1// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB 2/* 3 * Copyright 2018-2020 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 17enum { 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 27struct 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_STATS(op) \ 34 op(EFA_TX_BYTES, "tx_bytes") \ 35 op(EFA_TX_PKTS, "tx_pkts") \ 36 op(EFA_RX_BYTES, "rx_bytes") \ 37 op(EFA_RX_PKTS, "rx_pkts") \ 38 op(EFA_RX_DROPS, "rx_drops") \ 39 op(EFA_SEND_BYTES, "send_bytes") \ 40 op(EFA_SEND_WRS, "send_wrs") \ 41 op(EFA_RECV_BYTES, "recv_bytes") \ 42 op(EFA_RECV_WRS, "recv_wrs") \ 43 op(EFA_RDMA_READ_WRS, "rdma_read_wrs") \ 44 op(EFA_RDMA_READ_BYTES, "rdma_read_bytes") \ 45 op(EFA_RDMA_READ_WR_ERR, "rdma_read_wr_err") \ 46 op(EFA_RDMA_READ_RESP_BYTES, "rdma_read_resp_bytes") \ 47 op(EFA_SUBMITTED_CMDS, "submitted_cmds") \ 48 op(EFA_COMPLETED_CMDS, "completed_cmds") \ 49 op(EFA_CMDS_ERR, "cmds_err") \ 50 op(EFA_NO_COMPLETION_CMDS, "no_completion_cmds") \ 51 op(EFA_KEEP_ALIVE_RCVD, "keep_alive_rcvd") \ 52 op(EFA_ALLOC_PD_ERR, "alloc_pd_err") \ 53 op(EFA_CREATE_QP_ERR, "create_qp_err") \ 54 op(EFA_CREATE_CQ_ERR, "create_cq_err") \ 55 op(EFA_REG_MR_ERR, "reg_mr_err") \ 56 op(EFA_ALLOC_UCONTEXT_ERR, "alloc_ucontext_err") \ 57 op(EFA_CREATE_AH_ERR, "create_ah_err") \ 58 op(EFA_MMAP_ERR, "mmap_err") 59 60#define EFA_STATS_ENUM(ename, name) ename, 61#define EFA_STATS_STR(ename, name) [ename] = name, 62 63enum efa_hw_stats { 64 EFA_DEFINE_STATS(EFA_STATS_ENUM) 65}; 66 67static const char *const efa_stats_names[] = { 68 EFA_DEFINE_STATS(EFA_STATS_STR) 69}; 70 71#define EFA_CHUNK_PAYLOAD_SHIFT 12 72#define EFA_CHUNK_PAYLOAD_SIZE BIT(EFA_CHUNK_PAYLOAD_SHIFT) 73#define EFA_CHUNK_PAYLOAD_PTR_SIZE 8 74 75#define EFA_CHUNK_SHIFT 12 76#define EFA_CHUNK_SIZE BIT(EFA_CHUNK_SHIFT) 77#define EFA_CHUNK_PTR_SIZE sizeof(struct efa_com_ctrl_buff_info) 78 79#define EFA_PTRS_PER_CHUNK \ 80 ((EFA_CHUNK_SIZE - EFA_CHUNK_PTR_SIZE) / EFA_CHUNK_PAYLOAD_PTR_SIZE) 81 82#define EFA_CHUNK_USED_SIZE \ 83 ((EFA_PTRS_PER_CHUNK * EFA_CHUNK_PAYLOAD_PTR_SIZE) + EFA_CHUNK_PTR_SIZE) 84 85struct pbl_chunk { 86 dma_addr_t dma_addr; 87 u64 *buf; 88 u32 length; 89}; 90 91struct pbl_chunk_list { 92 struct pbl_chunk *chunks; 93 unsigned int size; 94}; 95 96struct pbl_context { 97 union { 98 struct { 99 dma_addr_t dma_addr; 100 } continuous; 101 struct { 102 u32 pbl_buf_size_in_pages; 103 struct scatterlist *sgl; 104 int sg_dma_cnt; 105 struct pbl_chunk_list chunk_list; 106 } indirect; 107 } phys; 108 u64 *pbl_buf; 109 u32 pbl_buf_size_in_bytes; 110 u8 physically_continuous; 111}; 112 113static inline struct efa_dev *to_edev(struct ib_device *ibdev) 114{ 115 return container_of(ibdev, struct efa_dev, ibdev); 116} 117 118static inline struct efa_ucontext *to_eucontext(struct ib_ucontext *ibucontext) 119{ 120 return container_of(ibucontext, struct efa_ucontext, ibucontext); 121} 122 123static inline struct efa_pd *to_epd(struct ib_pd *ibpd) 124{ 125 return container_of(ibpd, struct efa_pd, ibpd); 126} 127 128static inline struct efa_mr *to_emr(struct ib_mr *ibmr) 129{ 130 return container_of(ibmr, struct efa_mr, ibmr); 131} 132 133static inline struct efa_qp *to_eqp(struct ib_qp *ibqp) 134{ 135 return container_of(ibqp, struct efa_qp, ibqp); 136} 137 138static inline struct efa_cq *to_ecq(struct ib_cq *ibcq) 139{ 140 return container_of(ibcq, struct efa_cq, ibcq); 141} 142 143static inline struct efa_ah *to_eah(struct ib_ah *ibah) 144{ 145 return container_of(ibah, struct efa_ah, ibah); 146} 147 148static inline struct efa_user_mmap_entry * 149to_emmap(struct rdma_user_mmap_entry *rdma_entry) 150{ 151 return container_of(rdma_entry, struct efa_user_mmap_entry, rdma_entry); 152} 153 154#define EFA_DEV_CAP(dev, cap) \ 155 ((dev)->dev_attr.device_caps & \ 156 EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_##cap##_MASK) 157 158#define is_reserved_cleared(reserved) \ 159 !memchr_inv(reserved, 0, sizeof(reserved)) 160 161static void *efa_zalloc_mapped(struct efa_dev *dev, dma_addr_t *dma_addr, 162 size_t size, enum dma_data_direction dir) 163{ 164 void *addr; 165 166 addr = alloc_pages_exact(size, GFP_KERNEL | __GFP_ZERO); 167 if (!addr) 168 return NULL; 169 170 *dma_addr = dma_map_single(&dev->pdev->dev, addr, size, dir); 171 if (dma_mapping_error(&dev->pdev->dev, *dma_addr)) { 172 ibdev_err(&dev->ibdev, "Failed to map DMA address\n"); 173 free_pages_exact(addr, size); 174 return NULL; 175 } 176 177 return addr; 178} 179 180static void efa_free_mapped(struct efa_dev *dev, void *cpu_addr, 181 dma_addr_t dma_addr, 182 size_t size, enum dma_data_direction dir) 183{ 184 dma_unmap_single(&dev->pdev->dev, dma_addr, size, dir); 185 free_pages_exact(cpu_addr, size); 186} 187 188int efa_query_device(struct ib_device *ibdev, 189 struct ib_device_attr *props, 190 struct ib_udata *udata) 191{ 192 struct efa_com_get_device_attr_result *dev_attr; 193 struct efa_ibv_ex_query_device_resp resp = {}; 194 struct efa_dev *dev = to_edev(ibdev); 195 int err; 196 197 if (udata && udata->inlen && 198 !ib_is_udata_cleared(udata, 0, udata->inlen)) { 199 ibdev_dbg(ibdev, 200 "Incompatible ABI params, udata not cleared\n"); 201 return -EINVAL; 202 } 203 204 dev_attr = &dev->dev_attr; 205 206 memset(props, 0, sizeof(*props)); 207 props->max_mr_size = dev_attr->max_mr_pages * PAGE_SIZE; 208 props->page_size_cap = dev_attr->page_size_cap; 209 props->vendor_id = dev->pdev->vendor; 210 props->vendor_part_id = dev->pdev->device; 211 props->hw_ver = dev->pdev->subsystem_device; 212 props->max_qp = dev_attr->max_qp; 213 props->max_cq = dev_attr->max_cq; 214 props->max_pd = dev_attr->max_pd; 215 props->max_mr = dev_attr->max_mr; 216 props->max_ah = dev_attr->max_ah; 217 props->max_cqe = dev_attr->max_cq_depth; 218 props->max_qp_wr = min_t(u32, dev_attr->max_sq_depth, 219 dev_attr->max_rq_depth); 220 props->max_send_sge = dev_attr->max_sq_sge; 221 props->max_recv_sge = dev_attr->max_rq_sge; 222 props->max_sge_rd = dev_attr->max_wr_rdma_sge; 223 props->max_pkeys = 1; 224 225 if (udata && udata->outlen) { 226 resp.max_sq_sge = dev_attr->max_sq_sge; 227 resp.max_rq_sge = dev_attr->max_rq_sge; 228 resp.max_sq_wr = dev_attr->max_sq_depth; 229 resp.max_rq_wr = dev_attr->max_rq_depth; 230 resp.max_rdma_size = dev_attr->max_rdma_size; 231 232 if (EFA_DEV_CAP(dev, RDMA_READ)) 233 resp.device_caps |= EFA_QUERY_DEVICE_CAPS_RDMA_READ; 234 235 if (EFA_DEV_CAP(dev, RNR_RETRY)) 236 resp.device_caps |= EFA_QUERY_DEVICE_CAPS_RNR_RETRY; 237 238 err = ib_copy_to_udata(udata, &resp, 239 min(sizeof(resp), udata->outlen)); 240 if (err) { 241 ibdev_dbg(ibdev, 242 "Failed to copy udata for query_device\n"); 243 return err; 244 } 245 } 246 247 return 0; 248} 249 250int efa_query_port(struct ib_device *ibdev, u8 port, 251 struct ib_port_attr *props) 252{ 253 struct efa_dev *dev = to_edev(ibdev); 254 255 props->lmc = 1; 256 257 props->state = IB_PORT_ACTIVE; 258 props->phys_state = IB_PORT_PHYS_STATE_LINK_UP; 259 props->gid_tbl_len = 1; 260 props->pkey_tbl_len = 1; 261 props->active_speed = IB_SPEED_EDR; 262 props->active_width = IB_WIDTH_4X; 263 props->max_mtu = ib_mtu_int_to_enum(dev->dev_attr.mtu); 264 props->active_mtu = ib_mtu_int_to_enum(dev->dev_attr.mtu); 265 props->max_msg_sz = dev->dev_attr.mtu; 266 props->max_vl_num = 1; 267 268 return 0; 269} 270 271int efa_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, 272 int qp_attr_mask, 273 struct ib_qp_init_attr *qp_init_attr) 274{ 275 struct efa_dev *dev = to_edev(ibqp->device); 276 struct efa_com_query_qp_params params = {}; 277 struct efa_com_query_qp_result result; 278 struct efa_qp *qp = to_eqp(ibqp); 279 int err; 280 281#define EFA_QUERY_QP_SUPP_MASK \ 282 (IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT | \ 283 IB_QP_QKEY | IB_QP_SQ_PSN | IB_QP_CAP | IB_QP_RNR_RETRY) 284 285 if (qp_attr_mask & ~EFA_QUERY_QP_SUPP_MASK) { 286 ibdev_dbg(&dev->ibdev, 287 "Unsupported qp_attr_mask[%#x] supported[%#x]\n", 288 qp_attr_mask, EFA_QUERY_QP_SUPP_MASK); 289 return -EOPNOTSUPP; 290 } 291 292 memset(qp_attr, 0, sizeof(*qp_attr)); 293 memset(qp_init_attr, 0, sizeof(*qp_init_attr)); 294 295 params.qp_handle = qp->qp_handle; 296 err = efa_com_query_qp(&dev->edev, ¶ms, &result); 297 if (err) 298 return err; 299 300 qp_attr->qp_state = result.qp_state; 301 qp_attr->qkey = result.qkey; 302 qp_attr->sq_psn = result.sq_psn; 303 qp_attr->sq_draining = result.sq_draining; 304 qp_attr->port_num = 1; 305 qp_attr->rnr_retry = result.rnr_retry; 306 307 qp_attr->cap.max_send_wr = qp->max_send_wr; 308 qp_attr->cap.max_recv_wr = qp->max_recv_wr; 309 qp_attr->cap.max_send_sge = qp->max_send_sge; 310 qp_attr->cap.max_recv_sge = qp->max_recv_sge; 311 qp_attr->cap.max_inline_data = qp->max_inline_data; 312 313 qp_init_attr->qp_type = ibqp->qp_type; 314 qp_init_attr->recv_cq = ibqp->recv_cq; 315 qp_init_attr->send_cq = ibqp->send_cq; 316 qp_init_attr->qp_context = ibqp->qp_context; 317 qp_init_attr->cap = qp_attr->cap; 318 319 return 0; 320} 321 322int efa_query_gid(struct ib_device *ibdev, u8 port, int index, 323 union ib_gid *gid) 324{ 325 struct efa_dev *dev = to_edev(ibdev); 326 327 memcpy(gid->raw, dev->dev_attr.addr, sizeof(dev->dev_attr.addr)); 328 329 return 0; 330} 331 332int efa_query_pkey(struct ib_device *ibdev, u8 port, u16 index, 333 u16 *pkey) 334{ 335 if (index > 0) 336 return -EINVAL; 337 338 *pkey = 0xffff; 339 return 0; 340} 341 342static int efa_pd_dealloc(struct efa_dev *dev, u16 pdn) 343{ 344 struct efa_com_dealloc_pd_params params = { 345 .pdn = pdn, 346 }; 347 348 return efa_com_dealloc_pd(&dev->edev, ¶ms); 349} 350 351int efa_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata) 352{ 353 struct efa_dev *dev = to_edev(ibpd->device); 354 struct efa_ibv_alloc_pd_resp resp = {}; 355 struct efa_com_alloc_pd_result result; 356 struct efa_pd *pd = to_epd(ibpd); 357 int err; 358 359 if (udata->inlen && 360 !ib_is_udata_cleared(udata, 0, udata->inlen)) { 361 ibdev_dbg(&dev->ibdev, 362 "Incompatible ABI params, udata not cleared\n"); 363 err = -EINVAL; 364 goto err_out; 365 } 366 367 err = efa_com_alloc_pd(&dev->edev, &result); 368 if (err) 369 goto err_out; 370 371 pd->pdn = result.pdn; 372 resp.pdn = result.pdn; 373 374 if (udata->outlen) { 375 err = ib_copy_to_udata(udata, &resp, 376 min(sizeof(resp), udata->outlen)); 377 if (err) { 378 ibdev_dbg(&dev->ibdev, 379 "Failed to copy udata for alloc_pd\n"); 380 goto err_dealloc_pd; 381 } 382 } 383 384 ibdev_dbg(&dev->ibdev, "Allocated pd[%d]\n", pd->pdn); 385 386 return 0; 387 388err_dealloc_pd: 389 efa_pd_dealloc(dev, result.pdn); 390err_out: 391 atomic64_inc(&dev->stats.alloc_pd_err); 392 return err; 393} 394 395int efa_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata) 396{ 397 struct efa_dev *dev = to_edev(ibpd->device); 398 struct efa_pd *pd = to_epd(ibpd); 399 400 ibdev_dbg(&dev->ibdev, "Dealloc pd[%d]\n", pd->pdn); 401 efa_pd_dealloc(dev, pd->pdn); 402 return 0; 403} 404 405static int efa_destroy_qp_handle(struct efa_dev *dev, u32 qp_handle) 406{ 407 struct efa_com_destroy_qp_params params = { .qp_handle = qp_handle }; 408 409 return efa_com_destroy_qp(&dev->edev, ¶ms); 410} 411 412static void efa_qp_user_mmap_entries_remove(struct efa_qp *qp) 413{ 414 rdma_user_mmap_entry_remove(qp->rq_mmap_entry); 415 rdma_user_mmap_entry_remove(qp->rq_db_mmap_entry); 416 rdma_user_mmap_entry_remove(qp->llq_desc_mmap_entry); 417 rdma_user_mmap_entry_remove(qp->sq_db_mmap_entry); 418} 419 420int efa_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata) 421{ 422 struct efa_dev *dev = to_edev(ibqp->pd->device); 423 struct efa_qp *qp = to_eqp(ibqp); 424 int err; 425 426 ibdev_dbg(&dev->ibdev, "Destroy qp[%u]\n", ibqp->qp_num); 427 428 efa_qp_user_mmap_entries_remove(qp); 429 430 err = efa_destroy_qp_handle(dev, qp->qp_handle); 431 if (err) 432 return err; 433 434 if (qp->rq_cpu_addr) { 435 ibdev_dbg(&dev->ibdev, 436 "qp->cpu_addr[0x%p] freed: size[%lu], dma[%pad]\n", 437 qp->rq_cpu_addr, qp->rq_size, 438 &qp->rq_dma_addr); 439 efa_free_mapped(dev, qp->rq_cpu_addr, qp->rq_dma_addr, 440 qp->rq_size, DMA_TO_DEVICE); 441 } 442 443 kfree(qp); 444 return 0; 445} 446 447static struct rdma_user_mmap_entry* 448efa_user_mmap_entry_insert(struct ib_ucontext *ucontext, 449 u64 address, size_t length, 450 u8 mmap_flag, u64 *offset) 451{ 452 struct efa_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL); 453 int err; 454 455 if (!entry) 456 return NULL; 457 458 entry->address = address; 459 entry->mmap_flag = mmap_flag; 460 461 err = rdma_user_mmap_entry_insert(ucontext, &entry->rdma_entry, 462 length); 463 if (err) { 464 kfree(entry); 465 return NULL; 466 } 467 *offset = rdma_user_mmap_get_offset(&entry->rdma_entry); 468 469 return &entry->rdma_entry; 470} 471 472static int qp_mmap_entries_setup(struct efa_qp *qp, 473 struct efa_dev *dev, 474 struct efa_ucontext *ucontext, 475 struct efa_com_create_qp_params *params, 476 struct efa_ibv_create_qp_resp *resp) 477{ 478 size_t length; 479 u64 address; 480 481 address = dev->db_bar_addr + resp->sq_db_offset; 482 qp->sq_db_mmap_entry = 483 efa_user_mmap_entry_insert(&ucontext->ibucontext, 484 address, 485 PAGE_SIZE, EFA_MMAP_IO_NC, 486 &resp->sq_db_mmap_key); 487 if (!qp->sq_db_mmap_entry) 488 return -ENOMEM; 489 490 resp->sq_db_offset &= ~PAGE_MASK; 491 492 address = dev->mem_bar_addr + resp->llq_desc_offset; 493 length = PAGE_ALIGN(params->sq_ring_size_in_bytes + 494 (resp->llq_desc_offset & ~PAGE_MASK)); 495 496 qp->llq_desc_mmap_entry = 497 efa_user_mmap_entry_insert(&ucontext->ibucontext, 498 address, length, 499 EFA_MMAP_IO_WC, 500 &resp->llq_desc_mmap_key); 501 if (!qp->llq_desc_mmap_entry) 502 goto err_remove_mmap; 503 504 resp->llq_desc_offset &= ~PAGE_MASK; 505 506 if (qp->rq_size) { 507 address = dev->db_bar_addr + resp->rq_db_offset; 508 509 qp->rq_db_mmap_entry = 510 efa_user_mmap_entry_insert(&ucontext->ibucontext, 511 address, PAGE_SIZE, 512 EFA_MMAP_IO_NC, 513 &resp->rq_db_mmap_key); 514 if (!qp->rq_db_mmap_entry) 515 goto err_remove_mmap; 516 517 resp->rq_db_offset &= ~PAGE_MASK; 518 519 address = virt_to_phys(qp->rq_cpu_addr); 520 qp->rq_mmap_entry = 521 efa_user_mmap_entry_insert(&ucontext->ibucontext, 522 address, qp->rq_size, 523 EFA_MMAP_DMA_PAGE, 524 &resp->rq_mmap_key); 525 if (!qp->rq_mmap_entry) 526 goto err_remove_mmap; 527 528 resp->rq_mmap_size = qp->rq_size; 529 } 530 531 return 0; 532 533err_remove_mmap: 534 efa_qp_user_mmap_entries_remove(qp); 535 536 return -ENOMEM; 537} 538 539static int efa_qp_validate_cap(struct efa_dev *dev, 540 struct ib_qp_init_attr *init_attr) 541{ 542 if (init_attr->cap.max_send_wr > dev->dev_attr.max_sq_depth) { 543 ibdev_dbg(&dev->ibdev, 544 "qp: requested send wr[%u] exceeds the max[%u]\n", 545 init_attr->cap.max_send_wr, 546 dev->dev_attr.max_sq_depth); 547 return -EINVAL; 548 } 549 if (init_attr->cap.max_recv_wr > dev->dev_attr.max_rq_depth) { 550 ibdev_dbg(&dev->ibdev, 551 "qp: requested receive wr[%u] exceeds the max[%u]\n", 552 init_attr->cap.max_recv_wr, 553 dev->dev_attr.max_rq_depth); 554 return -EINVAL; 555 } 556 if (init_attr->cap.max_send_sge > dev->dev_attr.max_sq_sge) { 557 ibdev_dbg(&dev->ibdev, 558 "qp: requested sge send[%u] exceeds the max[%u]\n", 559 init_attr->cap.max_send_sge, dev->dev_attr.max_sq_sge); 560 return -EINVAL; 561 } 562 if (init_attr->cap.max_recv_sge > dev->dev_attr.max_rq_sge) { 563 ibdev_dbg(&dev->ibdev, 564 "qp: requested sge recv[%u] exceeds the max[%u]\n", 565 init_attr->cap.max_recv_sge, dev->dev_attr.max_rq_sge); 566 return -EINVAL; 567 } 568 if (init_attr->cap.max_inline_data > dev->dev_attr.inline_buf_size) { 569 ibdev_dbg(&dev->ibdev, 570 "qp: requested inline data[%u] exceeds the max[%u]\n", 571 init_attr->cap.max_inline_data, 572 dev->dev_attr.inline_buf_size); 573 return -EINVAL; 574 } 575 576 return 0; 577} 578 579static int efa_qp_validate_attr(struct efa_dev *dev, 580 struct ib_qp_init_attr *init_attr) 581{ 582 if (init_attr->qp_type != IB_QPT_DRIVER && 583 init_attr->qp_type != IB_QPT_UD) { 584 ibdev_dbg(&dev->ibdev, 585 "Unsupported qp type %d\n", init_attr->qp_type); 586 return -EOPNOTSUPP; 587 } 588 589 if (init_attr->srq) { 590 ibdev_dbg(&dev->ibdev, "SRQ is not supported\n"); 591 return -EOPNOTSUPP; 592 } 593 594 if (init_attr->create_flags) { 595 ibdev_dbg(&dev->ibdev, "Unsupported create flags\n"); 596 return -EOPNOTSUPP; 597 } 598 599 return 0; 600} 601 602struct ib_qp *efa_create_qp(struct ib_pd *ibpd, 603 struct ib_qp_init_attr *init_attr, 604 struct ib_udata *udata) 605{ 606 struct efa_com_create_qp_params create_qp_params = {}; 607 struct efa_com_create_qp_result create_qp_resp; 608 struct efa_dev *dev = to_edev(ibpd->device); 609 struct efa_ibv_create_qp_resp resp = {}; 610 struct efa_ibv_create_qp cmd = {}; 611 struct efa_ucontext *ucontext; 612 struct efa_qp *qp; 613 int err; 614 615 ucontext = rdma_udata_to_drv_context(udata, struct efa_ucontext, 616 ibucontext); 617 618 err = efa_qp_validate_cap(dev, init_attr); 619 if (err) 620 goto err_out; 621 622 err = efa_qp_validate_attr(dev, init_attr); 623 if (err) 624 goto err_out; 625 626 if (offsetofend(typeof(cmd), driver_qp_type) > udata->inlen) { 627 ibdev_dbg(&dev->ibdev, 628 "Incompatible ABI params, no input udata\n"); 629 err = -EINVAL; 630 goto err_out; 631 } 632 633 if (udata->inlen > sizeof(cmd) && 634 !ib_is_udata_cleared(udata, sizeof(cmd), 635 udata->inlen - sizeof(cmd))) { 636 ibdev_dbg(&dev->ibdev, 637 "Incompatible ABI params, unknown fields in udata\n"); 638 err = -EINVAL; 639 goto err_out; 640 } 641 642 err = ib_copy_from_udata(&cmd, udata, 643 min(sizeof(cmd), udata->inlen)); 644 if (err) { 645 ibdev_dbg(&dev->ibdev, 646 "Cannot copy udata for create_qp\n"); 647 goto err_out; 648 } 649 650 if (cmd.comp_mask) { 651 ibdev_dbg(&dev->ibdev, 652 "Incompatible ABI params, unknown fields in udata\n"); 653 err = -EINVAL; 654 goto err_out; 655 } 656 657 qp = kzalloc(sizeof(*qp), GFP_KERNEL); 658 if (!qp) { 659 err = -ENOMEM; 660 goto err_out; 661 } 662 663 create_qp_params.uarn = ucontext->uarn; 664 create_qp_params.pd = to_epd(ibpd)->pdn; 665 666 if (init_attr->qp_type == IB_QPT_UD) { 667 create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_UD; 668 } else if (cmd.driver_qp_type == EFA_QP_DRIVER_TYPE_SRD) { 669 create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_SRD; 670 } else { 671 ibdev_dbg(&dev->ibdev, 672 "Unsupported qp type %d driver qp type %d\n", 673 init_attr->qp_type, cmd.driver_qp_type); 674 err = -EOPNOTSUPP; 675 goto err_free_qp; 676 } 677 678 ibdev_dbg(&dev->ibdev, "Create QP: qp type %d driver qp type %#x\n", 679 init_attr->qp_type, cmd.driver_qp_type); 680 create_qp_params.send_cq_idx = to_ecq(init_attr->send_cq)->cq_idx; 681 create_qp_params.recv_cq_idx = to_ecq(init_attr->recv_cq)->cq_idx; 682 create_qp_params.sq_depth = init_attr->cap.max_send_wr; 683 create_qp_params.sq_ring_size_in_bytes = cmd.sq_ring_size; 684 685 create_qp_params.rq_depth = init_attr->cap.max_recv_wr; 686 create_qp_params.rq_ring_size_in_bytes = cmd.rq_ring_size; 687 qp->rq_size = PAGE_ALIGN(create_qp_params.rq_ring_size_in_bytes); 688 if (qp->rq_size) { 689 qp->rq_cpu_addr = efa_zalloc_mapped(dev, &qp->rq_dma_addr, 690 qp->rq_size, DMA_TO_DEVICE); 691 if (!qp->rq_cpu_addr) { 692 err = -ENOMEM; 693 goto err_free_qp; 694 } 695 696 ibdev_dbg(&dev->ibdev, 697 "qp->cpu_addr[0x%p] allocated: size[%lu], dma[%pad]\n", 698 qp->rq_cpu_addr, qp->rq_size, &qp->rq_dma_addr); 699 create_qp_params.rq_base_addr = qp->rq_dma_addr; 700 } 701 702 err = efa_com_create_qp(&dev->edev, &create_qp_params, 703 &create_qp_resp); 704 if (err) 705 goto err_free_mapped; 706 707 resp.sq_db_offset = create_qp_resp.sq_db_offset; 708 resp.rq_db_offset = create_qp_resp.rq_db_offset; 709 resp.llq_desc_offset = create_qp_resp.llq_descriptors_offset; 710 resp.send_sub_cq_idx = create_qp_resp.send_sub_cq_idx; 711 resp.recv_sub_cq_idx = create_qp_resp.recv_sub_cq_idx; 712 713 err = qp_mmap_entries_setup(qp, dev, ucontext, &create_qp_params, 714 &resp); 715 if (err) 716 goto err_destroy_qp; 717 718 qp->qp_handle = create_qp_resp.qp_handle; 719 qp->ibqp.qp_num = create_qp_resp.qp_num; 720 qp->max_send_wr = init_attr->cap.max_send_wr; 721 qp->max_recv_wr = init_attr->cap.max_recv_wr; 722 qp->max_send_sge = init_attr->cap.max_send_sge; 723 qp->max_recv_sge = init_attr->cap.max_recv_sge; 724 qp->max_inline_data = init_attr->cap.max_inline_data; 725 726 if (udata->outlen) { 727 err = ib_copy_to_udata(udata, &resp, 728 min(sizeof(resp), udata->outlen)); 729 if (err) { 730 ibdev_dbg(&dev->ibdev, 731 "Failed to copy udata for qp[%u]\n", 732 create_qp_resp.qp_num); 733 goto err_remove_mmap_entries; 734 } 735 } 736 737 ibdev_dbg(&dev->ibdev, "Created qp[%d]\n", qp->ibqp.qp_num); 738 739 return &qp->ibqp; 740 741err_remove_mmap_entries: 742 efa_qp_user_mmap_entries_remove(qp); 743err_destroy_qp: 744 efa_destroy_qp_handle(dev, create_qp_resp.qp_handle); 745err_free_mapped: 746 if (qp->rq_size) 747 efa_free_mapped(dev, qp->rq_cpu_addr, qp->rq_dma_addr, 748 qp->rq_size, DMA_TO_DEVICE); 749err_free_qp: 750 kfree(qp); 751err_out: 752 atomic64_inc(&dev->stats.create_qp_err); 753 return ERR_PTR(err); 754} 755 756static 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 839static 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 865static 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 909int 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 (udata->inlen && 920 !ib_is_udata_cleared(udata, 0, udata->inlen)) { 921 ibdev_dbg(&dev->ibdev, 922 "Incompatible ABI params, udata not cleared\n"); 923 return -EINVAL; 924 } 925 926 cur_state = qp_attr_mask & IB_QP_CUR_STATE ? qp_attr->cur_qp_state : 927 qp->state; 928 new_state = qp_attr_mask & IB_QP_STATE ? qp_attr->qp_state : cur_state; 929 930 err = efa_modify_qp_validate(dev, qp, qp_attr, qp_attr_mask, cur_state, 931 new_state); 932 if (err) 933 return err; 934 935 params.qp_handle = qp->qp_handle; 936 937 if (qp_attr_mask & IB_QP_STATE) { 938 EFA_SET(¶ms.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_QP_STATE, 939 1); 940 EFA_SET(¶ms.modify_mask, 941 EFA_ADMIN_MODIFY_QP_CMD_CUR_QP_STATE, 1); 942 params.cur_qp_state = cur_state; 943 params.qp_state = new_state; 944 } 945 946 if (qp_attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) { 947 EFA_SET(¶ms.modify_mask, 948 EFA_ADMIN_MODIFY_QP_CMD_SQ_DRAINED_ASYNC_NOTIFY, 1); 949 params.sq_drained_async_notify = qp_attr->en_sqd_async_notify; 950 } 951 952 if (qp_attr_mask & IB_QP_QKEY) { 953 EFA_SET(¶ms.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_QKEY, 1); 954 params.qkey = qp_attr->qkey; 955 } 956 957 if (qp_attr_mask & IB_QP_SQ_PSN) { 958 EFA_SET(¶ms.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_SQ_PSN, 1); 959 params.sq_psn = qp_attr->sq_psn; 960 } 961 962 if (qp_attr_mask & IB_QP_RNR_RETRY) { 963 EFA_SET(¶ms.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_RNR_RETRY, 964 1); 965 params.rnr_retry = qp_attr->rnr_retry; 966 } 967 968 err = efa_com_modify_qp(&dev->edev, ¶ms); 969 if (err) 970 return err; 971 972 qp->state = new_state; 973 974 return 0; 975} 976 977static int efa_destroy_cq_idx(struct efa_dev *dev, int cq_idx) 978{ 979 struct efa_com_destroy_cq_params params = { .cq_idx = cq_idx }; 980 981 return efa_com_destroy_cq(&dev->edev, ¶ms); 982} 983 984int efa_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata) 985{ 986 struct efa_dev *dev = to_edev(ibcq->device); 987 struct efa_cq *cq = to_ecq(ibcq); 988 989 ibdev_dbg(&dev->ibdev, 990 "Destroy cq[%d] virt[0x%p] freed: size[%lu], dma[%pad]\n", 991 cq->cq_idx, cq->cpu_addr, cq->size, &cq->dma_addr); 992 993 rdma_user_mmap_entry_remove(cq->mmap_entry); 994 efa_destroy_cq_idx(dev, cq->cq_idx); 995 efa_free_mapped(dev, cq->cpu_addr, cq->dma_addr, cq->size, 996 DMA_FROM_DEVICE); 997 return 0; 998} 999 1000static int cq_mmap_entries_setup(struct efa_dev *dev, struct efa_cq *cq, 1001 struct efa_ibv_create_cq_resp *resp) 1002{ 1003 resp->q_mmap_size = cq->size; 1004 cq->mmap_entry = efa_user_mmap_entry_insert(&cq->ucontext->ibucontext, 1005 virt_to_phys(cq->cpu_addr), 1006 cq->size, EFA_MMAP_DMA_PAGE, 1007 &resp->q_mmap_key); 1008 if (!cq->mmap_entry) 1009 return -ENOMEM; 1010 1011 return 0; 1012} 1013 1014int efa_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr, 1015 struct ib_udata *udata) 1016{ 1017 struct efa_ucontext *ucontext = rdma_udata_to_drv_context( 1018 udata, struct efa_ucontext, ibucontext); 1019 struct efa_ibv_create_cq_resp resp = {}; 1020 struct efa_com_create_cq_params params; 1021 struct efa_com_create_cq_result result; 1022 struct ib_device *ibdev = ibcq->device; 1023 struct efa_dev *dev = to_edev(ibdev); 1024 struct efa_ibv_create_cq cmd = {}; 1025 struct efa_cq *cq = to_ecq(ibcq); 1026 int entries = attr->cqe; 1027 int err; 1028 1029 ibdev_dbg(ibdev, "create_cq entries %d\n", entries); 1030 1031 if (entries < 1 || entries > dev->dev_attr.max_cq_depth) { 1032 ibdev_dbg(ibdev, 1033 "cq: requested entries[%u] non-positive or greater than max[%u]\n", 1034 entries, dev->dev_attr.max_cq_depth); 1035 err = -EINVAL; 1036 goto err_out; 1037 } 1038 1039 if (offsetofend(typeof(cmd), num_sub_cqs) > udata->inlen) { 1040 ibdev_dbg(ibdev, 1041 "Incompatible ABI params, no input udata\n"); 1042 err = -EINVAL; 1043 goto err_out; 1044 } 1045 1046 if (udata->inlen > sizeof(cmd) && 1047 !ib_is_udata_cleared(udata, sizeof(cmd), 1048 udata->inlen - sizeof(cmd))) { 1049 ibdev_dbg(ibdev, 1050 "Incompatible ABI params, unknown fields in udata\n"); 1051 err = -EINVAL; 1052 goto err_out; 1053 } 1054 1055 err = ib_copy_from_udata(&cmd, udata, 1056 min(sizeof(cmd), udata->inlen)); 1057 if (err) { 1058 ibdev_dbg(ibdev, "Cannot copy udata for create_cq\n"); 1059 goto err_out; 1060 } 1061 1062 if (cmd.comp_mask || !is_reserved_cleared(cmd.reserved_50)) { 1063 ibdev_dbg(ibdev, 1064 "Incompatible ABI params, unknown fields in udata\n"); 1065 err = -EINVAL; 1066 goto err_out; 1067 } 1068 1069 if (!cmd.cq_entry_size) { 1070 ibdev_dbg(ibdev, 1071 "Invalid entry size [%u]\n", cmd.cq_entry_size); 1072 err = -EINVAL; 1073 goto err_out; 1074 } 1075 1076 if (cmd.num_sub_cqs != dev->dev_attr.sub_cqs_per_cq) { 1077 ibdev_dbg(ibdev, 1078 "Invalid number of sub cqs[%u] expected[%u]\n", 1079 cmd.num_sub_cqs, dev->dev_attr.sub_cqs_per_cq); 1080 err = -EINVAL; 1081 goto err_out; 1082 } 1083 1084 cq->ucontext = ucontext; 1085 cq->size = PAGE_ALIGN(cmd.cq_entry_size * entries * cmd.num_sub_cqs); 1086 cq->cpu_addr = efa_zalloc_mapped(dev, &cq->dma_addr, cq->size, 1087 DMA_FROM_DEVICE); 1088 if (!cq->cpu_addr) { 1089 err = -ENOMEM; 1090 goto err_out; 1091 } 1092 1093 params.uarn = cq->ucontext->uarn; 1094 params.cq_depth = entries; 1095 params.dma_addr = cq->dma_addr; 1096 params.entry_size_in_bytes = cmd.cq_entry_size; 1097 params.num_sub_cqs = cmd.num_sub_cqs; 1098 err = efa_com_create_cq(&dev->edev, ¶ms, &result); 1099 if (err) 1100 goto err_free_mapped; 1101 1102 resp.cq_idx = result.cq_idx; 1103 cq->cq_idx = result.cq_idx; 1104 cq->ibcq.cqe = result.actual_depth; 1105 WARN_ON_ONCE(entries != result.actual_depth); 1106 1107 err = cq_mmap_entries_setup(dev, cq, &resp); 1108 if (err) { 1109 ibdev_dbg(ibdev, "Could not setup cq[%u] mmap entries\n", 1110 cq->cq_idx); 1111 goto err_destroy_cq; 1112 } 1113 1114 if (udata->outlen) { 1115 err = ib_copy_to_udata(udata, &resp, 1116 min(sizeof(resp), udata->outlen)); 1117 if (err) { 1118 ibdev_dbg(ibdev, 1119 "Failed to copy udata for create_cq\n"); 1120 goto err_remove_mmap; 1121 } 1122 } 1123 1124 ibdev_dbg(ibdev, "Created cq[%d], cq depth[%u]. dma[%pad] virt[0x%p]\n", 1125 cq->cq_idx, result.actual_depth, &cq->dma_addr, cq->cpu_addr); 1126 1127 return 0; 1128 1129err_remove_mmap: 1130 rdma_user_mmap_entry_remove(cq->mmap_entry); 1131err_destroy_cq: 1132 efa_destroy_cq_idx(dev, cq->cq_idx); 1133err_free_mapped: 1134 efa_free_mapped(dev, cq->cpu_addr, cq->dma_addr, cq->size, 1135 DMA_FROM_DEVICE); 1136 1137err_out: 1138 atomic64_inc(&dev->stats.create_cq_err); 1139 return err; 1140} 1141 1142static int umem_to_page_list(struct efa_dev *dev, 1143 struct ib_umem *umem, 1144 u64 *page_list, 1145 u32 hp_cnt, 1146 u8 hp_shift) 1147{ 1148 u32 pages_in_hp = BIT(hp_shift - PAGE_SHIFT); 1149 struct ib_block_iter biter; 1150 unsigned int hp_idx = 0; 1151 1152 ibdev_dbg(&dev->ibdev, "hp_cnt[%u], pages_in_hp[%u]\n", 1153 hp_cnt, pages_in_hp); 1154 1155 rdma_umem_for_each_dma_block(umem, &biter, BIT(hp_shift)) 1156 page_list[hp_idx++] = rdma_block_iter_dma_address(&biter); 1157 1158 return 0; 1159} 1160 1161static struct scatterlist *efa_vmalloc_buf_to_sg(u64 *buf, int page_cnt) 1162{ 1163 struct scatterlist *sglist; 1164 struct page *pg; 1165 int i; 1166 1167 sglist = kmalloc_array(page_cnt, sizeof(*sglist), GFP_KERNEL); 1168 if (!sglist) 1169 return NULL; 1170 sg_init_table(sglist, page_cnt); 1171 for (i = 0; i < page_cnt; i++) { 1172 pg = vmalloc_to_page(buf); 1173 if (!pg) 1174 goto err; 1175 sg_set_page(&sglist[i], pg, PAGE_SIZE, 0); 1176 buf += PAGE_SIZE / sizeof(*buf); 1177 } 1178 return sglist; 1179 1180err: 1181 kfree(sglist); 1182 return NULL; 1183} 1184 1185/* 1186 * create a chunk list of physical pages dma addresses from the supplied 1187 * scatter gather list 1188 */ 1189static int pbl_chunk_list_create(struct efa_dev *dev, struct pbl_context *pbl) 1190{ 1191 struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list; 1192 int page_cnt = pbl->phys.indirect.pbl_buf_size_in_pages; 1193 struct scatterlist *pages_sgl = pbl->phys.indirect.sgl; 1194 unsigned int chunk_list_size, chunk_idx, payload_idx; 1195 int sg_dma_cnt = pbl->phys.indirect.sg_dma_cnt; 1196 struct efa_com_ctrl_buff_info *ctrl_buf; 1197 u64 *cur_chunk_buf, *prev_chunk_buf; 1198 struct ib_block_iter biter; 1199 dma_addr_t dma_addr; 1200 int i; 1201 1202 /* allocate a chunk list that consists of 4KB chunks */ 1203 chunk_list_size = DIV_ROUND_UP(page_cnt, EFA_PTRS_PER_CHUNK); 1204 1205 chunk_list->size = chunk_list_size; 1206 chunk_list->chunks = kcalloc(chunk_list_size, 1207 sizeof(*chunk_list->chunks), 1208 GFP_KERNEL); 1209 if (!chunk_list->chunks) 1210 return -ENOMEM; 1211 1212 ibdev_dbg(&dev->ibdev, 1213 "chunk_list_size[%u] - pages[%u]\n", chunk_list_size, 1214 page_cnt); 1215 1216 /* allocate chunk buffers: */ 1217 for (i = 0; i < chunk_list_size; i++) { 1218 chunk_list->chunks[i].buf = kzalloc(EFA_CHUNK_SIZE, GFP_KERNEL); 1219 if (!chunk_list->chunks[i].buf) 1220 goto chunk_list_dealloc; 1221 1222 chunk_list->chunks[i].length = EFA_CHUNK_USED_SIZE; 1223 } 1224 chunk_list->chunks[chunk_list_size - 1].length = 1225 ((page_cnt % EFA_PTRS_PER_CHUNK) * EFA_CHUNK_PAYLOAD_PTR_SIZE) + 1226 EFA_CHUNK_PTR_SIZE; 1227 1228 /* fill the dma addresses of sg list pages to chunks: */ 1229 chunk_idx = 0; 1230 payload_idx = 0; 1231 cur_chunk_buf = chunk_list->chunks[0].buf; 1232 rdma_for_each_block(pages_sgl, &biter, sg_dma_cnt, 1233 EFA_CHUNK_PAYLOAD_SIZE) { 1234 cur_chunk_buf[payload_idx++] = 1235 rdma_block_iter_dma_address(&biter); 1236 1237 if (payload_idx == EFA_PTRS_PER_CHUNK) { 1238 chunk_idx++; 1239 cur_chunk_buf = chunk_list->chunks[chunk_idx].buf; 1240 payload_idx = 0; 1241 } 1242 } 1243 1244 /* map chunks to dma and fill chunks next ptrs */ 1245 for (i = chunk_list_size - 1; i >= 0; i--) { 1246 dma_addr = dma_map_single(&dev->pdev->dev, 1247 chunk_list->chunks[i].buf, 1248 chunk_list->chunks[i].length, 1249 DMA_TO_DEVICE); 1250 if (dma_mapping_error(&dev->pdev->dev, dma_addr)) { 1251 ibdev_err(&dev->ibdev, 1252 "chunk[%u] dma_map_failed\n", i); 1253 goto chunk_list_unmap; 1254 } 1255 1256 chunk_list->chunks[i].dma_addr = dma_addr; 1257 ibdev_dbg(&dev->ibdev, 1258 "chunk[%u] mapped at [%pad]\n", i, &dma_addr); 1259 1260 if (!i) 1261 break; 1262 1263 prev_chunk_buf = chunk_list->chunks[i - 1].buf; 1264 1265 ctrl_buf = (struct efa_com_ctrl_buff_info *) 1266 &prev_chunk_buf[EFA_PTRS_PER_CHUNK]; 1267 ctrl_buf->length = chunk_list->chunks[i].length; 1268 1269 efa_com_set_dma_addr(dma_addr, 1270 &ctrl_buf->address.mem_addr_high, 1271 &ctrl_buf->address.mem_addr_low); 1272 } 1273 1274 return 0; 1275 1276chunk_list_unmap: 1277 for (; i < chunk_list_size; i++) { 1278 dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr, 1279 chunk_list->chunks[i].length, DMA_TO_DEVICE); 1280 } 1281chunk_list_dealloc: 1282 for (i = 0; i < chunk_list_size; i++) 1283 kfree(chunk_list->chunks[i].buf); 1284 1285 kfree(chunk_list->chunks); 1286 return -ENOMEM; 1287} 1288 1289static void pbl_chunk_list_destroy(struct efa_dev *dev, struct pbl_context *pbl) 1290{ 1291 struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list; 1292 int i; 1293 1294 for (i = 0; i < chunk_list->size; i++) { 1295 dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr, 1296 chunk_list->chunks[i].length, DMA_TO_DEVICE); 1297 kfree(chunk_list->chunks[i].buf); 1298 } 1299 1300 kfree(chunk_list->chunks); 1301} 1302 1303/* initialize pbl continuous mode: map pbl buffer to a dma address. */ 1304static int pbl_continuous_initialize(struct efa_dev *dev, 1305 struct pbl_context *pbl) 1306{ 1307 dma_addr_t dma_addr; 1308 1309 dma_addr = dma_map_single(&dev->pdev->dev, pbl->pbl_buf, 1310 pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE); 1311 if (dma_mapping_error(&dev->pdev->dev, dma_addr)) { 1312 ibdev_err(&dev->ibdev, "Unable to map pbl to DMA address\n"); 1313 return -ENOMEM; 1314 } 1315 1316 pbl->phys.continuous.dma_addr = dma_addr; 1317 ibdev_dbg(&dev->ibdev, 1318 "pbl continuous - dma_addr = %pad, size[%u]\n", 1319 &dma_addr, pbl->pbl_buf_size_in_bytes); 1320 1321 return 0; 1322} 1323 1324/* 1325 * initialize pbl indirect mode: 1326 * create a chunk list out of the dma addresses of the physical pages of 1327 * pbl buffer. 1328 */ 1329static int pbl_indirect_initialize(struct efa_dev *dev, struct pbl_context *pbl) 1330{ 1331 u32 size_in_pages = DIV_ROUND_UP(pbl->pbl_buf_size_in_bytes, EFA_CHUNK_PAYLOAD_SIZE); 1332 struct scatterlist *sgl; 1333 int sg_dma_cnt, err; 1334 1335 BUILD_BUG_ON(EFA_CHUNK_PAYLOAD_SIZE > PAGE_SIZE); 1336 sgl = efa_vmalloc_buf_to_sg(pbl->pbl_buf, size_in_pages); 1337 if (!sgl) 1338 return -ENOMEM; 1339 1340 sg_dma_cnt = dma_map_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE); 1341 if (!sg_dma_cnt) { 1342 err = -EINVAL; 1343 goto err_map; 1344 } 1345 1346 pbl->phys.indirect.pbl_buf_size_in_pages = size_in_pages; 1347 pbl->phys.indirect.sgl = sgl; 1348 pbl->phys.indirect.sg_dma_cnt = sg_dma_cnt; 1349 err = pbl_chunk_list_create(dev, pbl); 1350 if (err) { 1351 ibdev_dbg(&dev->ibdev, 1352 "chunk_list creation failed[%d]\n", err); 1353 goto err_chunk; 1354 } 1355 1356 ibdev_dbg(&dev->ibdev, 1357 "pbl indirect - size[%u], chunks[%u]\n", 1358 pbl->pbl_buf_size_in_bytes, 1359 pbl->phys.indirect.chunk_list.size); 1360 1361 return 0; 1362 1363err_chunk: 1364 dma_unmap_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE); 1365err_map: 1366 kfree(sgl); 1367 return err; 1368} 1369 1370static void pbl_indirect_terminate(struct efa_dev *dev, struct pbl_context *pbl) 1371{ 1372 pbl_chunk_list_destroy(dev, pbl); 1373 dma_unmap_sg(&dev->pdev->dev, pbl->phys.indirect.sgl, 1374 pbl->phys.indirect.pbl_buf_size_in_pages, DMA_TO_DEVICE); 1375 kfree(pbl->phys.indirect.sgl); 1376} 1377 1378/* create a page buffer list from a mapped user memory region */ 1379static int pbl_create(struct efa_dev *dev, 1380 struct pbl_context *pbl, 1381 struct ib_umem *umem, 1382 int hp_cnt, 1383 u8 hp_shift) 1384{ 1385 int err; 1386 1387 pbl->pbl_buf_size_in_bytes = hp_cnt * EFA_CHUNK_PAYLOAD_PTR_SIZE; 1388 pbl->pbl_buf = kvzalloc(pbl->pbl_buf_size_in_bytes, GFP_KERNEL); 1389 if (!pbl->pbl_buf) 1390 return -ENOMEM; 1391 1392 if (is_vmalloc_addr(pbl->pbl_buf)) { 1393 pbl->physically_continuous = 0; 1394 err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt, 1395 hp_shift); 1396 if (err) 1397 goto err_free; 1398 1399 err = pbl_indirect_initialize(dev, pbl); 1400 if (err) 1401 goto err_free; 1402 } else { 1403 pbl->physically_continuous = 1; 1404 err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt, 1405 hp_shift); 1406 if (err) 1407 goto err_free; 1408 1409 err = pbl_continuous_initialize(dev, pbl); 1410 if (err) 1411 goto err_free; 1412 } 1413 1414 ibdev_dbg(&dev->ibdev, 1415 "user_pbl_created: user_pages[%u], continuous[%u]\n", 1416 hp_cnt, pbl->physically_continuous); 1417 1418 return 0; 1419 1420err_free: 1421 kvfree(pbl->pbl_buf); 1422 return err; 1423} 1424 1425static void pbl_destroy(struct efa_dev *dev, struct pbl_context *pbl) 1426{ 1427 if (pbl->physically_continuous) 1428 dma_unmap_single(&dev->pdev->dev, pbl->phys.continuous.dma_addr, 1429 pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE); 1430 else 1431 pbl_indirect_terminate(dev, pbl); 1432 1433 kvfree(pbl->pbl_buf); 1434} 1435 1436static int efa_create_inline_pbl(struct efa_dev *dev, struct efa_mr *mr, 1437 struct efa_com_reg_mr_params *params) 1438{ 1439 int err; 1440 1441 params->inline_pbl = 1; 1442 err = umem_to_page_list(dev, mr->umem, params->pbl.inline_pbl_array, 1443 params->page_num, params->page_shift); 1444 if (err) 1445 return err; 1446 1447 ibdev_dbg(&dev->ibdev, 1448 "inline_pbl_array - pages[%u]\n", params->page_num); 1449 1450 return 0; 1451} 1452 1453static int efa_create_pbl(struct efa_dev *dev, 1454 struct pbl_context *pbl, 1455 struct efa_mr *mr, 1456 struct efa_com_reg_mr_params *params) 1457{ 1458 int err; 1459 1460 err = pbl_create(dev, pbl, mr->umem, params->page_num, 1461 params->page_shift); 1462 if (err) { 1463 ibdev_dbg(&dev->ibdev, "Failed to create pbl[%d]\n", err); 1464 return err; 1465 } 1466 1467 params->inline_pbl = 0; 1468 params->indirect = !pbl->physically_continuous; 1469 if (pbl->physically_continuous) { 1470 params->pbl.pbl.length = pbl->pbl_buf_size_in_bytes; 1471 1472 efa_com_set_dma_addr(pbl->phys.continuous.dma_addr, 1473 ¶ms->pbl.pbl.address.mem_addr_high, 1474 ¶ms->pbl.pbl.address.mem_addr_low); 1475 } else { 1476 params->pbl.pbl.length = 1477 pbl->phys.indirect.chunk_list.chunks[0].length; 1478 1479 efa_com_set_dma_addr(pbl->phys.indirect.chunk_list.chunks[0].dma_addr, 1480 ¶ms->pbl.pbl.address.mem_addr_high, 1481 ¶ms->pbl.pbl.address.mem_addr_low); 1482 } 1483 1484 return 0; 1485} 1486 1487struct ib_mr *efa_reg_mr(struct ib_pd *ibpd, u64 start, u64 length, 1488 u64 virt_addr, int access_flags, 1489 struct ib_udata *udata) 1490{ 1491 struct efa_dev *dev = to_edev(ibpd->device); 1492 struct efa_com_reg_mr_params params = {}; 1493 struct efa_com_reg_mr_result result = {}; 1494 struct pbl_context pbl; 1495 int supp_access_flags; 1496 unsigned int pg_sz; 1497 struct efa_mr *mr; 1498 int inline_size; 1499 int err; 1500 1501 if (udata && udata->inlen && 1502 !ib_is_udata_cleared(udata, 0, sizeof(udata->inlen))) { 1503 ibdev_dbg(&dev->ibdev, 1504 "Incompatible ABI params, udata not cleared\n"); 1505 err = -EINVAL; 1506 goto err_out; 1507 } 1508 1509 supp_access_flags = 1510 IB_ACCESS_LOCAL_WRITE | 1511 (EFA_DEV_CAP(dev, RDMA_READ) ? IB_ACCESS_REMOTE_READ : 0); 1512 1513 access_flags &= ~IB_ACCESS_OPTIONAL; 1514 if (access_flags & ~supp_access_flags) { 1515 ibdev_dbg(&dev->ibdev, 1516 "Unsupported access flags[%#x], supported[%#x]\n", 1517 access_flags, supp_access_flags); 1518 err = -EOPNOTSUPP; 1519 goto err_out; 1520 } 1521 1522 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 1523 if (!mr) { 1524 err = -ENOMEM; 1525 goto err_out; 1526 } 1527 1528 mr->umem = ib_umem_get(ibpd->device, start, length, access_flags); 1529 if (IS_ERR(mr->umem)) { 1530 err = PTR_ERR(mr->umem); 1531 ibdev_dbg(&dev->ibdev, 1532 "Failed to pin and map user space memory[%d]\n", err); 1533 goto err_free; 1534 } 1535 1536 params.pd = to_epd(ibpd)->pdn; 1537 params.iova = virt_addr; 1538 params.mr_length_in_bytes = length; 1539 params.permissions = access_flags; 1540 1541 pg_sz = ib_umem_find_best_pgsz(mr->umem, 1542 dev->dev_attr.page_size_cap, 1543 virt_addr); 1544 if (!pg_sz) { 1545 err = -EOPNOTSUPP; 1546 ibdev_dbg(&dev->ibdev, "Failed to find a suitable page size in page_size_cap %#llx\n", 1547 dev->dev_attr.page_size_cap); 1548 goto err_unmap; 1549 } 1550 1551 params.page_shift = order_base_2(pg_sz); 1552 params.page_num = ib_umem_num_dma_blocks(mr->umem, pg_sz); 1553 1554 ibdev_dbg(&dev->ibdev, 1555 "start %#llx length %#llx params.page_shift %u params.page_num %u\n", 1556 start, length, params.page_shift, params.page_num); 1557 1558 inline_size = ARRAY_SIZE(params.pbl.inline_pbl_array); 1559 if (params.page_num <= inline_size) { 1560 err = efa_create_inline_pbl(dev, mr, ¶ms); 1561 if (err) 1562 goto err_unmap; 1563 1564 err = efa_com_register_mr(&dev->edev, ¶ms, &result); 1565 if (err) 1566 goto err_unmap; 1567 } else { 1568 err = efa_create_pbl(dev, &pbl, mr, ¶ms); 1569 if (err) 1570 goto err_unmap; 1571 1572 err = efa_com_register_mr(&dev->edev, ¶ms, &result); 1573 pbl_destroy(dev, &pbl); 1574 1575 if (err) 1576 goto err_unmap; 1577 } 1578 1579 mr->ibmr.lkey = result.l_key; 1580 mr->ibmr.rkey = result.r_key; 1581 mr->ibmr.length = length; 1582 ibdev_dbg(&dev->ibdev, "Registered mr[%d]\n", mr->ibmr.lkey); 1583 1584 return &mr->ibmr; 1585 1586err_unmap: 1587 ib_umem_release(mr->umem); 1588err_free: 1589 kfree(mr); 1590err_out: 1591 atomic64_inc(&dev->stats.reg_mr_err); 1592 return ERR_PTR(err); 1593} 1594 1595int efa_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata) 1596{ 1597 struct efa_dev *dev = to_edev(ibmr->device); 1598 struct efa_com_dereg_mr_params params; 1599 struct efa_mr *mr = to_emr(ibmr); 1600 int err; 1601 1602 ibdev_dbg(&dev->ibdev, "Deregister mr[%d]\n", ibmr->lkey); 1603 1604 params.l_key = mr->ibmr.lkey; 1605 err = efa_com_dereg_mr(&dev->edev, ¶ms); 1606 if (err) 1607 return err; 1608 1609 ib_umem_release(mr->umem); 1610 kfree(mr); 1611 1612 return 0; 1613} 1614 1615int efa_get_port_immutable(struct ib_device *ibdev, u8 port_num, 1616 struct ib_port_immutable *immutable) 1617{ 1618 struct ib_port_attr attr; 1619 int err; 1620 1621 err = ib_query_port(ibdev, port_num, &attr); 1622 if (err) { 1623 ibdev_dbg(ibdev, "Couldn't query port err[%d]\n", err); 1624 return err; 1625 } 1626 1627 immutable->pkey_tbl_len = attr.pkey_tbl_len; 1628 immutable->gid_tbl_len = attr.gid_tbl_len; 1629 1630 return 0; 1631} 1632 1633static int efa_dealloc_uar(struct efa_dev *dev, u16 uarn) 1634{ 1635 struct efa_com_dealloc_uar_params params = { 1636 .uarn = uarn, 1637 }; 1638 1639 return efa_com_dealloc_uar(&dev->edev, ¶ms); 1640} 1641 1642#define EFA_CHECK_USER_COMP(_dev, _comp_mask, _attr, _mask, _attr_str) \ 1643 (_attr_str = (!(_dev)->dev_attr._attr || ((_comp_mask) & (_mask))) ? \ 1644 NULL : #_attr) 1645 1646static int efa_user_comp_handshake(const struct ib_ucontext *ibucontext, 1647 const struct efa_ibv_alloc_ucontext_cmd *cmd) 1648{ 1649 struct efa_dev *dev = to_edev(ibucontext->device); 1650 char *attr_str; 1651 1652 if (EFA_CHECK_USER_COMP(dev, cmd->comp_mask, max_tx_batch, 1653 EFA_ALLOC_UCONTEXT_CMD_COMP_TX_BATCH, attr_str)) 1654 goto err; 1655 1656 if (EFA_CHECK_USER_COMP(dev, cmd->comp_mask, min_sq_depth, 1657 EFA_ALLOC_UCONTEXT_CMD_COMP_MIN_SQ_WR, 1658 attr_str)) 1659 goto err; 1660 1661 return 0; 1662 1663err: 1664 ibdev_dbg(&dev->ibdev, "Userspace handshake failed for %s attribute\n", 1665 attr_str); 1666 return -EOPNOTSUPP; 1667} 1668 1669int efa_alloc_ucontext(struct ib_ucontext *ibucontext, struct ib_udata *udata) 1670{ 1671 struct efa_ucontext *ucontext = to_eucontext(ibucontext); 1672 struct efa_dev *dev = to_edev(ibucontext->device); 1673 struct efa_ibv_alloc_ucontext_resp resp = {}; 1674 struct efa_ibv_alloc_ucontext_cmd cmd = {}; 1675 struct efa_com_alloc_uar_result result; 1676 int err; 1677 1678 /* 1679 * it's fine if the driver does not know all request fields, 1680 * we will ack input fields in our response. 1681 */ 1682 1683 err = ib_copy_from_udata(&cmd, udata, 1684 min(sizeof(cmd), udata->inlen)); 1685 if (err) { 1686 ibdev_dbg(&dev->ibdev, 1687 "Cannot copy udata for alloc_ucontext\n"); 1688 goto err_out; 1689 } 1690 1691 err = efa_user_comp_handshake(ibucontext, &cmd); 1692 if (err) 1693 goto err_out; 1694 1695 err = efa_com_alloc_uar(&dev->edev, &result); 1696 if (err) 1697 goto err_out; 1698 1699 ucontext->uarn = result.uarn; 1700 1701 resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_QUERY_DEVICE; 1702 resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_CREATE_AH; 1703 resp.sub_cqs_per_cq = dev->dev_attr.sub_cqs_per_cq; 1704 resp.inline_buf_size = dev->dev_attr.inline_buf_size; 1705 resp.max_llq_size = dev->dev_attr.max_llq_size; 1706 resp.max_tx_batch = dev->dev_attr.max_tx_batch; 1707 resp.min_sq_wr = dev->dev_attr.min_sq_depth; 1708 1709 err = ib_copy_to_udata(udata, &resp, 1710 min(sizeof(resp), udata->outlen)); 1711 if (err) 1712 goto err_dealloc_uar; 1713 1714 return 0; 1715 1716err_dealloc_uar: 1717 efa_dealloc_uar(dev, result.uarn); 1718err_out: 1719 atomic64_inc(&dev->stats.alloc_ucontext_err); 1720 return err; 1721} 1722 1723void efa_dealloc_ucontext(struct ib_ucontext *ibucontext) 1724{ 1725 struct efa_ucontext *ucontext = to_eucontext(ibucontext); 1726 struct efa_dev *dev = to_edev(ibucontext->device); 1727 1728 efa_dealloc_uar(dev, ucontext->uarn); 1729} 1730 1731void efa_mmap_free(struct rdma_user_mmap_entry *rdma_entry) 1732{ 1733 struct efa_user_mmap_entry *entry = to_emmap(rdma_entry); 1734 1735 kfree(entry); 1736} 1737 1738static int __efa_mmap(struct efa_dev *dev, struct efa_ucontext *ucontext, 1739 struct vm_area_struct *vma) 1740{ 1741 struct rdma_user_mmap_entry *rdma_entry; 1742 struct efa_user_mmap_entry *entry; 1743 unsigned long va; 1744 int err = 0; 1745 u64 pfn; 1746 1747 rdma_entry = rdma_user_mmap_entry_get(&ucontext->ibucontext, vma); 1748 if (!rdma_entry) { 1749 ibdev_dbg(&dev->ibdev, 1750 "pgoff[%#lx] does not have valid entry\n", 1751 vma->vm_pgoff); 1752 atomic64_inc(&dev->stats.mmap_err); 1753 return -EINVAL; 1754 } 1755 entry = to_emmap(rdma_entry); 1756 1757 ibdev_dbg(&dev->ibdev, 1758 "Mapping address[%#llx], length[%#zx], mmap_flag[%d]\n", 1759 entry->address, rdma_entry->npages * PAGE_SIZE, 1760 entry->mmap_flag); 1761 1762 pfn = entry->address >> PAGE_SHIFT; 1763 switch (entry->mmap_flag) { 1764 case EFA_MMAP_IO_NC: 1765 err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn, 1766 entry->rdma_entry.npages * PAGE_SIZE, 1767 pgprot_noncached(vma->vm_page_prot), 1768 rdma_entry); 1769 break; 1770 case EFA_MMAP_IO_WC: 1771 err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn, 1772 entry->rdma_entry.npages * PAGE_SIZE, 1773 pgprot_writecombine(vma->vm_page_prot), 1774 rdma_entry); 1775 break; 1776 case EFA_MMAP_DMA_PAGE: 1777 for (va = vma->vm_start; va < vma->vm_end; 1778 va += PAGE_SIZE, pfn++) { 1779 err = vm_insert_page(vma, va, pfn_to_page(pfn)); 1780 if (err) 1781 break; 1782 } 1783 break; 1784 default: 1785 err = -EINVAL; 1786 } 1787 1788 if (err) { 1789 ibdev_dbg( 1790 &dev->ibdev, 1791 "Couldn't mmap address[%#llx] length[%#zx] mmap_flag[%d] err[%d]\n", 1792 entry->address, rdma_entry->npages * PAGE_SIZE, 1793 entry->mmap_flag, err); 1794 atomic64_inc(&dev->stats.mmap_err); 1795 } 1796 1797 rdma_user_mmap_entry_put(rdma_entry); 1798 return err; 1799} 1800 1801int efa_mmap(struct ib_ucontext *ibucontext, 1802 struct vm_area_struct *vma) 1803{ 1804 struct efa_ucontext *ucontext = to_eucontext(ibucontext); 1805 struct efa_dev *dev = to_edev(ibucontext->device); 1806 size_t length = vma->vm_end - vma->vm_start; 1807 1808 ibdev_dbg(&dev->ibdev, 1809 "start %#lx, end %#lx, length = %#zx, pgoff = %#lx\n", 1810 vma->vm_start, vma->vm_end, length, vma->vm_pgoff); 1811 1812 return __efa_mmap(dev, ucontext, vma); 1813} 1814 1815static int efa_ah_destroy(struct efa_dev *dev, struct efa_ah *ah) 1816{ 1817 struct efa_com_destroy_ah_params params = { 1818 .ah = ah->ah, 1819 .pdn = to_epd(ah->ibah.pd)->pdn, 1820 }; 1821 1822 return efa_com_destroy_ah(&dev->edev, ¶ms); 1823} 1824 1825int efa_create_ah(struct ib_ah *ibah, 1826 struct rdma_ah_init_attr *init_attr, 1827 struct ib_udata *udata) 1828{ 1829 struct rdma_ah_attr *ah_attr = init_attr->ah_attr; 1830 struct efa_dev *dev = to_edev(ibah->device); 1831 struct efa_com_create_ah_params params = {}; 1832 struct efa_ibv_create_ah_resp resp = {}; 1833 struct efa_com_create_ah_result result; 1834 struct efa_ah *ah = to_eah(ibah); 1835 int err; 1836 1837 if (!(init_attr->flags & RDMA_CREATE_AH_SLEEPABLE)) { 1838 ibdev_dbg(&dev->ibdev, 1839 "Create address handle is not supported in atomic context\n"); 1840 err = -EOPNOTSUPP; 1841 goto err_out; 1842 } 1843 1844 if (udata->inlen && 1845 !ib_is_udata_cleared(udata, 0, udata->inlen)) { 1846 ibdev_dbg(&dev->ibdev, "Incompatible ABI params\n"); 1847 err = -EINVAL; 1848 goto err_out; 1849 } 1850 1851 memcpy(params.dest_addr, ah_attr->grh.dgid.raw, 1852 sizeof(params.dest_addr)); 1853 params.pdn = to_epd(ibah->pd)->pdn; 1854 err = efa_com_create_ah(&dev->edev, ¶ms, &result); 1855 if (err) 1856 goto err_out; 1857 1858 memcpy(ah->id, ah_attr->grh.dgid.raw, sizeof(ah->id)); 1859 ah->ah = result.ah; 1860 1861 resp.efa_address_handle = result.ah; 1862 1863 if (udata->outlen) { 1864 err = ib_copy_to_udata(udata, &resp, 1865 min(sizeof(resp), udata->outlen)); 1866 if (err) { 1867 ibdev_dbg(&dev->ibdev, 1868 "Failed to copy udata for create_ah response\n"); 1869 goto err_destroy_ah; 1870 } 1871 } 1872 ibdev_dbg(&dev->ibdev, "Created ah[%d]\n", ah->ah); 1873 1874 return 0; 1875 1876err_destroy_ah: 1877 efa_ah_destroy(dev, ah); 1878err_out: 1879 atomic64_inc(&dev->stats.create_ah_err); 1880 return err; 1881} 1882 1883int efa_destroy_ah(struct ib_ah *ibah, u32 flags) 1884{ 1885 struct efa_dev *dev = to_edev(ibah->pd->device); 1886 struct efa_ah *ah = to_eah(ibah); 1887 1888 ibdev_dbg(&dev->ibdev, "Destroy ah[%d]\n", ah->ah); 1889 1890 if (!(flags & RDMA_DESTROY_AH_SLEEPABLE)) { 1891 ibdev_dbg(&dev->ibdev, 1892 "Destroy address handle is not supported in atomic context\n"); 1893 return -EOPNOTSUPP; 1894 } 1895 1896 efa_ah_destroy(dev, ah); 1897 return 0; 1898} 1899 1900struct rdma_hw_stats *efa_alloc_hw_stats(struct ib_device *ibdev, u8 port_num) 1901{ 1902 return rdma_alloc_hw_stats_struct(efa_stats_names, 1903 ARRAY_SIZE(efa_stats_names), 1904 RDMA_HW_STATS_DEFAULT_LIFESPAN); 1905} 1906 1907int efa_get_hw_stats(struct ib_device *ibdev, struct rdma_hw_stats *stats, 1908 u8 port_num, int index) 1909{ 1910 struct efa_com_get_stats_params params = {}; 1911 union efa_com_get_stats_result result; 1912 struct efa_dev *dev = to_edev(ibdev); 1913 struct efa_com_rdma_read_stats *rrs; 1914 struct efa_com_messages_stats *ms; 1915 struct efa_com_basic_stats *bs; 1916 struct efa_com_stats_admin *as; 1917 struct efa_stats *s; 1918 int err; 1919 1920 params.scope = EFA_ADMIN_GET_STATS_SCOPE_ALL; 1921 params.type = EFA_ADMIN_GET_STATS_TYPE_BASIC; 1922 1923 err = efa_com_get_stats(&dev->edev, ¶ms, &result); 1924 if (err) 1925 return err; 1926 1927 bs = &result.basic_stats; 1928 stats->value[EFA_TX_BYTES] = bs->tx_bytes; 1929 stats->value[EFA_TX_PKTS] = bs->tx_pkts; 1930 stats->value[EFA_RX_BYTES] = bs->rx_bytes; 1931 stats->value[EFA_RX_PKTS] = bs->rx_pkts; 1932 stats->value[EFA_RX_DROPS] = bs->rx_drops; 1933 1934 params.type = EFA_ADMIN_GET_STATS_TYPE_MESSAGES; 1935 err = efa_com_get_stats(&dev->edev, ¶ms, &result); 1936 if (err) 1937 return err; 1938 1939 ms = &result.messages_stats; 1940 stats->value[EFA_SEND_BYTES] = ms->send_bytes; 1941 stats->value[EFA_SEND_WRS] = ms->send_wrs; 1942 stats->value[EFA_RECV_BYTES] = ms->recv_bytes; 1943 stats->value[EFA_RECV_WRS] = ms->recv_wrs; 1944 1945 params.type = EFA_ADMIN_GET_STATS_TYPE_RDMA_READ; 1946 err = efa_com_get_stats(&dev->edev, ¶ms, &result); 1947 if (err) 1948 return err; 1949 1950 rrs = &result.rdma_read_stats; 1951 stats->value[EFA_RDMA_READ_WRS] = rrs->read_wrs; 1952 stats->value[EFA_RDMA_READ_BYTES] = rrs->read_bytes; 1953 stats->value[EFA_RDMA_READ_WR_ERR] = rrs->read_wr_err; 1954 stats->value[EFA_RDMA_READ_RESP_BYTES] = rrs->read_resp_bytes; 1955 1956 as = &dev->edev.aq.stats; 1957 stats->value[EFA_SUBMITTED_CMDS] = atomic64_read(&as->submitted_cmd); 1958 stats->value[EFA_COMPLETED_CMDS] = atomic64_read(&as->completed_cmd); 1959 stats->value[EFA_CMDS_ERR] = atomic64_read(&as->cmd_err); 1960 stats->value[EFA_NO_COMPLETION_CMDS] = atomic64_read(&as->no_completion); 1961 1962 s = &dev->stats; 1963 stats->value[EFA_KEEP_ALIVE_RCVD] = atomic64_read(&s->keep_alive_rcvd); 1964 stats->value[EFA_ALLOC_PD_ERR] = atomic64_read(&s->alloc_pd_err); 1965 stats->value[EFA_CREATE_QP_ERR] = atomic64_read(&s->create_qp_err); 1966 stats->value[EFA_CREATE_CQ_ERR] = atomic64_read(&s->create_cq_err); 1967 stats->value[EFA_REG_MR_ERR] = atomic64_read(&s->reg_mr_err); 1968 stats->value[EFA_ALLOC_UCONTEXT_ERR] = 1969 atomic64_read(&s->alloc_ucontext_err); 1970 stats->value[EFA_CREATE_AH_ERR] = atomic64_read(&s->create_ah_err); 1971 stats->value[EFA_MMAP_ERR] = atomic64_read(&s->mmap_err); 1972 1973 return ARRAY_SIZE(efa_stats_names); 1974} 1975 1976enum rdma_link_layer efa_port_link_layer(struct ib_device *ibdev, 1977 u8 port_num) 1978{ 1979 return IB_LINK_LAYER_UNSPECIFIED; 1980} 1981 1982