drivers/infiniband/hw/efa/efa_main.c

   1 // SPDX-License-Identifier: GPL-2.0 OR BSD-2-Clause
   2 /*
   3  * Copyright 2018-2021 Amazon.com, Inc. or its affiliates. All rights reserved.
   4  */
   5
   6 #include <linux/module.h>
   7 #include <linux/pci.h>
   8 #include <linux/utsname.h>
   9 #include <linux/version.h>
  10
  11 #include <rdma/ib_user_verbs.h>
  12
  13 #include "efa.h"
  14
  15 #define PCI_DEV_ID_EFA0_VF 0xefa0
  16 #define PCI_DEV_ID_EFA1_VF 0xefa1
  17
  18 static const struct pci_device_id efa_pci_tbl[] = {
  19         { PCI_VDEVICE(AMAZON, PCI_DEV_ID_EFA0_VF) },
  20         { PCI_VDEVICE(AMAZON, PCI_DEV_ID_EFA1_VF) },
  21         { }
  22 };
  23
  24 MODULE_AUTHOR("Amazon.com, Inc. or its affiliates");
  25 MODULE_LICENSE("Dual BSD/GPL");
  26 MODULE_DESCRIPTION(DEVICE_NAME);
  27 MODULE_DEVICE_TABLE(pci, efa_pci_tbl);
  28
  29 #define EFA_REG_BAR 0
  30 #define EFA_MEM_BAR 2
  31 #define EFA_BASE_BAR_MASK (BIT(EFA_REG_BAR) | BIT(EFA_MEM_BAR))
  32
  33 #define EFA_AENQ_ENABLED_GROUPS \
  34         (BIT(EFA_ADMIN_FATAL_ERROR) | BIT(EFA_ADMIN_WARNING) | \
  35          BIT(EFA_ADMIN_NOTIFICATION) | BIT(EFA_ADMIN_KEEP_ALIVE))
  36
  37 /* This handler will called for unknown event group or unimplemented handlers */
  38 static void unimplemented_aenq_handler(void *data,
  39                                        struct efa_admin_aenq_entry *aenq_e)
  40 {
  41         struct efa_dev *dev = (struct efa_dev *)data;
  42
  43         ibdev_err(&dev->ibdev,
  44                   "Unknown event was received or event with unimplemented handler\n");
  45 }
  46
  47 static void efa_keep_alive(void *data, struct efa_admin_aenq_entry *aenq_e)
  48 {
  49         struct efa_dev *dev = (struct efa_dev *)data;
  50
  51         atomic64_inc(&dev->stats.keep_alive_rcvd);
  52 }
  53
  54 static struct efa_aenq_handlers aenq_handlers = {
  55         .handlers = {
  56                 [EFA_ADMIN_KEEP_ALIVE] = efa_keep_alive,
  57         },
  58         .unimplemented_handler = unimplemented_aenq_handler
  59 };
  60
  61 static void efa_release_bars(struct efa_dev *dev, int bars_mask)
  62 {
  63         struct pci_dev *pdev = dev->pdev;
  64         int release_bars;
  65
  66         release_bars = pci_select_bars(pdev, IORESOURCE_MEM) & bars_mask;
  67         pci_release_selected_regions(pdev, release_bars);
  68 }
  69
  70 static irqreturn_t efa_intr_msix_mgmnt(int irq, void *data)
  71 {
  72         struct efa_dev *dev = data;
  73
  74         efa_com_admin_q_comp_intr_handler(&dev->edev);
  75         efa_com_aenq_intr_handler(&dev->edev, data);
  76
  77         return IRQ_HANDLED;
  78 }
  79
  80 static int efa_request_mgmnt_irq(struct efa_dev *dev)
  81 {
  82         struct efa_irq *irq;
  83         int err;
  84
  85         irq = &dev->admin_irq;
  86         err = request_irq(irq->vector, irq->handler, 0, irq->name,
  87                           irq->data);
  88         if (err) {
  89                 dev_err(&dev->pdev->dev, "Failed to request admin irq (%d)\n",
  90                         err);
  91                 return err;
  92         }
  93
  94         dev_dbg(&dev->pdev->dev, "Set affinity hint of mgmnt irq to %*pbl (irq vector: %d)\n",
  95                 nr_cpumask_bits, &irq->affinity_hint_mask, irq->vector);
  96         irq_set_affinity_hint(irq->vector, &irq->affinity_hint_mask);
  97
  98         return 0;
  99 }
 100
 101 static void efa_setup_mgmnt_irq(struct efa_dev *dev)
 102 {
 103         u32 cpu;
 104
 105         snprintf(dev->admin_irq.name, EFA_IRQNAME_SIZE,
 106                  "efa-mgmnt@pci:%s", pci_name(dev->pdev));
 107         dev->admin_irq.handler = efa_intr_msix_mgmnt;
 108         dev->admin_irq.data = dev;
 109         dev->admin_irq.vector =
 110                 pci_irq_vector(dev->pdev, dev->admin_msix_vector_idx);
 111         cpu = cpumask_first(cpu_online_mask);
 112         dev->admin_irq.cpu = cpu;
 113         cpumask_set_cpu(cpu,
 114                         &dev->admin_irq.affinity_hint_mask);
 115         dev_info(&dev->pdev->dev, "Setup irq:0x%p vector:%d name:%s\n",
 116                  &dev->admin_irq,
 117                  dev->admin_irq.vector,
 118                  dev->admin_irq.name);
 119 }
 120
 121 static void efa_free_mgmnt_irq(struct efa_dev *dev)
 122 {
 123         struct efa_irq *irq;
 124
 125         irq = &dev->admin_irq;
 126         irq_set_affinity_hint(irq->vector, NULL);
 127         free_irq(irq->vector, irq->data);
 128 }
 129
 130 static int efa_set_mgmnt_irq(struct efa_dev *dev)
 131 {
 132         efa_setup_mgmnt_irq(dev);
 133
 134         return efa_request_mgmnt_irq(dev);
 135 }
 136
 137 static int efa_request_doorbell_bar(struct efa_dev *dev)
 138 {
 139         u8 db_bar_idx = dev->dev_attr.db_bar;
 140         struct pci_dev *pdev = dev->pdev;
 141         int bars;
 142         int err;
 143
 144         if (!(BIT(db_bar_idx) & EFA_BASE_BAR_MASK)) {
 145                 bars = pci_select_bars(pdev, IORESOURCE_MEM) & BIT(db_bar_idx);
 146
 147                 err = pci_request_selected_regions(pdev, bars, DRV_MODULE_NAME);
 148                 if (err) {
 149                         dev_err(&dev->pdev->dev,
 150                                 "pci_request_selected_regions for bar %d failed %d\n",
 151                                 db_bar_idx, err);
 152                         return err;
 153                 }
 154         }
 155
 156         dev->db_bar_addr = pci_resource_start(dev->pdev, db_bar_idx);
 157         dev->db_bar_len = pci_resource_len(dev->pdev, db_bar_idx);
 158
 159         return 0;
 160 }
 161
 162 static void efa_release_doorbell_bar(struct efa_dev *dev)
 163 {
 164         if (!(BIT(dev->dev_attr.db_bar) & EFA_BASE_BAR_MASK))
 165                 efa_release_bars(dev, BIT(dev->dev_attr.db_bar));
 166 }
 167
 168 static void efa_update_hw_hints(struct efa_dev *dev,
 169                                 struct efa_com_get_hw_hints_result *hw_hints)
 170 {
 171         struct efa_com_dev *edev = &dev->edev;
 172
 173         if (hw_hints->mmio_read_timeout)
 174                 edev->mmio_read.mmio_read_timeout =
 175                         hw_hints->mmio_read_timeout * 1000;
 176
 177         if (hw_hints->poll_interval)
 178                 edev->aq.poll_interval = hw_hints->poll_interval;
 179
 180         if (hw_hints->admin_completion_timeout)
 181                 edev->aq.completion_timeout =
 182                         hw_hints->admin_completion_timeout;
 183 }
 184
 185 static void efa_stats_init(struct efa_dev *dev)
 186 {
 187         atomic64_t *s = (atomic64_t *)&dev->stats;
 188         int i;
 189
 190         for (i = 0; i < sizeof(dev->stats) / sizeof(*s); i++, s++)
 191                 atomic64_set(s, 0);
 192 }
 193
 194 static void efa_set_host_info(struct efa_dev *dev)
 195 {
 196         struct efa_admin_set_feature_resp resp = {};
 197         struct efa_admin_set_feature_cmd cmd = {};
 198         struct efa_admin_host_info *hinf;
 199         u32 bufsz = sizeof(*hinf);
 200         dma_addr_t hinf_dma;
 201
 202         if (!efa_com_check_supported_feature_id(&dev->edev,
 203                                                 EFA_ADMIN_HOST_INFO))
 204                 return;
 205
 206         /* Failures in host info set shall not disturb probe */
 207         hinf = dma_alloc_coherent(&dev->pdev->dev, bufsz, &hinf_dma,
 208                                   GFP_KERNEL);
 209         if (!hinf)
 210                 return;
 211
 212         strscpy(hinf->os_dist_str, utsname()->release,
 213                 sizeof(hinf->os_dist_str));
 214         hinf->os_type = EFA_ADMIN_OS_LINUX;
 215         strscpy(hinf->kernel_ver_str, utsname()->version,
 216                 sizeof(hinf->kernel_ver_str));
 217         hinf->kernel_ver = LINUX_VERSION_CODE;
 218         EFA_SET(&hinf->driver_ver, EFA_ADMIN_HOST_INFO_DRIVER_MAJOR, 0);
 219         EFA_SET(&hinf->driver_ver, EFA_ADMIN_HOST_INFO_DRIVER_MINOR, 0);
 220         EFA_SET(&hinf->driver_ver, EFA_ADMIN_HOST_INFO_DRIVER_SUB_MINOR, 0);
 221         EFA_SET(&hinf->driver_ver, EFA_ADMIN_HOST_INFO_DRIVER_MODULE_TYPE, 0);
 222         EFA_SET(&hinf->bdf, EFA_ADMIN_HOST_INFO_BUS, dev->pdev->bus->number);
 223         EFA_SET(&hinf->bdf, EFA_ADMIN_HOST_INFO_DEVICE,
 224                 PCI_SLOT(dev->pdev->devfn));
 225         EFA_SET(&hinf->bdf, EFA_ADMIN_HOST_INFO_FUNCTION,
 226                 PCI_FUNC(dev->pdev->devfn));
 227         EFA_SET(&hinf->spec_ver, EFA_ADMIN_HOST_INFO_SPEC_MAJOR,
 228                 EFA_COMMON_SPEC_VERSION_MAJOR);
 229         EFA_SET(&hinf->spec_ver, EFA_ADMIN_HOST_INFO_SPEC_MINOR,
 230                 EFA_COMMON_SPEC_VERSION_MINOR);
 231         EFA_SET(&hinf->flags, EFA_ADMIN_HOST_INFO_INTREE, 1);
 232         EFA_SET(&hinf->flags, EFA_ADMIN_HOST_INFO_GDR, 0);
 233
 234         efa_com_set_feature_ex(&dev->edev, &resp, &cmd, EFA_ADMIN_HOST_INFO,
 235                                hinf_dma, bufsz);
 236
 237         dma_free_coherent(&dev->pdev->dev, bufsz, hinf, hinf_dma);
 238 }
 239
 240 static const struct ib_device_ops efa_dev_ops = {
 241         .owner = THIS_MODULE,
 242         .driver_id = RDMA_DRIVER_EFA,
 243         .uverbs_abi_ver = EFA_UVERBS_ABI_VERSION,
 244
 245         .alloc_hw_port_stats = efa_alloc_hw_port_stats,
 246         .alloc_hw_device_stats = efa_alloc_hw_device_stats,
 247         .alloc_pd = efa_alloc_pd,
 248         .alloc_ucontext = efa_alloc_ucontext,
 249         .create_cq = efa_create_cq,
 250         .create_qp = efa_create_qp,
 251         .create_user_ah = efa_create_ah,
 252         .dealloc_pd = efa_dealloc_pd,
 253         .dealloc_ucontext = efa_dealloc_ucontext,
 254         .dereg_mr = efa_dereg_mr,
 255         .destroy_ah = efa_destroy_ah,
 256         .destroy_cq = efa_destroy_cq,
 257         .destroy_qp = efa_destroy_qp,
 258         .get_hw_stats = efa_get_hw_stats,
 259         .get_link_layer = efa_port_link_layer,
 260         .get_port_immutable = efa_get_port_immutable,
 261         .mmap = efa_mmap,
 262         .mmap_free = efa_mmap_free,
 263         .modify_qp = efa_modify_qp,
 264         .query_device = efa_query_device,
 265         .query_gid = efa_query_gid,
 266         .query_pkey = efa_query_pkey,
 267         .query_port = efa_query_port,
 268         .query_qp = efa_query_qp,
 269         .reg_user_mr = efa_reg_mr,
 270
 271         INIT_RDMA_OBJ_SIZE(ib_ah, efa_ah, ibah),
 272         INIT_RDMA_OBJ_SIZE(ib_cq, efa_cq, ibcq),
 273         INIT_RDMA_OBJ_SIZE(ib_pd, efa_pd, ibpd),
 274         INIT_RDMA_OBJ_SIZE(ib_ucontext, efa_ucontext, ibucontext),
 275 };
 276
 277 static int efa_ib_device_add(struct efa_dev *dev)
 278 {
 279         struct efa_com_get_hw_hints_result hw_hints;
 280         struct pci_dev *pdev = dev->pdev;
 281         int err;
 282
 283         efa_stats_init(dev);
 284
 285         err = efa_com_get_device_attr(&dev->edev, &dev->dev_attr);
 286         if (err)
 287                 return err;
 288
 289         dev_dbg(&dev->pdev->dev, "Doorbells bar (%d)\n", dev->dev_attr.db_bar);
 290         err = efa_request_doorbell_bar(dev);
 291         if (err)
 292                 return err;
 293
 294         err = efa_com_get_hw_hints(&dev->edev, &hw_hints);
 295         if (err)
 296                 goto err_release_doorbell_bar;
 297
 298         efa_update_hw_hints(dev, &hw_hints);
 299
 300         /* Try to enable all the available aenq groups */
 301         err = efa_com_set_aenq_config(&dev->edev, EFA_AENQ_ENABLED_GROUPS);
 302         if (err)
 303                 goto err_release_doorbell_bar;
 304
 305         efa_set_host_info(dev);
 306
 307         dev->ibdev.node_type = RDMA_NODE_UNSPECIFIED;
 308         dev->ibdev.phys_port_cnt = 1;
 309         dev->ibdev.num_comp_vectors = 1;
 310         dev->ibdev.dev.parent = &pdev->dev;
 311
 312         ib_set_device_ops(&dev->ibdev, &efa_dev_ops);
 313
 314         err = ib_register_device(&dev->ibdev, "efa_%d", &pdev->dev);
 315         if (err)
 316                 goto err_release_doorbell_bar;
 317
 318         ibdev_info(&dev->ibdev, "IB device registered\n");
 319
 320         return 0;
 321
 322 err_release_doorbell_bar:
 323         efa_release_doorbell_bar(dev);
 324         return err;
 325 }
 326
 327 static void efa_ib_device_remove(struct efa_dev *dev)
 328 {
 329         efa_com_dev_reset(&dev->edev, EFA_REGS_RESET_NORMAL);
 330         ibdev_info(&dev->ibdev, "Unregister ib device\n");
 331         ib_unregister_device(&dev->ibdev);
 332         efa_release_doorbell_bar(dev);
 333 }
 334
 335 static void efa_disable_msix(struct efa_dev *dev)
 336 {
 337         pci_free_irq_vectors(dev->pdev);
 338 }
 339
 340 static int efa_enable_msix(struct efa_dev *dev)
 341 {
 342         int msix_vecs, irq_num;
 343
 344         /* Reserve the max msix vectors we might need */
 345         msix_vecs = EFA_NUM_MSIX_VEC;
 346         dev_dbg(&dev->pdev->dev, "Trying to enable MSI-X, vectors %d\n",
 347                 msix_vecs);
 348
 349         dev->admin_msix_vector_idx = EFA_MGMNT_MSIX_VEC_IDX;
 350         irq_num = pci_alloc_irq_vectors(dev->pdev, msix_vecs,
 351                                         msix_vecs, PCI_IRQ_MSIX);
 352
 353         if (irq_num < 0) {
 354                 dev_err(&dev->pdev->dev, "Failed to enable MSI-X. irq_num %d\n",
 355                         irq_num);
 356                 return -ENOSPC;
 357         }
 358
 359         if (irq_num != msix_vecs) {
 360                 dev_err(&dev->pdev->dev,
 361                         "Allocated %d MSI-X (out of %d requested)\n",
 362                         irq_num, msix_vecs);
 363                 return -ENOSPC;
 364         }
 365
 366         return 0;
 367 }
 368
 369 static int efa_device_init(struct efa_com_dev *edev, struct pci_dev *pdev)
 370 {
 371         int dma_width;
 372         int err;
 373
 374         err = efa_com_dev_reset(edev, EFA_REGS_RESET_NORMAL);
 375         if (err)
 376                 return err;
 377
 378         err = efa_com_validate_version(edev);
 379         if (err)
 380                 return err;
 381
 382         dma_width = efa_com_get_dma_width(edev);
 383         if (dma_width < 0) {
 384                 err = dma_width;
 385                 return err;
 386         }
 387
 388         err = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(dma_width));
 389         if (err) {
 390                 dev_err(&pdev->dev, "dma_set_mask_and_coherent failed %d\n", err);
 391                 return err;
 392         }
 393
 394         dma_set_max_seg_size(&pdev->dev, UINT_MAX);
 395         return 0;
 396 }
 397
 398 static struct efa_dev *efa_probe_device(struct pci_dev *pdev)
 399 {
 400         struct efa_com_dev *edev;
 401         struct efa_dev *dev;
 402         int bars;
 403         int err;
 404
 405         err = pci_enable_device_mem(pdev);
 406         if (err) {
 407                 dev_err(&pdev->dev, "pci_enable_device_mem() failed!\n");
 408                 return ERR_PTR(err);
 409         }
 410
 411         pci_set_master(pdev);
 412
 413         dev = ib_alloc_device(efa_dev, ibdev);
 414         if (!dev) {
 415                 dev_err(&pdev->dev, "Device alloc failed\n");
 416                 err = -ENOMEM;
 417                 goto err_disable_device;
 418         }
 419
 420         pci_set_drvdata(pdev, dev);
 421         edev = &dev->edev;
 422         edev->efa_dev = dev;
 423         edev->dmadev = &pdev->dev;
 424         dev->pdev = pdev;
 425
 426         bars = pci_select_bars(pdev, IORESOURCE_MEM) & EFA_BASE_BAR_MASK;
 427         err = pci_request_selected_regions(pdev, bars, DRV_MODULE_NAME);
 428         if (err) {
 429                 dev_err(&pdev->dev, "pci_request_selected_regions failed %d\n",
 430                         err);
 431                 goto err_ibdev_destroy;
 432         }
 433
 434         dev->reg_bar_addr = pci_resource_start(pdev, EFA_REG_BAR);
 435         dev->reg_bar_len = pci_resource_len(pdev, EFA_REG_BAR);
 436         dev->mem_bar_addr = pci_resource_start(pdev, EFA_MEM_BAR);
 437         dev->mem_bar_len = pci_resource_len(pdev, EFA_MEM_BAR);
 438
 439         edev->reg_bar = devm_ioremap(&pdev->dev,
 440                                      dev->reg_bar_addr,
 441                                      dev->reg_bar_len);
 442         if (!edev->reg_bar) {
 443                 dev_err(&pdev->dev, "Failed to remap register bar\n");
 444                 err = -EFAULT;
 445                 goto err_release_bars;
 446         }
 447
 448         err = efa_com_mmio_reg_read_init(edev);
 449         if (err) {
 450                 dev_err(&pdev->dev, "Failed to init readless MMIO\n");
 451                 goto err_iounmap;
 452         }
 453
 454         err = efa_device_init(edev, pdev);
 455         if (err) {
 456                 dev_err(&pdev->dev, "EFA device init failed\n");
 457                 if (err == -ETIME)
 458                         err = -EPROBE_DEFER;
 459                 goto err_reg_read_destroy;
 460         }
 461
 462         err = efa_enable_msix(dev);
 463         if (err)
 464                 goto err_reg_read_destroy;
 465
 466         edev->aq.msix_vector_idx = dev->admin_msix_vector_idx;
 467         edev->aenq.msix_vector_idx = dev->admin_msix_vector_idx;
 468
 469         err = efa_set_mgmnt_irq(dev);
 470         if (err)
 471                 goto err_disable_msix;
 472
 473         err = efa_com_admin_init(edev, &aenq_handlers);
 474         if (err)
 475                 goto err_free_mgmnt_irq;
 476
 477         return dev;
 478
 479 err_free_mgmnt_irq:
 480         efa_free_mgmnt_irq(dev);
 481 err_disable_msix:
 482         efa_disable_msix(dev);
 483 err_reg_read_destroy:
 484         efa_com_mmio_reg_read_destroy(edev);
 485 err_iounmap:
 486         devm_iounmap(&pdev->dev, edev->reg_bar);
 487 err_release_bars:
 488         efa_release_bars(dev, EFA_BASE_BAR_MASK);
 489 err_ibdev_destroy:
 490         ib_dealloc_device(&dev->ibdev);
 491 err_disable_device:
 492         pci_disable_device(pdev);
 493         return ERR_PTR(err);
 494 }
 495
 496 static void efa_remove_device(struct pci_dev *pdev)
 497 {
 498         struct efa_dev *dev = pci_get_drvdata(pdev);
 499         struct efa_com_dev *edev;
 500
 501         edev = &dev->edev;
 502         efa_com_admin_destroy(edev);
 503         efa_free_mgmnt_irq(dev);
 504         efa_disable_msix(dev);
 505         efa_com_mmio_reg_read_destroy(edev);
 506         devm_iounmap(&pdev->dev, edev->reg_bar);
 507         efa_release_bars(dev, EFA_BASE_BAR_MASK);
 508         ib_dealloc_device(&dev->ibdev);
 509         pci_disable_device(pdev);
 510 }
 511
 512 static int efa_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
 513 {
 514         struct efa_dev *dev;
 515         int err;
 516
 517         dev = efa_probe_device(pdev);
 518         if (IS_ERR(dev))
 519                 return PTR_ERR(dev);
 520
 521         err = efa_ib_device_add(dev);
 522         if (err)
 523                 goto err_remove_device;
 524
 525         return 0;
 526
 527 err_remove_device:
 528         efa_remove_device(pdev);
 529         return err;
 530 }
 531
 532 static void efa_remove(struct pci_dev *pdev)
 533 {
 534         struct efa_dev *dev = pci_get_drvdata(pdev);
 535
 536         efa_ib_device_remove(dev);
 537         efa_remove_device(pdev);
 538 }
 539
 540 static struct pci_driver efa_pci_driver = {
 541         .name           = DRV_MODULE_NAME,
 542         .id_table       = efa_pci_tbl,
 543         .probe          = efa_probe,
 544         .remove         = efa_remove,
 545 };
 546
 547 module_pci_driver(efa_pci_driver);