drivers/infiniband/hw/efa/efa_verbs.c

   1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
   2 /*
   3  * Copyright 2018-2023 Amazon.com, Inc. or its affiliates. All rights reserved.
   4  */
   5
   6 #include <linux/dma-buf.h>
   7 #include <linux/dma-resv.h>
   8 #include <linux/vmalloc.h>
   9 #include <linux/log2.h>
  10
  11 #include <rdma/ib_addr.h>
  12 #include <rdma/ib_umem.h>
  13 #include <rdma/ib_user_verbs.h>
  14 #include <rdma/ib_verbs.h>
  15 #include <rdma/uverbs_ioctl.h>
  16
  17 #include "efa.h"
  18 #include "efa_io_defs.h"
  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_user_mmap_entry {
  31         struct rdma_user_mmap_entry rdma_entry;
  32         u64 address;
  33         u8 mmap_flag;
  34 };
  35
  36 #define EFA_DEFINE_DEVICE_STATS(op) \
  37         op(EFA_SUBMITTED_CMDS, "submitted_cmds") \
  38         op(EFA_COMPLETED_CMDS, "completed_cmds") \
  39         op(EFA_CMDS_ERR, "cmds_err") \
  40         op(EFA_NO_COMPLETION_CMDS, "no_completion_cmds") \
  41         op(EFA_KEEP_ALIVE_RCVD, "keep_alive_rcvd") \
  42         op(EFA_ALLOC_PD_ERR, "alloc_pd_err") \
  43         op(EFA_CREATE_QP_ERR, "create_qp_err") \
  44         op(EFA_CREATE_CQ_ERR, "create_cq_err") \
  45         op(EFA_REG_MR_ERR, "reg_mr_err") \
  46         op(EFA_ALLOC_UCONTEXT_ERR, "alloc_ucontext_err") \
  47         op(EFA_CREATE_AH_ERR, "create_ah_err") \
  48         op(EFA_MMAP_ERR, "mmap_err")
  49
  50 #define EFA_DEFINE_PORT_STATS(op) \
  51         op(EFA_TX_BYTES, "tx_bytes") \
  52         op(EFA_TX_PKTS, "tx_pkts") \
  53         op(EFA_RX_BYTES, "rx_bytes") \
  54         op(EFA_RX_PKTS, "rx_pkts") \
  55         op(EFA_RX_DROPS, "rx_drops") \
  56         op(EFA_SEND_BYTES, "send_bytes") \
  57         op(EFA_SEND_WRS, "send_wrs") \
  58         op(EFA_RECV_BYTES, "recv_bytes") \
  59         op(EFA_RECV_WRS, "recv_wrs") \
  60         op(EFA_RDMA_READ_WRS, "rdma_read_wrs") \
  61         op(EFA_RDMA_READ_BYTES, "rdma_read_bytes") \
  62         op(EFA_RDMA_READ_WR_ERR, "rdma_read_wr_err") \
  63         op(EFA_RDMA_READ_RESP_BYTES, "rdma_read_resp_bytes") \
  64
  65 #define EFA_STATS_ENUM(ename, name) ename,
  66 #define EFA_STATS_STR(ename, nam) \
  67         [ename].name = nam,
  68
  69 enum efa_hw_device_stats {
  70         EFA_DEFINE_DEVICE_STATS(EFA_STATS_ENUM)
  71 };
  72
  73 static const struct rdma_stat_desc efa_device_stats_descs[] = {
  74         EFA_DEFINE_DEVICE_STATS(EFA_STATS_STR)
  75 };
  76
  77 enum efa_hw_port_stats {
  78         EFA_DEFINE_PORT_STATS(EFA_STATS_ENUM)
  79 };
  80
  81 static const struct rdma_stat_desc efa_port_stats_descs[] = {
  82         EFA_DEFINE_PORT_STATS(EFA_STATS_STR)
  83 };
  84
  85 #define EFA_CHUNK_PAYLOAD_SHIFT       12
  86 #define EFA_CHUNK_PAYLOAD_SIZE        BIT(EFA_CHUNK_PAYLOAD_SHIFT)
  87 #define EFA_CHUNK_PAYLOAD_PTR_SIZE    8
  88
  89 #define EFA_CHUNK_SHIFT               12
  90 #define EFA_CHUNK_SIZE                BIT(EFA_CHUNK_SHIFT)
  91 #define EFA_CHUNK_PTR_SIZE            sizeof(struct efa_com_ctrl_buff_info)
  92
  93 #define EFA_PTRS_PER_CHUNK \
  94         ((EFA_CHUNK_SIZE - EFA_CHUNK_PTR_SIZE) / EFA_CHUNK_PAYLOAD_PTR_SIZE)
  95
  96 #define EFA_CHUNK_USED_SIZE \
  97         ((EFA_PTRS_PER_CHUNK * EFA_CHUNK_PAYLOAD_PTR_SIZE) + EFA_CHUNK_PTR_SIZE)
  98
  99 struct pbl_chunk {
 100         dma_addr_t dma_addr;
 101         u64 *buf;
 102         u32 length;
 103 };
 104
 105 struct pbl_chunk_list {
 106         struct pbl_chunk *chunks;
 107         unsigned int size;
 108 };
 109
 110 struct pbl_context {
 111         union {
 112                 struct {
 113                         dma_addr_t dma_addr;
 114                 } continuous;
 115                 struct {
 116                         u32 pbl_buf_size_in_pages;
 117                         struct scatterlist *sgl;
 118                         int sg_dma_cnt;
 119                         struct pbl_chunk_list chunk_list;
 120                 } indirect;
 121         } phys;
 122         u64 *pbl_buf;
 123         u32 pbl_buf_size_in_bytes;
 124         u8 physically_continuous;
 125 };
 126
 127 static inline struct efa_dev *to_edev(struct ib_device *ibdev)
 128 {
 129         return container_of(ibdev, struct efa_dev, ibdev);
 130 }
 131
 132 static inline struct efa_ucontext *to_eucontext(struct ib_ucontext *ibucontext)
 133 {
 134         return container_of(ibucontext, struct efa_ucontext, ibucontext);
 135 }
 136
 137 static inline struct efa_pd *to_epd(struct ib_pd *ibpd)
 138 {
 139         return container_of(ibpd, struct efa_pd, ibpd);
 140 }
 141
 142 static inline struct efa_mr *to_emr(struct ib_mr *ibmr)
 143 {
 144         return container_of(ibmr, struct efa_mr, ibmr);
 145 }
 146
 147 static inline struct efa_qp *to_eqp(struct ib_qp *ibqp)
 148 {
 149         return container_of(ibqp, struct efa_qp, ibqp);
 150 }
 151
 152 static inline struct efa_cq *to_ecq(struct ib_cq *ibcq)
 153 {
 154         return container_of(ibcq, struct efa_cq, ibcq);
 155 }
 156
 157 static inline struct efa_ah *to_eah(struct ib_ah *ibah)
 158 {
 159         return container_of(ibah, struct efa_ah, ibah);
 160 }
 161
 162 static inline struct efa_user_mmap_entry *
 163 to_emmap(struct rdma_user_mmap_entry *rdma_entry)
 164 {
 165         return container_of(rdma_entry, struct efa_user_mmap_entry, rdma_entry);
 166 }
 167
 168 #define EFA_DEV_CAP(dev, cap) \
 169         ((dev)->dev_attr.device_caps & \
 170          EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_##cap##_MASK)
 171
 172 #define is_reserved_cleared(reserved) \
 173         !memchr_inv(reserved, 0, sizeof(reserved))
 174
 175 static void *efa_zalloc_mapped(struct efa_dev *dev, dma_addr_t *dma_addr,
 176                                size_t size, enum dma_data_direction dir)
 177 {
 178         void *addr;
 179
 180         addr = alloc_pages_exact(size, GFP_KERNEL | __GFP_ZERO);
 181         if (!addr)
 182                 return NULL;
 183
 184         *dma_addr = dma_map_single(&dev->pdev->dev, addr, size, dir);
 185         if (dma_mapping_error(&dev->pdev->dev, *dma_addr)) {
 186                 ibdev_err(&dev->ibdev, "Failed to map DMA address\n");
 187                 free_pages_exact(addr, size);
 188                 return NULL;
 189         }
 190
 191         return addr;
 192 }
 193
 194 static void efa_free_mapped(struct efa_dev *dev, void *cpu_addr,
 195                             dma_addr_t dma_addr,
 196                             size_t size, enum dma_data_direction dir)
 197 {
 198         dma_unmap_single(&dev->pdev->dev, dma_addr, size, dir);
 199         free_pages_exact(cpu_addr, size);
 200 }
 201
 202 int efa_query_device(struct ib_device *ibdev,
 203                      struct ib_device_attr *props,
 204                      struct ib_udata *udata)
 205 {
 206         struct efa_com_get_device_attr_result *dev_attr;
 207         struct efa_ibv_ex_query_device_resp resp = {};
 208         struct efa_dev *dev = to_edev(ibdev);
 209         int err;
 210
 211         if (udata && udata->inlen &&
 212             !ib_is_udata_cleared(udata, 0, udata->inlen)) {
 213                 ibdev_dbg(ibdev,
 214                           "Incompatible ABI params, udata not cleared\n");
 215                 return -EINVAL;
 216         }
 217
 218         dev_attr = &dev->dev_attr;
 219
 220         memset(props, 0, sizeof(*props));
 221         props->max_mr_size = dev_attr->max_mr_pages * PAGE_SIZE;
 222         props->page_size_cap = dev_attr->page_size_cap;
 223         props->vendor_id = dev->pdev->vendor;
 224         props->vendor_part_id = dev->pdev->device;
 225         props->hw_ver = dev->pdev->subsystem_device;
 226         props->max_qp = dev_attr->max_qp;
 227         props->max_cq = dev_attr->max_cq;
 228         props->max_pd = dev_attr->max_pd;
 229         props->max_mr = dev_attr->max_mr;
 230         props->max_ah = dev_attr->max_ah;
 231         props->max_cqe = dev_attr->max_cq_depth;
 232         props->max_qp_wr = min_t(u32, dev_attr->max_sq_depth,
 233                                  dev_attr->max_rq_depth);
 234         props->max_send_sge = dev_attr->max_sq_sge;
 235         props->max_recv_sge = dev_attr->max_rq_sge;
 236         props->max_sge_rd = dev_attr->max_wr_rdma_sge;
 237         props->max_pkeys = 1;
 238
 239         if (udata && udata->outlen) {
 240                 resp.max_sq_sge = dev_attr->max_sq_sge;
 241                 resp.max_rq_sge = dev_attr->max_rq_sge;
 242                 resp.max_sq_wr = dev_attr->max_sq_depth;
 243                 resp.max_rq_wr = dev_attr->max_rq_depth;
 244                 resp.max_rdma_size = dev_attr->max_rdma_size;
 245
 246                 resp.device_caps |= EFA_QUERY_DEVICE_CAPS_CQ_WITH_SGID;
 247                 if (EFA_DEV_CAP(dev, RDMA_READ))
 248                         resp.device_caps |= EFA_QUERY_DEVICE_CAPS_RDMA_READ;
 249
 250                 if (EFA_DEV_CAP(dev, RNR_RETRY))
 251                         resp.device_caps |= EFA_QUERY_DEVICE_CAPS_RNR_RETRY;
 252
 253                 if (EFA_DEV_CAP(dev, DATA_POLLING_128))
 254                         resp.device_caps |= EFA_QUERY_DEVICE_CAPS_DATA_POLLING_128;
 255
 256                 if (EFA_DEV_CAP(dev, RDMA_WRITE))
 257                         resp.device_caps |= EFA_QUERY_DEVICE_CAPS_RDMA_WRITE;
 258
 259                 if (dev->neqs)
 260                         resp.device_caps |= EFA_QUERY_DEVICE_CAPS_CQ_NOTIFICATIONS;
 261
 262                 err = ib_copy_to_udata(udata, &resp,
 263                                        min(sizeof(resp), udata->outlen));
 264                 if (err) {
 265                         ibdev_dbg(ibdev,
 266                                   "Failed to copy udata for query_device\n");
 267                         return err;
 268                 }
 269         }
 270
 271         return 0;
 272 }
 273
 274 int efa_query_port(struct ib_device *ibdev, u32 port,
 275                    struct ib_port_attr *props)
 276 {
 277         struct efa_dev *dev = to_edev(ibdev);
 278
 279         props->lmc = 1;
 280
 281         props->state = IB_PORT_ACTIVE;
 282         props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
 283         props->gid_tbl_len = 1;
 284         props->pkey_tbl_len = 1;
 285         props->active_speed = IB_SPEED_EDR;
 286         props->active_width = IB_WIDTH_4X;
 287         props->max_mtu = ib_mtu_int_to_enum(dev->dev_attr.mtu);
 288         props->active_mtu = ib_mtu_int_to_enum(dev->dev_attr.mtu);
 289         props->max_msg_sz = dev->dev_attr.mtu;
 290         props->max_vl_num = 1;
 291
 292         return 0;
 293 }
 294
 295 int efa_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
 296                  int qp_attr_mask,
 297                  struct ib_qp_init_attr *qp_init_attr)
 298 {
 299         struct efa_dev *dev = to_edev(ibqp->device);
 300         struct efa_com_query_qp_params params = {};
 301         struct efa_com_query_qp_result result;
 302         struct efa_qp *qp = to_eqp(ibqp);
 303         int err;
 304
 305 #define EFA_QUERY_QP_SUPP_MASK \
 306         (IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT | \
 307          IB_QP_QKEY | IB_QP_SQ_PSN | IB_QP_CAP | IB_QP_RNR_RETRY)
 308
 309         if (qp_attr_mask & ~EFA_QUERY_QP_SUPP_MASK) {
 310                 ibdev_dbg(&dev->ibdev,
 311                           "Unsupported qp_attr_mask[%#x] supported[%#x]\n",
 312                           qp_attr_mask, EFA_QUERY_QP_SUPP_MASK);
 313                 return -EOPNOTSUPP;
 314         }
 315
 316         memset(qp_attr, 0, sizeof(*qp_attr));
 317         memset(qp_init_attr, 0, sizeof(*qp_init_attr));
 318
 319         params.qp_handle = qp->qp_handle;
 320         err = efa_com_query_qp(&dev->edev, &params, &result);
 321         if (err)
 322                 return err;
 323
 324         qp_attr->qp_state = result.qp_state;
 325         qp_attr->qkey = result.qkey;
 326         qp_attr->sq_psn = result.sq_psn;
 327         qp_attr->sq_draining = result.sq_draining;
 328         qp_attr->port_num = 1;
 329         qp_attr->rnr_retry = result.rnr_retry;
 330
 331         qp_attr->cap.max_send_wr = qp->max_send_wr;
 332         qp_attr->cap.max_recv_wr = qp->max_recv_wr;
 333         qp_attr->cap.max_send_sge = qp->max_send_sge;
 334         qp_attr->cap.max_recv_sge = qp->max_recv_sge;
 335         qp_attr->cap.max_inline_data = qp->max_inline_data;
 336
 337         qp_init_attr->qp_type = ibqp->qp_type;
 338         qp_init_attr->recv_cq = ibqp->recv_cq;
 339         qp_init_attr->send_cq = ibqp->send_cq;
 340         qp_init_attr->qp_context = ibqp->qp_context;
 341         qp_init_attr->cap = qp_attr->cap;
 342
 343         return 0;
 344 }
 345
 346 int efa_query_gid(struct ib_device *ibdev, u32 port, int index,
 347                   union ib_gid *gid)
 348 {
 349         struct efa_dev *dev = to_edev(ibdev);
 350
 351         memcpy(gid->raw, dev->dev_attr.addr, sizeof(dev->dev_attr.addr));
 352
 353         return 0;
 354 }
 355
 356 int efa_query_pkey(struct ib_device *ibdev, u32 port, u16 index,
 357                    u16 *pkey)
 358 {
 359         if (index > 0)
 360                 return -EINVAL;
 361
 362         *pkey = 0xffff;
 363         return 0;
 364 }
 365
 366 static int efa_pd_dealloc(struct efa_dev *dev, u16 pdn)
 367 {
 368         struct efa_com_dealloc_pd_params params = {
 369                 .pdn = pdn,
 370         };
 371
 372         return efa_com_dealloc_pd(&dev->edev, &params);
 373 }
 374
 375 int efa_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 376 {
 377         struct efa_dev *dev = to_edev(ibpd->device);
 378         struct efa_ibv_alloc_pd_resp resp = {};
 379         struct efa_com_alloc_pd_result result;
 380         struct efa_pd *pd = to_epd(ibpd);
 381         int err;
 382
 383         if (udata->inlen &&
 384             !ib_is_udata_cleared(udata, 0, udata->inlen)) {
 385                 ibdev_dbg(&dev->ibdev,
 386                           "Incompatible ABI params, udata not cleared\n");
 387                 err = -EINVAL;
 388                 goto err_out;
 389         }
 390
 391         err = efa_com_alloc_pd(&dev->edev, &result);
 392         if (err)
 393                 goto err_out;
 394
 395         pd->pdn = result.pdn;
 396         resp.pdn = result.pdn;
 397
 398         if (udata->outlen) {
 399                 err = ib_copy_to_udata(udata, &resp,
 400                                        min(sizeof(resp), udata->outlen));
 401                 if (err) {
 402                         ibdev_dbg(&dev->ibdev,
 403                                   "Failed to copy udata for alloc_pd\n");
 404                         goto err_dealloc_pd;
 405                 }
 406         }
 407
 408         ibdev_dbg(&dev->ibdev, "Allocated pd[%d]\n", pd->pdn);
 409
 410         return 0;
 411
 412 err_dealloc_pd:
 413         efa_pd_dealloc(dev, result.pdn);
 414 err_out:
 415         atomic64_inc(&dev->stats.alloc_pd_err);
 416         return err;
 417 }
 418
 419 int efa_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 420 {
 421         struct efa_dev *dev = to_edev(ibpd->device);
 422         struct efa_pd *pd = to_epd(ibpd);
 423
 424         ibdev_dbg(&dev->ibdev, "Dealloc pd[%d]\n", pd->pdn);
 425         efa_pd_dealloc(dev, pd->pdn);
 426         return 0;
 427 }
 428
 429 static int efa_destroy_qp_handle(struct efa_dev *dev, u32 qp_handle)
 430 {
 431         struct efa_com_destroy_qp_params params = { .qp_handle = qp_handle };
 432
 433         return efa_com_destroy_qp(&dev->edev, &params);
 434 }
 435
 436 static void efa_qp_user_mmap_entries_remove(struct efa_qp *qp)
 437 {
 438         rdma_user_mmap_entry_remove(qp->rq_mmap_entry);
 439         rdma_user_mmap_entry_remove(qp->rq_db_mmap_entry);
 440         rdma_user_mmap_entry_remove(qp->llq_desc_mmap_entry);
 441         rdma_user_mmap_entry_remove(qp->sq_db_mmap_entry);
 442 }
 443
 444 int efa_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
 445 {
 446         struct efa_dev *dev = to_edev(ibqp->pd->device);
 447         struct efa_qp *qp = to_eqp(ibqp);
 448         int err;
 449
 450         ibdev_dbg(&dev->ibdev, "Destroy qp[%u]\n", ibqp->qp_num);
 451
 452         efa_qp_user_mmap_entries_remove(qp);
 453
 454         err = efa_destroy_qp_handle(dev, qp->qp_handle);
 455         if (err)
 456                 return err;
 457
 458         if (qp->rq_cpu_addr) {
 459                 ibdev_dbg(&dev->ibdev,
 460                           "qp->cpu_addr[0x%p] freed: size[%lu], dma[%pad]\n",
 461                           qp->rq_cpu_addr, qp->rq_size,
 462                           &qp->rq_dma_addr);
 463                 efa_free_mapped(dev, qp->rq_cpu_addr, qp->rq_dma_addr,
 464                                 qp->rq_size, DMA_TO_DEVICE);
 465         }
 466
 467         return 0;
 468 }
 469
 470 static struct rdma_user_mmap_entry*
 471 efa_user_mmap_entry_insert(struct ib_ucontext *ucontext,
 472                            u64 address, size_t length,
 473                            u8 mmap_flag, u64 *offset)
 474 {
 475         struct efa_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 476         int err;
 477
 478         if (!entry)
 479                 return NULL;
 480
 481         entry->address = address;
 482         entry->mmap_flag = mmap_flag;
 483
 484         err = rdma_user_mmap_entry_insert(ucontext, &entry->rdma_entry,
 485                                           length);
 486         if (err) {
 487                 kfree(entry);
 488                 return NULL;
 489         }
 490         *offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
 491
 492         return &entry->rdma_entry;
 493 }
 494
 495 static int qp_mmap_entries_setup(struct efa_qp *qp,
 496                                  struct efa_dev *dev,
 497                                  struct efa_ucontext *ucontext,
 498                                  struct efa_com_create_qp_params *params,
 499                                  struct efa_ibv_create_qp_resp *resp)
 500 {
 501         size_t length;
 502         u64 address;
 503
 504         address = dev->db_bar_addr + resp->sq_db_offset;
 505         qp->sq_db_mmap_entry =
 506                 efa_user_mmap_entry_insert(&ucontext->ibucontext,
 507                                            address,
 508                                            PAGE_SIZE, EFA_MMAP_IO_NC,
 509                                            &resp->sq_db_mmap_key);
 510         if (!qp->sq_db_mmap_entry)
 511                 return -ENOMEM;
 512
 513         resp->sq_db_offset &= ~PAGE_MASK;
 514
 515         address = dev->mem_bar_addr + resp->llq_desc_offset;
 516         length = PAGE_ALIGN(params->sq_ring_size_in_bytes +
 517                             (resp->llq_desc_offset & ~PAGE_MASK));
 518
 519         qp->llq_desc_mmap_entry =
 520                 efa_user_mmap_entry_insert(&ucontext->ibucontext,
 521                                            address, length,
 522                                            EFA_MMAP_IO_WC,
 523                                            &resp->llq_desc_mmap_key);
 524         if (!qp->llq_desc_mmap_entry)
 525                 goto err_remove_mmap;
 526
 527         resp->llq_desc_offset &= ~PAGE_MASK;
 528
 529         if (qp->rq_size) {
 530                 address = dev->db_bar_addr + resp->rq_db_offset;
 531
 532                 qp->rq_db_mmap_entry =
 533                         efa_user_mmap_entry_insert(&ucontext->ibucontext,
 534                                                    address, PAGE_SIZE,
 535                                                    EFA_MMAP_IO_NC,
 536                                                    &resp->rq_db_mmap_key);
 537                 if (!qp->rq_db_mmap_entry)
 538                         goto err_remove_mmap;
 539
 540                 resp->rq_db_offset &= ~PAGE_MASK;
 541
 542                 address = virt_to_phys(qp->rq_cpu_addr);
 543                 qp->rq_mmap_entry =
 544                         efa_user_mmap_entry_insert(&ucontext->ibucontext,
 545                                                    address, qp->rq_size,
 546                                                    EFA_MMAP_DMA_PAGE,
 547                                                    &resp->rq_mmap_key);
 548                 if (!qp->rq_mmap_entry)
 549                         goto err_remove_mmap;
 550
 551                 resp->rq_mmap_size = qp->rq_size;
 552         }
 553
 554         return 0;
 555
 556 err_remove_mmap:
 557         efa_qp_user_mmap_entries_remove(qp);
 558
 559         return -ENOMEM;
 560 }
 561
 562 static int efa_qp_validate_cap(struct efa_dev *dev,
 563                                struct ib_qp_init_attr *init_attr)
 564 {
 565         if (init_attr->cap.max_send_wr > dev->dev_attr.max_sq_depth) {
 566                 ibdev_dbg(&dev->ibdev,
 567                           "qp: requested send wr[%u] exceeds the max[%u]\n",
 568                           init_attr->cap.max_send_wr,
 569                           dev->dev_attr.max_sq_depth);
 570                 return -EINVAL;
 571         }
 572         if (init_attr->cap.max_recv_wr > dev->dev_attr.max_rq_depth) {
 573                 ibdev_dbg(&dev->ibdev,
 574                           "qp: requested receive wr[%u] exceeds the max[%u]\n",
 575                           init_attr->cap.max_recv_wr,
 576                           dev->dev_attr.max_rq_depth);
 577                 return -EINVAL;
 578         }
 579         if (init_attr->cap.max_send_sge > dev->dev_attr.max_sq_sge) {
 580                 ibdev_dbg(&dev->ibdev,
 581                           "qp: requested sge send[%u] exceeds the max[%u]\n",
 582                           init_attr->cap.max_send_sge, dev->dev_attr.max_sq_sge);
 583                 return -EINVAL;
 584         }
 585         if (init_attr->cap.max_recv_sge > dev->dev_attr.max_rq_sge) {
 586                 ibdev_dbg(&dev->ibdev,
 587                           "qp: requested sge recv[%u] exceeds the max[%u]\n",
 588                           init_attr->cap.max_recv_sge, dev->dev_attr.max_rq_sge);
 589                 return -EINVAL;
 590         }
 591         if (init_attr->cap.max_inline_data > dev->dev_attr.inline_buf_size) {
 592                 ibdev_dbg(&dev->ibdev,
 593                           "qp: requested inline data[%u] exceeds the max[%u]\n",
 594                           init_attr->cap.max_inline_data,
 595                           dev->dev_attr.inline_buf_size);
 596                 return -EINVAL;
 597         }
 598
 599         return 0;
 600 }
 601
 602 static int efa_qp_validate_attr(struct efa_dev *dev,
 603                                 struct ib_qp_init_attr *init_attr)
 604 {
 605         if (init_attr->qp_type != IB_QPT_DRIVER &&
 606             init_attr->qp_type != IB_QPT_UD) {
 607                 ibdev_dbg(&dev->ibdev,
 608                           "Unsupported qp type %d\n", init_attr->qp_type);
 609                 return -EOPNOTSUPP;
 610         }
 611
 612         if (init_attr->srq) {
 613                 ibdev_dbg(&dev->ibdev, "SRQ is not supported\n");
 614                 return -EOPNOTSUPP;
 615         }
 616
 617         if (init_attr->create_flags) {
 618                 ibdev_dbg(&dev->ibdev, "Unsupported create flags\n");
 619                 return -EOPNOTSUPP;
 620         }
 621
 622         return 0;
 623 }
 624
 625 int efa_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init_attr,
 626                   struct ib_udata *udata)
 627 {
 628         struct efa_com_create_qp_params create_qp_params = {};
 629         struct efa_com_create_qp_result create_qp_resp;
 630         struct efa_dev *dev = to_edev(ibqp->device);
 631         struct efa_ibv_create_qp_resp resp = {};
 632         struct efa_ibv_create_qp cmd = {};
 633         struct efa_qp *qp = to_eqp(ibqp);
 634         struct efa_ucontext *ucontext;
 635         int err;
 636
 637         ucontext = rdma_udata_to_drv_context(udata, struct efa_ucontext,
 638                                              ibucontext);
 639
 640         err = efa_qp_validate_cap(dev, init_attr);
 641         if (err)
 642                 goto err_out;
 643
 644         err = efa_qp_validate_attr(dev, init_attr);
 645         if (err)
 646                 goto err_out;
 647
 648         if (offsetofend(typeof(cmd), driver_qp_type) > udata->inlen) {
 649                 ibdev_dbg(&dev->ibdev,
 650                           "Incompatible ABI params, no input udata\n");
 651                 err = -EINVAL;
 652                 goto err_out;
 653         }
 654
 655         if (udata->inlen > sizeof(cmd) &&
 656             !ib_is_udata_cleared(udata, sizeof(cmd),
 657                                  udata->inlen - sizeof(cmd))) {
 658                 ibdev_dbg(&dev->ibdev,
 659                           "Incompatible ABI params, unknown fields in udata\n");
 660                 err = -EINVAL;
 661                 goto err_out;
 662         }
 663
 664         err = ib_copy_from_udata(&cmd, udata,
 665                                  min(sizeof(cmd), udata->inlen));
 666         if (err) {
 667                 ibdev_dbg(&dev->ibdev,
 668                           "Cannot copy udata for create_qp\n");
 669                 goto err_out;
 670         }
 671
 672         if (cmd.comp_mask) {
 673                 ibdev_dbg(&dev->ibdev,
 674                           "Incompatible ABI params, unknown fields in udata\n");
 675                 err = -EINVAL;
 676                 goto err_out;
 677         }
 678
 679         create_qp_params.uarn = ucontext->uarn;
 680         create_qp_params.pd = to_epd(ibqp->pd)->pdn;
 681
 682         if (init_attr->qp_type == IB_QPT_UD) {
 683                 create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_UD;
 684         } else if (cmd.driver_qp_type == EFA_QP_DRIVER_TYPE_SRD) {
 685                 create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_SRD;
 686         } else {
 687                 ibdev_dbg(&dev->ibdev,
 688                           "Unsupported qp type %d driver qp type %d\n",
 689                           init_attr->qp_type, cmd.driver_qp_type);
 690                 err = -EOPNOTSUPP;
 691                 goto err_out;
 692         }
 693
 694         ibdev_dbg(&dev->ibdev, "Create QP: qp type %d driver qp type %#x\n",
 695                   init_attr->qp_type, cmd.driver_qp_type);
 696         create_qp_params.send_cq_idx = to_ecq(init_attr->send_cq)->cq_idx;
 697         create_qp_params.recv_cq_idx = to_ecq(init_attr->recv_cq)->cq_idx;
 698         create_qp_params.sq_depth = init_attr->cap.max_send_wr;
 699         create_qp_params.sq_ring_size_in_bytes = cmd.sq_ring_size;
 700
 701         create_qp_params.rq_depth = init_attr->cap.max_recv_wr;
 702         create_qp_params.rq_ring_size_in_bytes = cmd.rq_ring_size;
 703         qp->rq_size = PAGE_ALIGN(create_qp_params.rq_ring_size_in_bytes);
 704         if (qp->rq_size) {
 705                 qp->rq_cpu_addr = efa_zalloc_mapped(dev, &qp->rq_dma_addr,
 706                                                     qp->rq_size, DMA_TO_DEVICE);
 707                 if (!qp->rq_cpu_addr) {
 708                         err = -ENOMEM;
 709                         goto err_out;
 710                 }
 711
 712                 ibdev_dbg(&dev->ibdev,
 713                           "qp->cpu_addr[0x%p] allocated: size[%lu], dma[%pad]\n",
 714                           qp->rq_cpu_addr, qp->rq_size, &qp->rq_dma_addr);
 715                 create_qp_params.rq_base_addr = qp->rq_dma_addr;
 716         }
 717
 718         err = efa_com_create_qp(&dev->edev, &create_qp_params,
 719                                 &create_qp_resp);
 720         if (err)
 721                 goto err_free_mapped;
 722
 723         resp.sq_db_offset = create_qp_resp.sq_db_offset;
 724         resp.rq_db_offset = create_qp_resp.rq_db_offset;
 725         resp.llq_desc_offset = create_qp_resp.llq_descriptors_offset;
 726         resp.send_sub_cq_idx = create_qp_resp.send_sub_cq_idx;
 727         resp.recv_sub_cq_idx = create_qp_resp.recv_sub_cq_idx;
 728
 729         err = qp_mmap_entries_setup(qp, dev, ucontext, &create_qp_params,
 730                                     &resp);
 731         if (err)
 732                 goto err_destroy_qp;
 733
 734         qp->qp_handle = create_qp_resp.qp_handle;
 735         qp->ibqp.qp_num = create_qp_resp.qp_num;
 736         qp->max_send_wr = init_attr->cap.max_send_wr;
 737         qp->max_recv_wr = init_attr->cap.max_recv_wr;
 738         qp->max_send_sge = init_attr->cap.max_send_sge;
 739         qp->max_recv_sge = init_attr->cap.max_recv_sge;
 740         qp->max_inline_data = init_attr->cap.max_inline_data;
 741
 742         if (udata->outlen) {
 743                 err = ib_copy_to_udata(udata, &resp,
 744                                        min(sizeof(resp), udata->outlen));
 745                 if (err) {
 746                         ibdev_dbg(&dev->ibdev,
 747                                   "Failed to copy udata for qp[%u]\n",
 748                                   create_qp_resp.qp_num);
 749                         goto err_remove_mmap_entries;
 750                 }
 751         }
 752
 753         ibdev_dbg(&dev->ibdev, "Created qp[%d]\n", qp->ibqp.qp_num);
 754
 755         return 0;
 756
 757 err_remove_mmap_entries:
 758         efa_qp_user_mmap_entries_remove(qp);
 759 err_destroy_qp:
 760         efa_destroy_qp_handle(dev, create_qp_resp.qp_handle);
 761 err_free_mapped:
 762         if (qp->rq_size)
 763                 efa_free_mapped(dev, qp->rq_cpu_addr, qp->rq_dma_addr,
 764                                 qp->rq_size, DMA_TO_DEVICE);
 765 err_out:
 766         atomic64_inc(&dev->stats.create_qp_err);
 767         return err;
 768 }
 769
 770 static const struct {
 771         int                     valid;
 772         enum ib_qp_attr_mask    req_param;
 773         enum ib_qp_attr_mask    opt_param;
 774 } srd_qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
 775         [IB_QPS_RESET] = {
 776                 [IB_QPS_RESET] = { .valid = 1 },
 777                 [IB_QPS_INIT]  = {
 778                         .valid = 1,
 779                         .req_param = IB_QP_PKEY_INDEX |
 780                                      IB_QP_PORT |
 781                                      IB_QP_QKEY,
 782                 },
 783         },
 784         [IB_QPS_INIT] = {
 785                 [IB_QPS_RESET] = { .valid = 1 },
 786                 [IB_QPS_ERR]   = { .valid = 1 },
 787                 [IB_QPS_INIT]  = {
 788                         .valid = 1,
 789                         .opt_param = IB_QP_PKEY_INDEX |
 790                                      IB_QP_PORT |
 791                                      IB_QP_QKEY,
 792                 },
 793                 [IB_QPS_RTR]   = {
 794                         .valid = 1,
 795                         .opt_param = IB_QP_PKEY_INDEX |
 796                                      IB_QP_QKEY,
 797                 },
 798         },
 799         [IB_QPS_RTR] = {
 800                 [IB_QPS_RESET] = { .valid = 1 },
 801                 [IB_QPS_ERR]   = { .valid = 1 },
 802                 [IB_QPS_RTS]   = {
 803                         .valid = 1,
 804                         .req_param = IB_QP_SQ_PSN,
 805                         .opt_param = IB_QP_CUR_STATE |
 806                                      IB_QP_QKEY |
 807                                      IB_QP_RNR_RETRY,
 808
 809                 }
 810         },
 811         [IB_QPS_RTS] = {
 812                 [IB_QPS_RESET] = { .valid = 1 },
 813                 [IB_QPS_ERR]   = { .valid = 1 },
 814                 [IB_QPS_RTS]   = {
 815                         .valid = 1,
 816                         .opt_param = IB_QP_CUR_STATE |
 817                                      IB_QP_QKEY,
 818                 },
 819                 [IB_QPS_SQD] = {
 820                         .valid = 1,
 821                         .opt_param = IB_QP_EN_SQD_ASYNC_NOTIFY,
 822                 },
 823         },
 824         [IB_QPS_SQD] = {
 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                 [IB_QPS_SQD] = {
 833                         .valid = 1,
 834                         .opt_param = IB_QP_PKEY_INDEX |
 835                                      IB_QP_QKEY,
 836                 }
 837         },
 838         [IB_QPS_SQE] = {
 839                 [IB_QPS_RESET] = { .valid = 1 },
 840                 [IB_QPS_ERR]   = { .valid = 1 },
 841                 [IB_QPS_RTS]   = {
 842                         .valid = 1,
 843                         .opt_param = IB_QP_CUR_STATE |
 844                                      IB_QP_QKEY,
 845                 }
 846         },
 847         [IB_QPS_ERR] = {
 848                 [IB_QPS_RESET] = { .valid = 1 },
 849                 [IB_QPS_ERR]   = { .valid = 1 },
 850         }
 851 };
 852
 853 static bool efa_modify_srd_qp_is_ok(enum ib_qp_state cur_state,
 854                                     enum ib_qp_state next_state,
 855                                     enum ib_qp_attr_mask mask)
 856 {
 857         enum ib_qp_attr_mask req_param, opt_param;
 858
 859         if (mask & IB_QP_CUR_STATE  &&
 860             cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
 861             cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
 862                 return false;
 863
 864         if (!srd_qp_state_table[cur_state][next_state].valid)
 865                 return false;
 866
 867         req_param = srd_qp_state_table[cur_state][next_state].req_param;
 868         opt_param = srd_qp_state_table[cur_state][next_state].opt_param;
 869
 870         if ((mask & req_param) != req_param)
 871                 return false;
 872
 873         if (mask & ~(req_param | opt_param | IB_QP_STATE))
 874                 return false;
 875
 876         return true;
 877 }
 878
 879 static int efa_modify_qp_validate(struct efa_dev *dev, struct efa_qp *qp,
 880                                   struct ib_qp_attr *qp_attr, int qp_attr_mask,
 881                                   enum ib_qp_state cur_state,
 882                                   enum ib_qp_state new_state)
 883 {
 884         int err;
 885
 886 #define EFA_MODIFY_QP_SUPP_MASK \
 887         (IB_QP_STATE | IB_QP_CUR_STATE | IB_QP_EN_SQD_ASYNC_NOTIFY | \
 888          IB_QP_PKEY_INDEX | IB_QP_PORT | IB_QP_QKEY | IB_QP_SQ_PSN | \
 889          IB_QP_RNR_RETRY)
 890
 891         if (qp_attr_mask & ~EFA_MODIFY_QP_SUPP_MASK) {
 892                 ibdev_dbg(&dev->ibdev,
 893                           "Unsupported qp_attr_mask[%#x] supported[%#x]\n",
 894                           qp_attr_mask, EFA_MODIFY_QP_SUPP_MASK);
 895                 return -EOPNOTSUPP;
 896         }
 897
 898         if (qp->ibqp.qp_type == IB_QPT_DRIVER)
 899                 err = !efa_modify_srd_qp_is_ok(cur_state, new_state,
 900                                                qp_attr_mask);
 901         else
 902                 err = !ib_modify_qp_is_ok(cur_state, new_state, IB_QPT_UD,
 903                                           qp_attr_mask);
 904
 905         if (err) {
 906                 ibdev_dbg(&dev->ibdev, "Invalid modify QP parameters\n");
 907                 return -EINVAL;
 908         }
 909
 910         if ((qp_attr_mask & IB_QP_PORT) && qp_attr->port_num != 1) {
 911                 ibdev_dbg(&dev->ibdev, "Can't change port num\n");
 912                 return -EOPNOTSUPP;
 913         }
 914
 915         if ((qp_attr_mask & IB_QP_PKEY_INDEX) && qp_attr->pkey_index) {
 916                 ibdev_dbg(&dev->ibdev, "Can't change pkey index\n");
 917                 return -EOPNOTSUPP;
 918         }
 919
 920         return 0;
 921 }
 922
 923 int efa_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
 924                   int qp_attr_mask, struct ib_udata *udata)
 925 {
 926         struct efa_dev *dev = to_edev(ibqp->device);
 927         struct efa_com_modify_qp_params params = {};
 928         struct efa_qp *qp = to_eqp(ibqp);
 929         enum ib_qp_state cur_state;
 930         enum ib_qp_state new_state;
 931         int err;
 932
 933         if (qp_attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
 934                 return -EOPNOTSUPP;
 935
 936         if (udata->inlen &&
 937             !ib_is_udata_cleared(udata, 0, udata->inlen)) {
 938                 ibdev_dbg(&dev->ibdev,
 939                           "Incompatible ABI params, udata not cleared\n");
 940                 return -EINVAL;
 941         }
 942
 943         cur_state = qp_attr_mask & IB_QP_CUR_STATE ? qp_attr->cur_qp_state :
 944                                                      qp->state;
 945         new_state = qp_attr_mask & IB_QP_STATE ? qp_attr->qp_state : cur_state;
 946
 947         err = efa_modify_qp_validate(dev, qp, qp_attr, qp_attr_mask, cur_state,
 948                                      new_state);
 949         if (err)
 950                 return err;
 951
 952         params.qp_handle = qp->qp_handle;
 953
 954         if (qp_attr_mask & IB_QP_STATE) {
 955                 EFA_SET(&params.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_QP_STATE,
 956                         1);
 957                 EFA_SET(&params.modify_mask,
 958                         EFA_ADMIN_MODIFY_QP_CMD_CUR_QP_STATE, 1);
 959                 params.cur_qp_state = cur_state;
 960                 params.qp_state = new_state;
 961         }
 962
 963         if (qp_attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) {
 964                 EFA_SET(&params.modify_mask,
 965                         EFA_ADMIN_MODIFY_QP_CMD_SQ_DRAINED_ASYNC_NOTIFY, 1);
 966                 params.sq_drained_async_notify = qp_attr->en_sqd_async_notify;
 967         }
 968
 969         if (qp_attr_mask & IB_QP_QKEY) {
 970                 EFA_SET(&params.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_QKEY, 1);
 971                 params.qkey = qp_attr->qkey;
 972         }
 973
 974         if (qp_attr_mask & IB_QP_SQ_PSN) {
 975                 EFA_SET(&params.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_SQ_PSN, 1);
 976                 params.sq_psn = qp_attr->sq_psn;
 977         }
 978
 979         if (qp_attr_mask & IB_QP_RNR_RETRY) {
 980                 EFA_SET(&params.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_RNR_RETRY,
 981                         1);
 982                 params.rnr_retry = qp_attr->rnr_retry;
 983         }
 984
 985         err = efa_com_modify_qp(&dev->edev, &params);
 986         if (err)
 987                 return err;
 988
 989         qp->state = new_state;
 990
 991         return 0;
 992 }
 993
 994 static int efa_destroy_cq_idx(struct efa_dev *dev, int cq_idx)
 995 {
 996         struct efa_com_destroy_cq_params params = { .cq_idx = cq_idx };
 997
 998         return efa_com_destroy_cq(&dev->edev, &params);
 999 }
1000
1001 static void efa_cq_user_mmap_entries_remove(struct efa_cq *cq)
1002 {
1003         rdma_user_mmap_entry_remove(cq->db_mmap_entry);
1004         rdma_user_mmap_entry_remove(cq->mmap_entry);
1005 }
1006
1007 int efa_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
1008 {
1009         struct efa_dev *dev = to_edev(ibcq->device);
1010         struct efa_cq *cq = to_ecq(ibcq);
1011
1012         ibdev_dbg(&dev->ibdev,
1013                   "Destroy cq[%d] virt[0x%p] freed: size[%lu], dma[%pad]\n",
1014                   cq->cq_idx, cq->cpu_addr, cq->size, &cq->dma_addr);
1015
1016         efa_cq_user_mmap_entries_remove(cq);
1017         efa_destroy_cq_idx(dev, cq->cq_idx);
1018         if (cq->eq) {
1019                 xa_erase(&dev->cqs_xa, cq->cq_idx);
1020                 synchronize_irq(cq->eq->irq.irqn);
1021         }
1022         efa_free_mapped(dev, cq->cpu_addr, cq->dma_addr, cq->size,
1023                         DMA_FROM_DEVICE);
1024         return 0;
1025 }
1026
1027 static struct efa_eq *efa_vec2eq(struct efa_dev *dev, int vec)
1028 {
1029         return &dev->eqs[vec];
1030 }
1031
1032 static int cq_mmap_entries_setup(struct efa_dev *dev, struct efa_cq *cq,
1033                                  struct efa_ibv_create_cq_resp *resp,
1034                                  bool db_valid)
1035 {
1036         resp->q_mmap_size = cq->size;
1037         cq->mmap_entry = efa_user_mmap_entry_insert(&cq->ucontext->ibucontext,
1038                                                     virt_to_phys(cq->cpu_addr),
1039                                                     cq->size, EFA_MMAP_DMA_PAGE,
1040                                                     &resp->q_mmap_key);
1041         if (!cq->mmap_entry)
1042                 return -ENOMEM;
1043
1044         if (db_valid) {
1045                 cq->db_mmap_entry =
1046                         efa_user_mmap_entry_insert(&cq->ucontext->ibucontext,
1047                                                    dev->db_bar_addr + resp->db_off,
1048                                                    PAGE_SIZE, EFA_MMAP_IO_NC,
1049                                                    &resp->db_mmap_key);
1050                 if (!cq->db_mmap_entry) {
1051                         rdma_user_mmap_entry_remove(cq->mmap_entry);
1052                         return -ENOMEM;
1053                 }
1054
1055                 resp->db_off &= ~PAGE_MASK;
1056                 resp->comp_mask |= EFA_CREATE_CQ_RESP_DB_OFF;
1057         }
1058
1059         return 0;
1060 }
1061
1062 int efa_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
1063                   struct ib_udata *udata)
1064 {
1065         struct efa_ucontext *ucontext = rdma_udata_to_drv_context(
1066                 udata, struct efa_ucontext, ibucontext);
1067         struct efa_com_create_cq_params params = {};
1068         struct efa_ibv_create_cq_resp resp = {};
1069         struct efa_com_create_cq_result result;
1070         struct ib_device *ibdev = ibcq->device;
1071         struct efa_dev *dev = to_edev(ibdev);
1072         struct efa_ibv_create_cq cmd = {};
1073         struct efa_cq *cq = to_ecq(ibcq);
1074         int entries = attr->cqe;
1075         bool set_src_addr;
1076         int err;
1077
1078         ibdev_dbg(ibdev, "create_cq entries %d\n", entries);
1079
1080         if (attr->flags)
1081                 return -EOPNOTSUPP;
1082
1083         if (entries < 1 || entries > dev->dev_attr.max_cq_depth) {
1084                 ibdev_dbg(ibdev,
1085                           "cq: requested entries[%u] non-positive or greater than max[%u]\n",
1086                           entries, dev->dev_attr.max_cq_depth);
1087                 err = -EINVAL;
1088                 goto err_out;
1089         }
1090
1091         if (offsetofend(typeof(cmd), num_sub_cqs) > udata->inlen) {
1092                 ibdev_dbg(ibdev,
1093                           "Incompatible ABI params, no input udata\n");
1094                 err = -EINVAL;
1095                 goto err_out;
1096         }
1097
1098         if (udata->inlen > sizeof(cmd) &&
1099             !ib_is_udata_cleared(udata, sizeof(cmd),
1100                                  udata->inlen - sizeof(cmd))) {
1101                 ibdev_dbg(ibdev,
1102                           "Incompatible ABI params, unknown fields in udata\n");
1103                 err = -EINVAL;
1104                 goto err_out;
1105         }
1106
1107         err = ib_copy_from_udata(&cmd, udata,
1108                                  min(sizeof(cmd), udata->inlen));
1109         if (err) {
1110                 ibdev_dbg(ibdev, "Cannot copy udata for create_cq\n");
1111                 goto err_out;
1112         }
1113
1114         if (cmd.comp_mask || !is_reserved_cleared(cmd.reserved_58)) {
1115                 ibdev_dbg(ibdev,
1116                           "Incompatible ABI params, unknown fields in udata\n");
1117                 err = -EINVAL;
1118                 goto err_out;
1119         }
1120
1121         set_src_addr = !!(cmd.flags & EFA_CREATE_CQ_WITH_SGID);
1122         if ((cmd.cq_entry_size != sizeof(struct efa_io_rx_cdesc_ex)) &&
1123             (set_src_addr ||
1124              cmd.cq_entry_size != sizeof(struct efa_io_rx_cdesc))) {
1125                 ibdev_dbg(ibdev,
1126                           "Invalid entry size [%u]\n", cmd.cq_entry_size);
1127                 err = -EINVAL;
1128                 goto err_out;
1129         }
1130
1131         if (cmd.num_sub_cqs != dev->dev_attr.sub_cqs_per_cq) {
1132                 ibdev_dbg(ibdev,
1133                           "Invalid number of sub cqs[%u] expected[%u]\n",
1134                           cmd.num_sub_cqs, dev->dev_attr.sub_cqs_per_cq);
1135                 err = -EINVAL;
1136                 goto err_out;
1137         }
1138
1139         cq->ucontext = ucontext;
1140         cq->size = PAGE_ALIGN(cmd.cq_entry_size * entries * cmd.num_sub_cqs);
1141         cq->cpu_addr = efa_zalloc_mapped(dev, &cq->dma_addr, cq->size,
1142                                          DMA_FROM_DEVICE);
1143         if (!cq->cpu_addr) {
1144                 err = -ENOMEM;
1145                 goto err_out;
1146         }
1147
1148         params.uarn = cq->ucontext->uarn;
1149         params.cq_depth = entries;
1150         params.dma_addr = cq->dma_addr;
1151         params.entry_size_in_bytes = cmd.cq_entry_size;
1152         params.num_sub_cqs = cmd.num_sub_cqs;
1153         params.set_src_addr = set_src_addr;
1154         if (cmd.flags & EFA_CREATE_CQ_WITH_COMPLETION_CHANNEL) {
1155                 cq->eq = efa_vec2eq(dev, attr->comp_vector);
1156                 params.eqn = cq->eq->eeq.eqn;
1157                 params.interrupt_mode_enabled = true;
1158         }
1159
1160         err = efa_com_create_cq(&dev->edev, &params, &result);
1161         if (err)
1162                 goto err_free_mapped;
1163
1164         resp.db_off = result.db_off;
1165         resp.cq_idx = result.cq_idx;
1166         cq->cq_idx = result.cq_idx;
1167         cq->ibcq.cqe = result.actual_depth;
1168         WARN_ON_ONCE(entries != result.actual_depth);
1169
1170         err = cq_mmap_entries_setup(dev, cq, &resp, result.db_valid);
1171         if (err) {
1172                 ibdev_dbg(ibdev, "Could not setup cq[%u] mmap entries\n",
1173                           cq->cq_idx);
1174                 goto err_destroy_cq;
1175         }
1176
1177         if (cq->eq) {
1178                 err = xa_err(xa_store(&dev->cqs_xa, cq->cq_idx, cq, GFP_KERNEL));
1179                 if (err) {
1180                         ibdev_dbg(ibdev, "Failed to store cq[%u] in xarray\n",
1181                                   cq->cq_idx);
1182                         goto err_remove_mmap;
1183                 }
1184         }
1185
1186         if (udata->outlen) {
1187                 err = ib_copy_to_udata(udata, &resp,
1188                                        min(sizeof(resp), udata->outlen));
1189                 if (err) {
1190                         ibdev_dbg(ibdev,
1191                                   "Failed to copy udata for create_cq\n");
1192                         goto err_xa_erase;
1193                 }
1194         }
1195
1196         ibdev_dbg(ibdev, "Created cq[%d], cq depth[%u]. dma[%pad] virt[0x%p]\n",
1197                   cq->cq_idx, result.actual_depth, &cq->dma_addr, cq->cpu_addr);
1198
1199         return 0;
1200
1201 err_xa_erase:
1202         if (cq->eq)
1203                 xa_erase(&dev->cqs_xa, cq->cq_idx);
1204 err_remove_mmap:
1205         efa_cq_user_mmap_entries_remove(cq);
1206 err_destroy_cq:
1207         efa_destroy_cq_idx(dev, cq->cq_idx);
1208 err_free_mapped:
1209         efa_free_mapped(dev, cq->cpu_addr, cq->dma_addr, cq->size,
1210                         DMA_FROM_DEVICE);
1211
1212 err_out:
1213         atomic64_inc(&dev->stats.create_cq_err);
1214         return err;
1215 }
1216
1217 static int umem_to_page_list(struct efa_dev *dev,
1218                              struct ib_umem *umem,
1219                              u64 *page_list,
1220                              u32 hp_cnt,
1221                              u8 hp_shift)
1222 {
1223         u32 pages_in_hp = BIT(hp_shift - PAGE_SHIFT);
1224         struct ib_block_iter biter;
1225         unsigned int hp_idx = 0;
1226
1227         ibdev_dbg(&dev->ibdev, "hp_cnt[%u], pages_in_hp[%u]\n",
1228                   hp_cnt, pages_in_hp);
1229
1230         rdma_umem_for_each_dma_block(umem, &biter, BIT(hp_shift))
1231                 page_list[hp_idx++] = rdma_block_iter_dma_address(&biter);
1232
1233         return 0;
1234 }
1235
1236 static struct scatterlist *efa_vmalloc_buf_to_sg(u64 *buf, int page_cnt)
1237 {
1238         struct scatterlist *sglist;
1239         struct page *pg;
1240         int i;
1241
1242         sglist = kmalloc_array(page_cnt, sizeof(*sglist), GFP_KERNEL);
1243         if (!sglist)
1244                 return NULL;
1245         sg_init_table(sglist, page_cnt);
1246         for (i = 0; i < page_cnt; i++) {
1247                 pg = vmalloc_to_page(buf);
1248                 if (!pg)
1249                         goto err;
1250                 sg_set_page(&sglist[i], pg, PAGE_SIZE, 0);
1251                 buf += PAGE_SIZE / sizeof(*buf);
1252         }
1253         return sglist;
1254
1255 err:
1256         kfree(sglist);
1257         return NULL;
1258 }
1259
1260 /*
1261  * create a chunk list of physical pages dma addresses from the supplied
1262  * scatter gather list
1263  */
1264 static int pbl_chunk_list_create(struct efa_dev *dev, struct pbl_context *pbl)
1265 {
1266         struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list;
1267         int page_cnt = pbl->phys.indirect.pbl_buf_size_in_pages;
1268         struct scatterlist *pages_sgl = pbl->phys.indirect.sgl;
1269         unsigned int chunk_list_size, chunk_idx, payload_idx;
1270         int sg_dma_cnt = pbl->phys.indirect.sg_dma_cnt;
1271         struct efa_com_ctrl_buff_info *ctrl_buf;
1272         u64 *cur_chunk_buf, *prev_chunk_buf;
1273         struct ib_block_iter biter;
1274         dma_addr_t dma_addr;
1275         int i;
1276
1277         /* allocate a chunk list that consists of 4KB chunks */
1278         chunk_list_size = DIV_ROUND_UP(page_cnt, EFA_PTRS_PER_CHUNK);
1279
1280         chunk_list->size = chunk_list_size;
1281         chunk_list->chunks = kcalloc(chunk_list_size,
1282                                      sizeof(*chunk_list->chunks),
1283                                      GFP_KERNEL);
1284         if (!chunk_list->chunks)
1285                 return -ENOMEM;
1286
1287         ibdev_dbg(&dev->ibdev,
1288                   "chunk_list_size[%u] - pages[%u]\n", chunk_list_size,
1289                   page_cnt);
1290
1291         /* allocate chunk buffers: */
1292         for (i = 0; i < chunk_list_size; i++) {
1293                 chunk_list->chunks[i].buf = kzalloc(EFA_CHUNK_SIZE, GFP_KERNEL);
1294                 if (!chunk_list->chunks[i].buf)
1295                         goto chunk_list_dealloc;
1296
1297                 chunk_list->chunks[i].length = EFA_CHUNK_USED_SIZE;
1298         }
1299         chunk_list->chunks[chunk_list_size - 1].length =
1300                 ((page_cnt % EFA_PTRS_PER_CHUNK) * EFA_CHUNK_PAYLOAD_PTR_SIZE) +
1301                         EFA_CHUNK_PTR_SIZE;
1302
1303         /* fill the dma addresses of sg list pages to chunks: */
1304         chunk_idx = 0;
1305         payload_idx = 0;
1306         cur_chunk_buf = chunk_list->chunks[0].buf;
1307         rdma_for_each_block(pages_sgl, &biter, sg_dma_cnt,
1308                             EFA_CHUNK_PAYLOAD_SIZE) {
1309                 cur_chunk_buf[payload_idx++] =
1310                         rdma_block_iter_dma_address(&biter);
1311
1312                 if (payload_idx == EFA_PTRS_PER_CHUNK) {
1313                         chunk_idx++;
1314                         cur_chunk_buf = chunk_list->chunks[chunk_idx].buf;
1315                         payload_idx = 0;
1316                 }
1317         }
1318
1319         /* map chunks to dma and fill chunks next ptrs */
1320         for (i = chunk_list_size - 1; i >= 0; i--) {
1321                 dma_addr = dma_map_single(&dev->pdev->dev,
1322                                           chunk_list->chunks[i].buf,
1323                                           chunk_list->chunks[i].length,
1324                                           DMA_TO_DEVICE);
1325                 if (dma_mapping_error(&dev->pdev->dev, dma_addr)) {
1326                         ibdev_err(&dev->ibdev,
1327                                   "chunk[%u] dma_map_failed\n", i);
1328                         goto chunk_list_unmap;
1329                 }
1330
1331                 chunk_list->chunks[i].dma_addr = dma_addr;
1332                 ibdev_dbg(&dev->ibdev,
1333                           "chunk[%u] mapped at [%pad]\n", i, &dma_addr);
1334
1335                 if (!i)
1336                         break;
1337
1338                 prev_chunk_buf = chunk_list->chunks[i - 1].buf;
1339
1340                 ctrl_buf = (struct efa_com_ctrl_buff_info *)
1341                                 &prev_chunk_buf[EFA_PTRS_PER_CHUNK];
1342                 ctrl_buf->length = chunk_list->chunks[i].length;
1343
1344                 efa_com_set_dma_addr(dma_addr,
1345                                      &ctrl_buf->address.mem_addr_high,
1346                                      &ctrl_buf->address.mem_addr_low);
1347         }
1348
1349         return 0;
1350
1351 chunk_list_unmap:
1352         for (; i < chunk_list_size; i++) {
1353                 dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr,
1354                                  chunk_list->chunks[i].length, DMA_TO_DEVICE);
1355         }
1356 chunk_list_dealloc:
1357         for (i = 0; i < chunk_list_size; i++)
1358                 kfree(chunk_list->chunks[i].buf);
1359
1360         kfree(chunk_list->chunks);
1361         return -ENOMEM;
1362 }
1363
1364 static void pbl_chunk_list_destroy(struct efa_dev *dev, struct pbl_context *pbl)
1365 {
1366         struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list;
1367         int i;
1368
1369         for (i = 0; i < chunk_list->size; i++) {
1370                 dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr,
1371                                  chunk_list->chunks[i].length, DMA_TO_DEVICE);
1372                 kfree(chunk_list->chunks[i].buf);
1373         }
1374
1375         kfree(chunk_list->chunks);
1376 }
1377
1378 /* initialize pbl continuous mode: map pbl buffer to a dma address. */
1379 static int pbl_continuous_initialize(struct efa_dev *dev,
1380                                      struct pbl_context *pbl)
1381 {
1382         dma_addr_t dma_addr;
1383
1384         dma_addr = dma_map_single(&dev->pdev->dev, pbl->pbl_buf,
1385                                   pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE);
1386         if (dma_mapping_error(&dev->pdev->dev, dma_addr)) {
1387                 ibdev_err(&dev->ibdev, "Unable to map pbl to DMA address\n");
1388                 return -ENOMEM;
1389         }
1390
1391         pbl->phys.continuous.dma_addr = dma_addr;
1392         ibdev_dbg(&dev->ibdev,
1393                   "pbl continuous - dma_addr = %pad, size[%u]\n",
1394                   &dma_addr, pbl->pbl_buf_size_in_bytes);
1395
1396         return 0;
1397 }
1398
1399 /*
1400  * initialize pbl indirect mode:
1401  * create a chunk list out of the dma addresses of the physical pages of
1402  * pbl buffer.
1403  */
1404 static int pbl_indirect_initialize(struct efa_dev *dev, struct pbl_context *pbl)
1405 {
1406         u32 size_in_pages = DIV_ROUND_UP(pbl->pbl_buf_size_in_bytes, PAGE_SIZE);
1407         struct scatterlist *sgl;
1408         int sg_dma_cnt, err;
1409
1410         BUILD_BUG_ON(EFA_CHUNK_PAYLOAD_SIZE > PAGE_SIZE);
1411         sgl = efa_vmalloc_buf_to_sg(pbl->pbl_buf, size_in_pages);
1412         if (!sgl)
1413                 return -ENOMEM;
1414
1415         sg_dma_cnt = dma_map_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE);
1416         if (!sg_dma_cnt) {
1417                 err = -EINVAL;
1418                 goto err_map;
1419         }
1420
1421         pbl->phys.indirect.pbl_buf_size_in_pages = size_in_pages;
1422         pbl->phys.indirect.sgl = sgl;
1423         pbl->phys.indirect.sg_dma_cnt = sg_dma_cnt;
1424         err = pbl_chunk_list_create(dev, pbl);
1425         if (err) {
1426                 ibdev_dbg(&dev->ibdev,
1427                           "chunk_list creation failed[%d]\n", err);
1428                 goto err_chunk;
1429         }
1430
1431         ibdev_dbg(&dev->ibdev,
1432                   "pbl indirect - size[%u], chunks[%u]\n",
1433                   pbl->pbl_buf_size_in_bytes,
1434                   pbl->phys.indirect.chunk_list.size);
1435
1436         return 0;
1437
1438 err_chunk:
1439         dma_unmap_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE);
1440 err_map:
1441         kfree(sgl);
1442         return err;
1443 }
1444
1445 static void pbl_indirect_terminate(struct efa_dev *dev, struct pbl_context *pbl)
1446 {
1447         pbl_chunk_list_destroy(dev, pbl);
1448         dma_unmap_sg(&dev->pdev->dev, pbl->phys.indirect.sgl,
1449                      pbl->phys.indirect.pbl_buf_size_in_pages, DMA_TO_DEVICE);
1450         kfree(pbl->phys.indirect.sgl);
1451 }
1452
1453 /* create a page buffer list from a mapped user memory region */
1454 static int pbl_create(struct efa_dev *dev,
1455                       struct pbl_context *pbl,
1456                       struct ib_umem *umem,
1457                       int hp_cnt,
1458                       u8 hp_shift)
1459 {
1460         int err;
1461
1462         pbl->pbl_buf_size_in_bytes = hp_cnt * EFA_CHUNK_PAYLOAD_PTR_SIZE;
1463         pbl->pbl_buf = kvzalloc(pbl->pbl_buf_size_in_bytes, GFP_KERNEL);
1464         if (!pbl->pbl_buf)
1465                 return -ENOMEM;
1466
1467         if (is_vmalloc_addr(pbl->pbl_buf)) {
1468                 pbl->physically_continuous = 0;
1469                 err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt,
1470                                         hp_shift);
1471                 if (err)
1472                         goto err_free;
1473
1474                 err = pbl_indirect_initialize(dev, pbl);
1475                 if (err)
1476                         goto err_free;
1477         } else {
1478                 pbl->physically_continuous = 1;
1479                 err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt,
1480                                         hp_shift);
1481                 if (err)
1482                         goto err_free;
1483
1484                 err = pbl_continuous_initialize(dev, pbl);
1485                 if (err)
1486                         goto err_free;
1487         }
1488
1489         ibdev_dbg(&dev->ibdev,
1490                   "user_pbl_created: user_pages[%u], continuous[%u]\n",
1491                   hp_cnt, pbl->physically_continuous);
1492
1493         return 0;
1494
1495 err_free:
1496         kvfree(pbl->pbl_buf);
1497         return err;
1498 }
1499
1500 static void pbl_destroy(struct efa_dev *dev, struct pbl_context *pbl)
1501 {
1502         if (pbl->physically_continuous)
1503                 dma_unmap_single(&dev->pdev->dev, pbl->phys.continuous.dma_addr,
1504                                  pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE);
1505         else
1506                 pbl_indirect_terminate(dev, pbl);
1507
1508         kvfree(pbl->pbl_buf);
1509 }
1510
1511 static int efa_create_inline_pbl(struct efa_dev *dev, struct efa_mr *mr,
1512                                  struct efa_com_reg_mr_params *params)
1513 {
1514         int err;
1515
1516         params->inline_pbl = 1;
1517         err = umem_to_page_list(dev, mr->umem, params->pbl.inline_pbl_array,
1518                                 params->page_num, params->page_shift);
1519         if (err)
1520                 return err;
1521
1522         ibdev_dbg(&dev->ibdev,
1523                   "inline_pbl_array - pages[%u]\n", params->page_num);
1524
1525         return 0;
1526 }
1527
1528 static int efa_create_pbl(struct efa_dev *dev,
1529                           struct pbl_context *pbl,
1530                           struct efa_mr *mr,
1531                           struct efa_com_reg_mr_params *params)
1532 {
1533         int err;
1534
1535         err = pbl_create(dev, pbl, mr->umem, params->page_num,
1536                          params->page_shift);
1537         if (err) {
1538                 ibdev_dbg(&dev->ibdev, "Failed to create pbl[%d]\n", err);
1539                 return err;
1540         }
1541
1542         params->inline_pbl = 0;
1543         params->indirect = !pbl->physically_continuous;
1544         if (pbl->physically_continuous) {
1545                 params->pbl.pbl.length = pbl->pbl_buf_size_in_bytes;
1546
1547                 efa_com_set_dma_addr(pbl->phys.continuous.dma_addr,
1548                                      &params->pbl.pbl.address.mem_addr_high,
1549                                      &params->pbl.pbl.address.mem_addr_low);
1550         } else {
1551                 params->pbl.pbl.length =
1552                         pbl->phys.indirect.chunk_list.chunks[0].length;
1553
1554                 efa_com_set_dma_addr(pbl->phys.indirect.chunk_list.chunks[0].dma_addr,
1555                                      &params->pbl.pbl.address.mem_addr_high,
1556                                      &params->pbl.pbl.address.mem_addr_low);
1557         }
1558
1559         return 0;
1560 }
1561
1562 static struct efa_mr *efa_alloc_mr(struct ib_pd *ibpd, int access_flags,
1563                                    struct ib_udata *udata)
1564 {
1565         struct efa_dev *dev = to_edev(ibpd->device);
1566         int supp_access_flags;
1567         struct efa_mr *mr;
1568
1569         if (udata && udata->inlen &&
1570             !ib_is_udata_cleared(udata, 0, sizeof(udata->inlen))) {
1571                 ibdev_dbg(&dev->ibdev,
1572                           "Incompatible ABI params, udata not cleared\n");
1573                 return ERR_PTR(-EINVAL);
1574         }
1575
1576         supp_access_flags =
1577                 IB_ACCESS_LOCAL_WRITE |
1578                 (EFA_DEV_CAP(dev, RDMA_READ) ? IB_ACCESS_REMOTE_READ : 0) |
1579                 (EFA_DEV_CAP(dev, RDMA_WRITE) ? IB_ACCESS_REMOTE_WRITE : 0);
1580
1581         access_flags &= ~IB_ACCESS_OPTIONAL;
1582         if (access_flags & ~supp_access_flags) {
1583                 ibdev_dbg(&dev->ibdev,
1584                           "Unsupported access flags[%#x], supported[%#x]\n",
1585                           access_flags, supp_access_flags);
1586                 return ERR_PTR(-EOPNOTSUPP);
1587         }
1588
1589         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1590         if (!mr)
1591                 return ERR_PTR(-ENOMEM);
1592
1593         return mr;
1594 }
1595
1596 static int efa_register_mr(struct ib_pd *ibpd, struct efa_mr *mr, u64 start,
1597                            u64 length, u64 virt_addr, int access_flags)
1598 {
1599         struct efa_dev *dev = to_edev(ibpd->device);
1600         struct efa_com_reg_mr_params params = {};
1601         struct efa_com_reg_mr_result result = {};
1602         struct pbl_context pbl;
1603         unsigned int pg_sz;
1604         int inline_size;
1605         int err;
1606
1607         params.pd = to_epd(ibpd)->pdn;
1608         params.iova = virt_addr;
1609         params.mr_length_in_bytes = length;
1610         params.permissions = access_flags;
1611
1612         pg_sz = ib_umem_find_best_pgsz(mr->umem,
1613                                        dev->dev_attr.page_size_cap,
1614                                        virt_addr);
1615         if (!pg_sz) {
1616                 ibdev_dbg(&dev->ibdev, "Failed to find a suitable page size in page_size_cap %#llx\n",
1617                           dev->dev_attr.page_size_cap);
1618                 return -EOPNOTSUPP;
1619         }
1620
1621         params.page_shift = order_base_2(pg_sz);
1622         params.page_num = ib_umem_num_dma_blocks(mr->umem, pg_sz);
1623
1624         ibdev_dbg(&dev->ibdev,
1625                   "start %#llx length %#llx params.page_shift %u params.page_num %u\n",
1626                   start, length, params.page_shift, params.page_num);
1627
1628         inline_size = ARRAY_SIZE(params.pbl.inline_pbl_array);
1629         if (params.page_num <= inline_size) {
1630                 err = efa_create_inline_pbl(dev, mr, &params);
1631                 if (err)
1632                         return err;
1633
1634                 err = efa_com_register_mr(&dev->edev, &params, &result);
1635                 if (err)
1636                         return err;
1637         } else {
1638                 err = efa_create_pbl(dev, &pbl, mr, &params);
1639                 if (err)
1640                         return err;
1641
1642                 err = efa_com_register_mr(&dev->edev, &params, &result);
1643                 pbl_destroy(dev, &pbl);
1644
1645                 if (err)
1646                         return err;
1647         }
1648
1649         mr->ibmr.lkey = result.l_key;
1650         mr->ibmr.rkey = result.r_key;
1651         mr->ibmr.length = length;
1652         ibdev_dbg(&dev->ibdev, "Registered mr[%d]\n", mr->ibmr.lkey);
1653
1654         return 0;
1655 }
1656
1657 struct ib_mr *efa_reg_user_mr_dmabuf(struct ib_pd *ibpd, u64 start,
1658                                      u64 length, u64 virt_addr,
1659                                      int fd, int access_flags,
1660                                      struct ib_udata *udata)
1661 {
1662         struct efa_dev *dev = to_edev(ibpd->device);
1663         struct ib_umem_dmabuf *umem_dmabuf;
1664         struct efa_mr *mr;
1665         int err;
1666
1667         mr = efa_alloc_mr(ibpd, access_flags, udata);
1668         if (IS_ERR(mr)) {
1669                 err = PTR_ERR(mr);
1670                 goto err_out;
1671         }
1672
1673         umem_dmabuf = ib_umem_dmabuf_get_pinned(ibpd->device, start, length, fd,
1674                                                 access_flags);
1675         if (IS_ERR(umem_dmabuf)) {
1676                 err = PTR_ERR(umem_dmabuf);
1677                 ibdev_dbg(&dev->ibdev, "Failed to get dmabuf umem[%d]\n", err);
1678                 goto err_free;
1679         }
1680
1681         mr->umem = &umem_dmabuf->umem;
1682         err = efa_register_mr(ibpd, mr, start, length, virt_addr, access_flags);
1683         if (err)
1684                 goto err_release;
1685
1686         return &mr->ibmr;
1687
1688 err_release:
1689         ib_umem_release(mr->umem);
1690 err_free:
1691         kfree(mr);
1692 err_out:
1693         atomic64_inc(&dev->stats.reg_mr_err);
1694         return ERR_PTR(err);
1695 }
1696
1697 struct ib_mr *efa_reg_mr(struct ib_pd *ibpd, u64 start, u64 length,
1698                          u64 virt_addr, int access_flags,
1699                          struct ib_udata *udata)
1700 {
1701         struct efa_dev *dev = to_edev(ibpd->device);
1702         struct efa_mr *mr;
1703         int err;
1704
1705         mr = efa_alloc_mr(ibpd, access_flags, udata);
1706         if (IS_ERR(mr)) {
1707                 err = PTR_ERR(mr);
1708                 goto err_out;
1709         }
1710
1711         mr->umem = ib_umem_get(ibpd->device, start, length, access_flags);
1712         if (IS_ERR(mr->umem)) {
1713                 err = PTR_ERR(mr->umem);
1714                 ibdev_dbg(&dev->ibdev,
1715                           "Failed to pin and map user space memory[%d]\n", err);
1716                 goto err_free;
1717         }
1718
1719         err = efa_register_mr(ibpd, mr, start, length, virt_addr, access_flags);
1720         if (err)
1721                 goto err_release;
1722
1723         return &mr->ibmr;
1724
1725 err_release:
1726         ib_umem_release(mr->umem);
1727 err_free:
1728         kfree(mr);
1729 err_out:
1730         atomic64_inc(&dev->stats.reg_mr_err);
1731         return ERR_PTR(err);
1732 }
1733
1734 int efa_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
1735 {
1736         struct efa_dev *dev = to_edev(ibmr->device);
1737         struct efa_com_dereg_mr_params params;
1738         struct efa_mr *mr = to_emr(ibmr);
1739         int err;
1740
1741         ibdev_dbg(&dev->ibdev, "Deregister mr[%d]\n", ibmr->lkey);
1742
1743         params.l_key = mr->ibmr.lkey;
1744         err = efa_com_dereg_mr(&dev->edev, &params);
1745         if (err)
1746                 return err;
1747
1748         ib_umem_release(mr->umem);
1749         kfree(mr);
1750
1751         return 0;
1752 }
1753
1754 int efa_get_port_immutable(struct ib_device *ibdev, u32 port_num,
1755                            struct ib_port_immutable *immutable)
1756 {
1757         struct ib_port_attr attr;
1758         int err;
1759
1760         err = ib_query_port(ibdev, port_num, &attr);
1761         if (err) {
1762                 ibdev_dbg(ibdev, "Couldn't query port err[%d]\n", err);
1763                 return err;
1764         }
1765
1766         immutable->pkey_tbl_len = attr.pkey_tbl_len;
1767         immutable->gid_tbl_len = attr.gid_tbl_len;
1768
1769         return 0;
1770 }
1771
1772 static int efa_dealloc_uar(struct efa_dev *dev, u16 uarn)
1773 {
1774         struct efa_com_dealloc_uar_params params = {
1775                 .uarn = uarn,
1776         };
1777
1778         return efa_com_dealloc_uar(&dev->edev, &params);
1779 }
1780
1781 #define EFA_CHECK_USER_COMP(_dev, _comp_mask, _attr, _mask, _attr_str) \
1782         (_attr_str = (!(_dev)->dev_attr._attr || ((_comp_mask) & (_mask))) ? \
1783                      NULL : #_attr)
1784
1785 static int efa_user_comp_handshake(const struct ib_ucontext *ibucontext,
1786                                    const struct efa_ibv_alloc_ucontext_cmd *cmd)
1787 {
1788         struct efa_dev *dev = to_edev(ibucontext->device);
1789         char *attr_str;
1790
1791         if (EFA_CHECK_USER_COMP(dev, cmd->comp_mask, max_tx_batch,
1792                                 EFA_ALLOC_UCONTEXT_CMD_COMP_TX_BATCH, attr_str))
1793                 goto err;
1794
1795         if (EFA_CHECK_USER_COMP(dev, cmd->comp_mask, min_sq_depth,
1796                                 EFA_ALLOC_UCONTEXT_CMD_COMP_MIN_SQ_WR,
1797                                 attr_str))
1798                 goto err;
1799
1800         return 0;
1801
1802 err:
1803         ibdev_dbg(&dev->ibdev, "Userspace handshake failed for %s attribute\n",
1804                   attr_str);
1805         return -EOPNOTSUPP;
1806 }
1807
1808 int efa_alloc_ucontext(struct ib_ucontext *ibucontext, struct ib_udata *udata)
1809 {
1810         struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1811         struct efa_dev *dev = to_edev(ibucontext->device);
1812         struct efa_ibv_alloc_ucontext_resp resp = {};
1813         struct efa_ibv_alloc_ucontext_cmd cmd = {};
1814         struct efa_com_alloc_uar_result result;
1815         int err;
1816
1817         /*
1818          * it's fine if the driver does not know all request fields,
1819          * we will ack input fields in our response.
1820          */
1821
1822         err = ib_copy_from_udata(&cmd, udata,
1823                                  min(sizeof(cmd), udata->inlen));
1824         if (err) {
1825                 ibdev_dbg(&dev->ibdev,
1826                           "Cannot copy udata for alloc_ucontext\n");
1827                 goto err_out;
1828         }
1829
1830         err = efa_user_comp_handshake(ibucontext, &cmd);
1831         if (err)
1832                 goto err_out;
1833
1834         err = efa_com_alloc_uar(&dev->edev, &result);
1835         if (err)
1836                 goto err_out;
1837
1838         ucontext->uarn = result.uarn;
1839
1840         resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_QUERY_DEVICE;
1841         resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_CREATE_AH;
1842         resp.sub_cqs_per_cq = dev->dev_attr.sub_cqs_per_cq;
1843         resp.inline_buf_size = dev->dev_attr.inline_buf_size;
1844         resp.max_llq_size = dev->dev_attr.max_llq_size;
1845         resp.max_tx_batch = dev->dev_attr.max_tx_batch;
1846         resp.min_sq_wr = dev->dev_attr.min_sq_depth;
1847
1848         err = ib_copy_to_udata(udata, &resp,
1849                                min(sizeof(resp), udata->outlen));
1850         if (err)
1851                 goto err_dealloc_uar;
1852
1853         return 0;
1854
1855 err_dealloc_uar:
1856         efa_dealloc_uar(dev, result.uarn);
1857 err_out:
1858         atomic64_inc(&dev->stats.alloc_ucontext_err);
1859         return err;
1860 }
1861
1862 void efa_dealloc_ucontext(struct ib_ucontext *ibucontext)
1863 {
1864         struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1865         struct efa_dev *dev = to_edev(ibucontext->device);
1866
1867         efa_dealloc_uar(dev, ucontext->uarn);
1868 }
1869
1870 void efa_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
1871 {
1872         struct efa_user_mmap_entry *entry = to_emmap(rdma_entry);
1873
1874         kfree(entry);
1875 }
1876
1877 static int __efa_mmap(struct efa_dev *dev, struct efa_ucontext *ucontext,
1878                       struct vm_area_struct *vma)
1879 {
1880         struct rdma_user_mmap_entry *rdma_entry;
1881         struct efa_user_mmap_entry *entry;
1882         unsigned long va;
1883         int err = 0;
1884         u64 pfn;
1885
1886         rdma_entry = rdma_user_mmap_entry_get(&ucontext->ibucontext, vma);
1887         if (!rdma_entry) {
1888                 ibdev_dbg(&dev->ibdev,
1889                           "pgoff[%#lx] does not have valid entry\n",
1890                           vma->vm_pgoff);
1891                 atomic64_inc(&dev->stats.mmap_err);
1892                 return -EINVAL;
1893         }
1894         entry = to_emmap(rdma_entry);
1895
1896         ibdev_dbg(&dev->ibdev,
1897                   "Mapping address[%#llx], length[%#zx], mmap_flag[%d]\n",
1898                   entry->address, rdma_entry->npages * PAGE_SIZE,
1899                   entry->mmap_flag);
1900
1901         pfn = entry->address >> PAGE_SHIFT;
1902         switch (entry->mmap_flag) {
1903         case EFA_MMAP_IO_NC:
1904                 err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn,
1905                                         entry->rdma_entry.npages * PAGE_SIZE,
1906                                         pgprot_noncached(vma->vm_page_prot),
1907                                         rdma_entry);
1908                 break;
1909         case EFA_MMAP_IO_WC:
1910                 err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn,
1911                                         entry->rdma_entry.npages * PAGE_SIZE,
1912                                         pgprot_writecombine(vma->vm_page_prot),
1913                                         rdma_entry);
1914                 break;
1915         case EFA_MMAP_DMA_PAGE:
1916                 for (va = vma->vm_start; va < vma->vm_end;
1917                      va += PAGE_SIZE, pfn++) {
1918                         err = vm_insert_page(vma, va, pfn_to_page(pfn));
1919                         if (err)
1920                                 break;
1921                 }
1922                 break;
1923         default:
1924                 err = -EINVAL;
1925         }
1926
1927         if (err) {
1928                 ibdev_dbg(
1929                         &dev->ibdev,
1930                         "Couldn't mmap address[%#llx] length[%#zx] mmap_flag[%d] err[%d]\n",
1931                         entry->address, rdma_entry->npages * PAGE_SIZE,
1932                         entry->mmap_flag, err);
1933                 atomic64_inc(&dev->stats.mmap_err);
1934         }
1935
1936         rdma_user_mmap_entry_put(rdma_entry);
1937         return err;
1938 }
1939
1940 int efa_mmap(struct ib_ucontext *ibucontext,
1941              struct vm_area_struct *vma)
1942 {
1943         struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1944         struct efa_dev *dev = to_edev(ibucontext->device);
1945         size_t length = vma->vm_end - vma->vm_start;
1946
1947         ibdev_dbg(&dev->ibdev,
1948                   "start %#lx, end %#lx, length = %#zx, pgoff = %#lx\n",
1949                   vma->vm_start, vma->vm_end, length, vma->vm_pgoff);
1950
1951         return __efa_mmap(dev, ucontext, vma);
1952 }
1953
1954 static int efa_ah_destroy(struct efa_dev *dev, struct efa_ah *ah)
1955 {
1956         struct efa_com_destroy_ah_params params = {
1957                 .ah = ah->ah,
1958                 .pdn = to_epd(ah->ibah.pd)->pdn,
1959         };
1960
1961         return efa_com_destroy_ah(&dev->edev, &params);
1962 }
1963
1964 int efa_create_ah(struct ib_ah *ibah,
1965                   struct rdma_ah_init_attr *init_attr,
1966                   struct ib_udata *udata)
1967 {
1968         struct rdma_ah_attr *ah_attr = init_attr->ah_attr;
1969         struct efa_dev *dev = to_edev(ibah->device);
1970         struct efa_com_create_ah_params params = {};
1971         struct efa_ibv_create_ah_resp resp = {};
1972         struct efa_com_create_ah_result result;
1973         struct efa_ah *ah = to_eah(ibah);
1974         int err;
1975
1976         if (!(init_attr->flags & RDMA_CREATE_AH_SLEEPABLE)) {
1977                 ibdev_dbg(&dev->ibdev,
1978                           "Create address handle is not supported in atomic context\n");
1979                 err = -EOPNOTSUPP;
1980                 goto err_out;
1981         }
1982
1983         if (udata->inlen &&
1984             !ib_is_udata_cleared(udata, 0, udata->inlen)) {
1985                 ibdev_dbg(&dev->ibdev, "Incompatible ABI params\n");
1986                 err = -EINVAL;
1987                 goto err_out;
1988         }
1989
1990         memcpy(params.dest_addr, ah_attr->grh.dgid.raw,
1991                sizeof(params.dest_addr));
1992         params.pdn = to_epd(ibah->pd)->pdn;
1993         err = efa_com_create_ah(&dev->edev, &params, &result);
1994         if (err)
1995                 goto err_out;
1996
1997         memcpy(ah->id, ah_attr->grh.dgid.raw, sizeof(ah->id));
1998         ah->ah = result.ah;
1999
2000         resp.efa_address_handle = result.ah;
2001
2002         if (udata->outlen) {
2003                 err = ib_copy_to_udata(udata, &resp,
2004                                        min(sizeof(resp), udata->outlen));
2005                 if (err) {
2006                         ibdev_dbg(&dev->ibdev,
2007                                   "Failed to copy udata for create_ah response\n");
2008                         goto err_destroy_ah;
2009                 }
2010         }
2011         ibdev_dbg(&dev->ibdev, "Created ah[%d]\n", ah->ah);
2012
2013         return 0;
2014
2015 err_destroy_ah:
2016         efa_ah_destroy(dev, ah);
2017 err_out:
2018         atomic64_inc(&dev->stats.create_ah_err);
2019         return err;
2020 }
2021
2022 int efa_destroy_ah(struct ib_ah *ibah, u32 flags)
2023 {
2024         struct efa_dev *dev = to_edev(ibah->pd->device);
2025         struct efa_ah *ah = to_eah(ibah);
2026
2027         ibdev_dbg(&dev->ibdev, "Destroy ah[%d]\n", ah->ah);
2028
2029         if (!(flags & RDMA_DESTROY_AH_SLEEPABLE)) {
2030                 ibdev_dbg(&dev->ibdev,
2031                           "Destroy address handle is not supported in atomic context\n");
2032                 return -EOPNOTSUPP;
2033         }
2034
2035         efa_ah_destroy(dev, ah);
2036         return 0;
2037 }
2038
2039 struct rdma_hw_stats *efa_alloc_hw_port_stats(struct ib_device *ibdev,
2040                                               u32 port_num)
2041 {
2042         return rdma_alloc_hw_stats_struct(efa_port_stats_descs,
2043                                           ARRAY_SIZE(efa_port_stats_descs),
2044                                           RDMA_HW_STATS_DEFAULT_LIFESPAN);
2045 }
2046
2047 struct rdma_hw_stats *efa_alloc_hw_device_stats(struct ib_device *ibdev)
2048 {
2049         return rdma_alloc_hw_stats_struct(efa_device_stats_descs,
2050                                           ARRAY_SIZE(efa_device_stats_descs),
2051                                           RDMA_HW_STATS_DEFAULT_LIFESPAN);
2052 }
2053
2054 static int efa_fill_device_stats(struct efa_dev *dev,
2055                                  struct rdma_hw_stats *stats)
2056 {
2057         struct efa_com_stats_admin *as = &dev->edev.aq.stats;
2058         struct efa_stats *s = &dev->stats;
2059
2060         stats->value[EFA_SUBMITTED_CMDS] = atomic64_read(&as->submitted_cmd);
2061         stats->value[EFA_COMPLETED_CMDS] = atomic64_read(&as->completed_cmd);
2062         stats->value[EFA_CMDS_ERR] = atomic64_read(&as->cmd_err);
2063         stats->value[EFA_NO_COMPLETION_CMDS] = atomic64_read(&as->no_completion);
2064
2065         stats->value[EFA_KEEP_ALIVE_RCVD] = atomic64_read(&s->keep_alive_rcvd);
2066         stats->value[EFA_ALLOC_PD_ERR] = atomic64_read(&s->alloc_pd_err);
2067         stats->value[EFA_CREATE_QP_ERR] = atomic64_read(&s->create_qp_err);
2068         stats->value[EFA_CREATE_CQ_ERR] = atomic64_read(&s->create_cq_err);
2069         stats->value[EFA_REG_MR_ERR] = atomic64_read(&s->reg_mr_err);
2070         stats->value[EFA_ALLOC_UCONTEXT_ERR] =
2071                 atomic64_read(&s->alloc_ucontext_err);
2072         stats->value[EFA_CREATE_AH_ERR] = atomic64_read(&s->create_ah_err);
2073         stats->value[EFA_MMAP_ERR] = atomic64_read(&s->mmap_err);
2074
2075         return ARRAY_SIZE(efa_device_stats_descs);
2076 }
2077
2078 static int efa_fill_port_stats(struct efa_dev *dev, struct rdma_hw_stats *stats,
2079                                u32 port_num)
2080 {
2081         struct efa_com_get_stats_params params = {};
2082         union efa_com_get_stats_result result;
2083         struct efa_com_rdma_read_stats *rrs;
2084         struct efa_com_messages_stats *ms;
2085         struct efa_com_basic_stats *bs;
2086         int err;
2087
2088         params.scope = EFA_ADMIN_GET_STATS_SCOPE_ALL;
2089         params.type = EFA_ADMIN_GET_STATS_TYPE_BASIC;
2090
2091         err = efa_com_get_stats(&dev->edev, &params, &result);
2092         if (err)
2093                 return err;
2094
2095         bs = &result.basic_stats;
2096         stats->value[EFA_TX_BYTES] = bs->tx_bytes;
2097         stats->value[EFA_TX_PKTS] = bs->tx_pkts;
2098         stats->value[EFA_RX_BYTES] = bs->rx_bytes;
2099         stats->value[EFA_RX_PKTS] = bs->rx_pkts;
2100         stats->value[EFA_RX_DROPS] = bs->rx_drops;
2101
2102         params.type = EFA_ADMIN_GET_STATS_TYPE_MESSAGES;
2103         err = efa_com_get_stats(&dev->edev, &params, &result);
2104         if (err)
2105                 return err;
2106
2107         ms = &result.messages_stats;
2108         stats->value[EFA_SEND_BYTES] = ms->send_bytes;
2109         stats->value[EFA_SEND_WRS] = ms->send_wrs;
2110         stats->value[EFA_RECV_BYTES] = ms->recv_bytes;
2111         stats->value[EFA_RECV_WRS] = ms->recv_wrs;
2112
2113         params.type = EFA_ADMIN_GET_STATS_TYPE_RDMA_READ;
2114         err = efa_com_get_stats(&dev->edev, &params, &result);
2115         if (err)
2116                 return err;
2117
2118         rrs = &result.rdma_read_stats;
2119         stats->value[EFA_RDMA_READ_WRS] = rrs->read_wrs;
2120         stats->value[EFA_RDMA_READ_BYTES] = rrs->read_bytes;
2121         stats->value[EFA_RDMA_READ_WR_ERR] = rrs->read_wr_err;
2122         stats->value[EFA_RDMA_READ_RESP_BYTES] = rrs->read_resp_bytes;
2123
2124         return ARRAY_SIZE(efa_port_stats_descs);
2125 }
2126
2127 int efa_get_hw_stats(struct ib_device *ibdev, struct rdma_hw_stats *stats,
2128                      u32 port_num, int index)
2129 {
2130         if (port_num)
2131                 return efa_fill_port_stats(to_edev(ibdev), stats, port_num);
2132         else
2133                 return efa_fill_device_stats(to_edev(ibdev), stats);
2134 }
2135
2136 enum rdma_link_layer efa_port_link_layer(struct ib_device *ibdev,
2137                                          u32 port_num)
2138 {
2139         return IB_LINK_LAYER_UNSPECIFIED;
2140 }
2141