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