1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause 2 3 /* Authors: Cheng Xu <chengyou@linux.alibaba.com> */ 4 /* Kai Shen <kaishen@linux.alibaba.com> */ 5 /* Copyright (c) 2020-2022, Alibaba Group. */ 6 7 #include <linux/module.h> 8 #include <net/addrconf.h> 9 #include <rdma/erdma-abi.h> 10 11 #include "erdma.h" 12 #include "erdma_cm.h" 13 #include "erdma_verbs.h" 14 15 MODULE_AUTHOR("Cheng Xu <chengyou@linux.alibaba.com>"); 16 MODULE_DESCRIPTION("Alibaba elasticRDMA adapter driver"); 17 MODULE_LICENSE("Dual BSD/GPL"); 18 19 static int erdma_netdev_event(struct notifier_block *nb, unsigned long event, 20 void *arg) 21 { 22 struct net_device *netdev = netdev_notifier_info_to_dev(arg); 23 struct erdma_dev *dev = container_of(nb, struct erdma_dev, netdev_nb); 24 25 if (dev->netdev == NULL || dev->netdev != netdev) 26 goto done; 27 28 switch (event) { 29 case NETDEV_UP: 30 dev->state = IB_PORT_ACTIVE; 31 erdma_port_event(dev, IB_EVENT_PORT_ACTIVE); 32 break; 33 case NETDEV_DOWN: 34 dev->state = IB_PORT_DOWN; 35 erdma_port_event(dev, IB_EVENT_PORT_ERR); 36 break; 37 case NETDEV_CHANGEMTU: 38 if (dev->mtu != netdev->mtu) { 39 erdma_set_mtu(dev, netdev->mtu); 40 dev->mtu = netdev->mtu; 41 } 42 break; 43 case NETDEV_REGISTER: 44 case NETDEV_UNREGISTER: 45 case NETDEV_CHANGEADDR: 46 case NETDEV_GOING_DOWN: 47 case NETDEV_CHANGE: 48 default: 49 break; 50 } 51 52 done: 53 return NOTIFY_OK; 54 } 55 56 static int erdma_enum_and_get_netdev(struct erdma_dev *dev) 57 { 58 struct net_device *netdev; 59 int ret = -EPROBE_DEFER; 60 61 /* Already binded to a net_device, so we skip. */ 62 if (dev->netdev) 63 return 0; 64 65 rtnl_lock(); 66 for_each_netdev(&init_net, netdev) { 67 /* 68 * In erdma, the paired netdev and ibdev should have the same 69 * MAC address. erdma can get the value from its PCIe bar 70 * registers. Since erdma can not get the paired netdev 71 * reference directly, we do a traverse here to get the paired 72 * netdev. 73 */ 74 if (ether_addr_equal_unaligned(netdev->perm_addr, 75 dev->attrs.peer_addr)) { 76 ret = ib_device_set_netdev(&dev->ibdev, netdev, 1); 77 if (ret) { 78 rtnl_unlock(); 79 ibdev_warn(&dev->ibdev, 80 "failed (%d) to link netdev", ret); 81 return ret; 82 } 83 84 dev->netdev = netdev; 85 break; 86 } 87 } 88 89 rtnl_unlock(); 90 91 return ret; 92 } 93 94 static int erdma_device_register(struct erdma_dev *dev) 95 { 96 struct ib_device *ibdev = &dev->ibdev; 97 int ret; 98 99 ret = erdma_enum_and_get_netdev(dev); 100 if (ret) 101 return ret; 102 103 dev->mtu = dev->netdev->mtu; 104 addrconf_addr_eui48((u8 *)&ibdev->node_guid, dev->netdev->dev_addr); 105 106 ret = ib_register_device(ibdev, "erdma_%d", &dev->pdev->dev); 107 if (ret) { 108 dev_err(&dev->pdev->dev, 109 "ib_register_device failed: ret = %d\n", ret); 110 return ret; 111 } 112 113 dev->netdev_nb.notifier_call = erdma_netdev_event; 114 ret = register_netdevice_notifier(&dev->netdev_nb); 115 if (ret) { 116 ibdev_err(&dev->ibdev, "failed to register notifier.\n"); 117 ib_unregister_device(ibdev); 118 } 119 120 return ret; 121 } 122 123 static irqreturn_t erdma_comm_irq_handler(int irq, void *data) 124 { 125 struct erdma_dev *dev = data; 126 127 erdma_cmdq_completion_handler(&dev->cmdq); 128 erdma_aeq_event_handler(dev); 129 130 return IRQ_HANDLED; 131 } 132 133 static void erdma_dwqe_resource_init(struct erdma_dev *dev) 134 { 135 int total_pages, type0, type1; 136 137 dev->attrs.grp_num = erdma_reg_read32(dev, ERDMA_REGS_GRP_NUM_REG); 138 139 if (dev->attrs.grp_num < 4) 140 dev->attrs.disable_dwqe = true; 141 else 142 dev->attrs.disable_dwqe = false; 143 144 /* One page contains 4 goups. */ 145 total_pages = dev->attrs.grp_num * 4; 146 147 if (dev->attrs.grp_num >= ERDMA_DWQE_MAX_GRP_CNT) { 148 dev->attrs.grp_num = ERDMA_DWQE_MAX_GRP_CNT; 149 type0 = ERDMA_DWQE_TYPE0_CNT; 150 type1 = ERDMA_DWQE_TYPE1_CNT / ERDMA_DWQE_TYPE1_CNT_PER_PAGE; 151 } else { 152 type1 = total_pages / 3; 153 type0 = total_pages - type1 - 1; 154 } 155 156 dev->attrs.dwqe_pages = type0; 157 dev->attrs.dwqe_entries = type1 * ERDMA_DWQE_TYPE1_CNT_PER_PAGE; 158 } 159 160 static int erdma_request_vectors(struct erdma_dev *dev) 161 { 162 int expect_irq_num = min(num_possible_cpus() + 1, ERDMA_NUM_MSIX_VEC); 163 int ret; 164 165 ret = pci_alloc_irq_vectors(dev->pdev, 1, expect_irq_num, PCI_IRQ_MSIX); 166 if (ret < 0) { 167 dev_err(&dev->pdev->dev, "request irq vectors failed(%d)\n", 168 ret); 169 return ret; 170 } 171 dev->attrs.irq_num = ret; 172 173 return 0; 174 } 175 176 static int erdma_comm_irq_init(struct erdma_dev *dev) 177 { 178 snprintf(dev->comm_irq.name, ERDMA_IRQNAME_SIZE, "erdma-common@pci:%s", 179 pci_name(dev->pdev)); 180 dev->comm_irq.msix_vector = 181 pci_irq_vector(dev->pdev, ERDMA_MSIX_VECTOR_CMDQ); 182 183 cpumask_set_cpu(cpumask_first(cpumask_of_pcibus(dev->pdev->bus)), 184 &dev->comm_irq.affinity_hint_mask); 185 irq_set_affinity_hint(dev->comm_irq.msix_vector, 186 &dev->comm_irq.affinity_hint_mask); 187 188 return request_irq(dev->comm_irq.msix_vector, erdma_comm_irq_handler, 0, 189 dev->comm_irq.name, dev); 190 } 191 192 static void erdma_comm_irq_uninit(struct erdma_dev *dev) 193 { 194 irq_set_affinity_hint(dev->comm_irq.msix_vector, NULL); 195 free_irq(dev->comm_irq.msix_vector, dev); 196 } 197 198 static int erdma_device_init(struct erdma_dev *dev, struct pci_dev *pdev) 199 { 200 int ret; 201 202 erdma_dwqe_resource_init(dev); 203 204 ret = dma_set_mask_and_coherent(&pdev->dev, 205 DMA_BIT_MASK(ERDMA_PCI_WIDTH)); 206 if (ret) 207 return ret; 208 209 dma_set_max_seg_size(&pdev->dev, UINT_MAX); 210 211 return 0; 212 } 213 214 static void erdma_hw_reset(struct erdma_dev *dev) 215 { 216 u32 ctrl = FIELD_PREP(ERDMA_REG_DEV_CTRL_RESET_MASK, 1); 217 218 erdma_reg_write32(dev, ERDMA_REGS_DEV_CTRL_REG, ctrl); 219 } 220 221 static int erdma_wait_hw_init_done(struct erdma_dev *dev) 222 { 223 int i; 224 225 erdma_reg_write32(dev, ERDMA_REGS_DEV_CTRL_REG, 226 FIELD_PREP(ERDMA_REG_DEV_CTRL_INIT_MASK, 1)); 227 228 for (i = 0; i < ERDMA_WAIT_DEV_DONE_CNT; i++) { 229 if (erdma_reg_read32_filed(dev, ERDMA_REGS_DEV_ST_REG, 230 ERDMA_REG_DEV_ST_INIT_DONE_MASK)) 231 break; 232 233 msleep(ERDMA_REG_ACCESS_WAIT_MS); 234 } 235 236 if (i == ERDMA_WAIT_DEV_DONE_CNT) { 237 dev_err(&dev->pdev->dev, "wait init done failed.\n"); 238 return -ETIMEDOUT; 239 } 240 241 return 0; 242 } 243 244 static const struct pci_device_id erdma_pci_tbl[] = { 245 { PCI_DEVICE(PCI_VENDOR_ID_ALIBABA, 0x107f) }, 246 {} 247 }; 248 249 static int erdma_probe_dev(struct pci_dev *pdev) 250 { 251 struct erdma_dev *dev; 252 int bars, err; 253 u32 version; 254 255 err = pci_enable_device(pdev); 256 if (err) { 257 dev_err(&pdev->dev, "pci_enable_device failed(%d)\n", err); 258 return err; 259 } 260 261 pci_set_master(pdev); 262 263 dev = ib_alloc_device(erdma_dev, ibdev); 264 if (!dev) { 265 dev_err(&pdev->dev, "ib_alloc_device failed\n"); 266 err = -ENOMEM; 267 goto err_disable_device; 268 } 269 270 pci_set_drvdata(pdev, dev); 271 dev->pdev = pdev; 272 dev->attrs.numa_node = dev_to_node(&pdev->dev); 273 274 bars = pci_select_bars(pdev, IORESOURCE_MEM); 275 err = pci_request_selected_regions(pdev, bars, DRV_MODULE_NAME); 276 if (bars != ERDMA_BAR_MASK || err) { 277 err = err ? err : -EINVAL; 278 goto err_ib_device_release; 279 } 280 281 dev->func_bar_addr = pci_resource_start(pdev, ERDMA_FUNC_BAR); 282 dev->func_bar_len = pci_resource_len(pdev, ERDMA_FUNC_BAR); 283 284 dev->func_bar = 285 devm_ioremap(&pdev->dev, dev->func_bar_addr, dev->func_bar_len); 286 if (!dev->func_bar) { 287 dev_err(&pdev->dev, "devm_ioremap failed.\n"); 288 err = -EFAULT; 289 goto err_release_bars; 290 } 291 292 version = erdma_reg_read32(dev, ERDMA_REGS_VERSION_REG); 293 if (version == 0) { 294 /* we knows that it is a non-functional function. */ 295 err = -ENODEV; 296 goto err_iounmap_func_bar; 297 } 298 299 err = erdma_device_init(dev, pdev); 300 if (err) 301 goto err_iounmap_func_bar; 302 303 err = erdma_request_vectors(dev); 304 if (err) 305 goto err_iounmap_func_bar; 306 307 err = erdma_comm_irq_init(dev); 308 if (err) 309 goto err_free_vectors; 310 311 err = erdma_aeq_init(dev); 312 if (err) 313 goto err_uninit_comm_irq; 314 315 err = erdma_cmdq_init(dev); 316 if (err) 317 goto err_uninit_aeq; 318 319 err = erdma_wait_hw_init_done(dev); 320 if (err) 321 goto err_uninit_cmdq; 322 323 err = erdma_ceqs_init(dev); 324 if (err) 325 goto err_reset_hw; 326 327 erdma_finish_cmdq_init(dev); 328 329 return 0; 330 331 err_reset_hw: 332 erdma_hw_reset(dev); 333 334 err_uninit_cmdq: 335 erdma_cmdq_destroy(dev); 336 337 err_uninit_aeq: 338 erdma_aeq_destroy(dev); 339 340 err_uninit_comm_irq: 341 erdma_comm_irq_uninit(dev); 342 343 err_free_vectors: 344 pci_free_irq_vectors(dev->pdev); 345 346 err_iounmap_func_bar: 347 devm_iounmap(&pdev->dev, dev->func_bar); 348 349 err_release_bars: 350 pci_release_selected_regions(pdev, bars); 351 352 err_ib_device_release: 353 ib_dealloc_device(&dev->ibdev); 354 355 err_disable_device: 356 pci_disable_device(pdev); 357 358 return err; 359 } 360 361 static void erdma_remove_dev(struct pci_dev *pdev) 362 { 363 struct erdma_dev *dev = pci_get_drvdata(pdev); 364 365 erdma_ceqs_uninit(dev); 366 erdma_hw_reset(dev); 367 erdma_cmdq_destroy(dev); 368 erdma_aeq_destroy(dev); 369 erdma_comm_irq_uninit(dev); 370 pci_free_irq_vectors(dev->pdev); 371 372 devm_iounmap(&pdev->dev, dev->func_bar); 373 pci_release_selected_regions(pdev, ERDMA_BAR_MASK); 374 375 ib_dealloc_device(&dev->ibdev); 376 377 pci_disable_device(pdev); 378 } 379 380 #define ERDMA_GET_CAP(name, cap) FIELD_GET(ERDMA_CMD_DEV_CAP_##name##_MASK, cap) 381 382 static int erdma_dev_attrs_init(struct erdma_dev *dev) 383 { 384 int err; 385 u64 req_hdr, cap0, cap1; 386 387 erdma_cmdq_build_reqhdr(&req_hdr, CMDQ_SUBMOD_RDMA, 388 CMDQ_OPCODE_QUERY_DEVICE); 389 390 err = erdma_post_cmd_wait(&dev->cmdq, &req_hdr, sizeof(req_hdr), &cap0, 391 &cap1); 392 if (err) 393 return err; 394 395 dev->attrs.max_cqe = 1 << ERDMA_GET_CAP(MAX_CQE, cap0); 396 dev->attrs.max_mr_size = 1ULL << ERDMA_GET_CAP(MAX_MR_SIZE, cap0); 397 dev->attrs.max_mw = 1 << ERDMA_GET_CAP(MAX_MW, cap1); 398 dev->attrs.max_recv_wr = 1 << ERDMA_GET_CAP(MAX_RECV_WR, cap0); 399 dev->attrs.local_dma_key = ERDMA_GET_CAP(DMA_LOCAL_KEY, cap1); 400 dev->attrs.cc = ERDMA_GET_CAP(DEFAULT_CC, cap1); 401 dev->attrs.max_qp = ERDMA_NQP_PER_QBLOCK * ERDMA_GET_CAP(QBLOCK, cap1); 402 dev->attrs.max_mr = dev->attrs.max_qp << 1; 403 dev->attrs.max_cq = dev->attrs.max_qp << 1; 404 dev->attrs.cap_flags = ERDMA_GET_CAP(FLAGS, cap0); 405 406 dev->attrs.max_send_wr = ERDMA_MAX_SEND_WR; 407 dev->attrs.max_ord = ERDMA_MAX_ORD; 408 dev->attrs.max_ird = ERDMA_MAX_IRD; 409 dev->attrs.max_send_sge = ERDMA_MAX_SEND_SGE; 410 dev->attrs.max_recv_sge = ERDMA_MAX_RECV_SGE; 411 dev->attrs.max_sge_rd = ERDMA_MAX_SGE_RD; 412 dev->attrs.max_pd = ERDMA_MAX_PD; 413 414 dev->res_cb[ERDMA_RES_TYPE_PD].max_cap = ERDMA_MAX_PD; 415 dev->res_cb[ERDMA_RES_TYPE_STAG_IDX].max_cap = dev->attrs.max_mr; 416 417 erdma_cmdq_build_reqhdr(&req_hdr, CMDQ_SUBMOD_COMMON, 418 CMDQ_OPCODE_QUERY_FW_INFO); 419 420 err = erdma_post_cmd_wait(&dev->cmdq, &req_hdr, sizeof(req_hdr), &cap0, 421 &cap1); 422 if (!err) 423 dev->attrs.fw_version = 424 FIELD_GET(ERDMA_CMD_INFO0_FW_VER_MASK, cap0); 425 426 return err; 427 } 428 429 static int erdma_res_cb_init(struct erdma_dev *dev) 430 { 431 int i, j; 432 433 for (i = 0; i < ERDMA_RES_CNT; i++) { 434 dev->res_cb[i].next_alloc_idx = 1; 435 spin_lock_init(&dev->res_cb[i].lock); 436 dev->res_cb[i].bitmap = 437 bitmap_zalloc(dev->res_cb[i].max_cap, GFP_KERNEL); 438 if (!dev->res_cb[i].bitmap) 439 goto err; 440 } 441 442 return 0; 443 444 err: 445 for (j = 0; j < i; j++) 446 bitmap_free(dev->res_cb[j].bitmap); 447 448 return -ENOMEM; 449 } 450 451 static void erdma_res_cb_free(struct erdma_dev *dev) 452 { 453 int i; 454 455 for (i = 0; i < ERDMA_RES_CNT; i++) 456 bitmap_free(dev->res_cb[i].bitmap); 457 } 458 459 static const struct ib_device_ops erdma_device_ops = { 460 .owner = THIS_MODULE, 461 .driver_id = RDMA_DRIVER_ERDMA, 462 .uverbs_abi_ver = ERDMA_ABI_VERSION, 463 464 .alloc_mr = erdma_ib_alloc_mr, 465 .alloc_pd = erdma_alloc_pd, 466 .alloc_ucontext = erdma_alloc_ucontext, 467 .create_cq = erdma_create_cq, 468 .create_qp = erdma_create_qp, 469 .dealloc_pd = erdma_dealloc_pd, 470 .dealloc_ucontext = erdma_dealloc_ucontext, 471 .dereg_mr = erdma_dereg_mr, 472 .destroy_cq = erdma_destroy_cq, 473 .destroy_qp = erdma_destroy_qp, 474 .get_dma_mr = erdma_get_dma_mr, 475 .get_port_immutable = erdma_get_port_immutable, 476 .iw_accept = erdma_accept, 477 .iw_add_ref = erdma_qp_get_ref, 478 .iw_connect = erdma_connect, 479 .iw_create_listen = erdma_create_listen, 480 .iw_destroy_listen = erdma_destroy_listen, 481 .iw_get_qp = erdma_get_ibqp, 482 .iw_reject = erdma_reject, 483 .iw_rem_ref = erdma_qp_put_ref, 484 .map_mr_sg = erdma_map_mr_sg, 485 .mmap = erdma_mmap, 486 .mmap_free = erdma_mmap_free, 487 .modify_qp = erdma_modify_qp, 488 .post_recv = erdma_post_recv, 489 .post_send = erdma_post_send, 490 .poll_cq = erdma_poll_cq, 491 .query_device = erdma_query_device, 492 .query_gid = erdma_query_gid, 493 .query_port = erdma_query_port, 494 .query_qp = erdma_query_qp, 495 .req_notify_cq = erdma_req_notify_cq, 496 .reg_user_mr = erdma_reg_user_mr, 497 498 INIT_RDMA_OBJ_SIZE(ib_cq, erdma_cq, ibcq), 499 INIT_RDMA_OBJ_SIZE(ib_pd, erdma_pd, ibpd), 500 INIT_RDMA_OBJ_SIZE(ib_ucontext, erdma_ucontext, ibucontext), 501 INIT_RDMA_OBJ_SIZE(ib_qp, erdma_qp, ibqp), 502 }; 503 504 static int erdma_ib_device_add(struct pci_dev *pdev) 505 { 506 struct erdma_dev *dev = pci_get_drvdata(pdev); 507 struct ib_device *ibdev = &dev->ibdev; 508 u64 mac; 509 int ret; 510 511 ret = erdma_dev_attrs_init(dev); 512 if (ret) 513 return ret; 514 515 ibdev->node_type = RDMA_NODE_RNIC; 516 memcpy(ibdev->node_desc, ERDMA_NODE_DESC, sizeof(ERDMA_NODE_DESC)); 517 518 /* 519 * Current model (one-to-one device association): 520 * One ERDMA device per net_device or, equivalently, 521 * per physical port. 522 */ 523 ibdev->phys_port_cnt = 1; 524 ibdev->num_comp_vectors = dev->attrs.irq_num - 1; 525 526 ib_set_device_ops(ibdev, &erdma_device_ops); 527 528 INIT_LIST_HEAD(&dev->cep_list); 529 530 spin_lock_init(&dev->lock); 531 xa_init_flags(&dev->qp_xa, XA_FLAGS_ALLOC1); 532 xa_init_flags(&dev->cq_xa, XA_FLAGS_ALLOC1); 533 dev->next_alloc_cqn = 1; 534 dev->next_alloc_qpn = 1; 535 536 ret = erdma_res_cb_init(dev); 537 if (ret) 538 return ret; 539 540 spin_lock_init(&dev->db_bitmap_lock); 541 bitmap_zero(dev->sdb_page, ERDMA_DWQE_TYPE0_CNT); 542 bitmap_zero(dev->sdb_entry, ERDMA_DWQE_TYPE1_CNT); 543 544 atomic_set(&dev->num_ctx, 0); 545 546 mac = erdma_reg_read32(dev, ERDMA_REGS_NETDEV_MAC_L_REG); 547 mac |= (u64)erdma_reg_read32(dev, ERDMA_REGS_NETDEV_MAC_H_REG) << 32; 548 549 u64_to_ether_addr(mac, dev->attrs.peer_addr); 550 551 dev->reflush_wq = alloc_workqueue("erdma-reflush-wq", WQ_UNBOUND, 552 WQ_UNBOUND_MAX_ACTIVE); 553 if (!dev->reflush_wq) { 554 ret = -ENOMEM; 555 goto err_alloc_workqueue; 556 } 557 558 ret = erdma_device_register(dev); 559 if (ret) 560 goto err_register; 561 562 return 0; 563 564 err_register: 565 destroy_workqueue(dev->reflush_wq); 566 err_alloc_workqueue: 567 xa_destroy(&dev->qp_xa); 568 xa_destroy(&dev->cq_xa); 569 570 erdma_res_cb_free(dev); 571 572 return ret; 573 } 574 575 static void erdma_ib_device_remove(struct pci_dev *pdev) 576 { 577 struct erdma_dev *dev = pci_get_drvdata(pdev); 578 579 unregister_netdevice_notifier(&dev->netdev_nb); 580 ib_unregister_device(&dev->ibdev); 581 582 destroy_workqueue(dev->reflush_wq); 583 erdma_res_cb_free(dev); 584 xa_destroy(&dev->qp_xa); 585 xa_destroy(&dev->cq_xa); 586 } 587 588 static int erdma_probe(struct pci_dev *pdev, const struct pci_device_id *ent) 589 { 590 int ret; 591 592 ret = erdma_probe_dev(pdev); 593 if (ret) 594 return ret; 595 596 ret = erdma_ib_device_add(pdev); 597 if (ret) { 598 erdma_remove_dev(pdev); 599 return ret; 600 } 601 602 return 0; 603 } 604 605 static void erdma_remove(struct pci_dev *pdev) 606 { 607 erdma_ib_device_remove(pdev); 608 erdma_remove_dev(pdev); 609 } 610 611 static struct pci_driver erdma_pci_driver = { 612 .name = DRV_MODULE_NAME, 613 .id_table = erdma_pci_tbl, 614 .probe = erdma_probe, 615 .remove = erdma_remove 616 }; 617 618 MODULE_DEVICE_TABLE(pci, erdma_pci_tbl); 619 620 static __init int erdma_init_module(void) 621 { 622 int ret; 623 624 ret = erdma_cm_init(); 625 if (ret) 626 return ret; 627 628 ret = pci_register_driver(&erdma_pci_driver); 629 if (ret) 630 erdma_cm_exit(); 631 632 return ret; 633 } 634 635 static void __exit erdma_exit_module(void) 636 { 637 pci_unregister_driver(&erdma_pci_driver); 638 639 erdma_cm_exit(); 640 } 641 642 module_init(erdma_init_module); 643 module_exit(erdma_exit_module); 644