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 = -ENODEV; 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_device_uninit(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 const struct pci_device_id erdma_pci_tbl[] = { 222 { PCI_DEVICE(PCI_VENDOR_ID_ALIBABA, 0x107f) }, 223 {} 224 }; 225 226 static int erdma_probe_dev(struct pci_dev *pdev) 227 { 228 struct erdma_dev *dev; 229 int bars, err; 230 u32 version; 231 232 err = pci_enable_device(pdev); 233 if (err) { 234 dev_err(&pdev->dev, "pci_enable_device failed(%d)\n", err); 235 return err; 236 } 237 238 pci_set_master(pdev); 239 240 dev = ib_alloc_device(erdma_dev, ibdev); 241 if (!dev) { 242 dev_err(&pdev->dev, "ib_alloc_device failed\n"); 243 err = -ENOMEM; 244 goto err_disable_device; 245 } 246 247 pci_set_drvdata(pdev, dev); 248 dev->pdev = pdev; 249 dev->attrs.numa_node = dev_to_node(&pdev->dev); 250 251 bars = pci_select_bars(pdev, IORESOURCE_MEM); 252 err = pci_request_selected_regions(pdev, bars, DRV_MODULE_NAME); 253 if (bars != ERDMA_BAR_MASK || err) { 254 err = err ? err : -EINVAL; 255 goto err_ib_device_release; 256 } 257 258 dev->func_bar_addr = pci_resource_start(pdev, ERDMA_FUNC_BAR); 259 dev->func_bar_len = pci_resource_len(pdev, ERDMA_FUNC_BAR); 260 261 dev->func_bar = 262 devm_ioremap(&pdev->dev, dev->func_bar_addr, dev->func_bar_len); 263 if (!dev->func_bar) { 264 dev_err(&pdev->dev, "devm_ioremap failed.\n"); 265 err = -EFAULT; 266 goto err_release_bars; 267 } 268 269 version = erdma_reg_read32(dev, ERDMA_REGS_VERSION_REG); 270 if (version == 0) { 271 /* we knows that it is a non-functional function. */ 272 err = -ENODEV; 273 goto err_iounmap_func_bar; 274 } 275 276 err = erdma_device_init(dev, pdev); 277 if (err) 278 goto err_iounmap_func_bar; 279 280 err = erdma_request_vectors(dev); 281 if (err) 282 goto err_iounmap_func_bar; 283 284 err = erdma_comm_irq_init(dev); 285 if (err) 286 goto err_free_vectors; 287 288 err = erdma_aeq_init(dev); 289 if (err) 290 goto err_uninit_comm_irq; 291 292 err = erdma_cmdq_init(dev); 293 if (err) 294 goto err_uninit_aeq; 295 296 err = erdma_ceqs_init(dev); 297 if (err) 298 goto err_uninit_cmdq; 299 300 erdma_finish_cmdq_init(dev); 301 302 return 0; 303 304 err_uninit_cmdq: 305 erdma_device_uninit(dev); 306 erdma_cmdq_destroy(dev); 307 308 err_uninit_aeq: 309 erdma_aeq_destroy(dev); 310 311 err_uninit_comm_irq: 312 erdma_comm_irq_uninit(dev); 313 314 err_free_vectors: 315 pci_free_irq_vectors(dev->pdev); 316 317 err_iounmap_func_bar: 318 devm_iounmap(&pdev->dev, dev->func_bar); 319 320 err_release_bars: 321 pci_release_selected_regions(pdev, bars); 322 323 err_ib_device_release: 324 ib_dealloc_device(&dev->ibdev); 325 326 err_disable_device: 327 pci_disable_device(pdev); 328 329 return err; 330 } 331 332 static void erdma_remove_dev(struct pci_dev *pdev) 333 { 334 struct erdma_dev *dev = pci_get_drvdata(pdev); 335 336 erdma_ceqs_uninit(dev); 337 338 erdma_device_uninit(dev); 339 340 erdma_cmdq_destroy(dev); 341 erdma_aeq_destroy(dev); 342 erdma_comm_irq_uninit(dev); 343 pci_free_irq_vectors(dev->pdev); 344 345 devm_iounmap(&pdev->dev, dev->func_bar); 346 pci_release_selected_regions(pdev, ERDMA_BAR_MASK); 347 348 ib_dealloc_device(&dev->ibdev); 349 350 pci_disable_device(pdev); 351 } 352 353 #define ERDMA_GET_CAP(name, cap) FIELD_GET(ERDMA_CMD_DEV_CAP_##name##_MASK, cap) 354 355 static int erdma_dev_attrs_init(struct erdma_dev *dev) 356 { 357 int err; 358 u64 req_hdr, cap0, cap1; 359 360 erdma_cmdq_build_reqhdr(&req_hdr, CMDQ_SUBMOD_RDMA, 361 CMDQ_OPCODE_QUERY_DEVICE); 362 363 err = erdma_post_cmd_wait(&dev->cmdq, &req_hdr, sizeof(req_hdr), &cap0, 364 &cap1); 365 if (err) 366 return err; 367 368 dev->attrs.max_cqe = 1 << ERDMA_GET_CAP(MAX_CQE, cap0); 369 dev->attrs.max_mr_size = 1ULL << ERDMA_GET_CAP(MAX_MR_SIZE, cap0); 370 dev->attrs.max_mw = 1 << ERDMA_GET_CAP(MAX_MW, cap1); 371 dev->attrs.max_recv_wr = 1 << ERDMA_GET_CAP(MAX_RECV_WR, cap0); 372 dev->attrs.local_dma_key = ERDMA_GET_CAP(DMA_LOCAL_KEY, cap1); 373 dev->attrs.cc = ERDMA_GET_CAP(DEFAULT_CC, cap1); 374 dev->attrs.max_qp = ERDMA_NQP_PER_QBLOCK * ERDMA_GET_CAP(QBLOCK, cap1); 375 dev->attrs.max_mr = dev->attrs.max_qp << 1; 376 dev->attrs.max_cq = dev->attrs.max_qp << 1; 377 dev->attrs.cap_flags = ERDMA_GET_CAP(FLAGS, cap0); 378 379 dev->attrs.max_send_wr = ERDMA_MAX_SEND_WR; 380 dev->attrs.max_ord = ERDMA_MAX_ORD; 381 dev->attrs.max_ird = ERDMA_MAX_IRD; 382 dev->attrs.max_send_sge = ERDMA_MAX_SEND_SGE; 383 dev->attrs.max_recv_sge = ERDMA_MAX_RECV_SGE; 384 dev->attrs.max_sge_rd = ERDMA_MAX_SGE_RD; 385 dev->attrs.max_pd = ERDMA_MAX_PD; 386 387 dev->res_cb[ERDMA_RES_TYPE_PD].max_cap = ERDMA_MAX_PD; 388 dev->res_cb[ERDMA_RES_TYPE_STAG_IDX].max_cap = dev->attrs.max_mr; 389 390 erdma_cmdq_build_reqhdr(&req_hdr, CMDQ_SUBMOD_COMMON, 391 CMDQ_OPCODE_QUERY_FW_INFO); 392 393 err = erdma_post_cmd_wait(&dev->cmdq, &req_hdr, sizeof(req_hdr), &cap0, 394 &cap1); 395 if (!err) 396 dev->attrs.fw_version = 397 FIELD_GET(ERDMA_CMD_INFO0_FW_VER_MASK, cap0); 398 399 return err; 400 } 401 402 static int erdma_res_cb_init(struct erdma_dev *dev) 403 { 404 int i, j; 405 406 for (i = 0; i < ERDMA_RES_CNT; i++) { 407 dev->res_cb[i].next_alloc_idx = 1; 408 spin_lock_init(&dev->res_cb[i].lock); 409 dev->res_cb[i].bitmap = 410 bitmap_zalloc(dev->res_cb[i].max_cap, GFP_KERNEL); 411 if (!dev->res_cb[i].bitmap) 412 goto err; 413 } 414 415 return 0; 416 417 err: 418 for (j = 0; j < i; j++) 419 bitmap_free(dev->res_cb[j].bitmap); 420 421 return -ENOMEM; 422 } 423 424 static void erdma_res_cb_free(struct erdma_dev *dev) 425 { 426 int i; 427 428 for (i = 0; i < ERDMA_RES_CNT; i++) 429 bitmap_free(dev->res_cb[i].bitmap); 430 } 431 432 static const struct ib_device_ops erdma_device_ops = { 433 .owner = THIS_MODULE, 434 .driver_id = RDMA_DRIVER_ERDMA, 435 .uverbs_abi_ver = ERDMA_ABI_VERSION, 436 437 .alloc_mr = erdma_ib_alloc_mr, 438 .alloc_pd = erdma_alloc_pd, 439 .alloc_ucontext = erdma_alloc_ucontext, 440 .create_cq = erdma_create_cq, 441 .create_qp = erdma_create_qp, 442 .dealloc_pd = erdma_dealloc_pd, 443 .dealloc_ucontext = erdma_dealloc_ucontext, 444 .dereg_mr = erdma_dereg_mr, 445 .destroy_cq = erdma_destroy_cq, 446 .destroy_qp = erdma_destroy_qp, 447 .get_dma_mr = erdma_get_dma_mr, 448 .get_port_immutable = erdma_get_port_immutable, 449 .iw_accept = erdma_accept, 450 .iw_add_ref = erdma_qp_get_ref, 451 .iw_connect = erdma_connect, 452 .iw_create_listen = erdma_create_listen, 453 .iw_destroy_listen = erdma_destroy_listen, 454 .iw_get_qp = erdma_get_ibqp, 455 .iw_reject = erdma_reject, 456 .iw_rem_ref = erdma_qp_put_ref, 457 .map_mr_sg = erdma_map_mr_sg, 458 .mmap = erdma_mmap, 459 .mmap_free = erdma_mmap_free, 460 .modify_qp = erdma_modify_qp, 461 .post_recv = erdma_post_recv, 462 .post_send = erdma_post_send, 463 .poll_cq = erdma_poll_cq, 464 .query_device = erdma_query_device, 465 .query_gid = erdma_query_gid, 466 .query_port = erdma_query_port, 467 .query_qp = erdma_query_qp, 468 .req_notify_cq = erdma_req_notify_cq, 469 .reg_user_mr = erdma_reg_user_mr, 470 471 INIT_RDMA_OBJ_SIZE(ib_cq, erdma_cq, ibcq), 472 INIT_RDMA_OBJ_SIZE(ib_pd, erdma_pd, ibpd), 473 INIT_RDMA_OBJ_SIZE(ib_ucontext, erdma_ucontext, ibucontext), 474 INIT_RDMA_OBJ_SIZE(ib_qp, erdma_qp, ibqp), 475 }; 476 477 static int erdma_ib_device_add(struct pci_dev *pdev) 478 { 479 struct erdma_dev *dev = pci_get_drvdata(pdev); 480 struct ib_device *ibdev = &dev->ibdev; 481 u64 mac; 482 int ret; 483 484 ret = erdma_dev_attrs_init(dev); 485 if (ret) 486 return ret; 487 488 ibdev->node_type = RDMA_NODE_RNIC; 489 memcpy(ibdev->node_desc, ERDMA_NODE_DESC, sizeof(ERDMA_NODE_DESC)); 490 491 /* 492 * Current model (one-to-one device association): 493 * One ERDMA device per net_device or, equivalently, 494 * per physical port. 495 */ 496 ibdev->phys_port_cnt = 1; 497 ibdev->num_comp_vectors = dev->attrs.irq_num - 1; 498 499 ib_set_device_ops(ibdev, &erdma_device_ops); 500 501 INIT_LIST_HEAD(&dev->cep_list); 502 503 spin_lock_init(&dev->lock); 504 xa_init_flags(&dev->qp_xa, XA_FLAGS_ALLOC1); 505 xa_init_flags(&dev->cq_xa, XA_FLAGS_ALLOC1); 506 dev->next_alloc_cqn = 1; 507 dev->next_alloc_qpn = 1; 508 509 ret = erdma_res_cb_init(dev); 510 if (ret) 511 return ret; 512 513 spin_lock_init(&dev->db_bitmap_lock); 514 bitmap_zero(dev->sdb_page, ERDMA_DWQE_TYPE0_CNT); 515 bitmap_zero(dev->sdb_entry, ERDMA_DWQE_TYPE1_CNT); 516 517 atomic_set(&dev->num_ctx, 0); 518 519 mac = erdma_reg_read32(dev, ERDMA_REGS_NETDEV_MAC_L_REG); 520 mac |= (u64)erdma_reg_read32(dev, ERDMA_REGS_NETDEV_MAC_H_REG) << 32; 521 522 u64_to_ether_addr(mac, dev->attrs.peer_addr); 523 524 dev->reflush_wq = alloc_workqueue("erdma-reflush-wq", WQ_UNBOUND, 525 WQ_UNBOUND_MAX_ACTIVE); 526 if (!dev->reflush_wq) { 527 ret = -ENOMEM; 528 goto err_alloc_workqueue; 529 } 530 531 ret = erdma_device_register(dev); 532 if (ret) 533 goto err_register; 534 535 return 0; 536 537 err_register: 538 destroy_workqueue(dev->reflush_wq); 539 err_alloc_workqueue: 540 xa_destroy(&dev->qp_xa); 541 xa_destroy(&dev->cq_xa); 542 543 erdma_res_cb_free(dev); 544 545 return ret; 546 } 547 548 static void erdma_ib_device_remove(struct pci_dev *pdev) 549 { 550 struct erdma_dev *dev = pci_get_drvdata(pdev); 551 552 unregister_netdevice_notifier(&dev->netdev_nb); 553 ib_unregister_device(&dev->ibdev); 554 555 destroy_workqueue(dev->reflush_wq); 556 erdma_res_cb_free(dev); 557 xa_destroy(&dev->qp_xa); 558 xa_destroy(&dev->cq_xa); 559 } 560 561 static int erdma_probe(struct pci_dev *pdev, const struct pci_device_id *ent) 562 { 563 int ret; 564 565 ret = erdma_probe_dev(pdev); 566 if (ret) 567 return ret; 568 569 ret = erdma_ib_device_add(pdev); 570 if (ret) { 571 erdma_remove_dev(pdev); 572 return ret; 573 } 574 575 return 0; 576 } 577 578 static void erdma_remove(struct pci_dev *pdev) 579 { 580 erdma_ib_device_remove(pdev); 581 erdma_remove_dev(pdev); 582 } 583 584 static struct pci_driver erdma_pci_driver = { 585 .name = DRV_MODULE_NAME, 586 .id_table = erdma_pci_tbl, 587 .probe = erdma_probe, 588 .remove = erdma_remove 589 }; 590 591 MODULE_DEVICE_TABLE(pci, erdma_pci_tbl); 592 593 static __init int erdma_init_module(void) 594 { 595 int ret; 596 597 ret = erdma_cm_init(); 598 if (ret) 599 return ret; 600 601 ret = pci_register_driver(&erdma_pci_driver); 602 if (ret) 603 erdma_cm_exit(); 604 605 return ret; 606 } 607 608 static void __exit erdma_exit_module(void) 609 { 610 pci_unregister_driver(&erdma_pci_driver); 611 612 erdma_cm_exit(); 613 } 614 615 module_init(erdma_init_module); 616 module_exit(erdma_exit_module); 617