drivers/infiniband/hw/efa/efa_verbs.c

   1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
   2 /*
   3  * Copyright 2018-2021 Amazon.com, Inc. or its affiliates. All rights reserved.
   4  */
   5
   6 #include <linux/vmalloc.h>
   7 #include <linux/log2.h>
   8
   9 #include <rdma/ib_addr.h>
  10 #include <rdma/ib_umem.h>
  11 #include <rdma/ib_user_verbs.h>
  12 #include <rdma/ib_verbs.h>
  13 #include <rdma/uverbs_ioctl.h>
  14
  15 #include "efa.h"
  16
  17 enum {
  18         EFA_MMAP_DMA_PAGE = 0,
  19         EFA_MMAP_IO_WC,
  20         EFA_MMAP_IO_NC,
  21 };
  22
  23 #define EFA_AENQ_ENABLED_GROUPS \
  24         (BIT(EFA_ADMIN_FATAL_ERROR) | BIT(EFA_ADMIN_WARNING) | \
  25          BIT(EFA_ADMIN_NOTIFICATION) | BIT(EFA_ADMIN_KEEP_ALIVE))
  26
  27 struct efa_user_mmap_entry {
  28         struct rdma_user_mmap_entry rdma_entry;
  29         u64 address;
  30         u8 mmap_flag;
  31 };
  32
  33 #define EFA_DEFINE_DEVICE_STATS(op) \
  34         op(EFA_SUBMITTED_CMDS, "submitted_cmds") \
  35         op(EFA_COMPLETED_CMDS, "completed_cmds") \
  36         op(EFA_CMDS_ERR, "cmds_err") \
  37         op(EFA_NO_COMPLETION_CMDS, "no_completion_cmds") \
  38         op(EFA_KEEP_ALIVE_RCVD, "keep_alive_rcvd") \
  39         op(EFA_ALLOC_PD_ERR, "alloc_pd_err") \
  40         op(EFA_CREATE_QP_ERR, "create_qp_err") \
  41         op(EFA_CREATE_CQ_ERR, "create_cq_err") \
  42         op(EFA_REG_MR_ERR, "reg_mr_err") \
  43         op(EFA_ALLOC_UCONTEXT_ERR, "alloc_ucontext_err") \
  44         op(EFA_CREATE_AH_ERR, "create_ah_err") \
  45         op(EFA_MMAP_ERR, "mmap_err")
  46
  47 #define EFA_DEFINE_PORT_STATS(op) \
  48         op(EFA_TX_BYTES, "tx_bytes") \
  49         op(EFA_TX_PKTS, "tx_pkts") \
  50         op(EFA_RX_BYTES, "rx_bytes") \
  51         op(EFA_RX_PKTS, "rx_pkts") \
  52         op(EFA_RX_DROPS, "rx_drops") \
  53         op(EFA_SEND_BYTES, "send_bytes") \
  54         op(EFA_SEND_WRS, "send_wrs") \
  55         op(EFA_RECV_BYTES, "recv_bytes") \
  56         op(EFA_RECV_WRS, "recv_wrs") \
  57         op(EFA_RDMA_READ_WRS, "rdma_read_wrs") \
  58         op(EFA_RDMA_READ_BYTES, "rdma_read_bytes") \
  59         op(EFA_RDMA_READ_WR_ERR, "rdma_read_wr_err") \
  60         op(EFA_RDMA_READ_RESP_BYTES, "rdma_read_resp_bytes") \
  61
  62 #define EFA_STATS_ENUM(ename, name) ename,
  63 #define EFA_STATS_STR(ename, name) [ename] = name,
  64
  65 enum efa_hw_device_stats {
  66         EFA_DEFINE_DEVICE_STATS(EFA_STATS_ENUM)
  67 };
  68
  69 static const char *const efa_device_stats_names[] = {
  70         EFA_DEFINE_DEVICE_STATS(EFA_STATS_STR)
  71 };
  72
  73 enum efa_hw_port_stats {
  74         EFA_DEFINE_PORT_STATS(EFA_STATS_ENUM)
  75 };
  76
  77 static const char *const efa_port_stats_names[] = {
  78         EFA_DEFINE_PORT_STATS(EFA_STATS_STR)
  79 };
  80
  81 #define EFA_CHUNK_PAYLOAD_SHIFT       12
  82 #define EFA_CHUNK_PAYLOAD_SIZE        BIT(EFA_CHUNK_PAYLOAD_SHIFT)
  83 #define EFA_CHUNK_PAYLOAD_PTR_SIZE    8
  84
  85 #define EFA_CHUNK_SHIFT               12
  86 #define EFA_CHUNK_SIZE                BIT(EFA_CHUNK_SHIFT)
  87 #define EFA_CHUNK_PTR_SIZE            sizeof(struct efa_com_ctrl_buff_info)
  88
  89 #define EFA_PTRS_PER_CHUNK \
  90         ((EFA_CHUNK_SIZE - EFA_CHUNK_PTR_SIZE) / EFA_CHUNK_PAYLOAD_PTR_SIZE)
  91
  92 #define EFA_CHUNK_USED_SIZE \
  93         ((EFA_PTRS_PER_CHUNK * EFA_CHUNK_PAYLOAD_PTR_SIZE) + EFA_CHUNK_PTR_SIZE)
  94
  95 struct pbl_chunk {
  96         dma_addr_t dma_addr;
  97         u64 *buf;
  98         u32 length;
  99 };
 100
 101 struct pbl_chunk_list {
 102         struct pbl_chunk *chunks;
 103         unsigned int size;
 104 };
 105
 106 struct pbl_context {
 107         union {
 108                 struct {
 109                         dma_addr_t dma_addr;
 110                 } continuous;
 111                 struct {
 112                         u32 pbl_buf_size_in_pages;
 113                         struct scatterlist *sgl;
 114                         int sg_dma_cnt;
 115                         struct pbl_chunk_list chunk_list;
 116                 } indirect;
 117         } phys;
 118         u64 *pbl_buf;
 119         u32 pbl_buf_size_in_bytes;
 120         u8 physically_continuous;
 121 };
 122
 123 static inline struct efa_dev *to_edev(struct ib_device *ibdev)
 124 {
 125         return container_of(ibdev, struct efa_dev, ibdev);
 126 }
 127
 128 static inline struct efa_ucontext *to_eucontext(struct ib_ucontext *ibucontext)
 129 {
 130         return container_of(ibucontext, struct efa_ucontext, ibucontext);
 131 }
 132
 133 static inline struct efa_pd *to_epd(struct ib_pd *ibpd)
 134 {
 135         return container_of(ibpd, struct efa_pd, ibpd);
 136 }
 137
 138 static inline struct efa_mr *to_emr(struct ib_mr *ibmr)
 139 {
 140         return container_of(ibmr, struct efa_mr, ibmr);
 141 }
 142
 143 static inline struct efa_qp *to_eqp(struct ib_qp *ibqp)
 144 {
 145         return container_of(ibqp, struct efa_qp, ibqp);
 146 }
 147
 148 static inline struct efa_cq *to_ecq(struct ib_cq *ibcq)
 149 {
 150         return container_of(ibcq, struct efa_cq, ibcq);
 151 }
 152
 153 static inline struct efa_ah *to_eah(struct ib_ah *ibah)
 154 {
 155         return container_of(ibah, struct efa_ah, ibah);
 156 }
 157
 158 static inline struct efa_user_mmap_entry *
 159 to_emmap(struct rdma_user_mmap_entry *rdma_entry)
 160 {
 161         return container_of(rdma_entry, struct efa_user_mmap_entry, rdma_entry);
 162 }
 163
 164 #define EFA_DEV_CAP(dev, cap) \
 165         ((dev)->dev_attr.device_caps & \
 166          EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_##cap##_MASK)
 167
 168 #define is_reserved_cleared(reserved) \
 169         !memchr_inv(reserved, 0, sizeof(reserved))
 170
 171 static void *efa_zalloc_mapped(struct efa_dev *dev, dma_addr_t *dma_addr,
 172                                size_t size, enum dma_data_direction dir)
 173 {
 174         void *addr;
 175
 176         addr = alloc_pages_exact(size, GFP_KERNEL | __GFP_ZERO);
 177         if (!addr)
 178                 return NULL;
 179
 180         *dma_addr = dma_map_single(&dev->pdev->dev, addr, size, dir);
 181         if (dma_mapping_error(&dev->pdev->dev, *dma_addr)) {
 182                 ibdev_err(&dev->ibdev, "Failed to map DMA address\n");
 183                 free_pages_exact(addr, size);
 184                 return NULL;
 185         }
 186
 187         return addr;
 188 }
 189
 190 static void efa_free_mapped(struct efa_dev *dev, void *cpu_addr,
 191                             dma_addr_t dma_addr,
 192                             size_t size, enum dma_data_direction dir)
 193 {
 194         dma_unmap_single(&dev->pdev->dev, dma_addr, size, dir);
 195         free_pages_exact(cpu_addr, size);
 196 }
 197
 198 int efa_query_device(struct ib_device *ibdev,
 199                      struct ib_device_attr *props,
 200                      struct ib_udata *udata)
 201 {
 202         struct efa_com_get_device_attr_result *dev_attr;
 203         struct efa_ibv_ex_query_device_resp resp = {};
 204         struct efa_dev *dev = to_edev(ibdev);
 205         int err;
 206
 207         if (udata && udata->inlen &&
 208             !ib_is_udata_cleared(udata, 0, udata->inlen)) {
 209                 ibdev_dbg(ibdev,
 210                           "Incompatible ABI params, udata not cleared\n");
 211                 return -EINVAL;
 212         }
 213
 214         dev_attr = &dev->dev_attr;
 215
 216         memset(props, 0, sizeof(*props));
 217         props->max_mr_size = dev_attr->max_mr_pages * PAGE_SIZE;
 218         props->page_size_cap = dev_attr->page_size_cap;
 219         props->vendor_id = dev->pdev->vendor;
 220         props->vendor_part_id = dev->pdev->device;
 221         props->hw_ver = dev->pdev->subsystem_device;
 222         props->max_qp = dev_attr->max_qp;
 223         props->max_cq = dev_attr->max_cq;
 224         props->max_pd = dev_attr->max_pd;
 225         props->max_mr = dev_attr->max_mr;
 226         props->max_ah = dev_attr->max_ah;
 227         props->max_cqe = dev_attr->max_cq_depth;
 228         props->max_qp_wr = min_t(u32, dev_attr->max_sq_depth,
 229                                  dev_attr->max_rq_depth);
 230         props->max_send_sge = dev_attr->max_sq_sge;
 231         props->max_recv_sge = dev_attr->max_rq_sge;
 232         props->max_sge_rd = dev_attr->max_wr_rdma_sge;
 233         props->max_pkeys = 1;
 234
 235         if (udata && udata->outlen) {
 236                 resp.max_sq_sge = dev_attr->max_sq_sge;
 237                 resp.max_rq_sge = dev_attr->max_rq_sge;
 238                 resp.max_sq_wr = dev_attr->max_sq_depth;
 239                 resp.max_rq_wr = dev_attr->max_rq_depth;
 240                 resp.max_rdma_size = dev_attr->max_rdma_size;
 241
 242                 if (EFA_DEV_CAP(dev, RDMA_READ))
 243                         resp.device_caps |= EFA_QUERY_DEVICE_CAPS_RDMA_READ;
 244
 245                 if (EFA_DEV_CAP(dev, RNR_RETRY))
 246                         resp.device_caps |= EFA_QUERY_DEVICE_CAPS_RNR_RETRY;
 247
 248                 err = ib_copy_to_udata(udata, &resp,
 249                                        min(sizeof(resp), udata->outlen));
 250                 if (err) {
 251                         ibdev_dbg(ibdev,
 252                                   "Failed to copy udata for query_device\n");
 253                         return err;
 254                 }
 255         }
 256
 257         return 0;
 258 }
 259
 260 int efa_query_port(struct ib_device *ibdev, u32 port,
 261                    struct ib_port_attr *props)
 262 {
 263         struct efa_dev *dev = to_edev(ibdev);
 264
 265         props->lmc = 1;
 266
 267         props->state = IB_PORT_ACTIVE;
 268         props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
 269         props->gid_tbl_len = 1;
 270         props->pkey_tbl_len = 1;
 271         props->active_speed = IB_SPEED_EDR;
 272         props->active_width = IB_WIDTH_4X;
 273         props->max_mtu = ib_mtu_int_to_enum(dev->dev_attr.mtu);
 274         props->active_mtu = ib_mtu_int_to_enum(dev->dev_attr.mtu);
 275         props->max_msg_sz = dev->dev_attr.mtu;
 276         props->max_vl_num = 1;
 277
 278         return 0;
 279 }
 280
 281 int efa_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
 282                  int qp_attr_mask,
 283                  struct ib_qp_init_attr *qp_init_attr)
 284 {
 285         struct efa_dev *dev = to_edev(ibqp->device);
 286         struct efa_com_query_qp_params params = {};
 287         struct efa_com_query_qp_result result;
 288         struct efa_qp *qp = to_eqp(ibqp);
 289         int err;
 290
 291 #define EFA_QUERY_QP_SUPP_MASK \
 292         (IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT | \
 293          IB_QP_QKEY | IB_QP_SQ_PSN | IB_QP_CAP | IB_QP_RNR_RETRY)
 294
 295         if (qp_attr_mask & ~EFA_QUERY_QP_SUPP_MASK) {
 296                 ibdev_dbg(&dev->ibdev,
 297                           "Unsupported qp_attr_mask[%#x] supported[%#x]\n",
 298                           qp_attr_mask, EFA_QUERY_QP_SUPP_MASK);
 299                 return -EOPNOTSUPP;
 300         }
 301
 302         memset(qp_attr, 0, sizeof(*qp_attr));
 303         memset(qp_init_attr, 0, sizeof(*qp_init_attr));
 304
 305         params.qp_handle = qp->qp_handle;
 306         err = efa_com_query_qp(&dev->edev, &params, &result);
 307         if (err)
 308                 return err;
 309
 310         qp_attr->qp_state = result.qp_state;
 311         qp_attr->qkey = result.qkey;
 312         qp_attr->sq_psn = result.sq_psn;
 313         qp_attr->sq_draining = result.sq_draining;
 314         qp_attr->port_num = 1;
 315         qp_attr->rnr_retry = result.rnr_retry;
 316
 317         qp_attr->cap.max_send_wr = qp->max_send_wr;
 318         qp_attr->cap.max_recv_wr = qp->max_recv_wr;
 319         qp_attr->cap.max_send_sge = qp->max_send_sge;
 320         qp_attr->cap.max_recv_sge = qp->max_recv_sge;
 321         qp_attr->cap.max_inline_data = qp->max_inline_data;
 322
 323         qp_init_attr->qp_type = ibqp->qp_type;
 324         qp_init_attr->recv_cq = ibqp->recv_cq;
 325         qp_init_attr->send_cq = ibqp->send_cq;
 326         qp_init_attr->qp_context = ibqp->qp_context;
 327         qp_init_attr->cap = qp_attr->cap;
 328
 329         return 0;
 330 }
 331
 332 int efa_query_gid(struct ib_device *ibdev, u32 port, int index,
 333                   union ib_gid *gid)
 334 {
 335         struct efa_dev *dev = to_edev(ibdev);
 336
 337         memcpy(gid->raw, dev->dev_attr.addr, sizeof(dev->dev_attr.addr));
 338
 339         return 0;
 340 }
 341
 342 int efa_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
 343                    u16 *pkey)
 344 {
 345         if (index > 0)
 346                 return -EINVAL;
 347
 348         *pkey = 0xffff;
 349         return 0;
 350 }
 351
 352 static int efa_pd_dealloc(struct efa_dev *dev, u16 pdn)
 353 {
 354         struct efa_com_dealloc_pd_params params = {
 355                 .pdn = pdn,
 356         };
 357
 358         return efa_com_dealloc_pd(&dev->edev, &params);
 359 }
 360
 361 int efa_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 362 {
 363         struct efa_dev *dev = to_edev(ibpd->device);
 364         struct efa_ibv_alloc_pd_resp resp = {};
 365         struct efa_com_alloc_pd_result result;
 366         struct efa_pd *pd = to_epd(ibpd);
 367         int err;
 368
 369         if (udata->inlen &&
 370             !ib_is_udata_cleared(udata, 0, udata->inlen)) {
 371                 ibdev_dbg(&dev->ibdev,
 372                           "Incompatible ABI params, udata not cleared\n");
 373                 err = -EINVAL;
 374                 goto err_out;
 375         }
 376
 377         err = efa_com_alloc_pd(&dev->edev, &result);
 378         if (err)
 379                 goto err_out;
 380
 381         pd->pdn = result.pdn;
 382         resp.pdn = result.pdn;
 383
 384         if (udata->outlen) {
 385                 err = ib_copy_to_udata(udata, &resp,
 386                                        min(sizeof(resp), udata->outlen));
 387                 if (err) {
 388                         ibdev_dbg(&dev->ibdev,
 389                                   "Failed to copy udata for alloc_pd\n");
 390                         goto err_dealloc_pd;
 391                 }
 392         }
 393
 394         ibdev_dbg(&dev->ibdev, "Allocated pd[%d]\n", pd->pdn);
 395
 396         return 0;
 397
 398 err_dealloc_pd:
 399         efa_pd_dealloc(dev, result.pdn);
 400 err_out:
 401         atomic64_inc(&dev->stats.alloc_pd_err);
 402         return err;
 403 }
 404
 405 int efa_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 406 {
 407         struct efa_dev *dev = to_edev(ibpd->device);
 408         struct efa_pd *pd = to_epd(ibpd);
 409
 410         ibdev_dbg(&dev->ibdev, "Dealloc pd[%d]\n", pd->pdn);
 411         efa_pd_dealloc(dev, pd->pdn);
 412         return 0;
 413 }
 414
 415 static int efa_destroy_qp_handle(struct efa_dev *dev, u32 qp_handle)
 416 {
 417         struct efa_com_destroy_qp_params params = { .qp_handle = qp_handle };
 418
 419         return efa_com_destroy_qp(&dev->edev, &params);
 420 }
 421
 422 static void efa_qp_user_mmap_entries_remove(struct efa_qp *qp)
 423 {
 424         rdma_user_mmap_entry_remove(qp->rq_mmap_entry);
 425         rdma_user_mmap_entry_remove(qp->rq_db_mmap_entry);
 426         rdma_user_mmap_entry_remove(qp->llq_desc_mmap_entry);
 427         rdma_user_mmap_entry_remove(qp->sq_db_mmap_entry);
 428 }
 429
 430 int efa_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
 431 {
 432         struct efa_dev *dev = to_edev(ibqp->pd->device);
 433         struct efa_qp *qp = to_eqp(ibqp);
 434         int err;
 435
 436         ibdev_dbg(&dev->ibdev, "Destroy qp[%u]\n", ibqp->qp_num);
 437
 438         efa_qp_user_mmap_entries_remove(qp);
 439
 440         err = efa_destroy_qp_handle(dev, qp->qp_handle);
 441         if (err)
 442                 return err;
 443
 444         if (qp->rq_cpu_addr) {
 445                 ibdev_dbg(&dev->ibdev,
 446                           "qp->cpu_addr[0x%p] freed: size[%lu], dma[%pad]\n",
 447                           qp->rq_cpu_addr, qp->rq_size,
 448                           &qp->rq_dma_addr);
 449                 efa_free_mapped(dev, qp->rq_cpu_addr, qp->rq_dma_addr,
 450                                 qp->rq_size, DMA_TO_DEVICE);
 451         }
 452
 453         return 0;
 454 }
 455
 456 static struct rdma_user_mmap_entry*
 457 efa_user_mmap_entry_insert(struct ib_ucontext *ucontext,
 458                            u64 address, size_t length,
 459                            u8 mmap_flag, u64 *offset)
 460 {
 461         struct efa_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 462         int err;
 463
 464         if (!entry)
 465                 return NULL;
 466
 467         entry->address = address;
 468         entry->mmap_flag = mmap_flag;
 469
 470         err = rdma_user_mmap_entry_insert(ucontext, &entry->rdma_entry,
 471                                           length);
 472         if (err) {
 473                 kfree(entry);
 474                 return NULL;
 475         }
 476         *offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
 477
 478         return &entry->rdma_entry;
 479 }
 480
 481 static int qp_mmap_entries_setup(struct efa_qp *qp,
 482                                  struct efa_dev *dev,
 483                                  struct efa_ucontext *ucontext,
 484                                  struct efa_com_create_qp_params *params,
 485                                  struct efa_ibv_create_qp_resp *resp)
 486 {
 487         size_t length;
 488         u64 address;
 489
 490         address = dev->db_bar_addr + resp->sq_db_offset;
 491         qp->sq_db_mmap_entry =
 492                 efa_user_mmap_entry_insert(&ucontext->ibucontext,
 493                                            address,
 494                                            PAGE_SIZE, EFA_MMAP_IO_NC,
 495                                            &resp->sq_db_mmap_key);
 496         if (!qp->sq_db_mmap_entry)
 497                 return -ENOMEM;
 498
 499         resp->sq_db_offset &= ~PAGE_MASK;
 500
 501         address = dev->mem_bar_addr + resp->llq_desc_offset;
 502         length = PAGE_ALIGN(params->sq_ring_size_in_bytes +
 503                             (resp->llq_desc_offset & ~PAGE_MASK));
 504
 505         qp->llq_desc_mmap_entry =
 506                 efa_user_mmap_entry_insert(&ucontext->ibucontext,
 507                                            address, length,
 508                                            EFA_MMAP_IO_WC,
 509                                            &resp->llq_desc_mmap_key);
 510         if (!qp->llq_desc_mmap_entry)
 511                 goto err_remove_mmap;
 512
 513         resp->llq_desc_offset &= ~PAGE_MASK;
 514
 515         if (qp->rq_size) {
 516                 address = dev->db_bar_addr + resp->rq_db_offset;
 517
 518                 qp->rq_db_mmap_entry =
 519                         efa_user_mmap_entry_insert(&ucontext->ibucontext,
 520                                                    address, PAGE_SIZE,
 521                                                    EFA_MMAP_IO_NC,
 522                                                    &resp->rq_db_mmap_key);
 523                 if (!qp->rq_db_mmap_entry)
 524                         goto err_remove_mmap;
 525
 526                 resp->rq_db_offset &= ~PAGE_MASK;
 527
 528                 address = virt_to_phys(qp->rq_cpu_addr);
 529                 qp->rq_mmap_entry =
 530                         efa_user_mmap_entry_insert(&ucontext->ibucontext,
 531                                                    address, qp->rq_size,
 532                                                    EFA_MMAP_DMA_PAGE,
 533                                                    &resp->rq_mmap_key);
 534                 if (!qp->rq_mmap_entry)
 535                         goto err_remove_mmap;
 536
 537                 resp->rq_mmap_size = qp->rq_size;
 538         }
 539
 540         return 0;
 541
 542 err_remove_mmap:
 543         efa_qp_user_mmap_entries_remove(qp);
 544
 545         return -ENOMEM;
 546 }
 547
 548 static int efa_qp_validate_cap(struct efa_dev *dev,
 549                                struct ib_qp_init_attr *init_attr)
 550 {
 551         if (init_attr->cap.max_send_wr > dev->dev_attr.max_sq_depth) {
 552                 ibdev_dbg(&dev->ibdev,
 553                           "qp: requested send wr[%u] exceeds the max[%u]\n",
 554                           init_attr->cap.max_send_wr,
 555                           dev->dev_attr.max_sq_depth);
 556                 return -EINVAL;
 557         }
 558         if (init_attr->cap.max_recv_wr > dev->dev_attr.max_rq_depth) {
 559                 ibdev_dbg(&dev->ibdev,
 560                           "qp: requested receive wr[%u] exceeds the max[%u]\n",
 561                           init_attr->cap.max_recv_wr,
 562                           dev->dev_attr.max_rq_depth);
 563                 return -EINVAL;
 564         }
 565         if (init_attr->cap.max_send_sge > dev->dev_attr.max_sq_sge) {
 566                 ibdev_dbg(&dev->ibdev,
 567                           "qp: requested sge send[%u] exceeds the max[%u]\n",
 568                           init_attr->cap.max_send_sge, dev->dev_attr.max_sq_sge);
 569                 return -EINVAL;
 570         }
 571         if (init_attr->cap.max_recv_sge > dev->dev_attr.max_rq_sge) {
 572                 ibdev_dbg(&dev->ibdev,
 573                           "qp: requested sge recv[%u] exceeds the max[%u]\n",
 574                           init_attr->cap.max_recv_sge, dev->dev_attr.max_rq_sge);
 575                 return -EINVAL;
 576         }
 577         if (init_attr->cap.max_inline_data > dev->dev_attr.inline_buf_size) {
 578                 ibdev_dbg(&dev->ibdev,
 579                           "qp: requested inline data[%u] exceeds the max[%u]\n",
 580                           init_attr->cap.max_inline_data,
 581                           dev->dev_attr.inline_buf_size);
 582                 return -EINVAL;
 583         }
 584
 585         return 0;
 586 }
 587
 588 static int efa_qp_validate_attr(struct efa_dev *dev,
 589                                 struct ib_qp_init_attr *init_attr)
 590 {
 591         if (init_attr->qp_type != IB_QPT_DRIVER &&
 592             init_attr->qp_type != IB_QPT_UD) {
 593                 ibdev_dbg(&dev->ibdev,
 594                           "Unsupported qp type %d\n", init_attr->qp_type);
 595                 return -EOPNOTSUPP;
 596         }
 597
 598         if (init_attr->srq) {
 599                 ibdev_dbg(&dev->ibdev, "SRQ is not supported\n");
 600                 return -EOPNOTSUPP;
 601         }
 602
 603         if (init_attr->create_flags) {
 604                 ibdev_dbg(&dev->ibdev, "Unsupported create flags\n");
 605                 return -EOPNOTSUPP;
 606         }
 607
 608         return 0;
 609 }
 610
 611 int efa_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init_attr,
 612                   struct ib_udata *udata)
 613 {
 614         struct efa_com_create_qp_params create_qp_params = {};
 615         struct efa_com_create_qp_result create_qp_resp;
 616         struct efa_dev *dev = to_edev(ibqp->device);
 617         struct efa_ibv_create_qp_resp resp = {};
 618         struct efa_ibv_create_qp cmd = {};
 619         struct efa_qp *qp = to_eqp(ibqp);
 620         struct efa_ucontext *ucontext;
 621         int err;
 622
 623         ucontext = rdma_udata_to_drv_context(udata, struct efa_ucontext,
 624                                              ibucontext);
 625
 626         err = efa_qp_validate_cap(dev, init_attr);
 627         if (err)
 628                 goto err_out;
 629
 630         err = efa_qp_validate_attr(dev, init_attr);
 631         if (err)
 632                 goto err_out;
 633
 634         if (offsetofend(typeof(cmd), driver_qp_type) > udata->inlen) {
 635                 ibdev_dbg(&dev->ibdev,
 636                           "Incompatible ABI params, no input udata\n");
 637                 err = -EINVAL;
 638                 goto err_out;
 639         }
 640
 641         if (udata->inlen > sizeof(cmd) &&
 642             !ib_is_udata_cleared(udata, sizeof(cmd),
 643                                  udata->inlen - sizeof(cmd))) {
 644                 ibdev_dbg(&dev->ibdev,
 645                           "Incompatible ABI params, unknown fields in udata\n");
 646                 err = -EINVAL;
 647                 goto err_out;
 648         }
 649
 650         err = ib_copy_from_udata(&cmd, udata,
 651                                  min(sizeof(cmd), udata->inlen));
 652         if (err) {
 653                 ibdev_dbg(&dev->ibdev,
 654                           "Cannot copy udata for create_qp\n");
 655                 goto err_out;
 656         }
 657
 658         if (cmd.comp_mask) {
 659                 ibdev_dbg(&dev->ibdev,
 660                           "Incompatible ABI params, unknown fields in udata\n");
 661                 err = -EINVAL;
 662                 goto err_out;
 663         }
 664
 665         create_qp_params.uarn = ucontext->uarn;
 666         create_qp_params.pd = to_epd(ibqp->pd)->pdn;
 667
 668         if (init_attr->qp_type == IB_QPT_UD) {
 669                 create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_UD;
 670         } else if (cmd.driver_qp_type == EFA_QP_DRIVER_TYPE_SRD) {
 671                 create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_SRD;
 672         } else {
 673                 ibdev_dbg(&dev->ibdev,
 674                           "Unsupported qp type %d driver qp type %d\n",
 675                           init_attr->qp_type, cmd.driver_qp_type);
 676                 err = -EOPNOTSUPP;
 677                 goto err_out;
 678         }
 679
 680         ibdev_dbg(&dev->ibdev, "Create QP: qp type %d driver qp type %#x\n",
 681                   init_attr->qp_type, cmd.driver_qp_type);
 682         create_qp_params.send_cq_idx = to_ecq(init_attr->send_cq)->cq_idx;
 683         create_qp_params.recv_cq_idx = to_ecq(init_attr->recv_cq)->cq_idx;
 684         create_qp_params.sq_depth = init_attr->cap.max_send_wr;
 685         create_qp_params.sq_ring_size_in_bytes = cmd.sq_ring_size;
 686
 687         create_qp_params.rq_depth = init_attr->cap.max_recv_wr;
 688         create_qp_params.rq_ring_size_in_bytes = cmd.rq_ring_size;
 689         qp->rq_size = PAGE_ALIGN(create_qp_params.rq_ring_size_in_bytes);
 690         if (qp->rq_size) {
 691                 qp->rq_cpu_addr = efa_zalloc_mapped(dev, &qp->rq_dma_addr,
 692                                                     qp->rq_size, DMA_TO_DEVICE);
 693                 if (!qp->rq_cpu_addr) {
 694                         err = -ENOMEM;
 695                         goto err_out;
 696                 }
 697
 698                 ibdev_dbg(&dev->ibdev,
 699                           "qp->cpu_addr[0x%p] allocated: size[%lu], dma[%pad]\n",
 700                           qp->rq_cpu_addr, qp->rq_size, &qp->rq_dma_addr);
 701                 create_qp_params.rq_base_addr = qp->rq_dma_addr;
 702         }
 703
 704         err = efa_com_create_qp(&dev->edev, &create_qp_params,
 705                                 &create_qp_resp);
 706         if (err)
 707                 goto err_free_mapped;
 708
 709         resp.sq_db_offset = create_qp_resp.sq_db_offset;
 710         resp.rq_db_offset = create_qp_resp.rq_db_offset;
 711         resp.llq_desc_offset = create_qp_resp.llq_descriptors_offset;
 712         resp.send_sub_cq_idx = create_qp_resp.send_sub_cq_idx;
 713         resp.recv_sub_cq_idx = create_qp_resp.recv_sub_cq_idx;
 714
 715         err = qp_mmap_entries_setup(qp, dev, ucontext, &create_qp_params,
 716                                     &resp);
 717         if (err)
 718                 goto err_destroy_qp;
 719
 720         qp->qp_handle = create_qp_resp.qp_handle;
 721         qp->ibqp.qp_num = create_qp_resp.qp_num;
 722         qp->max_send_wr = init_attr->cap.max_send_wr;
 723         qp->max_recv_wr = init_attr->cap.max_recv_wr;
 724         qp->max_send_sge = init_attr->cap.max_send_sge;
 725         qp->max_recv_sge = init_attr->cap.max_recv_sge;
 726         qp->max_inline_data = init_attr->cap.max_inline_data;
 727
 728         if (udata->outlen) {
 729                 err = ib_copy_to_udata(udata, &resp,
 730                                        min(sizeof(resp), udata->outlen));
 731                 if (err) {
 732                         ibdev_dbg(&dev->ibdev,
 733                                   "Failed to copy udata for qp[%u]\n",
 734                                   create_qp_resp.qp_num);
 735                         goto err_remove_mmap_entries;
 736                 }
 737         }
 738
 739         ibdev_dbg(&dev->ibdev, "Created qp[%d]\n", qp->ibqp.qp_num);
 740
 741         return 0;
 742
 743 err_remove_mmap_entries:
 744         efa_qp_user_mmap_entries_remove(qp);
 745 err_destroy_qp:
 746         efa_destroy_qp_handle(dev, create_qp_resp.qp_handle);
 747 err_free_mapped:
 748         if (qp->rq_size)
 749                 efa_free_mapped(dev, qp->rq_cpu_addr, qp->rq_dma_addr,
 750                                 qp->rq_size, DMA_TO_DEVICE);
 751 err_out:
 752         atomic64_inc(&dev->stats.create_qp_err);
 753         return err;
 754 }
 755
 756 static const struct {
 757         int                     valid;
 758         enum ib_qp_attr_mask    req_param;
 759         enum ib_qp_attr_mask    opt_param;
 760 } srd_qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
 761         [IB_QPS_RESET] = {
 762                 [IB_QPS_RESET] = { .valid = 1 },
 763                 [IB_QPS_INIT]  = {
 764                         .valid = 1,
 765                         .req_param = IB_QP_PKEY_INDEX |
 766                                      IB_QP_PORT |
 767                                      IB_QP_QKEY,
 768                 },
 769         },
 770         [IB_QPS_INIT] = {
 771                 [IB_QPS_RESET] = { .valid = 1 },
 772                 [IB_QPS_ERR]   = { .valid = 1 },
 773                 [IB_QPS_INIT]  = {
 774                         .valid = 1,
 775                         .opt_param = IB_QP_PKEY_INDEX |
 776                                      IB_QP_PORT |
 777                                      IB_QP_QKEY,
 778                 },
 779                 [IB_QPS_RTR]   = {
 780                         .valid = 1,
 781                         .opt_param = IB_QP_PKEY_INDEX |
 782                                      IB_QP_QKEY,
 783                 },
 784         },
 785         [IB_QPS_RTR] = {
 786                 [IB_QPS_RESET] = { .valid = 1 },
 787                 [IB_QPS_ERR]   = { .valid = 1 },
 788                 [IB_QPS_RTS]   = {
 789                         .valid = 1,
 790                         .req_param = IB_QP_SQ_PSN,
 791                         .opt_param = IB_QP_CUR_STATE |
 792                                      IB_QP_QKEY |
 793                                      IB_QP_RNR_RETRY,
 794
 795                 }
 796         },
 797         [IB_QPS_RTS] = {
 798                 [IB_QPS_RESET] = { .valid = 1 },
 799                 [IB_QPS_ERR]   = { .valid = 1 },
 800                 [IB_QPS_RTS]   = {
 801                         .valid = 1,
 802                         .opt_param = IB_QP_CUR_STATE |
 803                                      IB_QP_QKEY,
 804                 },
 805                 [IB_QPS_SQD] = {
 806                         .valid = 1,
 807                         .opt_param = IB_QP_EN_SQD_ASYNC_NOTIFY,
 808                 },
 809         },
 810         [IB_QPS_SQD] = {
 811                 [IB_QPS_RESET] = { .valid = 1 },
 812                 [IB_QPS_ERR]   = { .valid = 1 },
 813                 [IB_QPS_RTS]   = {
 814                         .valid = 1,
 815                         .opt_param = IB_QP_CUR_STATE |
 816                                      IB_QP_QKEY,
 817                 },
 818                 [IB_QPS_SQD] = {
 819                         .valid = 1,
 820                         .opt_param = IB_QP_PKEY_INDEX |
 821                                      IB_QP_QKEY,
 822                 }
 823         },
 824         [IB_QPS_SQE] = {
 825                 [IB_QPS_RESET] = { .valid = 1 },
 826                 [IB_QPS_ERR]   = { .valid = 1 },
 827                 [IB_QPS_RTS]   = {
 828                         .valid = 1,
 829                         .opt_param = IB_QP_CUR_STATE |
 830                                      IB_QP_QKEY,
 831                 }
 832         },
 833         [IB_QPS_ERR] = {
 834                 [IB_QPS_RESET] = { .valid = 1 },
 835                 [IB_QPS_ERR]   = { .valid = 1 },
 836         }
 837 };
 838
 839 static bool efa_modify_srd_qp_is_ok(enum ib_qp_state cur_state,
 840                                     enum ib_qp_state next_state,
 841                                     enum ib_qp_attr_mask mask)
 842 {
 843         enum ib_qp_attr_mask req_param, opt_param;
 844
 845         if (mask & IB_QP_CUR_STATE  &&
 846             cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
 847             cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
 848                 return false;
 849
 850         if (!srd_qp_state_table[cur_state][next_state].valid)
 851                 return false;
 852
 853         req_param = srd_qp_state_table[cur_state][next_state].req_param;
 854         opt_param = srd_qp_state_table[cur_state][next_state].opt_param;
 855
 856         if ((mask & req_param) != req_param)
 857                 return false;
 858
 859         if (mask & ~(req_param | opt_param | IB_QP_STATE))
 860                 return false;
 861
 862         return true;
 863 }
 864
 865 static int efa_modify_qp_validate(struct efa_dev *dev, struct efa_qp *qp,
 866                                   struct ib_qp_attr *qp_attr, int qp_attr_mask,
 867                                   enum ib_qp_state cur_state,
 868                                   enum ib_qp_state new_state)
 869 {
 870         int err;
 871
 872 #define EFA_MODIFY_QP_SUPP_MASK \
 873         (IB_QP_STATE | IB_QP_CUR_STATE | IB_QP_EN_SQD_ASYNC_NOTIFY | \
 874          IB_QP_PKEY_INDEX | IB_QP_PORT | IB_QP_QKEY | IB_QP_SQ_PSN | \
 875          IB_QP_RNR_RETRY)
 876
 877         if (qp_attr_mask & ~EFA_MODIFY_QP_SUPP_MASK) {
 878                 ibdev_dbg(&dev->ibdev,
 879                           "Unsupported qp_attr_mask[%#x] supported[%#x]\n",
 880                           qp_attr_mask, EFA_MODIFY_QP_SUPP_MASK);
 881                 return -EOPNOTSUPP;
 882         }
 883
 884         if (qp->ibqp.qp_type == IB_QPT_DRIVER)
 885                 err = !efa_modify_srd_qp_is_ok(cur_state, new_state,
 886                                                qp_attr_mask);
 887         else
 888                 err = !ib_modify_qp_is_ok(cur_state, new_state, IB_QPT_UD,
 889                                           qp_attr_mask);
 890
 891         if (err) {
 892                 ibdev_dbg(&dev->ibdev, "Invalid modify QP parameters\n");
 893                 return -EINVAL;
 894         }
 895
 896         if ((qp_attr_mask & IB_QP_PORT) && qp_attr->port_num != 1) {
 897                 ibdev_dbg(&dev->ibdev, "Can't change port num\n");
 898                 return -EOPNOTSUPP;
 899         }
 900
 901         if ((qp_attr_mask & IB_QP_PKEY_INDEX) && qp_attr->pkey_index) {
 902                 ibdev_dbg(&dev->ibdev, "Can't change pkey index\n");
 903                 return -EOPNOTSUPP;
 904         }
 905
 906         return 0;
 907 }
 908
 909 int efa_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
 910                   int qp_attr_mask, struct ib_udata *udata)
 911 {
 912         struct efa_dev *dev = to_edev(ibqp->device);
 913         struct efa_com_modify_qp_params params = {};
 914         struct efa_qp *qp = to_eqp(ibqp);
 915         enum ib_qp_state cur_state;
 916         enum ib_qp_state new_state;
 917         int err;
 918
 919         if (qp_attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
 920                 return -EOPNOTSUPP;
 921
 922         if (udata->inlen &&
 923             !ib_is_udata_cleared(udata, 0, udata->inlen)) {
 924                 ibdev_dbg(&dev->ibdev,
 925                           "Incompatible ABI params, udata not cleared\n");
 926                 return -EINVAL;
 927         }
 928
 929         cur_state = qp_attr_mask & IB_QP_CUR_STATE ? qp_attr->cur_qp_state :
 930                                                      qp->state;
 931         new_state = qp_attr_mask & IB_QP_STATE ? qp_attr->qp_state : cur_state;
 932
 933         err = efa_modify_qp_validate(dev, qp, qp_attr, qp_attr_mask, cur_state,
 934                                      new_state);
 935         if (err)
 936                 return err;
 937
 938         params.qp_handle = qp->qp_handle;
 939
 940         if (qp_attr_mask & IB_QP_STATE) {
 941                 EFA_SET(&params.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_QP_STATE,
 942                         1);
 943                 EFA_SET(&params.modify_mask,
 944                         EFA_ADMIN_MODIFY_QP_CMD_CUR_QP_STATE, 1);
 945                 params.cur_qp_state = cur_state;
 946                 params.qp_state = new_state;
 947         }
 948
 949         if (qp_attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) {
 950                 EFA_SET(&params.modify_mask,
 951                         EFA_ADMIN_MODIFY_QP_CMD_SQ_DRAINED_ASYNC_NOTIFY, 1);
 952                 params.sq_drained_async_notify = qp_attr->en_sqd_async_notify;
 953         }
 954
 955         if (qp_attr_mask & IB_QP_QKEY) {
 956                 EFA_SET(&params.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_QKEY, 1);
 957                 params.qkey = qp_attr->qkey;
 958         }
 959
 960         if (qp_attr_mask & IB_QP_SQ_PSN) {
 961                 EFA_SET(&params.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_SQ_PSN, 1);
 962                 params.sq_psn = qp_attr->sq_psn;
 963         }
 964
 965         if (qp_attr_mask & IB_QP_RNR_RETRY) {
 966                 EFA_SET(&params.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_RNR_RETRY,
 967                         1);
 968                 params.rnr_retry = qp_attr->rnr_retry;
 969         }
 970
 971         err = efa_com_modify_qp(&dev->edev, &params);
 972         if (err)
 973                 return err;
 974
 975         qp->state = new_state;
 976
 977         return 0;
 978 }
 979
 980 static int efa_destroy_cq_idx(struct efa_dev *dev, int cq_idx)
 981 {
 982         struct efa_com_destroy_cq_params params = { .cq_idx = cq_idx };
 983
 984         return efa_com_destroy_cq(&dev->edev, &params);
 985 }
 986
 987 int efa_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
 988 {
 989         struct efa_dev *dev = to_edev(ibcq->device);
 990         struct efa_cq *cq = to_ecq(ibcq);
 991
 992         ibdev_dbg(&dev->ibdev,
 993                   "Destroy cq[%d] virt[0x%p] freed: size[%lu], dma[%pad]\n",
 994                   cq->cq_idx, cq->cpu_addr, cq->size, &cq->dma_addr);
 995
 996         rdma_user_mmap_entry_remove(cq->mmap_entry);
 997         efa_destroy_cq_idx(dev, cq->cq_idx);
 998         efa_free_mapped(dev, cq->cpu_addr, cq->dma_addr, cq->size,
 999                         DMA_FROM_DEVICE);
1000         return 0;
1001 }
1002
1003 static int cq_mmap_entries_setup(struct efa_dev *dev, struct efa_cq *cq,
1004                                  struct efa_ibv_create_cq_resp *resp)
1005 {
1006         resp->q_mmap_size = cq->size;
1007         cq->mmap_entry = efa_user_mmap_entry_insert(&cq->ucontext->ibucontext,
1008                                                     virt_to_phys(cq->cpu_addr),
1009                                                     cq->size, EFA_MMAP_DMA_PAGE,
1010                                                     &resp->q_mmap_key);
1011         if (!cq->mmap_entry)
1012                 return -ENOMEM;
1013
1014         return 0;
1015 }
1016
1017 int efa_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
1018                   struct ib_udata *udata)
1019 {
1020         struct efa_ucontext *ucontext = rdma_udata_to_drv_context(
1021                 udata, struct efa_ucontext, ibucontext);
1022         struct efa_ibv_create_cq_resp resp = {};
1023         struct efa_com_create_cq_params params;
1024         struct efa_com_create_cq_result result;
1025         struct ib_device *ibdev = ibcq->device;
1026         struct efa_dev *dev = to_edev(ibdev);
1027         struct efa_ibv_create_cq cmd = {};
1028         struct efa_cq *cq = to_ecq(ibcq);
1029         int entries = attr->cqe;
1030         int err;
1031
1032         ibdev_dbg(ibdev, "create_cq entries %d\n", entries);
1033
1034         if (attr->flags)
1035                 return -EOPNOTSUPP;
1036
1037         if (entries < 1 || entries > dev->dev_attr.max_cq_depth) {
1038                 ibdev_dbg(ibdev,
1039                           "cq: requested entries[%u] non-positive or greater than max[%u]\n",
1040                           entries, dev->dev_attr.max_cq_depth);
1041                 err = -EINVAL;
1042                 goto err_out;
1043         }
1044
1045         if (offsetofend(typeof(cmd), num_sub_cqs) > udata->inlen) {
1046                 ibdev_dbg(ibdev,
1047                           "Incompatible ABI params, no input udata\n");
1048                 err = -EINVAL;
1049                 goto err_out;
1050         }
1051
1052         if (udata->inlen > sizeof(cmd) &&
1053             !ib_is_udata_cleared(udata, sizeof(cmd),
1054                                  udata->inlen - sizeof(cmd))) {
1055                 ibdev_dbg(ibdev,
1056                           "Incompatible ABI params, unknown fields in udata\n");
1057                 err = -EINVAL;
1058                 goto err_out;
1059         }
1060
1061         err = ib_copy_from_udata(&cmd, udata,
1062                                  min(sizeof(cmd), udata->inlen));
1063         if (err) {
1064                 ibdev_dbg(ibdev, "Cannot copy udata for create_cq\n");
1065                 goto err_out;
1066         }
1067
1068         if (cmd.comp_mask || !is_reserved_cleared(cmd.reserved_50)) {
1069                 ibdev_dbg(ibdev,
1070                           "Incompatible ABI params, unknown fields in udata\n");
1071                 err = -EINVAL;
1072                 goto err_out;
1073         }
1074
1075         if (!cmd.cq_entry_size) {
1076                 ibdev_dbg(ibdev,
1077                           "Invalid entry size [%u]\n", cmd.cq_entry_size);
1078                 err = -EINVAL;
1079                 goto err_out;
1080         }
1081
1082         if (cmd.num_sub_cqs != dev->dev_attr.sub_cqs_per_cq) {
1083                 ibdev_dbg(ibdev,
1084                           "Invalid number of sub cqs[%u] expected[%u]\n",
1085                           cmd.num_sub_cqs, dev->dev_attr.sub_cqs_per_cq);
1086                 err = -EINVAL;
1087                 goto err_out;
1088         }
1089
1090         cq->ucontext = ucontext;
1091         cq->size = PAGE_ALIGN(cmd.cq_entry_size * entries * cmd.num_sub_cqs);
1092         cq->cpu_addr = efa_zalloc_mapped(dev, &cq->dma_addr, cq->size,
1093                                          DMA_FROM_DEVICE);
1094         if (!cq->cpu_addr) {
1095                 err = -ENOMEM;
1096                 goto err_out;
1097         }
1098
1099         params.uarn = cq->ucontext->uarn;
1100         params.cq_depth = entries;
1101         params.dma_addr = cq->dma_addr;
1102         params.entry_size_in_bytes = cmd.cq_entry_size;
1103         params.num_sub_cqs = cmd.num_sub_cqs;
1104         err = efa_com_create_cq(&dev->edev, &params, &result);
1105         if (err)
1106                 goto err_free_mapped;
1107
1108         resp.cq_idx = result.cq_idx;
1109         cq->cq_idx = result.cq_idx;
1110         cq->ibcq.cqe = result.actual_depth;
1111         WARN_ON_ONCE(entries != result.actual_depth);
1112
1113         err = cq_mmap_entries_setup(dev, cq, &resp);
1114         if (err) {
1115                 ibdev_dbg(ibdev, "Could not setup cq[%u] mmap entries\n",
1116                           cq->cq_idx);
1117                 goto err_destroy_cq;
1118         }
1119
1120         if (udata->outlen) {
1121                 err = ib_copy_to_udata(udata, &resp,
1122                                        min(sizeof(resp), udata->outlen));
1123                 if (err) {
1124                         ibdev_dbg(ibdev,
1125                                   "Failed to copy udata for create_cq\n");
1126                         goto err_remove_mmap;
1127                 }
1128         }
1129
1130         ibdev_dbg(ibdev, "Created cq[%d], cq depth[%u]. dma[%pad] virt[0x%p]\n",
1131                   cq->cq_idx, result.actual_depth, &cq->dma_addr, cq->cpu_addr);
1132
1133         return 0;
1134
1135 err_remove_mmap:
1136         rdma_user_mmap_entry_remove(cq->mmap_entry);
1137 err_destroy_cq:
1138         efa_destroy_cq_idx(dev, cq->cq_idx);
1139 err_free_mapped:
1140         efa_free_mapped(dev, cq->cpu_addr, cq->dma_addr, cq->size,
1141                         DMA_FROM_DEVICE);
1142
1143 err_out:
1144         atomic64_inc(&dev->stats.create_cq_err);
1145         return err;
1146 }
1147
1148 static int umem_to_page_list(struct efa_dev *dev,
1149                              struct ib_umem *umem,
1150                              u64 *page_list,
1151                              u32 hp_cnt,
1152                              u8 hp_shift)
1153 {
1154         u32 pages_in_hp = BIT(hp_shift - PAGE_SHIFT);
1155         struct ib_block_iter biter;
1156         unsigned int hp_idx = 0;
1157
1158         ibdev_dbg(&dev->ibdev, "hp_cnt[%u], pages_in_hp[%u]\n",
1159                   hp_cnt, pages_in_hp);
1160
1161         rdma_umem_for_each_dma_block(umem, &biter, BIT(hp_shift))
1162                 page_list[hp_idx++] = rdma_block_iter_dma_address(&biter);
1163
1164         return 0;
1165 }
1166
1167 static struct scatterlist *efa_vmalloc_buf_to_sg(u64 *buf, int page_cnt)
1168 {
1169         struct scatterlist *sglist;
1170         struct page *pg;
1171         int i;
1172
1173         sglist = kmalloc_array(page_cnt, sizeof(*sglist), GFP_KERNEL);
1174         if (!sglist)
1175                 return NULL;
1176         sg_init_table(sglist, page_cnt);
1177         for (i = 0; i < page_cnt; i++) {
1178                 pg = vmalloc_to_page(buf);
1179                 if (!pg)
1180                         goto err;
1181                 sg_set_page(&sglist[i], pg, PAGE_SIZE, 0);
1182                 buf += PAGE_SIZE / sizeof(*buf);
1183         }
1184         return sglist;
1185
1186 err:
1187         kfree(sglist);
1188         return NULL;
1189 }
1190
1191 /*
1192  * create a chunk list of physical pages dma addresses from the supplied
1193  * scatter gather list
1194  */
1195 static int pbl_chunk_list_create(struct efa_dev *dev, struct pbl_context *pbl)
1196 {
1197         struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list;
1198         int page_cnt = pbl->phys.indirect.pbl_buf_size_in_pages;
1199         struct scatterlist *pages_sgl = pbl->phys.indirect.sgl;
1200         unsigned int chunk_list_size, chunk_idx, payload_idx;
1201         int sg_dma_cnt = pbl->phys.indirect.sg_dma_cnt;
1202         struct efa_com_ctrl_buff_info *ctrl_buf;
1203         u64 *cur_chunk_buf, *prev_chunk_buf;
1204         struct ib_block_iter biter;
1205         dma_addr_t dma_addr;
1206         int i;
1207
1208         /* allocate a chunk list that consists of 4KB chunks */
1209         chunk_list_size = DIV_ROUND_UP(page_cnt, EFA_PTRS_PER_CHUNK);
1210
1211         chunk_list->size = chunk_list_size;
1212         chunk_list->chunks = kcalloc(chunk_list_size,
1213                                      sizeof(*chunk_list->chunks),
1214                                      GFP_KERNEL);
1215         if (!chunk_list->chunks)
1216                 return -ENOMEM;
1217
1218         ibdev_dbg(&dev->ibdev,
1219                   "chunk_list_size[%u] - pages[%u]\n", chunk_list_size,
1220                   page_cnt);
1221
1222         /* allocate chunk buffers: */
1223         for (i = 0; i < chunk_list_size; i++) {
1224                 chunk_list->chunks[i].buf = kzalloc(EFA_CHUNK_SIZE, GFP_KERNEL);
1225                 if (!chunk_list->chunks[i].buf)
1226                         goto chunk_list_dealloc;
1227
1228                 chunk_list->chunks[i].length = EFA_CHUNK_USED_SIZE;
1229         }
1230         chunk_list->chunks[chunk_list_size - 1].length =
1231                 ((page_cnt % EFA_PTRS_PER_CHUNK) * EFA_CHUNK_PAYLOAD_PTR_SIZE) +
1232                         EFA_CHUNK_PTR_SIZE;
1233
1234         /* fill the dma addresses of sg list pages to chunks: */
1235         chunk_idx = 0;
1236         payload_idx = 0;
1237         cur_chunk_buf = chunk_list->chunks[0].buf;
1238         rdma_for_each_block(pages_sgl, &biter, sg_dma_cnt,
1239                             EFA_CHUNK_PAYLOAD_SIZE) {
1240                 cur_chunk_buf[payload_idx++] =
1241                         rdma_block_iter_dma_address(&biter);
1242
1243                 if (payload_idx == EFA_PTRS_PER_CHUNK) {
1244                         chunk_idx++;
1245                         cur_chunk_buf = chunk_list->chunks[chunk_idx].buf;
1246                         payload_idx = 0;
1247                 }
1248         }
1249
1250         /* map chunks to dma and fill chunks next ptrs */
1251         for (i = chunk_list_size - 1; i >= 0; i--) {
1252                 dma_addr = dma_map_single(&dev->pdev->dev,
1253                                           chunk_list->chunks[i].buf,
1254                                           chunk_list->chunks[i].length,
1255                                           DMA_TO_DEVICE);
1256                 if (dma_mapping_error(&dev->pdev->dev, dma_addr)) {
1257                         ibdev_err(&dev->ibdev,
1258                                   "chunk[%u] dma_map_failed\n", i);
1259                         goto chunk_list_unmap;
1260                 }
1261
1262                 chunk_list->chunks[i].dma_addr = dma_addr;
1263                 ibdev_dbg(&dev->ibdev,
1264                           "chunk[%u] mapped at [%pad]\n", i, &dma_addr);
1265
1266                 if (!i)
1267                         break;
1268
1269                 prev_chunk_buf = chunk_list->chunks[i - 1].buf;
1270
1271                 ctrl_buf = (struct efa_com_ctrl_buff_info *)
1272                                 &prev_chunk_buf[EFA_PTRS_PER_CHUNK];
1273                 ctrl_buf->length = chunk_list->chunks[i].length;
1274
1275                 efa_com_set_dma_addr(dma_addr,
1276                                      &ctrl_buf->address.mem_addr_high,
1277                                      &ctrl_buf->address.mem_addr_low);
1278         }
1279
1280         return 0;
1281
1282 chunk_list_unmap:
1283         for (; i < chunk_list_size; i++) {
1284                 dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr,
1285                                  chunk_list->chunks[i].length, DMA_TO_DEVICE);
1286         }
1287 chunk_list_dealloc:
1288         for (i = 0; i < chunk_list_size; i++)
1289                 kfree(chunk_list->chunks[i].buf);
1290
1291         kfree(chunk_list->chunks);
1292         return -ENOMEM;
1293 }
1294
1295 static void pbl_chunk_list_destroy(struct efa_dev *dev, struct pbl_context *pbl)
1296 {
1297         struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list;
1298         int i;
1299
1300         for (i = 0; i < chunk_list->size; i++) {
1301                 dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr,
1302                                  chunk_list->chunks[i].length, DMA_TO_DEVICE);
1303                 kfree(chunk_list->chunks[i].buf);
1304         }
1305
1306         kfree(chunk_list->chunks);
1307 }
1308
1309 /* initialize pbl continuous mode: map pbl buffer to a dma address. */
1310 static int pbl_continuous_initialize(struct efa_dev *dev,
1311                                      struct pbl_context *pbl)
1312 {
1313         dma_addr_t dma_addr;
1314
1315         dma_addr = dma_map_single(&dev->pdev->dev, pbl->pbl_buf,
1316                                   pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE);
1317         if (dma_mapping_error(&dev->pdev->dev, dma_addr)) {
1318                 ibdev_err(&dev->ibdev, "Unable to map pbl to DMA address\n");
1319                 return -ENOMEM;
1320         }
1321
1322         pbl->phys.continuous.dma_addr = dma_addr;
1323         ibdev_dbg(&dev->ibdev,
1324                   "pbl continuous - dma_addr = %pad, size[%u]\n",
1325                   &dma_addr, pbl->pbl_buf_size_in_bytes);
1326
1327         return 0;
1328 }
1329
1330 /*
1331  * initialize pbl indirect mode:
1332  * create a chunk list out of the dma addresses of the physical pages of
1333  * pbl buffer.
1334  */
1335 static int pbl_indirect_initialize(struct efa_dev *dev, struct pbl_context *pbl)
1336 {
1337         u32 size_in_pages = DIV_ROUND_UP(pbl->pbl_buf_size_in_bytes, PAGE_SIZE);
1338         struct scatterlist *sgl;
1339         int sg_dma_cnt, err;
1340
1341         BUILD_BUG_ON(EFA_CHUNK_PAYLOAD_SIZE > PAGE_SIZE);
1342         sgl = efa_vmalloc_buf_to_sg(pbl->pbl_buf, size_in_pages);
1343         if (!sgl)
1344                 return -ENOMEM;
1345
1346         sg_dma_cnt = dma_map_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE);
1347         if (!sg_dma_cnt) {
1348                 err = -EINVAL;
1349                 goto err_map;
1350         }
1351
1352         pbl->phys.indirect.pbl_buf_size_in_pages = size_in_pages;
1353         pbl->phys.indirect.sgl = sgl;
1354         pbl->phys.indirect.sg_dma_cnt = sg_dma_cnt;
1355         err = pbl_chunk_list_create(dev, pbl);
1356         if (err) {
1357                 ibdev_dbg(&dev->ibdev,
1358                           "chunk_list creation failed[%d]\n", err);
1359                 goto err_chunk;
1360         }
1361
1362         ibdev_dbg(&dev->ibdev,
1363                   "pbl indirect - size[%u], chunks[%u]\n",
1364                   pbl->pbl_buf_size_in_bytes,
1365                   pbl->phys.indirect.chunk_list.size);
1366
1367         return 0;
1368
1369 err_chunk:
1370         dma_unmap_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE);
1371 err_map:
1372         kfree(sgl);
1373         return err;
1374 }
1375
1376 static void pbl_indirect_terminate(struct efa_dev *dev, struct pbl_context *pbl)
1377 {
1378         pbl_chunk_list_destroy(dev, pbl);
1379         dma_unmap_sg(&dev->pdev->dev, pbl->phys.indirect.sgl,
1380                      pbl->phys.indirect.pbl_buf_size_in_pages, DMA_TO_DEVICE);
1381         kfree(pbl->phys.indirect.sgl);
1382 }
1383
1384 /* create a page buffer list from a mapped user memory region */
1385 static int pbl_create(struct efa_dev *dev,
1386                       struct pbl_context *pbl,
1387                       struct ib_umem *umem,
1388                       int hp_cnt,
1389                       u8 hp_shift)
1390 {
1391         int err;
1392
1393         pbl->pbl_buf_size_in_bytes = hp_cnt * EFA_CHUNK_PAYLOAD_PTR_SIZE;
1394         pbl->pbl_buf = kvzalloc(pbl->pbl_buf_size_in_bytes, GFP_KERNEL);
1395         if (!pbl->pbl_buf)
1396                 return -ENOMEM;
1397
1398         if (is_vmalloc_addr(pbl->pbl_buf)) {
1399                 pbl->physically_continuous = 0;
1400                 err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt,
1401                                         hp_shift);
1402                 if (err)
1403                         goto err_free;
1404
1405                 err = pbl_indirect_initialize(dev, pbl);
1406                 if (err)
1407                         goto err_free;
1408         } else {
1409                 pbl->physically_continuous = 1;
1410                 err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt,
1411                                         hp_shift);
1412                 if (err)
1413                         goto err_free;
1414
1415                 err = pbl_continuous_initialize(dev, pbl);
1416                 if (err)
1417                         goto err_free;
1418         }
1419
1420         ibdev_dbg(&dev->ibdev,
1421                   "user_pbl_created: user_pages[%u], continuous[%u]\n",
1422                   hp_cnt, pbl->physically_continuous);
1423
1424         return 0;
1425
1426 err_free:
1427         kvfree(pbl->pbl_buf);
1428         return err;
1429 }
1430
1431 static void pbl_destroy(struct efa_dev *dev, struct pbl_context *pbl)
1432 {
1433         if (pbl->physically_continuous)
1434                 dma_unmap_single(&dev->pdev->dev, pbl->phys.continuous.dma_addr,
1435                                  pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE);
1436         else
1437                 pbl_indirect_terminate(dev, pbl);
1438
1439         kvfree(pbl->pbl_buf);
1440 }
1441
1442 static int efa_create_inline_pbl(struct efa_dev *dev, struct efa_mr *mr,
1443                                  struct efa_com_reg_mr_params *params)
1444 {
1445         int err;
1446
1447         params->inline_pbl = 1;
1448         err = umem_to_page_list(dev, mr->umem, params->pbl.inline_pbl_array,
1449                                 params->page_num, params->page_shift);
1450         if (err)
1451                 return err;
1452
1453         ibdev_dbg(&dev->ibdev,
1454                   "inline_pbl_array - pages[%u]\n", params->page_num);
1455
1456         return 0;
1457 }
1458
1459 static int efa_create_pbl(struct efa_dev *dev,
1460                           struct pbl_context *pbl,
1461                           struct efa_mr *mr,
1462                           struct efa_com_reg_mr_params *params)
1463 {
1464         int err;
1465
1466         err = pbl_create(dev, pbl, mr->umem, params->page_num,
1467                          params->page_shift);
1468         if (err) {
1469                 ibdev_dbg(&dev->ibdev, "Failed to create pbl[%d]\n", err);
1470                 return err;
1471         }
1472
1473         params->inline_pbl = 0;
1474         params->indirect = !pbl->physically_continuous;
1475         if (pbl->physically_continuous) {
1476                 params->pbl.pbl.length = pbl->pbl_buf_size_in_bytes;
1477
1478                 efa_com_set_dma_addr(pbl->phys.continuous.dma_addr,
1479                                      &params->pbl.pbl.address.mem_addr_high,
1480                                      &params->pbl.pbl.address.mem_addr_low);
1481         } else {
1482                 params->pbl.pbl.length =
1483                         pbl->phys.indirect.chunk_list.chunks[0].length;
1484
1485                 efa_com_set_dma_addr(pbl->phys.indirect.chunk_list.chunks[0].dma_addr,
1486                                      &params->pbl.pbl.address.mem_addr_high,
1487                                      &params->pbl.pbl.address.mem_addr_low);
1488         }
1489
1490         return 0;
1491 }
1492
1493 struct ib_mr *efa_reg_mr(struct ib_pd *ibpd, u64 start, u64 length,
1494                          u64 virt_addr, int access_flags,
1495                          struct ib_udata *udata)
1496 {
1497         struct efa_dev *dev = to_edev(ibpd->device);
1498         struct efa_com_reg_mr_params params = {};
1499         struct efa_com_reg_mr_result result = {};
1500         struct pbl_context pbl;
1501         int supp_access_flags;
1502         unsigned int pg_sz;
1503         struct efa_mr *mr;
1504         int inline_size;
1505         int err;
1506
1507         if (udata && udata->inlen &&
1508             !ib_is_udata_cleared(udata, 0, sizeof(udata->inlen))) {
1509                 ibdev_dbg(&dev->ibdev,
1510                           "Incompatible ABI params, udata not cleared\n");
1511                 err = -EINVAL;
1512                 goto err_out;
1513         }
1514
1515         supp_access_flags =
1516                 IB_ACCESS_LOCAL_WRITE |
1517                 (EFA_DEV_CAP(dev, RDMA_READ) ? IB_ACCESS_REMOTE_READ : 0);
1518
1519         access_flags &= ~IB_ACCESS_OPTIONAL;
1520         if (access_flags & ~supp_access_flags) {
1521                 ibdev_dbg(&dev->ibdev,
1522                           "Unsupported access flags[%#x], supported[%#x]\n",
1523                           access_flags, supp_access_flags);
1524                 err = -EOPNOTSUPP;
1525                 goto err_out;
1526         }
1527
1528         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1529         if (!mr) {
1530                 err = -ENOMEM;
1531                 goto err_out;
1532         }
1533
1534         mr->umem = ib_umem_get(ibpd->device, start, length, access_flags);
1535         if (IS_ERR(mr->umem)) {
1536                 err = PTR_ERR(mr->umem);
1537                 ibdev_dbg(&dev->ibdev,
1538                           "Failed to pin and map user space memory[%d]\n", err);
1539                 goto err_free;
1540         }
1541
1542         params.pd = to_epd(ibpd)->pdn;
1543         params.iova = virt_addr;
1544         params.mr_length_in_bytes = length;
1545         params.permissions = access_flags;
1546
1547         pg_sz = ib_umem_find_best_pgsz(mr->umem,
1548                                        dev->dev_attr.page_size_cap,
1549                                        virt_addr);
1550         if (!pg_sz) {
1551                 err = -EOPNOTSUPP;
1552                 ibdev_dbg(&dev->ibdev, "Failed to find a suitable page size in page_size_cap %#llx\n",
1553                           dev->dev_attr.page_size_cap);
1554                 goto err_unmap;
1555         }
1556
1557         params.page_shift = order_base_2(pg_sz);
1558         params.page_num = ib_umem_num_dma_blocks(mr->umem, pg_sz);
1559
1560         ibdev_dbg(&dev->ibdev,
1561                   "start %#llx length %#llx params.page_shift %u params.page_num %u\n",
1562                   start, length, params.page_shift, params.page_num);
1563
1564         inline_size = ARRAY_SIZE(params.pbl.inline_pbl_array);
1565         if (params.page_num <= inline_size) {
1566                 err = efa_create_inline_pbl(dev, mr, &params);
1567                 if (err)
1568                         goto err_unmap;
1569
1570                 err = efa_com_register_mr(&dev->edev, &params, &result);
1571                 if (err)
1572                         goto err_unmap;
1573         } else {
1574                 err = efa_create_pbl(dev, &pbl, mr, &params);
1575                 if (err)
1576                         goto err_unmap;
1577
1578                 err = efa_com_register_mr(&dev->edev, &params, &result);
1579                 pbl_destroy(dev, &pbl);
1580
1581                 if (err)
1582                         goto err_unmap;
1583         }
1584
1585         mr->ibmr.lkey = result.l_key;
1586         mr->ibmr.rkey = result.r_key;
1587         mr->ibmr.length = length;
1588         ibdev_dbg(&dev->ibdev, "Registered mr[%d]\n", mr->ibmr.lkey);
1589
1590         return &mr->ibmr;
1591
1592 err_unmap:
1593         ib_umem_release(mr->umem);
1594 err_free:
1595         kfree(mr);
1596 err_out:
1597         atomic64_inc(&dev->stats.reg_mr_err);
1598         return ERR_PTR(err);
1599 }
1600
1601 int efa_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
1602 {
1603         struct efa_dev *dev = to_edev(ibmr->device);
1604         struct efa_com_dereg_mr_params params;
1605         struct efa_mr *mr = to_emr(ibmr);
1606         int err;
1607
1608         ibdev_dbg(&dev->ibdev, "Deregister mr[%d]\n", ibmr->lkey);
1609
1610         params.l_key = mr->ibmr.lkey;
1611         err = efa_com_dereg_mr(&dev->edev, &params);
1612         if (err)
1613                 return err;
1614
1615         ib_umem_release(mr->umem);
1616         kfree(mr);
1617
1618         return 0;
1619 }
1620
1621 int efa_get_port_immutable(struct ib_device *ibdev, u32 port_num,
1622                            struct ib_port_immutable *immutable)
1623 {
1624         struct ib_port_attr attr;
1625         int err;
1626
1627         err = ib_query_port(ibdev, port_num, &attr);
1628         if (err) {
1629                 ibdev_dbg(ibdev, "Couldn't query port err[%d]\n", err);
1630                 return err;
1631         }
1632
1633         immutable->pkey_tbl_len = attr.pkey_tbl_len;
1634         immutable->gid_tbl_len = attr.gid_tbl_len;
1635
1636         return 0;
1637 }
1638
1639 static int efa_dealloc_uar(struct efa_dev *dev, u16 uarn)
1640 {
1641         struct efa_com_dealloc_uar_params params = {
1642                 .uarn = uarn,
1643         };
1644
1645         return efa_com_dealloc_uar(&dev->edev, &params);
1646 }
1647
1648 #define EFA_CHECK_USER_COMP(_dev, _comp_mask, _attr, _mask, _attr_str) \
1649         (_attr_str = (!(_dev)->dev_attr._attr || ((_comp_mask) & (_mask))) ? \
1650                      NULL : #_attr)
1651
1652 static int efa_user_comp_handshake(const struct ib_ucontext *ibucontext,
1653                                    const struct efa_ibv_alloc_ucontext_cmd *cmd)
1654 {
1655         struct efa_dev *dev = to_edev(ibucontext->device);
1656         char *attr_str;
1657
1658         if (EFA_CHECK_USER_COMP(dev, cmd->comp_mask, max_tx_batch,
1659                                 EFA_ALLOC_UCONTEXT_CMD_COMP_TX_BATCH, attr_str))
1660                 goto err;
1661
1662         if (EFA_CHECK_USER_COMP(dev, cmd->comp_mask, min_sq_depth,
1663                                 EFA_ALLOC_UCONTEXT_CMD_COMP_MIN_SQ_WR,
1664                                 attr_str))
1665                 goto err;
1666
1667         return 0;
1668
1669 err:
1670         ibdev_dbg(&dev->ibdev, "Userspace handshake failed for %s attribute\n",
1671                   attr_str);
1672         return -EOPNOTSUPP;
1673 }
1674
1675 int efa_alloc_ucontext(struct ib_ucontext *ibucontext, struct ib_udata *udata)
1676 {
1677         struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1678         struct efa_dev *dev = to_edev(ibucontext->device);
1679         struct efa_ibv_alloc_ucontext_resp resp = {};
1680         struct efa_ibv_alloc_ucontext_cmd cmd = {};
1681         struct efa_com_alloc_uar_result result;
1682         int err;
1683
1684         /*
1685          * it's fine if the driver does not know all request fields,
1686          * we will ack input fields in our response.
1687          */
1688
1689         err = ib_copy_from_udata(&cmd, udata,
1690                                  min(sizeof(cmd), udata->inlen));
1691         if (err) {
1692                 ibdev_dbg(&dev->ibdev,
1693                           "Cannot copy udata for alloc_ucontext\n");
1694                 goto err_out;
1695         }
1696
1697         err = efa_user_comp_handshake(ibucontext, &cmd);
1698         if (err)
1699                 goto err_out;
1700
1701         err = efa_com_alloc_uar(&dev->edev, &result);
1702         if (err)
1703                 goto err_out;
1704
1705         ucontext->uarn = result.uarn;
1706
1707         resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_QUERY_DEVICE;
1708         resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_CREATE_AH;
1709         resp.sub_cqs_per_cq = dev->dev_attr.sub_cqs_per_cq;
1710         resp.inline_buf_size = dev->dev_attr.inline_buf_size;
1711         resp.max_llq_size = dev->dev_attr.max_llq_size;
1712         resp.max_tx_batch = dev->dev_attr.max_tx_batch;
1713         resp.min_sq_wr = dev->dev_attr.min_sq_depth;
1714
1715         err = ib_copy_to_udata(udata, &resp,
1716                                min(sizeof(resp), udata->outlen));
1717         if (err)
1718                 goto err_dealloc_uar;
1719
1720         return 0;
1721
1722 err_dealloc_uar:
1723         efa_dealloc_uar(dev, result.uarn);
1724 err_out:
1725         atomic64_inc(&dev->stats.alloc_ucontext_err);
1726         return err;
1727 }
1728
1729 void efa_dealloc_ucontext(struct ib_ucontext *ibucontext)
1730 {
1731         struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1732         struct efa_dev *dev = to_edev(ibucontext->device);
1733
1734         efa_dealloc_uar(dev, ucontext->uarn);
1735 }
1736
1737 void efa_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
1738 {
1739         struct efa_user_mmap_entry *entry = to_emmap(rdma_entry);
1740
1741         kfree(entry);
1742 }
1743
1744 static int __efa_mmap(struct efa_dev *dev, struct efa_ucontext *ucontext,
1745                       struct vm_area_struct *vma)
1746 {
1747         struct rdma_user_mmap_entry *rdma_entry;
1748         struct efa_user_mmap_entry *entry;
1749         unsigned long va;
1750         int err = 0;
1751         u64 pfn;
1752
1753         rdma_entry = rdma_user_mmap_entry_get(&ucontext->ibucontext, vma);
1754         if (!rdma_entry) {
1755                 ibdev_dbg(&dev->ibdev,
1756                           "pgoff[%#lx] does not have valid entry\n",
1757                           vma->vm_pgoff);
1758                 atomic64_inc(&dev->stats.mmap_err);
1759                 return -EINVAL;
1760         }
1761         entry = to_emmap(rdma_entry);
1762
1763         ibdev_dbg(&dev->ibdev,
1764                   "Mapping address[%#llx], length[%#zx], mmap_flag[%d]\n",
1765                   entry->address, rdma_entry->npages * PAGE_SIZE,
1766                   entry->mmap_flag);
1767
1768         pfn = entry->address >> PAGE_SHIFT;
1769         switch (entry->mmap_flag) {
1770         case EFA_MMAP_IO_NC:
1771                 err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn,
1772                                         entry->rdma_entry.npages * PAGE_SIZE,
1773                                         pgprot_noncached(vma->vm_page_prot),
1774                                         rdma_entry);
1775                 break;
1776         case EFA_MMAP_IO_WC:
1777                 err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn,
1778                                         entry->rdma_entry.npages * PAGE_SIZE,
1779                                         pgprot_writecombine(vma->vm_page_prot),
1780                                         rdma_entry);
1781                 break;
1782         case EFA_MMAP_DMA_PAGE:
1783                 for (va = vma->vm_start; va < vma->vm_end;
1784                      va += PAGE_SIZE, pfn++) {
1785                         err = vm_insert_page(vma, va, pfn_to_page(pfn));
1786                         if (err)
1787                                 break;
1788                 }
1789                 break;
1790         default:
1791                 err = -EINVAL;
1792         }
1793
1794         if (err) {
1795                 ibdev_dbg(
1796                         &dev->ibdev,
1797                         "Couldn't mmap address[%#llx] length[%#zx] mmap_flag[%d] err[%d]\n",
1798                         entry->address, rdma_entry->npages * PAGE_SIZE,
1799                         entry->mmap_flag, err);
1800                 atomic64_inc(&dev->stats.mmap_err);
1801         }
1802
1803         rdma_user_mmap_entry_put(rdma_entry);
1804         return err;
1805 }
1806
1807 int efa_mmap(struct ib_ucontext *ibucontext,
1808              struct vm_area_struct *vma)
1809 {
1810         struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1811         struct efa_dev *dev = to_edev(ibucontext->device);
1812         size_t length = vma->vm_end - vma->vm_start;
1813
1814         ibdev_dbg(&dev->ibdev,
1815                   "start %#lx, end %#lx, length = %#zx, pgoff = %#lx\n",
1816                   vma->vm_start, vma->vm_end, length, vma->vm_pgoff);
1817
1818         return __efa_mmap(dev, ucontext, vma);
1819 }
1820
1821 static int efa_ah_destroy(struct efa_dev *dev, struct efa_ah *ah)
1822 {
1823         struct efa_com_destroy_ah_params params = {
1824                 .ah = ah->ah,
1825                 .pdn = to_epd(ah->ibah.pd)->pdn,
1826         };
1827
1828         return efa_com_destroy_ah(&dev->edev, &params);
1829 }
1830
1831 int efa_create_ah(struct ib_ah *ibah,
1832                   struct rdma_ah_init_attr *init_attr,
1833                   struct ib_udata *udata)
1834 {
1835         struct rdma_ah_attr *ah_attr = init_attr->ah_attr;
1836         struct efa_dev *dev = to_edev(ibah->device);
1837         struct efa_com_create_ah_params params = {};
1838         struct efa_ibv_create_ah_resp resp = {};
1839         struct efa_com_create_ah_result result;
1840         struct efa_ah *ah = to_eah(ibah);
1841         int err;
1842
1843         if (!(init_attr->flags & RDMA_CREATE_AH_SLEEPABLE)) {
1844                 ibdev_dbg(&dev->ibdev,
1845                           "Create address handle is not supported in atomic context\n");
1846                 err = -EOPNOTSUPP;
1847                 goto err_out;
1848         }
1849
1850         if (udata->inlen &&
1851             !ib_is_udata_cleared(udata, 0, udata->inlen)) {
1852                 ibdev_dbg(&dev->ibdev, "Incompatible ABI params\n");
1853                 err = -EINVAL;
1854                 goto err_out;
1855         }
1856
1857         memcpy(params.dest_addr, ah_attr->grh.dgid.raw,
1858                sizeof(params.dest_addr));
1859         params.pdn = to_epd(ibah->pd)->pdn;
1860         err = efa_com_create_ah(&dev->edev, &params, &result);
1861         if (err)
1862                 goto err_out;
1863
1864         memcpy(ah->id, ah_attr->grh.dgid.raw, sizeof(ah->id));
1865         ah->ah = result.ah;
1866
1867         resp.efa_address_handle = result.ah;
1868
1869         if (udata->outlen) {
1870                 err = ib_copy_to_udata(udata, &resp,
1871                                        min(sizeof(resp), udata->outlen));
1872                 if (err) {
1873                         ibdev_dbg(&dev->ibdev,
1874                                   "Failed to copy udata for create_ah response\n");
1875                         goto err_destroy_ah;
1876                 }
1877         }
1878         ibdev_dbg(&dev->ibdev, "Created ah[%d]\n", ah->ah);
1879
1880         return 0;
1881
1882 err_destroy_ah:
1883         efa_ah_destroy(dev, ah);
1884 err_out:
1885         atomic64_inc(&dev->stats.create_ah_err);
1886         return err;
1887 }
1888
1889 int efa_destroy_ah(struct ib_ah *ibah, u32 flags)
1890 {
1891         struct efa_dev *dev = to_edev(ibah->pd->device);
1892         struct efa_ah *ah = to_eah(ibah);
1893
1894         ibdev_dbg(&dev->ibdev, "Destroy ah[%d]\n", ah->ah);
1895
1896         if (!(flags & RDMA_DESTROY_AH_SLEEPABLE)) {
1897                 ibdev_dbg(&dev->ibdev,
1898                           "Destroy address handle is not supported in atomic context\n");
1899                 return -EOPNOTSUPP;
1900         }
1901
1902         efa_ah_destroy(dev, ah);
1903         return 0;
1904 }
1905
1906 struct rdma_hw_stats *efa_alloc_hw_port_stats(struct ib_device *ibdev,
1907                                               u32 port_num)
1908 {
1909         return rdma_alloc_hw_stats_struct(efa_port_stats_names,
1910                                           ARRAY_SIZE(efa_port_stats_names),
1911                                           RDMA_HW_STATS_DEFAULT_LIFESPAN);
1912 }
1913
1914 struct rdma_hw_stats *efa_alloc_hw_device_stats(struct ib_device *ibdev)
1915 {
1916         return rdma_alloc_hw_stats_struct(efa_device_stats_names,
1917                                           ARRAY_SIZE(efa_device_stats_names),
1918                                           RDMA_HW_STATS_DEFAULT_LIFESPAN);
1919 }
1920
1921 static int efa_fill_device_stats(struct efa_dev *dev,
1922                                  struct rdma_hw_stats *stats)
1923 {
1924         struct efa_com_stats_admin *as = &dev->edev.aq.stats;
1925         struct efa_stats *s = &dev->stats;
1926
1927         stats->value[EFA_SUBMITTED_CMDS] = atomic64_read(&as->submitted_cmd);
1928         stats->value[EFA_COMPLETED_CMDS] = atomic64_read(&as->completed_cmd);
1929         stats->value[EFA_CMDS_ERR] = atomic64_read(&as->cmd_err);
1930         stats->value[EFA_NO_COMPLETION_CMDS] = atomic64_read(&as->no_completion);
1931
1932         stats->value[EFA_KEEP_ALIVE_RCVD] = atomic64_read(&s->keep_alive_rcvd);
1933         stats->value[EFA_ALLOC_PD_ERR] = atomic64_read(&s->alloc_pd_err);
1934         stats->value[EFA_CREATE_QP_ERR] = atomic64_read(&s->create_qp_err);
1935         stats->value[EFA_CREATE_CQ_ERR] = atomic64_read(&s->create_cq_err);
1936         stats->value[EFA_REG_MR_ERR] = atomic64_read(&s->reg_mr_err);
1937         stats->value[EFA_ALLOC_UCONTEXT_ERR] =
1938                 atomic64_read(&s->alloc_ucontext_err);
1939         stats->value[EFA_CREATE_AH_ERR] = atomic64_read(&s->create_ah_err);
1940         stats->value[EFA_MMAP_ERR] = atomic64_read(&s->mmap_err);
1941
1942         return ARRAY_SIZE(efa_device_stats_names);
1943 }
1944
1945 static int efa_fill_port_stats(struct efa_dev *dev, struct rdma_hw_stats *stats,
1946                                u32 port_num)
1947 {
1948         struct efa_com_get_stats_params params = {};
1949         union efa_com_get_stats_result result;
1950         struct efa_com_rdma_read_stats *rrs;
1951         struct efa_com_messages_stats *ms;
1952         struct efa_com_basic_stats *bs;
1953         int err;
1954
1955         params.scope = EFA_ADMIN_GET_STATS_SCOPE_ALL;
1956         params.type = EFA_ADMIN_GET_STATS_TYPE_BASIC;
1957
1958         err = efa_com_get_stats(&dev->edev, &params, &result);
1959         if (err)
1960                 return err;
1961
1962         bs = &result.basic_stats;
1963         stats->value[EFA_TX_BYTES] = bs->tx_bytes;
1964         stats->value[EFA_TX_PKTS] = bs->tx_pkts;
1965         stats->value[EFA_RX_BYTES] = bs->rx_bytes;
1966         stats->value[EFA_RX_PKTS] = bs->rx_pkts;
1967         stats->value[EFA_RX_DROPS] = bs->rx_drops;
1968
1969         params.type = EFA_ADMIN_GET_STATS_TYPE_MESSAGES;
1970         err = efa_com_get_stats(&dev->edev, &params, &result);
1971         if (err)
1972                 return err;
1973
1974         ms = &result.messages_stats;
1975         stats->value[EFA_SEND_BYTES] = ms->send_bytes;
1976         stats->value[EFA_SEND_WRS] = ms->send_wrs;
1977         stats->value[EFA_RECV_BYTES] = ms->recv_bytes;
1978         stats->value[EFA_RECV_WRS] = ms->recv_wrs;
1979
1980         params.type = EFA_ADMIN_GET_STATS_TYPE_RDMA_READ;
1981         err = efa_com_get_stats(&dev->edev, &params, &result);
1982         if (err)
1983                 return err;
1984
1985         rrs = &result.rdma_read_stats;
1986         stats->value[EFA_RDMA_READ_WRS] = rrs->read_wrs;
1987         stats->value[EFA_RDMA_READ_BYTES] = rrs->read_bytes;
1988         stats->value[EFA_RDMA_READ_WR_ERR] = rrs->read_wr_err;
1989         stats->value[EFA_RDMA_READ_RESP_BYTES] = rrs->read_resp_bytes;
1990
1991         return ARRAY_SIZE(efa_port_stats_names);
1992 }
1993
1994 int efa_get_hw_stats(struct ib_device *ibdev, struct rdma_hw_stats *stats,
1995                      u32 port_num, int index)
1996 {
1997         if (port_num)
1998                 return efa_fill_port_stats(to_edev(ibdev), stats, port_num);
1999         else
2000                 return efa_fill_device_stats(to_edev(ibdev), stats);
2001 }
2002
2003 enum rdma_link_layer efa_port_link_layer(struct ib_device *ibdev,
2004                                          u32 port_num)
2005 {
2006         return IB_LINK_LAYER_UNSPECIFIED;
2007 }
2008