return ERR_PTR(err);
1454 }
1455
1456 int efa_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
1457 {
1458         struct efa_dev *dev = to_edev(ibmr->device);
1459         struct efa_com_dereg_mr_params params;
1460         struct efa_mr *mr = to_emr(ibmr);
1461         int err;
1462
1463         ibdev_dbg(&dev->ibdev, "Deregister mr[%d]\n", ibmr->lkey);
1464
1465         if (mr->umem) {
1466                 params.l_key = mr->ibmr.lkey;
1467                 err = efa_com_dereg_mr(&dev->edev, &params);
1468                 if (err)
1469                         return err;
1470         }
1471         ib_umem_release(mr->umem);
1472
1473         kfree(mr);
1474
1475         return 0;
1476 }
1477
1478 int efa_get_port_immutable(struct ib_device *ibdev, u8 port_num,
1479                            struct ib_port_immutable *immutable)
1480 {
1481         struct ib_port_attr attr;
1482         int err;
1483
1484         err = ib_query_port(ibdev, port_num, &attr);
1485         if (err) {
1486                 ibdev_dbg(ibdev, "Couldn't query port err[%d]\n", err);
1487                 return err;
1488         }
1489
1490         immutable->pkey_tbl_len = attr.pkey_tbl_len;
1491         immutable->gid_tbl_len = attr.gid_tbl_len;
1492
1493         return 0;
1494 }
1495
1496 static int efa_dealloc_uar(struct efa_dev *dev, u16 uarn)
1497 {
1498         struct efa_com_dealloc_uar_params params = {
1499                 .uarn = uarn,
1500         };
1501
1502         return efa_com_dealloc_uar(&dev->edev, &params);
1503 }
1504
1505 int efa_alloc_ucontext(struct ib_ucontext *ibucontext, struct ib_udata *udata)
1506 {
1507         struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1508         struct efa_dev *dev = to_edev(ibucontext->device);
1509         struct efa_ibv_alloc_ucontext_resp resp = {};
1510         struct efa_com_alloc_uar_result result;
1511         int err;
1512
1513         /*
1514          * it's fine if the driver does not know all request fields,
1515          * we will ack input fields in our response.
1516          */
1517
1518         err = efa_com_alloc_uar(&dev->edev, &result);
1519         if (err)
1520                 goto err_out;
1521
1522         ucontext->uarn = result.uarn;
1523         xa_init(&ucontext->mmap_xa);
1524
1525         resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_QUERY_DEVICE;
1526         resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_CREATE_AH;
1527         resp.sub_cqs_per_cq = dev->dev_attr.sub_cqs_per_cq;
1528         resp.inline_buf_size = dev->dev_attr.inline_buf_size;
1529         resp.max_llq_size = dev->dev_attr.max_llq_size;
1530
1531         if (udata && udata->outlen) {
1532                 err = ib_copy_to_udata(udata, &resp,
1533                                        min(sizeof(resp), udata->outlen));
1534                 if (err)
1535                         goto err_dealloc_uar;
1536         }
1537
1538         return 0;
1539
1540 err_dealloc_uar:
1541         efa_dealloc_uar(dev, result.uarn);
1542 err_out:
1543         atomic64_inc(&dev->stats.sw_stats.alloc_ucontext_err);
1544         return err;
1545 }
1546
1547 void efa_dealloc_ucontext(struct ib_ucontext *ibucontext)
1548 {
1549         struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1550         struct efa_dev *dev = to_edev(ibucontext->device);
1551
1552         mmap_entries_remove_free(dev, ucontext);
1553         efa_dealloc_uar(dev, ucontext->uarn);
1554 }
1555
1556 static int __efa_mmap(struct efa_dev *dev, struct efa_ucontext *ucontext,
1557                       struct vm_area_struct *vma, u64 key, u64 length)
1558 {
1559         struct efa_mmap_entry *entry;
1560         unsigned long va;
1561         u64 pfn;
1562         int err;
1563
1564         entry = mmap_entry_get(dev, ucontext, key, length);
1565         if (!entry) {
1566                 ibdev_dbg(&dev->ibdev, "key[%#llx] does not have valid entry\n",
1567                           key);
1568                 return -EINVAL;
1569         }
1570
1571         ibdev_dbg(&dev->ibdev,
1572                   "Mapping address[%#llx], length[%#llx], mmap_flag[%d]\n",
1573                   entry->address, length, entry->mmap_flag);
1574
1575         pfn = entry->address >> PAGE_SHIFT;
1576         switch (entry->mmap_flag) {
1577         case EFA_MMAP_IO_NC:
1578                 err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn, length,
1579                                         pgprot_noncached(vma->vm_page_prot));
1580                 break;
1581         case EFA_MMAP_IO_WC:
1582                 err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn, length,
1583                                         pgprot_writecombine(vma->vm_page_prot));
1584                 break;
1585         case EFA_MMAP_DMA_PAGE:
1586                 for (va = vma->vm_start; va < vma->vm_end;
1587                      va += PAGE_SIZE, pfn++) {
1588                         err = vm_insert_page(vma, va, pfn_to_page(pfn));
1589                         if (err)
1590                                 break;
1591                 }
1592                 break;
1593         default:
1594                 err = -EINVAL;
1595         }
1596
1597         if (err)
1598                 ibdev_dbg(
1599                         &dev->ibdev,
1600                         "Couldn't mmap address[%#llx] length[%#llx] mmap_flag[%d] err[%d]\n",
1601                         entry->address, length, entry->mmap_flag, err);
1602
1603         return err;
1604 }
1605
1606 int efa_mmap(struct ib_ucontext *ibucontext,
1607              struct vm_area_struct *vma)
1608 {
1609         struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1610         struct efa_dev *dev = to_edev(ibucontext->device);
1611         u64 length = vma->vm_end - vma->vm_start;
1612         u64 key = vma->vm_pgoff << PAGE_SHIFT;
1613
1614         ibdev_dbg(&dev->ibdev,
1615                   "start %#lx, end %#lx, length = %#llx, key = %#llx\n",
1616                   vma->vm_start, vma->vm_end, length, key);
1617
1618         if (length % PAGE_SIZE != 0 || !(vma->vm_flags & VM_SHARED)) {
1619                 ibdev_dbg(&dev->ibdev,
1620                           "length[%#llx] is not page size aligned[%#lx] or VM_SHARED is not set [%#lx]\n",
1621                           length, PAGE_SIZE, vma->vm_flags);
1622                 return -EINVAL;
1623         }
1624
1625         if (vma->vm_flags & VM_EXEC) {
1626                 ibdev_dbg(&dev->ibdev, "Mapping executable pages is not permitted\n");
1627                 return -EPERM;
1628         }
1629         vma->vm_flags &= ~VM_MAYEXEC;
1630
1631         return __efa_mmap(dev, ucontext, vma, key, length);
1632 }
1633
1634 static int efa_ah_destroy(struct efa_dev *dev, struct efa_ah *ah)
1635 {
1636         struct efa_com_destroy_ah_params params = {
1637                 .ah = ah->ah,
1638                 .pdn = to_epd(ah->ibah.pd)->pdn,
1639         };
1640
1641         return efa_com_destroy_ah(&dev->edev, &params);
1642 }
1643
1644 int efa_create_ah(struct ib_ah *ibah,
1645                   struct rdma_ah_attr *ah_attr,
1646                   u32 flags,
1647                   struct ib_udata *udata)
1648 {
1649         struct efa_dev *dev = to_edev(ibah->device);
1650         struct efa_com_create_ah_params params = {};
1651         struct efa_ibv_create_ah_resp resp = {};
1652         struct efa_com_create_ah_result result;
1653         struct efa_ah *ah = to_eah(ibah);
1654         int err;
1655
1656         if (!(flags & RDMA_CREATE_AH_SLEEPABLE)) {
1657                 ibdev_dbg(&dev->ibdev,
1658                           "Create address handle is not supported in atomic context\n");
1659                 err = -EOPNOTSUPP;
1660                 goto err_out;
1661         }
1662
1663         if (udata->inlen &&
1664             !ib_is_udata_cleared(udata, 0, udata->inlen)) {
1665                 ibdev_dbg(&dev->ibdev, "Incompatible ABI params\n");
1666                 err = -EINVAL;
1667                 goto err_out;
1668         }
1669
1670         memcpy(params.dest_addr, ah_attr->grh.dgid.raw,
1671                sizeof(params.dest_addr));
1672         params.pdn = to_epd(ibah->pd)->pdn;
1673         err = efa_com_create_ah(&dev->edev, &params, &result);
1674         if (err)
1675                 goto err_out;
1676
1677         memcpy(ah->id, ah_attr->grh.dgid.raw, sizeof(ah->id));
1678         ah->ah = result.ah;
1679
1680         resp.efa_address_handle = result.ah;
1681
1682         if (udata->outlen) {
1683                 err = ib_copy_to_udata(udata, &resp,
1684                                        min(sizeof(resp), udata->outlen));
1685                 if (err) {
1686                         ibdev_dbg(&dev->ibdev,
1687                                   "Failed to copy udata for create_ah response\n");
1688                         goto err_destroy_ah;
1689                 }
1690         }
1691         ibdev_dbg(&dev->ibdev, "Created ah[%d]\n", ah->ah);
1692
1693         return 0;
1694
1695 err_destroy_ah:
1696         efa_ah_destroy(dev, ah);
1697 err_out:
1698         atomic64_inc(&dev->stats.sw_stats.create_ah_err);
1699         return err;
1700 }
1701
1702 void efa_destroy_ah(struct ib_ah *ibah, u32 flags)
1703 {
1704         struct efa_dev *dev = to_edev(ibah->pd->device);
1705         struct efa_ah *ah = to_eah(ibah);
1706
1707         ibdev_dbg(&dev->ibdev, "Destroy ah[%d]\n", ah->ah);
1708
1709         if (!(flags & RDMA_DESTROY_AH_SLEEPABLE)) {
1710                 ibdev_dbg(&dev->ibdev,
1711                           "Destroy address handle is not supported in atomic context\n");
1712                 return;
1713         }
1714
1715         efa_ah_destroy(dev, ah);
1716 }
1717
1718 enum rdma_link_layer efa_port_link_layer(struct ib_device *ibdev,
1719                                          u8 port_num)
1720 {
1721         return IB_LINK_LAYER_UNSPECIFIED;
1722 }
1723