linux/drivers/vfio/pci/vfio_pci.c
<<
>>
Prefs
   1/*
   2 * Copyright (C) 2012 Red Hat, Inc.  All rights reserved.
   3 *     Author: Alex Williamson <alex.williamson@redhat.com>
   4 *
   5 * This program is free software; you can redistribute it and/or modify
   6 * it under the terms of the GNU General Public License version 2 as
   7 * published by the Free Software Foundation.
   8 *
   9 * Derived from original vfio:
  10 * Copyright 2010 Cisco Systems, Inc.  All rights reserved.
  11 * Author: Tom Lyon, pugs@cisco.com
  12 */
  13
  14#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  15
  16#include <linux/device.h>
  17#include <linux/eventfd.h>
  18#include <linux/file.h>
  19#include <linux/interrupt.h>
  20#include <linux/iommu.h>
  21#include <linux/module.h>
  22#include <linux/mutex.h>
  23#include <linux/notifier.h>
  24#include <linux/pci.h>
  25#include <linux/pm_runtime.h>
  26#include <linux/slab.h>
  27#include <linux/types.h>
  28#include <linux/uaccess.h>
  29#include <linux/vfio.h>
  30#include <linux/vgaarb.h>
  31
  32#include "vfio_pci_private.h"
  33
  34#define DRIVER_VERSION  "0.2"
  35#define DRIVER_AUTHOR   "Alex Williamson <alex.williamson@redhat.com>"
  36#define DRIVER_DESC     "VFIO PCI - User Level meta-driver"
  37
  38static char ids[1024] __initdata;
  39module_param_string(ids, ids, sizeof(ids), 0);
  40MODULE_PARM_DESC(ids, "Initial PCI IDs to add to the vfio driver, format is \"vendor:device[:subvendor[:subdevice[:class[:class_mask]]]]\" and multiple comma separated entries can be specified");
  41
  42static bool nointxmask;
  43module_param_named(nointxmask, nointxmask, bool, S_IRUGO | S_IWUSR);
  44MODULE_PARM_DESC(nointxmask,
  45                  "Disable support for PCI 2.3 style INTx masking.  If this resolves problems for specific devices, report lspci -vvvxxx to linux-pci@vger.kernel.org so the device can be fixed automatically via the broken_intx_masking flag.");
  46
  47#ifdef CONFIG_VFIO_PCI_VGA
  48static bool disable_vga;
  49module_param(disable_vga, bool, S_IRUGO);
  50MODULE_PARM_DESC(disable_vga, "Disable VGA resource access through vfio-pci");
  51#endif
  52
  53static bool disable_idle_d3;
  54module_param(disable_idle_d3, bool, S_IRUGO | S_IWUSR);
  55MODULE_PARM_DESC(disable_idle_d3,
  56                 "Disable using the PCI D3 low power state for idle, unused devices");
  57
  58static DEFINE_MUTEX(driver_lock);
  59
  60static inline bool vfio_vga_disabled(void)
  61{
  62#ifdef CONFIG_VFIO_PCI_VGA
  63        return disable_vga;
  64#else
  65        return true;
  66#endif
  67}
  68
  69/*
  70 * Our VGA arbiter participation is limited since we don't know anything
  71 * about the device itself.  However, if the device is the only VGA device
  72 * downstream of a bridge and VFIO VGA support is disabled, then we can
  73 * safely return legacy VGA IO and memory as not decoded since the user
  74 * has no way to get to it and routing can be disabled externally at the
  75 * bridge.
  76 */
  77static unsigned int vfio_pci_set_vga_decode(void *opaque, bool single_vga)
  78{
  79        struct vfio_pci_device *vdev = opaque;
  80        struct pci_dev *tmp = NULL, *pdev = vdev->pdev;
  81        unsigned char max_busnr;
  82        unsigned int decodes;
  83
  84        if (single_vga || !vfio_vga_disabled() || pci_is_root_bus(pdev->bus))
  85                return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
  86                       VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
  87
  88        max_busnr = pci_bus_max_busnr(pdev->bus);
  89        decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
  90
  91        while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
  92                if (tmp == pdev ||
  93                    pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
  94                    pci_is_root_bus(tmp->bus))
  95                        continue;
  96
  97                if (tmp->bus->number >= pdev->bus->number &&
  98                    tmp->bus->number <= max_busnr) {
  99                        pci_dev_put(tmp);
 100                        decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
 101                        break;
 102                }
 103        }
 104
 105        return decodes;
 106}
 107
 108static inline bool vfio_pci_is_vga(struct pci_dev *pdev)
 109{
 110        return (pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA;
 111}
 112
 113static void vfio_pci_probe_mmaps(struct vfio_pci_device *vdev)
 114{
 115        struct resource *res;
 116        int bar;
 117        struct vfio_pci_dummy_resource *dummy_res;
 118
 119        INIT_LIST_HEAD(&vdev->dummy_resources_list);
 120
 121        for (bar = PCI_STD_RESOURCES; bar <= PCI_STD_RESOURCE_END; bar++) {
 122                res = vdev->pdev->resource + bar;
 123
 124                if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
 125                        goto no_mmap;
 126
 127                if (!(res->flags & IORESOURCE_MEM))
 128                        goto no_mmap;
 129
 130                /*
 131                 * The PCI core shouldn't set up a resource with a
 132                 * type but zero size. But there may be bugs that
 133                 * cause us to do that.
 134                 */
 135                if (!resource_size(res))
 136                        goto no_mmap;
 137
 138                if (resource_size(res) >= PAGE_SIZE) {
 139                        vdev->bar_mmap_supported[bar] = true;
 140                        continue;
 141                }
 142
 143                if (!(res->start & ~PAGE_MASK)) {
 144                        /*
 145                         * Add a dummy resource to reserve the remainder
 146                         * of the exclusive page in case that hot-add
 147                         * device's bar is assigned into it.
 148                         */
 149                        dummy_res = kzalloc(sizeof(*dummy_res), GFP_KERNEL);
 150                        if (dummy_res == NULL)
 151                                goto no_mmap;
 152
 153                        dummy_res->resource.name = "vfio sub-page reserved";
 154                        dummy_res->resource.start = res->end + 1;
 155                        dummy_res->resource.end = res->start + PAGE_SIZE - 1;
 156                        dummy_res->resource.flags = res->flags;
 157                        if (request_resource(res->parent,
 158                                                &dummy_res->resource)) {
 159                                kfree(dummy_res);
 160                                goto no_mmap;
 161                        }
 162                        dummy_res->index = bar;
 163                        list_add(&dummy_res->res_next,
 164                                        &vdev->dummy_resources_list);
 165                        vdev->bar_mmap_supported[bar] = true;
 166                        continue;
 167                }
 168                /*
 169                 * Here we don't handle the case when the BAR is not page
 170                 * aligned because we can't expect the BAR will be
 171                 * assigned into the same location in a page in guest
 172                 * when we passthrough the BAR. And it's hard to access
 173                 * this BAR in userspace because we have no way to get
 174                 * the BAR's location in a page.
 175                 */
 176no_mmap:
 177                vdev->bar_mmap_supported[bar] = false;
 178        }
 179}
 180
 181static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev);
 182static void vfio_pci_disable(struct vfio_pci_device *vdev);
 183
 184/*
 185 * INTx masking requires the ability to disable INTx signaling via PCI_COMMAND
 186 * _and_ the ability detect when the device is asserting INTx via PCI_STATUS.
 187 * If a device implements the former but not the latter we would typically
 188 * expect broken_intx_masking be set and require an exclusive interrupt.
 189 * However since we do have control of the device's ability to assert INTx,
 190 * we can instead pretend that the device does not implement INTx, virtualizing
 191 * the pin register to report zero and maintaining DisINTx set on the host.
 192 */
 193static bool vfio_pci_nointx(struct pci_dev *pdev)
 194{
 195        switch (pdev->vendor) {
 196        case PCI_VENDOR_ID_INTEL:
 197                switch (pdev->device) {
 198                /* All i40e (XL710/X710/XXV710) 10/20/25/40GbE NICs */
 199                case 0x1572:
 200                case 0x1574:
 201                case 0x1580 ... 0x1581:
 202                case 0x1583 ... 0x158b:
 203                case 0x37d0 ... 0x37d2:
 204                        return true;
 205                default:
 206                        return false;
 207                }
 208        }
 209
 210        return false;
 211}
 212
 213static int vfio_pci_enable(struct vfio_pci_device *vdev)
 214{
 215        struct pci_dev *pdev = vdev->pdev;
 216        int ret;
 217        u16 cmd;
 218        u8 msix_pos;
 219
 220        pci_set_power_state(pdev, PCI_D0);
 221
 222        /* Don't allow our initial saved state to include busmaster */
 223        pci_clear_master(pdev);
 224
 225        ret = pci_enable_device(pdev);
 226        if (ret)
 227                return ret;
 228
 229        /* If reset fails because of the device lock, fail this path entirely */
 230        ret = pci_try_reset_function(pdev);
 231        if (ret == -EAGAIN) {
 232                pci_disable_device(pdev);
 233                return ret;
 234        }
 235
 236        vdev->reset_works = !ret;
 237        pci_save_state(pdev);
 238        vdev->pci_saved_state = pci_store_saved_state(pdev);
 239        if (!vdev->pci_saved_state)
 240                pr_debug("%s: Couldn't store %s saved state\n",
 241                         __func__, dev_name(&pdev->dev));
 242
 243        if (likely(!nointxmask)) {
 244                if (vfio_pci_nointx(pdev)) {
 245                        dev_info(&pdev->dev, "Masking broken INTx support\n");
 246                        vdev->nointx = true;
 247                        pci_intx(pdev, 0);
 248                } else
 249                        vdev->pci_2_3 = pci_intx_mask_supported(pdev);
 250        }
 251
 252        pci_read_config_word(pdev, PCI_COMMAND, &cmd);
 253        if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
 254                cmd &= ~PCI_COMMAND_INTX_DISABLE;
 255                pci_write_config_word(pdev, PCI_COMMAND, cmd);
 256        }
 257
 258        ret = vfio_config_init(vdev);
 259        if (ret) {
 260                kfree(vdev->pci_saved_state);
 261                vdev->pci_saved_state = NULL;
 262                pci_disable_device(pdev);
 263                return ret;
 264        }
 265
 266        msix_pos = pdev->msix_cap;
 267        if (msix_pos) {
 268                u16 flags;
 269                u32 table;
 270
 271                pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
 272                pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
 273
 274                vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
 275                vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
 276                vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
 277        } else
 278                vdev->msix_bar = 0xFF;
 279
 280        if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
 281                vdev->has_vga = true;
 282
 283
 284        if (vfio_pci_is_vga(pdev) &&
 285            pdev->vendor == PCI_VENDOR_ID_INTEL &&
 286            IS_ENABLED(CONFIG_VFIO_PCI_IGD)) {
 287                ret = vfio_pci_igd_init(vdev);
 288                if (ret) {
 289                        dev_warn(&vdev->pdev->dev,
 290                                 "Failed to setup Intel IGD regions\n");
 291                        vfio_pci_disable(vdev);
 292                        return ret;
 293                }
 294        }
 295
 296        vfio_pci_probe_mmaps(vdev);
 297
 298        return 0;
 299}
 300
 301static void vfio_pci_disable(struct vfio_pci_device *vdev)
 302{
 303        struct pci_dev *pdev = vdev->pdev;
 304        struct vfio_pci_dummy_resource *dummy_res, *tmp;
 305        int i, bar;
 306
 307        /* Stop the device from further DMA */
 308        pci_clear_master(pdev);
 309
 310        vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
 311                                VFIO_IRQ_SET_ACTION_TRIGGER,
 312                                vdev->irq_type, 0, 0, NULL);
 313
 314        vdev->virq_disabled = false;
 315
 316        for (i = 0; i < vdev->num_regions; i++)
 317                vdev->region[i].ops->release(vdev, &vdev->region[i]);
 318
 319        vdev->num_regions = 0;
 320        kfree(vdev->region);
 321        vdev->region = NULL; /* don't krealloc a freed pointer */
 322
 323        vfio_config_free(vdev);
 324
 325        for (bar = PCI_STD_RESOURCES; bar <= PCI_STD_RESOURCE_END; bar++) {
 326                if (!vdev->barmap[bar])
 327                        continue;
 328                pci_iounmap(pdev, vdev->barmap[bar]);
 329                pci_release_selected_regions(pdev, 1 << bar);
 330                vdev->barmap[bar] = NULL;
 331        }
 332
 333        list_for_each_entry_safe(dummy_res, tmp,
 334                                 &vdev->dummy_resources_list, res_next) {
 335                list_del(&dummy_res->res_next);
 336                release_resource(&dummy_res->resource);
 337                kfree(dummy_res);
 338        }
 339
 340        vdev->needs_reset = true;
 341
 342        /*
 343         * If we have saved state, restore it.  If we can reset the device,
 344         * even better.  Resetting with current state seems better than
 345         * nothing, but saving and restoring current state without reset
 346         * is just busy work.
 347         */
 348        if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
 349                pr_info("%s: Couldn't reload %s saved state\n",
 350                        __func__, dev_name(&pdev->dev));
 351
 352                if (!vdev->reset_works)
 353                        goto out;
 354
 355                pci_save_state(pdev);
 356        }
 357
 358        /*
 359         * Disable INTx and MSI, presumably to avoid spurious interrupts
 360         * during reset.  Stolen from pci_reset_function()
 361         */
 362        pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
 363
 364        /*
 365         * Try to reset the device.  The success of this is dependent on
 366         * being able to lock the device, which is not always possible.
 367         */
 368        if (vdev->reset_works && !pci_try_reset_function(pdev))
 369                vdev->needs_reset = false;
 370
 371        pci_restore_state(pdev);
 372out:
 373        pci_disable_device(pdev);
 374
 375        vfio_pci_try_bus_reset(vdev);
 376
 377        if (!disable_idle_d3)
 378                pci_set_power_state(pdev, PCI_D3hot);
 379}
 380
 381static void vfio_pci_release(void *device_data)
 382{
 383        struct vfio_pci_device *vdev = device_data;
 384
 385        mutex_lock(&driver_lock);
 386
 387        if (!(--vdev->refcnt)) {
 388                vfio_spapr_pci_eeh_release(vdev->pdev);
 389                vfio_pci_disable(vdev);
 390        }
 391
 392        mutex_unlock(&driver_lock);
 393
 394        module_put(THIS_MODULE);
 395}
 396
 397static int vfio_pci_open(void *device_data)
 398{
 399        struct vfio_pci_device *vdev = device_data;
 400        int ret = 0;
 401
 402        if (!try_module_get(THIS_MODULE))
 403                return -ENODEV;
 404
 405        mutex_lock(&driver_lock);
 406
 407        if (!vdev->refcnt) {
 408                ret = vfio_pci_enable(vdev);
 409                if (ret)
 410                        goto error;
 411
 412                vfio_spapr_pci_eeh_open(vdev->pdev);
 413        }
 414        vdev->refcnt++;
 415error:
 416        mutex_unlock(&driver_lock);
 417        if (ret)
 418                module_put(THIS_MODULE);
 419        return ret;
 420}
 421
 422static int vfio_pci_get_irq_count(struct vfio_pci_device *vdev, int irq_type)
 423{
 424        if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
 425                u8 pin;
 426                pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
 427                if (IS_ENABLED(CONFIG_VFIO_PCI_INTX) && !vdev->nointx && pin)
 428                        return 1;
 429
 430        } else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
 431                u8 pos;
 432                u16 flags;
 433
 434                pos = vdev->pdev->msi_cap;
 435                if (pos) {
 436                        pci_read_config_word(vdev->pdev,
 437                                             pos + PCI_MSI_FLAGS, &flags);
 438                        return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
 439                }
 440        } else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
 441                u8 pos;
 442                u16 flags;
 443
 444                pos = vdev->pdev->msix_cap;
 445                if (pos) {
 446                        pci_read_config_word(vdev->pdev,
 447                                             pos + PCI_MSIX_FLAGS, &flags);
 448
 449                        return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
 450                }
 451        } else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
 452                if (pci_is_pcie(vdev->pdev))
 453                        return 1;
 454        } else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
 455                return 1;
 456        }
 457
 458        return 0;
 459}
 460
 461static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
 462{
 463        (*(int *)data)++;
 464        return 0;
 465}
 466
 467struct vfio_pci_fill_info {
 468        int max;
 469        int cur;
 470        struct vfio_pci_dependent_device *devices;
 471};
 472
 473static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
 474{
 475        struct vfio_pci_fill_info *fill = data;
 476        struct iommu_group *iommu_group;
 477
 478        if (fill->cur == fill->max)
 479                return -EAGAIN; /* Something changed, try again */
 480
 481        iommu_group = iommu_group_get(&pdev->dev);
 482        if (!iommu_group)
 483                return -EPERM; /* Cannot reset non-isolated devices */
 484
 485        fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
 486        fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
 487        fill->devices[fill->cur].bus = pdev->bus->number;
 488        fill->devices[fill->cur].devfn = pdev->devfn;
 489        fill->cur++;
 490        iommu_group_put(iommu_group);
 491        return 0;
 492}
 493
 494struct vfio_pci_group_entry {
 495        struct vfio_group *group;
 496        int id;
 497};
 498
 499struct vfio_pci_group_info {
 500        int count;
 501        struct vfio_pci_group_entry *groups;
 502};
 503
 504static int vfio_pci_validate_devs(struct pci_dev *pdev, void *data)
 505{
 506        struct vfio_pci_group_info *info = data;
 507        struct iommu_group *group;
 508        int id, i;
 509
 510        group = iommu_group_get(&pdev->dev);
 511        if (!group)
 512                return -EPERM;
 513
 514        id = iommu_group_id(group);
 515
 516        for (i = 0; i < info->count; i++)
 517                if (info->groups[i].id == id)
 518                        break;
 519
 520        iommu_group_put(group);
 521
 522        return (i == info->count) ? -EINVAL : 0;
 523}
 524
 525static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
 526{
 527        for (; pdev; pdev = pdev->bus->self)
 528                if (pdev->bus == slot->bus)
 529                        return (pdev->slot == slot);
 530        return false;
 531}
 532
 533struct vfio_pci_walk_info {
 534        int (*fn)(struct pci_dev *, void *data);
 535        void *data;
 536        struct pci_dev *pdev;
 537        bool slot;
 538        int ret;
 539};
 540
 541static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
 542{
 543        struct vfio_pci_walk_info *walk = data;
 544
 545        if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
 546                walk->ret = walk->fn(pdev, walk->data);
 547
 548        return walk->ret;
 549}
 550
 551static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
 552                                         int (*fn)(struct pci_dev *,
 553                                                   void *data), void *data,
 554                                         bool slot)
 555{
 556        struct vfio_pci_walk_info walk = {
 557                .fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
 558        };
 559
 560        pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
 561
 562        return walk.ret;
 563}
 564
 565static int msix_sparse_mmap_cap(struct vfio_pci_device *vdev,
 566                                struct vfio_info_cap *caps)
 567{
 568        struct vfio_region_info_cap_sparse_mmap *sparse;
 569        size_t end, size;
 570        int nr_areas = 2, i = 0, ret;
 571
 572        end = pci_resource_len(vdev->pdev, vdev->msix_bar);
 573
 574        /* If MSI-X table is aligned to the start or end, only one area */
 575        if (((vdev->msix_offset & PAGE_MASK) == 0) ||
 576            (PAGE_ALIGN(vdev->msix_offset + vdev->msix_size) >= end))
 577                nr_areas = 1;
 578
 579        size = sizeof(*sparse) + (nr_areas * sizeof(*sparse->areas));
 580
 581        sparse = kzalloc(size, GFP_KERNEL);
 582        if (!sparse)
 583                return -ENOMEM;
 584
 585        sparse->nr_areas = nr_areas;
 586
 587        if (vdev->msix_offset & PAGE_MASK) {
 588                sparse->areas[i].offset = 0;
 589                sparse->areas[i].size = vdev->msix_offset & PAGE_MASK;
 590                i++;
 591        }
 592
 593        if (PAGE_ALIGN(vdev->msix_offset + vdev->msix_size) < end) {
 594                sparse->areas[i].offset = PAGE_ALIGN(vdev->msix_offset +
 595                                                     vdev->msix_size);
 596                sparse->areas[i].size = end - sparse->areas[i].offset;
 597                i++;
 598        }
 599
 600        ret = vfio_info_add_capability(caps, VFIO_REGION_INFO_CAP_SPARSE_MMAP,
 601                                       sparse);
 602        kfree(sparse);
 603
 604        return ret;
 605}
 606
 607int vfio_pci_register_dev_region(struct vfio_pci_device *vdev,
 608                                 unsigned int type, unsigned int subtype,
 609                                 const struct vfio_pci_regops *ops,
 610                                 size_t size, u32 flags, void *data)
 611{
 612        struct vfio_pci_region *region;
 613
 614        region = krealloc(vdev->region,
 615                          (vdev->num_regions + 1) * sizeof(*region),
 616                          GFP_KERNEL);
 617        if (!region)
 618                return -ENOMEM;
 619
 620        vdev->region = region;
 621        vdev->region[vdev->num_regions].type = type;
 622        vdev->region[vdev->num_regions].subtype = subtype;
 623        vdev->region[vdev->num_regions].ops = ops;
 624        vdev->region[vdev->num_regions].size = size;
 625        vdev->region[vdev->num_regions].flags = flags;
 626        vdev->region[vdev->num_regions].data = data;
 627
 628        vdev->num_regions++;
 629
 630        return 0;
 631}
 632
 633static long vfio_pci_ioctl(void *device_data,
 634                           unsigned int cmd, unsigned long arg)
 635{
 636        struct vfio_pci_device *vdev = device_data;
 637        unsigned long minsz;
 638
 639        if (cmd == VFIO_DEVICE_GET_INFO) {
 640                struct vfio_device_info info;
 641
 642                minsz = offsetofend(struct vfio_device_info, num_irqs);
 643
 644                if (copy_from_user(&info, (void __user *)arg, minsz))
 645                        return -EFAULT;
 646
 647                if (info.argsz < minsz)
 648                        return -EINVAL;
 649
 650                info.flags = VFIO_DEVICE_FLAGS_PCI;
 651
 652                if (vdev->reset_works)
 653                        info.flags |= VFIO_DEVICE_FLAGS_RESET;
 654
 655                info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
 656                info.num_irqs = VFIO_PCI_NUM_IRQS;
 657
 658                return copy_to_user((void __user *)arg, &info, minsz) ?
 659                        -EFAULT : 0;
 660
 661        } else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
 662                struct pci_dev *pdev = vdev->pdev;
 663                struct vfio_region_info info;
 664                struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
 665                int i, ret;
 666
 667                minsz = offsetofend(struct vfio_region_info, offset);
 668
 669                if (copy_from_user(&info, (void __user *)arg, minsz))
 670                        return -EFAULT;
 671
 672                if (info.argsz < minsz)
 673                        return -EINVAL;
 674
 675                switch (info.index) {
 676                case VFIO_PCI_CONFIG_REGION_INDEX:
 677                        info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 678                        info.size = pdev->cfg_size;
 679                        info.flags = VFIO_REGION_INFO_FLAG_READ |
 680                                     VFIO_REGION_INFO_FLAG_WRITE;
 681                        break;
 682                case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
 683                        info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 684                        info.size = pci_resource_len(pdev, info.index);
 685                        if (!info.size) {
 686                                info.flags = 0;
 687                                break;
 688                        }
 689
 690                        info.flags = VFIO_REGION_INFO_FLAG_READ |
 691                                     VFIO_REGION_INFO_FLAG_WRITE;
 692                        if (vdev->bar_mmap_supported[info.index]) {
 693                                info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
 694                                if (info.index == vdev->msix_bar) {
 695                                        ret = msix_sparse_mmap_cap(vdev, &caps);
 696                                        if (ret)
 697                                                return ret;
 698                                }
 699                        }
 700
 701                        break;
 702                case VFIO_PCI_ROM_REGION_INDEX:
 703                {
 704                        void __iomem *io;
 705                        size_t size;
 706
 707                        info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 708                        info.flags = 0;
 709
 710                        /* Report the BAR size, not the ROM size */
 711                        info.size = pci_resource_len(pdev, info.index);
 712                        if (!info.size) {
 713                                /* Shadow ROMs appear as PCI option ROMs */
 714                                if (pdev->resource[PCI_ROM_RESOURCE].flags &
 715                                                        IORESOURCE_ROM_SHADOW)
 716                                        info.size = 0x20000;
 717                                else
 718                                        break;
 719                        }
 720
 721                        /* Is it really there? */
 722                        io = pci_map_rom(pdev, &size);
 723                        if (!io || !size) {
 724                                info.size = 0;
 725                                break;
 726                        }
 727                        pci_unmap_rom(pdev, io);
 728
 729                        info.flags = VFIO_REGION_INFO_FLAG_READ;
 730                        break;
 731                }
 732                case VFIO_PCI_VGA_REGION_INDEX:
 733                        if (!vdev->has_vga)
 734                                return -EINVAL;
 735
 736                        info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 737                        info.size = 0xc0000;
 738                        info.flags = VFIO_REGION_INFO_FLAG_READ |
 739                                     VFIO_REGION_INFO_FLAG_WRITE;
 740
 741                        break;
 742                default:
 743                {
 744                        struct vfio_region_info_cap_type cap_type;
 745
 746                        if (info.index >=
 747                            VFIO_PCI_NUM_REGIONS + vdev->num_regions)
 748                                return -EINVAL;
 749
 750                        i = info.index - VFIO_PCI_NUM_REGIONS;
 751
 752                        info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 753                        info.size = vdev->region[i].size;
 754                        info.flags = vdev->region[i].flags;
 755
 756                        cap_type.type = vdev->region[i].type;
 757                        cap_type.subtype = vdev->region[i].subtype;
 758
 759                        ret = vfio_info_add_capability(&caps,
 760                                                      VFIO_REGION_INFO_CAP_TYPE,
 761                                                      &cap_type);
 762                        if (ret)
 763                                return ret;
 764
 765                }
 766                }
 767
 768                if (caps.size) {
 769                        info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
 770                        if (info.argsz < sizeof(info) + caps.size) {
 771                                info.argsz = sizeof(info) + caps.size;
 772                                info.cap_offset = 0;
 773                        } else {
 774                                vfio_info_cap_shift(&caps, sizeof(info));
 775                                if (copy_to_user((void __user *)arg +
 776                                                  sizeof(info), caps.buf,
 777                                                  caps.size)) {
 778                                        kfree(caps.buf);
 779                                        return -EFAULT;
 780                                }
 781                                info.cap_offset = sizeof(info);
 782                        }
 783
 784                        kfree(caps.buf);
 785                }
 786
 787                return copy_to_user((void __user *)arg, &info, minsz) ?
 788                        -EFAULT : 0;
 789
 790        } else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
 791                struct vfio_irq_info info;
 792
 793                minsz = offsetofend(struct vfio_irq_info, count);
 794
 795                if (copy_from_user(&info, (void __user *)arg, minsz))
 796                        return -EFAULT;
 797
 798                if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
 799                        return -EINVAL;
 800
 801                switch (info.index) {
 802                case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
 803                case VFIO_PCI_REQ_IRQ_INDEX:
 804                        break;
 805                case VFIO_PCI_ERR_IRQ_INDEX:
 806                        if (pci_is_pcie(vdev->pdev))
 807                                break;
 808                /* pass thru to return error */
 809                default:
 810                        return -EINVAL;
 811                }
 812
 813                info.flags = VFIO_IRQ_INFO_EVENTFD;
 814
 815                info.count = vfio_pci_get_irq_count(vdev, info.index);
 816
 817                if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
 818                        info.flags |= (VFIO_IRQ_INFO_MASKABLE |
 819                                       VFIO_IRQ_INFO_AUTOMASKED);
 820                else
 821                        info.flags |= VFIO_IRQ_INFO_NORESIZE;
 822
 823                return copy_to_user((void __user *)arg, &info, minsz) ?
 824                        -EFAULT : 0;
 825
 826        } else if (cmd == VFIO_DEVICE_SET_IRQS) {
 827                struct vfio_irq_set hdr;
 828                u8 *data = NULL;
 829                int max, ret = 0;
 830                size_t data_size = 0;
 831
 832                minsz = offsetofend(struct vfio_irq_set, count);
 833
 834                if (copy_from_user(&hdr, (void __user *)arg, minsz))
 835                        return -EFAULT;
 836
 837                max = vfio_pci_get_irq_count(vdev, hdr.index);
 838
 839                ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
 840                                                 VFIO_PCI_NUM_IRQS, &data_size);
 841                if (ret)
 842                        return ret;
 843
 844                if (data_size) {
 845                        data = memdup_user((void __user *)(arg + minsz),
 846                                            data_size);
 847                        if (IS_ERR(data))
 848                                return PTR_ERR(data);
 849                }
 850
 851                mutex_lock(&vdev->igate);
 852
 853                ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
 854                                              hdr.start, hdr.count, data);
 855
 856                mutex_unlock(&vdev->igate);
 857                kfree(data);
 858
 859                return ret;
 860
 861        } else if (cmd == VFIO_DEVICE_RESET) {
 862                return vdev->reset_works ?
 863                        pci_try_reset_function(vdev->pdev) : -EINVAL;
 864
 865        } else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
 866                struct vfio_pci_hot_reset_info hdr;
 867                struct vfio_pci_fill_info fill = { 0 };
 868                struct vfio_pci_dependent_device *devices = NULL;
 869                bool slot = false;
 870                int ret = 0;
 871
 872                minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
 873
 874                if (copy_from_user(&hdr, (void __user *)arg, minsz))
 875                        return -EFAULT;
 876
 877                if (hdr.argsz < minsz)
 878                        return -EINVAL;
 879
 880                hdr.flags = 0;
 881
 882                /* Can we do a slot or bus reset or neither? */
 883                if (!pci_probe_reset_slot(vdev->pdev->slot))
 884                        slot = true;
 885                else if (pci_probe_reset_bus(vdev->pdev->bus))
 886                        return -ENODEV;
 887
 888                /* How many devices are affected? */
 889                ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
 890                                                    vfio_pci_count_devs,
 891                                                    &fill.max, slot);
 892                if (ret)
 893                        return ret;
 894
 895                WARN_ON(!fill.max); /* Should always be at least one */
 896
 897                /*
 898                 * If there's enough space, fill it now, otherwise return
 899                 * -ENOSPC and the number of devices affected.
 900                 */
 901                if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
 902                        ret = -ENOSPC;
 903                        hdr.count = fill.max;
 904                        goto reset_info_exit;
 905                }
 906
 907                devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
 908                if (!devices)
 909                        return -ENOMEM;
 910
 911                fill.devices = devices;
 912
 913                ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
 914                                                    vfio_pci_fill_devs,
 915                                                    &fill, slot);
 916
 917                /*
 918                 * If a device was removed between counting and filling,
 919                 * we may come up short of fill.max.  If a device was
 920                 * added, we'll have a return of -EAGAIN above.
 921                 */
 922                if (!ret)
 923                        hdr.count = fill.cur;
 924
 925reset_info_exit:
 926                if (copy_to_user((void __user *)arg, &hdr, minsz))
 927                        ret = -EFAULT;
 928
 929                if (!ret) {
 930                        if (copy_to_user((void __user *)(arg + minsz), devices,
 931                                         hdr.count * sizeof(*devices)))
 932                                ret = -EFAULT;
 933                }
 934
 935                kfree(devices);
 936                return ret;
 937
 938        } else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
 939                struct vfio_pci_hot_reset hdr;
 940                int32_t *group_fds;
 941                struct vfio_pci_group_entry *groups;
 942                struct vfio_pci_group_info info;
 943                bool slot = false;
 944                int i, count = 0, ret = 0;
 945
 946                minsz = offsetofend(struct vfio_pci_hot_reset, count);
 947
 948                if (copy_from_user(&hdr, (void __user *)arg, minsz))
 949                        return -EFAULT;
 950
 951                if (hdr.argsz < minsz || hdr.flags)
 952                        return -EINVAL;
 953
 954                /* Can we do a slot or bus reset or neither? */
 955                if (!pci_probe_reset_slot(vdev->pdev->slot))
 956                        slot = true;
 957                else if (pci_probe_reset_bus(vdev->pdev->bus))
 958                        return -ENODEV;
 959
 960                /*
 961                 * We can't let userspace give us an arbitrarily large
 962                 * buffer to copy, so verify how many we think there
 963                 * could be.  Note groups can have multiple devices so
 964                 * one group per device is the max.
 965                 */
 966                ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
 967                                                    vfio_pci_count_devs,
 968                                                    &count, slot);
 969                if (ret)
 970                        return ret;
 971
 972                /* Somewhere between 1 and count is OK */
 973                if (!hdr.count || hdr.count > count)
 974                        return -EINVAL;
 975
 976                group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
 977                groups = kcalloc(hdr.count, sizeof(*groups), GFP_KERNEL);
 978                if (!group_fds || !groups) {
 979                        kfree(group_fds);
 980                        kfree(groups);
 981                        return -ENOMEM;
 982                }
 983
 984                if (copy_from_user(group_fds, (void __user *)(arg + minsz),
 985                                   hdr.count * sizeof(*group_fds))) {
 986                        kfree(group_fds);
 987                        kfree(groups);
 988                        return -EFAULT;
 989                }
 990
 991                /*
 992                 * For each group_fd, get the group through the vfio external
 993                 * user interface and store the group and iommu ID.  This
 994                 * ensures the group is held across the reset.
 995                 */
 996                for (i = 0; i < hdr.count; i++) {
 997                        struct vfio_group *group;
 998                        struct fd f = fdget(group_fds[i]);
 999                        if (!f.file) {
1000                                ret = -EBADF;
1001                                break;
1002                        }
1003
1004                        group = vfio_group_get_external_user(f.file);
1005                        fdput(f);
1006                        if (IS_ERR(group)) {
1007                                ret = PTR_ERR(group);
1008                                break;
1009                        }
1010
1011                        groups[i].group = group;
1012                        groups[i].id = vfio_external_user_iommu_id(group);
1013                }
1014
1015                kfree(group_fds);
1016
1017                /* release reference to groups on error */
1018                if (ret)
1019                        goto hot_reset_release;
1020
1021                info.count = hdr.count;
1022                info.groups = groups;
1023
1024                /*
1025                 * Test whether all the affected devices are contained
1026                 * by the set of groups provided by the user.
1027                 */
1028                ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1029                                                    vfio_pci_validate_devs,
1030                                                    &info, slot);
1031                if (!ret)
1032                        /* User has access, do the reset */
1033                        ret = slot ? pci_try_reset_slot(vdev->pdev->slot) :
1034                                     pci_try_reset_bus(vdev->pdev->bus);
1035
1036hot_reset_release:
1037                for (i--; i >= 0; i--)
1038                        vfio_group_put_external_user(groups[i].group);
1039
1040                kfree(groups);
1041                return ret;
1042        }
1043
1044        return -ENOTTY;
1045}
1046
1047static ssize_t vfio_pci_rw(void *device_data, char __user *buf,
1048                           size_t count, loff_t *ppos, bool iswrite)
1049{
1050        unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1051        struct vfio_pci_device *vdev = device_data;
1052
1053        if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1054                return -EINVAL;
1055
1056        switch (index) {
1057        case VFIO_PCI_CONFIG_REGION_INDEX:
1058                return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1059
1060        case VFIO_PCI_ROM_REGION_INDEX:
1061                if (iswrite)
1062                        return -EINVAL;
1063                return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1064
1065        case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1066                return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1067
1068        case VFIO_PCI_VGA_REGION_INDEX:
1069                return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1070        default:
1071                index -= VFIO_PCI_NUM_REGIONS;
1072                return vdev->region[index].ops->rw(vdev, buf,
1073                                                   count, ppos, iswrite);
1074        }
1075
1076        return -EINVAL;
1077}
1078
1079static ssize_t vfio_pci_read(void *device_data, char __user *buf,
1080                             size_t count, loff_t *ppos)
1081{
1082        if (!count)
1083                return 0;
1084
1085        return vfio_pci_rw(device_data, buf, count, ppos, false);
1086}
1087
1088static ssize_t vfio_pci_write(void *device_data, const char __user *buf,
1089                              size_t count, loff_t *ppos)
1090{
1091        if (!count)
1092                return 0;
1093
1094        return vfio_pci_rw(device_data, (char __user *)buf, count, ppos, true);
1095}
1096
1097static int vfio_pci_mmap(void *device_data, struct vm_area_struct *vma)
1098{
1099        struct vfio_pci_device *vdev = device_data;
1100        struct pci_dev *pdev = vdev->pdev;
1101        unsigned int index;
1102        u64 phys_len, req_len, pgoff, req_start;
1103        int ret;
1104
1105        index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1106
1107        if (vma->vm_end < vma->vm_start)
1108                return -EINVAL;
1109        if ((vma->vm_flags & VM_SHARED) == 0)
1110                return -EINVAL;
1111        if (index >= VFIO_PCI_ROM_REGION_INDEX)
1112                return -EINVAL;
1113        if (!vdev->bar_mmap_supported[index])
1114                return -EINVAL;
1115
1116        phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1117        req_len = vma->vm_end - vma->vm_start;
1118        pgoff = vma->vm_pgoff &
1119                ((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1120        req_start = pgoff << PAGE_SHIFT;
1121
1122        if (req_start + req_len > phys_len)
1123                return -EINVAL;
1124
1125        if (index == vdev->msix_bar) {
1126                /*
1127                 * Disallow mmaps overlapping the MSI-X table; users don't
1128                 * get to touch this directly.  We could find somewhere
1129                 * else to map the overlap, but page granularity is only
1130                 * a recommendation, not a requirement, so the user needs
1131                 * to know which bits are real.  Requiring them to mmap
1132                 * around the table makes that clear.
1133                 */
1134
1135                /* If neither entirely above nor below, then it overlaps */
1136                if (!(req_start >= vdev->msix_offset + vdev->msix_size ||
1137                      req_start + req_len <= vdev->msix_offset))
1138                        return -EINVAL;
1139        }
1140
1141        /*
1142         * Even though we don't make use of the barmap for the mmap,
1143         * we need to request the region and the barmap tracks that.
1144         */
1145        if (!vdev->barmap[index]) {
1146                ret = pci_request_selected_regions(pdev,
1147                                                   1 << index, "vfio-pci");
1148                if (ret)
1149                        return ret;
1150
1151                vdev->barmap[index] = pci_iomap(pdev, index, 0);
1152                if (!vdev->barmap[index]) {
1153                        pci_release_selected_regions(pdev, 1 << index);
1154                        return -ENOMEM;
1155                }
1156        }
1157
1158        vma->vm_private_data = vdev;
1159        vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1160        vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1161
1162        return remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1163                               req_len, vma->vm_page_prot);
1164}
1165
1166static void vfio_pci_request(void *device_data, unsigned int count)
1167{
1168        struct vfio_pci_device *vdev = device_data;
1169
1170        mutex_lock(&vdev->igate);
1171
1172        if (vdev->req_trigger) {
1173                if (!(count % 10))
1174                        dev_notice_ratelimited(&vdev->pdev->dev,
1175                                "Relaying device request to user (#%u)\n",
1176                                count);
1177                eventfd_signal(vdev->req_trigger, 1);
1178        } else if (count == 0) {
1179                dev_warn(&vdev->pdev->dev,
1180                        "No device request channel registered, blocked until released by user\n");
1181        }
1182
1183        mutex_unlock(&vdev->igate);
1184}
1185
1186static const struct vfio_device_ops vfio_pci_ops = {
1187        .name           = "vfio-pci",
1188        .open           = vfio_pci_open,
1189        .release        = vfio_pci_release,
1190        .ioctl          = vfio_pci_ioctl,
1191        .read           = vfio_pci_read,
1192        .write          = vfio_pci_write,
1193        .mmap           = vfio_pci_mmap,
1194        .request        = vfio_pci_request,
1195};
1196
1197static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1198{
1199        struct vfio_pci_device *vdev;
1200        struct iommu_group *group;
1201        int ret;
1202
1203        if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1204                return -EINVAL;
1205
1206        group = vfio_iommu_group_get(&pdev->dev);
1207        if (!group)
1208                return -EINVAL;
1209
1210        vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
1211        if (!vdev) {
1212                vfio_iommu_group_put(group, &pdev->dev);
1213                return -ENOMEM;
1214        }
1215
1216        vdev->pdev = pdev;
1217        vdev->irq_type = VFIO_PCI_NUM_IRQS;
1218        mutex_init(&vdev->igate);
1219        spin_lock_init(&vdev->irqlock);
1220
1221        ret = vfio_add_group_dev(&pdev->dev, &vfio_pci_ops, vdev);
1222        if (ret) {
1223                vfio_iommu_group_put(group, &pdev->dev);
1224                kfree(vdev);
1225                return ret;
1226        }
1227
1228        if (vfio_pci_is_vga(pdev)) {
1229                vga_client_register(pdev, vdev, NULL, vfio_pci_set_vga_decode);
1230                vga_set_legacy_decoding(pdev,
1231                                        vfio_pci_set_vga_decode(vdev, false));
1232        }
1233
1234        if (!disable_idle_d3) {
1235                /*
1236                 * pci-core sets the device power state to an unknown value at
1237                 * bootup and after being removed from a driver.  The only
1238                 * transition it allows from this unknown state is to D0, which
1239                 * typically happens when a driver calls pci_enable_device().
1240                 * We're not ready to enable the device yet, but we do want to
1241                 * be able to get to D3.  Therefore first do a D0 transition
1242                 * before going to D3.
1243                 */
1244                pci_set_power_state(pdev, PCI_D0);
1245                pci_set_power_state(pdev, PCI_D3hot);
1246        }
1247
1248        return ret;
1249}
1250
1251static void vfio_pci_remove(struct pci_dev *pdev)
1252{
1253        struct vfio_pci_device *vdev;
1254
1255        vdev = vfio_del_group_dev(&pdev->dev);
1256        if (!vdev)
1257                return;
1258
1259        vfio_iommu_group_put(pdev->dev.iommu_group, &pdev->dev);
1260        kfree(vdev->region);
1261        kfree(vdev);
1262
1263        if (vfio_pci_is_vga(pdev)) {
1264                vga_client_register(pdev, NULL, NULL, NULL);
1265                vga_set_legacy_decoding(pdev,
1266                                VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
1267                                VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM);
1268        }
1269
1270        if (!disable_idle_d3)
1271                pci_set_power_state(pdev, PCI_D0);
1272}
1273
1274static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
1275                                                  pci_channel_state_t state)
1276{
1277        struct vfio_pci_device *vdev;
1278        struct vfio_device *device;
1279
1280        device = vfio_device_get_from_dev(&pdev->dev);
1281        if (device == NULL)
1282                return PCI_ERS_RESULT_DISCONNECT;
1283
1284        vdev = vfio_device_data(device);
1285        if (vdev == NULL) {
1286                vfio_device_put(device);
1287                return PCI_ERS_RESULT_DISCONNECT;
1288        }
1289
1290        mutex_lock(&vdev->igate);
1291
1292        if (vdev->err_trigger)
1293                eventfd_signal(vdev->err_trigger, 1);
1294
1295        mutex_unlock(&vdev->igate);
1296
1297        vfio_device_put(device);
1298
1299        return PCI_ERS_RESULT_CAN_RECOVER;
1300}
1301
1302static const struct pci_error_handlers vfio_err_handlers = {
1303        .error_detected = vfio_pci_aer_err_detected,
1304};
1305
1306static struct pci_driver vfio_pci_driver = {
1307        .name           = "vfio-pci",
1308        .id_table       = NULL, /* only dynamic ids */
1309        .probe          = vfio_pci_probe,
1310        .remove         = vfio_pci_remove,
1311        .err_handler    = &vfio_err_handlers,
1312};
1313
1314struct vfio_devices {
1315        struct vfio_device **devices;
1316        int cur_index;
1317        int max_index;
1318};
1319
1320static int vfio_pci_get_devs(struct pci_dev *pdev, void *data)
1321{
1322        struct vfio_devices *devs = data;
1323        struct vfio_device *device;
1324
1325        if (devs->cur_index == devs->max_index)
1326                return -ENOSPC;
1327
1328        device = vfio_device_get_from_dev(&pdev->dev);
1329        if (!device)
1330                return -EINVAL;
1331
1332        if (pci_dev_driver(pdev) != &vfio_pci_driver) {
1333                vfio_device_put(device);
1334                return -EBUSY;
1335        }
1336
1337        devs->devices[devs->cur_index++] = device;
1338        return 0;
1339}
1340
1341/*
1342 * Attempt to do a bus/slot reset if there are devices affected by a reset for
1343 * this device that are needs_reset and all of the affected devices are unused
1344 * (!refcnt).  Callers are required to hold driver_lock when calling this to
1345 * prevent device opens and concurrent bus reset attempts.  We prevent device
1346 * unbinds by acquiring and holding a reference to the vfio_device.
1347 *
1348 * NB: vfio-core considers a group to be viable even if some devices are
1349 * bound to drivers like pci-stub or pcieport.  Here we require all devices
1350 * to be bound to vfio_pci since that's the only way we can be sure they
1351 * stay put.
1352 */
1353static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev)
1354{
1355        struct vfio_devices devs = { .cur_index = 0 };
1356        int i = 0, ret = -EINVAL;
1357        bool needs_reset = false, slot = false;
1358        struct vfio_pci_device *tmp;
1359
1360        if (!pci_probe_reset_slot(vdev->pdev->slot))
1361                slot = true;
1362        else if (pci_probe_reset_bus(vdev->pdev->bus))
1363                return;
1364
1365        if (vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
1366                                          &i, slot) || !i)
1367                return;
1368
1369        devs.max_index = i;
1370        devs.devices = kcalloc(i, sizeof(struct vfio_device *), GFP_KERNEL);
1371        if (!devs.devices)
1372                return;
1373
1374        if (vfio_pci_for_each_slot_or_bus(vdev->pdev,
1375                                          vfio_pci_get_devs, &devs, slot))
1376                goto put_devs;
1377
1378        for (i = 0; i < devs.cur_index; i++) {
1379                tmp = vfio_device_data(devs.devices[i]);
1380                if (tmp->needs_reset)
1381                        needs_reset = true;
1382                if (tmp->refcnt)
1383                        goto put_devs;
1384        }
1385
1386        if (needs_reset)
1387                ret = slot ? pci_try_reset_slot(vdev->pdev->slot) :
1388                             pci_try_reset_bus(vdev->pdev->bus);
1389
1390put_devs:
1391        for (i = 0; i < devs.cur_index; i++) {
1392                tmp = vfio_device_data(devs.devices[i]);
1393                if (!ret)
1394                        tmp->needs_reset = false;
1395
1396                if (!tmp->refcnt && !disable_idle_d3)
1397                        pci_set_power_state(tmp->pdev, PCI_D3hot);
1398
1399                vfio_device_put(devs.devices[i]);
1400        }
1401
1402        kfree(devs.devices);
1403}
1404
1405static void __exit vfio_pci_cleanup(void)
1406{
1407        pci_unregister_driver(&vfio_pci_driver);
1408        vfio_pci_uninit_perm_bits();
1409}
1410
1411static void __init vfio_pci_fill_ids(void)
1412{
1413        char *p, *id;
1414        int rc;
1415
1416        /* no ids passed actually */
1417        if (ids[0] == '\0')
1418                return;
1419
1420        /* add ids specified in the module parameter */
1421        p = ids;
1422        while ((id = strsep(&p, ","))) {
1423                unsigned int vendor, device, subvendor = PCI_ANY_ID,
1424                        subdevice = PCI_ANY_ID, class = 0, class_mask = 0;
1425                int fields;
1426
1427                if (!strlen(id))
1428                        continue;
1429
1430                fields = sscanf(id, "%x:%x:%x:%x:%x:%x",
1431                                &vendor, &device, &subvendor, &subdevice,
1432                                &class, &class_mask);
1433
1434                if (fields < 2) {
1435                        pr_warn("invalid id string \"%s\"\n", id);
1436                        continue;
1437                }
1438
1439                rc = pci_add_dynid(&vfio_pci_driver, vendor, device,
1440                                   subvendor, subdevice, class, class_mask, 0);
1441                if (rc)
1442                        pr_warn("failed to add dynamic id [%04hx:%04hx[%04hx:%04hx]] class %#08x/%08x (%d)\n",
1443                                vendor, device, subvendor, subdevice,
1444                                class, class_mask, rc);
1445                else
1446                        pr_info("add [%04hx:%04hx[%04hx:%04hx]] class %#08x/%08x\n",
1447                                vendor, device, subvendor, subdevice,
1448                                class, class_mask);
1449        }
1450}
1451
1452static int __init vfio_pci_init(void)
1453{
1454        int ret;
1455
1456        /* Allocate shared config space permision data used by all devices */
1457        ret = vfio_pci_init_perm_bits();
1458        if (ret)
1459                return ret;
1460
1461        /* Register and scan for devices */
1462        ret = pci_register_driver(&vfio_pci_driver);
1463        if (ret)
1464                goto out_driver;
1465
1466        vfio_pci_fill_ids();
1467
1468        return 0;
1469
1470out_driver:
1471        vfio_pci_uninit_perm_bits();
1472        return ret;
1473}
1474
1475module_init(vfio_pci_init);
1476module_exit(vfio_pci_cleanup);
1477
1478MODULE_VERSION(DRIVER_VERSION);
1479MODULE_LICENSE("GPL v2");
1480MODULE_AUTHOR(DRIVER_AUTHOR);
1481MODULE_DESCRIPTION(DRIVER_DESC);
1482