linux/drivers/vfio/pci/vfio_pci.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (C) 2012 Red Hat, Inc.  All rights reserved.
   4 *     Author: Alex Williamson <alex.williamson@redhat.com>
   5 *
   6 * Derived from original vfio:
   7 * Copyright 2010 Cisco Systems, Inc.  All rights reserved.
   8 * Author: Tom Lyon, pugs@cisco.com
   9 */
  10
  11#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  12
  13#include <linux/device.h>
  14#include <linux/eventfd.h>
  15#include <linux/file.h>
  16#include <linux/interrupt.h>
  17#include <linux/iommu.h>
  18#include <linux/module.h>
  19#include <linux/mutex.h>
  20#include <linux/notifier.h>
  21#include <linux/pci.h>
  22#include <linux/pm_runtime.h>
  23#include <linux/slab.h>
  24#include <linux/types.h>
  25#include <linux/uaccess.h>
  26#include <linux/vfio.h>
  27#include <linux/vgaarb.h>
  28#include <linux/nospec.h>
  29#include <linux/sched/mm.h>
  30
  31#include "vfio_pci_private.h"
  32
  33#define DRIVER_VERSION  "0.2"
  34#define DRIVER_AUTHOR   "Alex Williamson <alex.williamson@redhat.com>"
  35#define DRIVER_DESC     "VFIO PCI - User Level meta-driver"
  36
  37static char ids[1024] __initdata;
  38module_param_string(ids, ids, sizeof(ids), 0);
  39MODULE_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");
  40
  41static bool nointxmask;
  42module_param_named(nointxmask, nointxmask, bool, S_IRUGO | S_IWUSR);
  43MODULE_PARM_DESC(nointxmask,
  44                  "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.");
  45
  46#ifdef CONFIG_VFIO_PCI_VGA
  47static bool disable_vga;
  48module_param(disable_vga, bool, S_IRUGO);
  49MODULE_PARM_DESC(disable_vga, "Disable VGA resource access through vfio-pci");
  50#endif
  51
  52static bool disable_idle_d3;
  53module_param(disable_idle_d3, bool, S_IRUGO | S_IWUSR);
  54MODULE_PARM_DESC(disable_idle_d3,
  55                 "Disable using the PCI D3 low power state for idle, unused devices");
  56
  57static bool enable_sriov;
  58#ifdef CONFIG_PCI_IOV
  59module_param(enable_sriov, bool, 0644);
  60MODULE_PARM_DESC(enable_sriov, "Enable support for SR-IOV configuration.  Enabling SR-IOV on a PF typically requires support of the userspace PF driver, enabling VFs without such support may result in non-functional VFs or PF.");
  61#endif
  62
  63static bool disable_denylist;
  64module_param(disable_denylist, bool, 0444);
  65MODULE_PARM_DESC(disable_denylist, "Disable use of device denylist. Disabling the denylist allows binding to devices with known errata that may lead to exploitable stability or security issues when accessed by untrusted users.");
  66
  67static inline bool vfio_vga_disabled(void)
  68{
  69#ifdef CONFIG_VFIO_PCI_VGA
  70        return disable_vga;
  71#else
  72        return true;
  73#endif
  74}
  75
  76static bool vfio_pci_dev_in_denylist(struct pci_dev *pdev)
  77{
  78        switch (pdev->vendor) {
  79        case PCI_VENDOR_ID_INTEL:
  80                switch (pdev->device) {
  81                case PCI_DEVICE_ID_INTEL_QAT_C3XXX:
  82                case PCI_DEVICE_ID_INTEL_QAT_C3XXX_VF:
  83                case PCI_DEVICE_ID_INTEL_QAT_C62X:
  84                case PCI_DEVICE_ID_INTEL_QAT_C62X_VF:
  85                case PCI_DEVICE_ID_INTEL_QAT_DH895XCC:
  86                case PCI_DEVICE_ID_INTEL_QAT_DH895XCC_VF:
  87                        return true;
  88                default:
  89                        return false;
  90                }
  91        }
  92
  93        return false;
  94}
  95
  96static bool vfio_pci_is_denylisted(struct pci_dev *pdev)
  97{
  98        if (!vfio_pci_dev_in_denylist(pdev))
  99                return false;
 100
 101        if (disable_denylist) {
 102                pci_warn(pdev,
 103                         "device denylist disabled - allowing device %04x:%04x.\n",
 104                         pdev->vendor, pdev->device);
 105                return false;
 106        }
 107
 108        pci_warn(pdev, "%04x:%04x exists in vfio-pci device denylist, driver probing disallowed.\n",
 109                 pdev->vendor, pdev->device);
 110
 111        return true;
 112}
 113
 114/*
 115 * Our VGA arbiter participation is limited since we don't know anything
 116 * about the device itself.  However, if the device is the only VGA device
 117 * downstream of a bridge and VFIO VGA support is disabled, then we can
 118 * safely return legacy VGA IO and memory as not decoded since the user
 119 * has no way to get to it and routing can be disabled externally at the
 120 * bridge.
 121 */
 122static unsigned int vfio_pci_set_vga_decode(void *opaque, bool single_vga)
 123{
 124        struct vfio_pci_device *vdev = opaque;
 125        struct pci_dev *tmp = NULL, *pdev = vdev->pdev;
 126        unsigned char max_busnr;
 127        unsigned int decodes;
 128
 129        if (single_vga || !vfio_vga_disabled() || pci_is_root_bus(pdev->bus))
 130                return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
 131                       VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
 132
 133        max_busnr = pci_bus_max_busnr(pdev->bus);
 134        decodes = VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
 135
 136        while ((tmp = pci_get_class(PCI_CLASS_DISPLAY_VGA << 8, tmp)) != NULL) {
 137                if (tmp == pdev ||
 138                    pci_domain_nr(tmp->bus) != pci_domain_nr(pdev->bus) ||
 139                    pci_is_root_bus(tmp->bus))
 140                        continue;
 141
 142                if (tmp->bus->number >= pdev->bus->number &&
 143                    tmp->bus->number <= max_busnr) {
 144                        pci_dev_put(tmp);
 145                        decodes |= VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM;
 146                        break;
 147                }
 148        }
 149
 150        return decodes;
 151}
 152
 153static inline bool vfio_pci_is_vga(struct pci_dev *pdev)
 154{
 155        return (pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA;
 156}
 157
 158static void vfio_pci_probe_mmaps(struct vfio_pci_device *vdev)
 159{
 160        struct resource *res;
 161        int i;
 162        struct vfio_pci_dummy_resource *dummy_res;
 163
 164        for (i = 0; i < PCI_STD_NUM_BARS; i++) {
 165                int bar = i + PCI_STD_RESOURCES;
 166
 167                res = &vdev->pdev->resource[bar];
 168
 169                if (!IS_ENABLED(CONFIG_VFIO_PCI_MMAP))
 170                        goto no_mmap;
 171
 172                if (!(res->flags & IORESOURCE_MEM))
 173                        goto no_mmap;
 174
 175                /*
 176                 * The PCI core shouldn't set up a resource with a
 177                 * type but zero size. But there may be bugs that
 178                 * cause us to do that.
 179                 */
 180                if (!resource_size(res))
 181                        goto no_mmap;
 182
 183                if (resource_size(res) >= PAGE_SIZE) {
 184                        vdev->bar_mmap_supported[bar] = true;
 185                        continue;
 186                }
 187
 188                if (!(res->start & ~PAGE_MASK)) {
 189                        /*
 190                         * Add a dummy resource to reserve the remainder
 191                         * of the exclusive page in case that hot-add
 192                         * device's bar is assigned into it.
 193                         */
 194                        dummy_res = kzalloc(sizeof(*dummy_res), GFP_KERNEL);
 195                        if (dummy_res == NULL)
 196                                goto no_mmap;
 197
 198                        dummy_res->resource.name = "vfio sub-page reserved";
 199                        dummy_res->resource.start = res->end + 1;
 200                        dummy_res->resource.end = res->start + PAGE_SIZE - 1;
 201                        dummy_res->resource.flags = res->flags;
 202                        if (request_resource(res->parent,
 203                                                &dummy_res->resource)) {
 204                                kfree(dummy_res);
 205                                goto no_mmap;
 206                        }
 207                        dummy_res->index = bar;
 208                        list_add(&dummy_res->res_next,
 209                                        &vdev->dummy_resources_list);
 210                        vdev->bar_mmap_supported[bar] = true;
 211                        continue;
 212                }
 213                /*
 214                 * Here we don't handle the case when the BAR is not page
 215                 * aligned because we can't expect the BAR will be
 216                 * assigned into the same location in a page in guest
 217                 * when we passthrough the BAR. And it's hard to access
 218                 * this BAR in userspace because we have no way to get
 219                 * the BAR's location in a page.
 220                 */
 221no_mmap:
 222                vdev->bar_mmap_supported[bar] = false;
 223        }
 224}
 225
 226static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev);
 227static void vfio_pci_disable(struct vfio_pci_device *vdev);
 228static int vfio_pci_try_zap_and_vma_lock_cb(struct pci_dev *pdev, void *data);
 229
 230/*
 231 * INTx masking requires the ability to disable INTx signaling via PCI_COMMAND
 232 * _and_ the ability detect when the device is asserting INTx via PCI_STATUS.
 233 * If a device implements the former but not the latter we would typically
 234 * expect broken_intx_masking be set and require an exclusive interrupt.
 235 * However since we do have control of the device's ability to assert INTx,
 236 * we can instead pretend that the device does not implement INTx, virtualizing
 237 * the pin register to report zero and maintaining DisINTx set on the host.
 238 */
 239static bool vfio_pci_nointx(struct pci_dev *pdev)
 240{
 241        switch (pdev->vendor) {
 242        case PCI_VENDOR_ID_INTEL:
 243                switch (pdev->device) {
 244                /* All i40e (XL710/X710/XXV710) 10/20/25/40GbE NICs */
 245                case 0x1572:
 246                case 0x1574:
 247                case 0x1580 ... 0x1581:
 248                case 0x1583 ... 0x158b:
 249                case 0x37d0 ... 0x37d2:
 250                /* X550 */
 251                case 0x1563:
 252                        return true;
 253                default:
 254                        return false;
 255                }
 256        }
 257
 258        return false;
 259}
 260
 261static void vfio_pci_probe_power_state(struct vfio_pci_device *vdev)
 262{
 263        struct pci_dev *pdev = vdev->pdev;
 264        u16 pmcsr;
 265
 266        if (!pdev->pm_cap)
 267                return;
 268
 269        pci_read_config_word(pdev, pdev->pm_cap + PCI_PM_CTRL, &pmcsr);
 270
 271        vdev->needs_pm_restore = !(pmcsr & PCI_PM_CTRL_NO_SOFT_RESET);
 272}
 273
 274/*
 275 * pci_set_power_state() wrapper handling devices which perform a soft reset on
 276 * D3->D0 transition.  Save state prior to D0/1/2->D3, stash it on the vdev,
 277 * restore when returned to D0.  Saved separately from pci_saved_state for use
 278 * by PM capability emulation and separately from pci_dev internal saved state
 279 * to avoid it being overwritten and consumed around other resets.
 280 */
 281int vfio_pci_set_power_state(struct vfio_pci_device *vdev, pci_power_t state)
 282{
 283        struct pci_dev *pdev = vdev->pdev;
 284        bool needs_restore = false, needs_save = false;
 285        int ret;
 286
 287        if (vdev->needs_pm_restore) {
 288                if (pdev->current_state < PCI_D3hot && state >= PCI_D3hot) {
 289                        pci_save_state(pdev);
 290                        needs_save = true;
 291                }
 292
 293                if (pdev->current_state >= PCI_D3hot && state <= PCI_D0)
 294                        needs_restore = true;
 295        }
 296
 297        ret = pci_set_power_state(pdev, state);
 298
 299        if (!ret) {
 300                /* D3 might be unsupported via quirk, skip unless in D3 */
 301                if (needs_save && pdev->current_state >= PCI_D3hot) {
 302                        vdev->pm_save = pci_store_saved_state(pdev);
 303                } else if (needs_restore) {
 304                        pci_load_and_free_saved_state(pdev, &vdev->pm_save);
 305                        pci_restore_state(pdev);
 306                }
 307        }
 308
 309        return ret;
 310}
 311
 312static int vfio_pci_enable(struct vfio_pci_device *vdev)
 313{
 314        struct pci_dev *pdev = vdev->pdev;
 315        int ret;
 316        u16 cmd;
 317        u8 msix_pos;
 318
 319        vfio_pci_set_power_state(vdev, PCI_D0);
 320
 321        /* Don't allow our initial saved state to include busmaster */
 322        pci_clear_master(pdev);
 323
 324        ret = pci_enable_device(pdev);
 325        if (ret)
 326                return ret;
 327
 328        /* If reset fails because of the device lock, fail this path entirely */
 329        ret = pci_try_reset_function(pdev);
 330        if (ret == -EAGAIN) {
 331                pci_disable_device(pdev);
 332                return ret;
 333        }
 334
 335        vdev->reset_works = !ret;
 336        pci_save_state(pdev);
 337        vdev->pci_saved_state = pci_store_saved_state(pdev);
 338        if (!vdev->pci_saved_state)
 339                pci_dbg(pdev, "%s: Couldn't store saved state\n", __func__);
 340
 341        if (likely(!nointxmask)) {
 342                if (vfio_pci_nointx(pdev)) {
 343                        pci_info(pdev, "Masking broken INTx support\n");
 344                        vdev->nointx = true;
 345                        pci_intx(pdev, 0);
 346                } else
 347                        vdev->pci_2_3 = pci_intx_mask_supported(pdev);
 348        }
 349
 350        pci_read_config_word(pdev, PCI_COMMAND, &cmd);
 351        if (vdev->pci_2_3 && (cmd & PCI_COMMAND_INTX_DISABLE)) {
 352                cmd &= ~PCI_COMMAND_INTX_DISABLE;
 353                pci_write_config_word(pdev, PCI_COMMAND, cmd);
 354        }
 355
 356        ret = vfio_config_init(vdev);
 357        if (ret) {
 358                kfree(vdev->pci_saved_state);
 359                vdev->pci_saved_state = NULL;
 360                pci_disable_device(pdev);
 361                return ret;
 362        }
 363
 364        msix_pos = pdev->msix_cap;
 365        if (msix_pos) {
 366                u16 flags;
 367                u32 table;
 368
 369                pci_read_config_word(pdev, msix_pos + PCI_MSIX_FLAGS, &flags);
 370                pci_read_config_dword(pdev, msix_pos + PCI_MSIX_TABLE, &table);
 371
 372                vdev->msix_bar = table & PCI_MSIX_TABLE_BIR;
 373                vdev->msix_offset = table & PCI_MSIX_TABLE_OFFSET;
 374                vdev->msix_size = ((flags & PCI_MSIX_FLAGS_QSIZE) + 1) * 16;
 375        } else
 376                vdev->msix_bar = 0xFF;
 377
 378        if (!vfio_vga_disabled() && vfio_pci_is_vga(pdev))
 379                vdev->has_vga = true;
 380
 381        if (vfio_pci_is_vga(pdev) &&
 382            pdev->vendor == PCI_VENDOR_ID_INTEL &&
 383            IS_ENABLED(CONFIG_VFIO_PCI_IGD)) {
 384                ret = vfio_pci_igd_init(vdev);
 385                if (ret && ret != -ENODEV) {
 386                        pci_warn(pdev, "Failed to setup Intel IGD regions\n");
 387                        goto disable_exit;
 388                }
 389        }
 390
 391        vfio_pci_probe_mmaps(vdev);
 392
 393        return 0;
 394
 395disable_exit:
 396        vfio_pci_disable(vdev);
 397        return ret;
 398}
 399
 400static void vfio_pci_disable(struct vfio_pci_device *vdev)
 401{
 402        struct pci_dev *pdev = vdev->pdev;
 403        struct vfio_pci_dummy_resource *dummy_res, *tmp;
 404        struct vfio_pci_ioeventfd *ioeventfd, *ioeventfd_tmp;
 405        int i, bar;
 406
 407        /* Stop the device from further DMA */
 408        pci_clear_master(pdev);
 409
 410        vfio_pci_set_irqs_ioctl(vdev, VFIO_IRQ_SET_DATA_NONE |
 411                                VFIO_IRQ_SET_ACTION_TRIGGER,
 412                                vdev->irq_type, 0, 0, NULL);
 413
 414        /* Device closed, don't need mutex here */
 415        list_for_each_entry_safe(ioeventfd, ioeventfd_tmp,
 416                                 &vdev->ioeventfds_list, next) {
 417                vfio_virqfd_disable(&ioeventfd->virqfd);
 418                list_del(&ioeventfd->next);
 419                kfree(ioeventfd);
 420        }
 421        vdev->ioeventfds_nr = 0;
 422
 423        vdev->virq_disabled = false;
 424
 425        for (i = 0; i < vdev->num_regions; i++)
 426                vdev->region[i].ops->release(vdev, &vdev->region[i]);
 427
 428        vdev->num_regions = 0;
 429        kfree(vdev->region);
 430        vdev->region = NULL; /* don't krealloc a freed pointer */
 431
 432        vfio_config_free(vdev);
 433
 434        for (i = 0; i < PCI_STD_NUM_BARS; i++) {
 435                bar = i + PCI_STD_RESOURCES;
 436                if (!vdev->barmap[bar])
 437                        continue;
 438                pci_iounmap(pdev, vdev->barmap[bar]);
 439                pci_release_selected_regions(pdev, 1 << bar);
 440                vdev->barmap[bar] = NULL;
 441        }
 442
 443        list_for_each_entry_safe(dummy_res, tmp,
 444                                 &vdev->dummy_resources_list, res_next) {
 445                list_del(&dummy_res->res_next);
 446                release_resource(&dummy_res->resource);
 447                kfree(dummy_res);
 448        }
 449
 450        vdev->needs_reset = true;
 451
 452        /*
 453         * If we have saved state, restore it.  If we can reset the device,
 454         * even better.  Resetting with current state seems better than
 455         * nothing, but saving and restoring current state without reset
 456         * is just busy work.
 457         */
 458        if (pci_load_and_free_saved_state(pdev, &vdev->pci_saved_state)) {
 459                pci_info(pdev, "%s: Couldn't reload saved state\n", __func__);
 460
 461                if (!vdev->reset_works)
 462                        goto out;
 463
 464                pci_save_state(pdev);
 465        }
 466
 467        /*
 468         * Disable INTx and MSI, presumably to avoid spurious interrupts
 469         * during reset.  Stolen from pci_reset_function()
 470         */
 471        pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
 472
 473        /*
 474         * Try to get the locks ourselves to prevent a deadlock. The
 475         * success of this is dependent on being able to lock the device,
 476         * which is not always possible.
 477         * We can not use the "try" reset interface here, which will
 478         * overwrite the previously restored configuration information.
 479         */
 480        if (vdev->reset_works && pci_dev_trylock(pdev)) {
 481                if (!__pci_reset_function_locked(pdev))
 482                        vdev->needs_reset = false;
 483                pci_dev_unlock(pdev);
 484        }
 485
 486        pci_restore_state(pdev);
 487out:
 488        pci_disable_device(pdev);
 489
 490        vfio_pci_try_bus_reset(vdev);
 491
 492        if (!disable_idle_d3)
 493                vfio_pci_set_power_state(vdev, PCI_D3hot);
 494}
 495
 496static struct pci_driver vfio_pci_driver;
 497
 498static struct vfio_pci_device *get_pf_vdev(struct vfio_pci_device *vdev)
 499{
 500        struct pci_dev *physfn = pci_physfn(vdev->pdev);
 501        struct vfio_device *pf_dev;
 502
 503        if (!vdev->pdev->is_virtfn)
 504                return NULL;
 505
 506        pf_dev = vfio_device_get_from_dev(&physfn->dev);
 507        if (!pf_dev)
 508                return NULL;
 509
 510        if (pci_dev_driver(physfn) != &vfio_pci_driver) {
 511                vfio_device_put(pf_dev);
 512                return NULL;
 513        }
 514
 515        return container_of(pf_dev, struct vfio_pci_device, vdev);
 516}
 517
 518static void vfio_pci_vf_token_user_add(struct vfio_pci_device *vdev, int val)
 519{
 520        struct vfio_pci_device *pf_vdev = get_pf_vdev(vdev);
 521
 522        if (!pf_vdev)
 523                return;
 524
 525        mutex_lock(&pf_vdev->vf_token->lock);
 526        pf_vdev->vf_token->users += val;
 527        WARN_ON(pf_vdev->vf_token->users < 0);
 528        mutex_unlock(&pf_vdev->vf_token->lock);
 529
 530        vfio_device_put(&pf_vdev->vdev);
 531}
 532
 533static void vfio_pci_release(struct vfio_device *core_vdev)
 534{
 535        struct vfio_pci_device *vdev =
 536                container_of(core_vdev, struct vfio_pci_device, vdev);
 537
 538        mutex_lock(&vdev->reflck->lock);
 539
 540        if (!(--vdev->refcnt)) {
 541                vfio_pci_vf_token_user_add(vdev, -1);
 542                vfio_spapr_pci_eeh_release(vdev->pdev);
 543                vfio_pci_disable(vdev);
 544
 545                mutex_lock(&vdev->igate);
 546                if (vdev->err_trigger) {
 547                        eventfd_ctx_put(vdev->err_trigger);
 548                        vdev->err_trigger = NULL;
 549                }
 550                if (vdev->req_trigger) {
 551                        eventfd_ctx_put(vdev->req_trigger);
 552                        vdev->req_trigger = NULL;
 553                }
 554                mutex_unlock(&vdev->igate);
 555        }
 556
 557        mutex_unlock(&vdev->reflck->lock);
 558}
 559
 560static int vfio_pci_open(struct vfio_device *core_vdev)
 561{
 562        struct vfio_pci_device *vdev =
 563                container_of(core_vdev, struct vfio_pci_device, vdev);
 564        int ret = 0;
 565
 566        mutex_lock(&vdev->reflck->lock);
 567
 568        if (!vdev->refcnt) {
 569                ret = vfio_pci_enable(vdev);
 570                if (ret)
 571                        goto error;
 572
 573                vfio_spapr_pci_eeh_open(vdev->pdev);
 574                vfio_pci_vf_token_user_add(vdev, 1);
 575        }
 576        vdev->refcnt++;
 577error:
 578        mutex_unlock(&vdev->reflck->lock);
 579        return ret;
 580}
 581
 582static int vfio_pci_get_irq_count(struct vfio_pci_device *vdev, int irq_type)
 583{
 584        if (irq_type == VFIO_PCI_INTX_IRQ_INDEX) {
 585                u8 pin;
 586
 587                if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) ||
 588                    vdev->nointx || vdev->pdev->is_virtfn)
 589                        return 0;
 590
 591                pci_read_config_byte(vdev->pdev, PCI_INTERRUPT_PIN, &pin);
 592
 593                return pin ? 1 : 0;
 594        } else if (irq_type == VFIO_PCI_MSI_IRQ_INDEX) {
 595                u8 pos;
 596                u16 flags;
 597
 598                pos = vdev->pdev->msi_cap;
 599                if (pos) {
 600                        pci_read_config_word(vdev->pdev,
 601                                             pos + PCI_MSI_FLAGS, &flags);
 602                        return 1 << ((flags & PCI_MSI_FLAGS_QMASK) >> 1);
 603                }
 604        } else if (irq_type == VFIO_PCI_MSIX_IRQ_INDEX) {
 605                u8 pos;
 606                u16 flags;
 607
 608                pos = vdev->pdev->msix_cap;
 609                if (pos) {
 610                        pci_read_config_word(vdev->pdev,
 611                                             pos + PCI_MSIX_FLAGS, &flags);
 612
 613                        return (flags & PCI_MSIX_FLAGS_QSIZE) + 1;
 614                }
 615        } else if (irq_type == VFIO_PCI_ERR_IRQ_INDEX) {
 616                if (pci_is_pcie(vdev->pdev))
 617                        return 1;
 618        } else if (irq_type == VFIO_PCI_REQ_IRQ_INDEX) {
 619                return 1;
 620        }
 621
 622        return 0;
 623}
 624
 625static int vfio_pci_count_devs(struct pci_dev *pdev, void *data)
 626{
 627        (*(int *)data)++;
 628        return 0;
 629}
 630
 631struct vfio_pci_fill_info {
 632        int max;
 633        int cur;
 634        struct vfio_pci_dependent_device *devices;
 635};
 636
 637static int vfio_pci_fill_devs(struct pci_dev *pdev, void *data)
 638{
 639        struct vfio_pci_fill_info *fill = data;
 640        struct iommu_group *iommu_group;
 641
 642        if (fill->cur == fill->max)
 643                return -EAGAIN; /* Something changed, try again */
 644
 645        iommu_group = iommu_group_get(&pdev->dev);
 646        if (!iommu_group)
 647                return -EPERM; /* Cannot reset non-isolated devices */
 648
 649        fill->devices[fill->cur].group_id = iommu_group_id(iommu_group);
 650        fill->devices[fill->cur].segment = pci_domain_nr(pdev->bus);
 651        fill->devices[fill->cur].bus = pdev->bus->number;
 652        fill->devices[fill->cur].devfn = pdev->devfn;
 653        fill->cur++;
 654        iommu_group_put(iommu_group);
 655        return 0;
 656}
 657
 658struct vfio_pci_group_entry {
 659        struct vfio_group *group;
 660        int id;
 661};
 662
 663struct vfio_pci_group_info {
 664        int count;
 665        struct vfio_pci_group_entry *groups;
 666};
 667
 668static int vfio_pci_validate_devs(struct pci_dev *pdev, void *data)
 669{
 670        struct vfio_pci_group_info *info = data;
 671        struct iommu_group *group;
 672        int id, i;
 673
 674        group = iommu_group_get(&pdev->dev);
 675        if (!group)
 676                return -EPERM;
 677
 678        id = iommu_group_id(group);
 679
 680        for (i = 0; i < info->count; i++)
 681                if (info->groups[i].id == id)
 682                        break;
 683
 684        iommu_group_put(group);
 685
 686        return (i == info->count) ? -EINVAL : 0;
 687}
 688
 689static bool vfio_pci_dev_below_slot(struct pci_dev *pdev, struct pci_slot *slot)
 690{
 691        for (; pdev; pdev = pdev->bus->self)
 692                if (pdev->bus == slot->bus)
 693                        return (pdev->slot == slot);
 694        return false;
 695}
 696
 697struct vfio_pci_walk_info {
 698        int (*fn)(struct pci_dev *, void *data);
 699        void *data;
 700        struct pci_dev *pdev;
 701        bool slot;
 702        int ret;
 703};
 704
 705static int vfio_pci_walk_wrapper(struct pci_dev *pdev, void *data)
 706{
 707        struct vfio_pci_walk_info *walk = data;
 708
 709        if (!walk->slot || vfio_pci_dev_below_slot(pdev, walk->pdev->slot))
 710                walk->ret = walk->fn(pdev, walk->data);
 711
 712        return walk->ret;
 713}
 714
 715static int vfio_pci_for_each_slot_or_bus(struct pci_dev *pdev,
 716                                         int (*fn)(struct pci_dev *,
 717                                                   void *data), void *data,
 718                                         bool slot)
 719{
 720        struct vfio_pci_walk_info walk = {
 721                .fn = fn, .data = data, .pdev = pdev, .slot = slot, .ret = 0,
 722        };
 723
 724        pci_walk_bus(pdev->bus, vfio_pci_walk_wrapper, &walk);
 725
 726        return walk.ret;
 727}
 728
 729static int msix_mmappable_cap(struct vfio_pci_device *vdev,
 730                              struct vfio_info_cap *caps)
 731{
 732        struct vfio_info_cap_header header = {
 733                .id = VFIO_REGION_INFO_CAP_MSIX_MAPPABLE,
 734                .version = 1
 735        };
 736
 737        return vfio_info_add_capability(caps, &header, sizeof(header));
 738}
 739
 740int vfio_pci_register_dev_region(struct vfio_pci_device *vdev,
 741                                 unsigned int type, unsigned int subtype,
 742                                 const struct vfio_pci_regops *ops,
 743                                 size_t size, u32 flags, void *data)
 744{
 745        struct vfio_pci_region *region;
 746
 747        region = krealloc(vdev->region,
 748                          (vdev->num_regions + 1) * sizeof(*region),
 749                          GFP_KERNEL);
 750        if (!region)
 751                return -ENOMEM;
 752
 753        vdev->region = region;
 754        vdev->region[vdev->num_regions].type = type;
 755        vdev->region[vdev->num_regions].subtype = subtype;
 756        vdev->region[vdev->num_regions].ops = ops;
 757        vdev->region[vdev->num_regions].size = size;
 758        vdev->region[vdev->num_regions].flags = flags;
 759        vdev->region[vdev->num_regions].data = data;
 760
 761        vdev->num_regions++;
 762
 763        return 0;
 764}
 765
 766struct vfio_devices {
 767        struct vfio_pci_device **devices;
 768        int cur_index;
 769        int max_index;
 770};
 771
 772static long vfio_pci_ioctl(struct vfio_device *core_vdev,
 773                           unsigned int cmd, unsigned long arg)
 774{
 775        struct vfio_pci_device *vdev =
 776                container_of(core_vdev, struct vfio_pci_device, vdev);
 777        unsigned long minsz;
 778
 779        if (cmd == VFIO_DEVICE_GET_INFO) {
 780                struct vfio_device_info info;
 781                struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
 782                unsigned long capsz;
 783                int ret;
 784
 785                minsz = offsetofend(struct vfio_device_info, num_irqs);
 786
 787                /* For backward compatibility, cannot require this */
 788                capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
 789
 790                if (copy_from_user(&info, (void __user *)arg, minsz))
 791                        return -EFAULT;
 792
 793                if (info.argsz < minsz)
 794                        return -EINVAL;
 795
 796                if (info.argsz >= capsz) {
 797                        minsz = capsz;
 798                        info.cap_offset = 0;
 799                }
 800
 801                info.flags = VFIO_DEVICE_FLAGS_PCI;
 802
 803                if (vdev->reset_works)
 804                        info.flags |= VFIO_DEVICE_FLAGS_RESET;
 805
 806                info.num_regions = VFIO_PCI_NUM_REGIONS + vdev->num_regions;
 807                info.num_irqs = VFIO_PCI_NUM_IRQS;
 808
 809                ret = vfio_pci_info_zdev_add_caps(vdev, &caps);
 810                if (ret && ret != -ENODEV) {
 811                        pci_warn(vdev->pdev, "Failed to setup zPCI info capabilities\n");
 812                        return ret;
 813                }
 814
 815                if (caps.size) {
 816                        info.flags |= VFIO_DEVICE_FLAGS_CAPS;
 817                        if (info.argsz < sizeof(info) + caps.size) {
 818                                info.argsz = sizeof(info) + caps.size;
 819                        } else {
 820                                vfio_info_cap_shift(&caps, sizeof(info));
 821                                if (copy_to_user((void __user *)arg +
 822                                                  sizeof(info), caps.buf,
 823                                                  caps.size)) {
 824                                        kfree(caps.buf);
 825                                        return -EFAULT;
 826                                }
 827                                info.cap_offset = sizeof(info);
 828                        }
 829
 830                        kfree(caps.buf);
 831                }
 832
 833                return copy_to_user((void __user *)arg, &info, minsz) ?
 834                        -EFAULT : 0;
 835
 836        } else if (cmd == VFIO_DEVICE_GET_REGION_INFO) {
 837                struct pci_dev *pdev = vdev->pdev;
 838                struct vfio_region_info info;
 839                struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
 840                int i, ret;
 841
 842                minsz = offsetofend(struct vfio_region_info, offset);
 843
 844                if (copy_from_user(&info, (void __user *)arg, minsz))
 845                        return -EFAULT;
 846
 847                if (info.argsz < minsz)
 848                        return -EINVAL;
 849
 850                switch (info.index) {
 851                case VFIO_PCI_CONFIG_REGION_INDEX:
 852                        info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 853                        info.size = pdev->cfg_size;
 854                        info.flags = VFIO_REGION_INFO_FLAG_READ |
 855                                     VFIO_REGION_INFO_FLAG_WRITE;
 856                        break;
 857                case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
 858                        info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 859                        info.size = pci_resource_len(pdev, info.index);
 860                        if (!info.size) {
 861                                info.flags = 0;
 862                                break;
 863                        }
 864
 865                        info.flags = VFIO_REGION_INFO_FLAG_READ |
 866                                     VFIO_REGION_INFO_FLAG_WRITE;
 867                        if (vdev->bar_mmap_supported[info.index]) {
 868                                info.flags |= VFIO_REGION_INFO_FLAG_MMAP;
 869                                if (info.index == vdev->msix_bar) {
 870                                        ret = msix_mmappable_cap(vdev, &caps);
 871                                        if (ret)
 872                                                return ret;
 873                                }
 874                        }
 875
 876                        break;
 877                case VFIO_PCI_ROM_REGION_INDEX:
 878                {
 879                        void __iomem *io;
 880                        size_t size;
 881                        u16 cmd;
 882
 883                        info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 884                        info.flags = 0;
 885
 886                        /* Report the BAR size, not the ROM size */
 887                        info.size = pci_resource_len(pdev, info.index);
 888                        if (!info.size) {
 889                                /* Shadow ROMs appear as PCI option ROMs */
 890                                if (pdev->resource[PCI_ROM_RESOURCE].flags &
 891                                                        IORESOURCE_ROM_SHADOW)
 892                                        info.size = 0x20000;
 893                                else
 894                                        break;
 895                        }
 896
 897                        /*
 898                         * Is it really there?  Enable memory decode for
 899                         * implicit access in pci_map_rom().
 900                         */
 901                        cmd = vfio_pci_memory_lock_and_enable(vdev);
 902                        io = pci_map_rom(pdev, &size);
 903                        if (io) {
 904                                info.flags = VFIO_REGION_INFO_FLAG_READ;
 905                                pci_unmap_rom(pdev, io);
 906                        } else {
 907                                info.size = 0;
 908                        }
 909                        vfio_pci_memory_unlock_and_restore(vdev, cmd);
 910
 911                        break;
 912                }
 913                case VFIO_PCI_VGA_REGION_INDEX:
 914                        if (!vdev->has_vga)
 915                                return -EINVAL;
 916
 917                        info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 918                        info.size = 0xc0000;
 919                        info.flags = VFIO_REGION_INFO_FLAG_READ |
 920                                     VFIO_REGION_INFO_FLAG_WRITE;
 921
 922                        break;
 923                default:
 924                {
 925                        struct vfio_region_info_cap_type cap_type = {
 926                                        .header.id = VFIO_REGION_INFO_CAP_TYPE,
 927                                        .header.version = 1 };
 928
 929                        if (info.index >=
 930                            VFIO_PCI_NUM_REGIONS + vdev->num_regions)
 931                                return -EINVAL;
 932                        info.index = array_index_nospec(info.index,
 933                                                        VFIO_PCI_NUM_REGIONS +
 934                                                        vdev->num_regions);
 935
 936                        i = info.index - VFIO_PCI_NUM_REGIONS;
 937
 938                        info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
 939                        info.size = vdev->region[i].size;
 940                        info.flags = vdev->region[i].flags;
 941
 942                        cap_type.type = vdev->region[i].type;
 943                        cap_type.subtype = vdev->region[i].subtype;
 944
 945                        ret = vfio_info_add_capability(&caps, &cap_type.header,
 946                                                       sizeof(cap_type));
 947                        if (ret)
 948                                return ret;
 949
 950                        if (vdev->region[i].ops->add_capability) {
 951                                ret = vdev->region[i].ops->add_capability(vdev,
 952                                                &vdev->region[i], &caps);
 953                                if (ret)
 954                                        return ret;
 955                        }
 956                }
 957                }
 958
 959                if (caps.size) {
 960                        info.flags |= VFIO_REGION_INFO_FLAG_CAPS;
 961                        if (info.argsz < sizeof(info) + caps.size) {
 962                                info.argsz = sizeof(info) + caps.size;
 963                                info.cap_offset = 0;
 964                        } else {
 965                                vfio_info_cap_shift(&caps, sizeof(info));
 966                                if (copy_to_user((void __user *)arg +
 967                                                  sizeof(info), caps.buf,
 968                                                  caps.size)) {
 969                                        kfree(caps.buf);
 970                                        return -EFAULT;
 971                                }
 972                                info.cap_offset = sizeof(info);
 973                        }
 974
 975                        kfree(caps.buf);
 976                }
 977
 978                return copy_to_user((void __user *)arg, &info, minsz) ?
 979                        -EFAULT : 0;
 980
 981        } else if (cmd == VFIO_DEVICE_GET_IRQ_INFO) {
 982                struct vfio_irq_info info;
 983
 984                minsz = offsetofend(struct vfio_irq_info, count);
 985
 986                if (copy_from_user(&info, (void __user *)arg, minsz))
 987                        return -EFAULT;
 988
 989                if (info.argsz < minsz || info.index >= VFIO_PCI_NUM_IRQS)
 990                        return -EINVAL;
 991
 992                switch (info.index) {
 993                case VFIO_PCI_INTX_IRQ_INDEX ... VFIO_PCI_MSIX_IRQ_INDEX:
 994                case VFIO_PCI_REQ_IRQ_INDEX:
 995                        break;
 996                case VFIO_PCI_ERR_IRQ_INDEX:
 997                        if (pci_is_pcie(vdev->pdev))
 998                                break;
 999                        fallthrough;
1000                default:
1001                        return -EINVAL;
1002                }
1003
1004                info.flags = VFIO_IRQ_INFO_EVENTFD;
1005
1006                info.count = vfio_pci_get_irq_count(vdev, info.index);
1007
1008                if (info.index == VFIO_PCI_INTX_IRQ_INDEX)
1009                        info.flags |= (VFIO_IRQ_INFO_MASKABLE |
1010                                       VFIO_IRQ_INFO_AUTOMASKED);
1011                else
1012                        info.flags |= VFIO_IRQ_INFO_NORESIZE;
1013
1014                return copy_to_user((void __user *)arg, &info, minsz) ?
1015                        -EFAULT : 0;
1016
1017        } else if (cmd == VFIO_DEVICE_SET_IRQS) {
1018                struct vfio_irq_set hdr;
1019                u8 *data = NULL;
1020                int max, ret = 0;
1021                size_t data_size = 0;
1022
1023                minsz = offsetofend(struct vfio_irq_set, count);
1024
1025                if (copy_from_user(&hdr, (void __user *)arg, minsz))
1026                        return -EFAULT;
1027
1028                max = vfio_pci_get_irq_count(vdev, hdr.index);
1029
1030                ret = vfio_set_irqs_validate_and_prepare(&hdr, max,
1031                                                 VFIO_PCI_NUM_IRQS, &data_size);
1032                if (ret)
1033                        return ret;
1034
1035                if (data_size) {
1036                        data = memdup_user((void __user *)(arg + minsz),
1037                                            data_size);
1038                        if (IS_ERR(data))
1039                                return PTR_ERR(data);
1040                }
1041
1042                mutex_lock(&vdev->igate);
1043
1044                ret = vfio_pci_set_irqs_ioctl(vdev, hdr.flags, hdr.index,
1045                                              hdr.start, hdr.count, data);
1046
1047                mutex_unlock(&vdev->igate);
1048                kfree(data);
1049
1050                return ret;
1051
1052        } else if (cmd == VFIO_DEVICE_RESET) {
1053                int ret;
1054
1055                if (!vdev->reset_works)
1056                        return -EINVAL;
1057
1058                vfio_pci_zap_and_down_write_memory_lock(vdev);
1059                ret = pci_try_reset_function(vdev->pdev);
1060                up_write(&vdev->memory_lock);
1061
1062                return ret;
1063
1064        } else if (cmd == VFIO_DEVICE_GET_PCI_HOT_RESET_INFO) {
1065                struct vfio_pci_hot_reset_info hdr;
1066                struct vfio_pci_fill_info fill = { 0 };
1067                struct vfio_pci_dependent_device *devices = NULL;
1068                bool slot = false;
1069                int ret = 0;
1070
1071                minsz = offsetofend(struct vfio_pci_hot_reset_info, count);
1072
1073                if (copy_from_user(&hdr, (void __user *)arg, minsz))
1074                        return -EFAULT;
1075
1076                if (hdr.argsz < minsz)
1077                        return -EINVAL;
1078
1079                hdr.flags = 0;
1080
1081                /* Can we do a slot or bus reset or neither? */
1082                if (!pci_probe_reset_slot(vdev->pdev->slot))
1083                        slot = true;
1084                else if (pci_probe_reset_bus(vdev->pdev->bus))
1085                        return -ENODEV;
1086
1087                /* How many devices are affected? */
1088                ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1089                                                    vfio_pci_count_devs,
1090                                                    &fill.max, slot);
1091                if (ret)
1092                        return ret;
1093
1094                WARN_ON(!fill.max); /* Should always be at least one */
1095
1096                /*
1097                 * If there's enough space, fill it now, otherwise return
1098                 * -ENOSPC and the number of devices affected.
1099                 */
1100                if (hdr.argsz < sizeof(hdr) + (fill.max * sizeof(*devices))) {
1101                        ret = -ENOSPC;
1102                        hdr.count = fill.max;
1103                        goto reset_info_exit;
1104                }
1105
1106                devices = kcalloc(fill.max, sizeof(*devices), GFP_KERNEL);
1107                if (!devices)
1108                        return -ENOMEM;
1109
1110                fill.devices = devices;
1111
1112                ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1113                                                    vfio_pci_fill_devs,
1114                                                    &fill, slot);
1115
1116                /*
1117                 * If a device was removed between counting and filling,
1118                 * we may come up short of fill.max.  If a device was
1119                 * added, we'll have a return of -EAGAIN above.
1120                 */
1121                if (!ret)
1122                        hdr.count = fill.cur;
1123
1124reset_info_exit:
1125                if (copy_to_user((void __user *)arg, &hdr, minsz))
1126                        ret = -EFAULT;
1127
1128                if (!ret) {
1129                        if (copy_to_user((void __user *)(arg + minsz), devices,
1130                                         hdr.count * sizeof(*devices)))
1131                                ret = -EFAULT;
1132                }
1133
1134                kfree(devices);
1135                return ret;
1136
1137        } else if (cmd == VFIO_DEVICE_PCI_HOT_RESET) {
1138                struct vfio_pci_hot_reset hdr;
1139                int32_t *group_fds;
1140                struct vfio_pci_group_entry *groups;
1141                struct vfio_pci_group_info info;
1142                struct vfio_devices devs = { .cur_index = 0 };
1143                bool slot = false;
1144                int i, group_idx, mem_idx = 0, count = 0, ret = 0;
1145
1146                minsz = offsetofend(struct vfio_pci_hot_reset, count);
1147
1148                if (copy_from_user(&hdr, (void __user *)arg, minsz))
1149                        return -EFAULT;
1150
1151                if (hdr.argsz < minsz || hdr.flags)
1152                        return -EINVAL;
1153
1154                /* Can we do a slot or bus reset or neither? */
1155                if (!pci_probe_reset_slot(vdev->pdev->slot))
1156                        slot = true;
1157                else if (pci_probe_reset_bus(vdev->pdev->bus))
1158                        return -ENODEV;
1159
1160                /*
1161                 * We can't let userspace give us an arbitrarily large
1162                 * buffer to copy, so verify how many we think there
1163                 * could be.  Note groups can have multiple devices so
1164                 * one group per device is the max.
1165                 */
1166                ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1167                                                    vfio_pci_count_devs,
1168                                                    &count, slot);
1169                if (ret)
1170                        return ret;
1171
1172                /* Somewhere between 1 and count is OK */
1173                if (!hdr.count || hdr.count > count)
1174                        return -EINVAL;
1175
1176                group_fds = kcalloc(hdr.count, sizeof(*group_fds), GFP_KERNEL);
1177                groups = kcalloc(hdr.count, sizeof(*groups), GFP_KERNEL);
1178                if (!group_fds || !groups) {
1179                        kfree(group_fds);
1180                        kfree(groups);
1181                        return -ENOMEM;
1182                }
1183
1184                if (copy_from_user(group_fds, (void __user *)(arg + minsz),
1185                                   hdr.count * sizeof(*group_fds))) {
1186                        kfree(group_fds);
1187                        kfree(groups);
1188                        return -EFAULT;
1189                }
1190
1191                /*
1192                 * For each group_fd, get the group through the vfio external
1193                 * user interface and store the group and iommu ID.  This
1194                 * ensures the group is held across the reset.
1195                 */
1196                for (group_idx = 0; group_idx < hdr.count; group_idx++) {
1197                        struct vfio_group *group;
1198                        struct fd f = fdget(group_fds[group_idx]);
1199                        if (!f.file) {
1200                                ret = -EBADF;
1201                                break;
1202                        }
1203
1204                        group = vfio_group_get_external_user(f.file);
1205                        fdput(f);
1206                        if (IS_ERR(group)) {
1207                                ret = PTR_ERR(group);
1208                                break;
1209                        }
1210
1211                        groups[group_idx].group = group;
1212                        groups[group_idx].id =
1213                                        vfio_external_user_iommu_id(group);
1214                }
1215
1216                kfree(group_fds);
1217
1218                /* release reference to groups on error */
1219                if (ret)
1220                        goto hot_reset_release;
1221
1222                info.count = hdr.count;
1223                info.groups = groups;
1224
1225                /*
1226                 * Test whether all the affected devices are contained
1227                 * by the set of groups provided by the user.
1228                 */
1229                ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1230                                                    vfio_pci_validate_devs,
1231                                                    &info, slot);
1232                if (ret)
1233                        goto hot_reset_release;
1234
1235                devs.max_index = count;
1236                devs.devices = kcalloc(count, sizeof(struct vfio_device *),
1237                                       GFP_KERNEL);
1238                if (!devs.devices) {
1239                        ret = -ENOMEM;
1240                        goto hot_reset_release;
1241                }
1242
1243                /*
1244                 * We need to get memory_lock for each device, but devices
1245                 * can share mmap_lock, therefore we need to zap and hold
1246                 * the vma_lock for each device, and only then get each
1247                 * memory_lock.
1248                 */
1249                ret = vfio_pci_for_each_slot_or_bus(vdev->pdev,
1250                                            vfio_pci_try_zap_and_vma_lock_cb,
1251                                            &devs, slot);
1252                if (ret)
1253                        goto hot_reset_release;
1254
1255                for (; mem_idx < devs.cur_index; mem_idx++) {
1256                        struct vfio_pci_device *tmp = devs.devices[mem_idx];
1257
1258                        ret = down_write_trylock(&tmp->memory_lock);
1259                        if (!ret) {
1260                                ret = -EBUSY;
1261                                goto hot_reset_release;
1262                        }
1263                        mutex_unlock(&tmp->vma_lock);
1264                }
1265
1266                /* User has access, do the reset */
1267                ret = pci_reset_bus(vdev->pdev);
1268
1269hot_reset_release:
1270                for (i = 0; i < devs.cur_index; i++) {
1271                        struct vfio_pci_device *tmp = devs.devices[i];
1272
1273                        if (i < mem_idx)
1274                                up_write(&tmp->memory_lock);
1275                        else
1276                                mutex_unlock(&tmp->vma_lock);
1277                        vfio_device_put(&tmp->vdev);
1278                }
1279                kfree(devs.devices);
1280
1281                for (group_idx--; group_idx >= 0; group_idx--)
1282                        vfio_group_put_external_user(groups[group_idx].group);
1283
1284                kfree(groups);
1285                return ret;
1286        } else if (cmd == VFIO_DEVICE_IOEVENTFD) {
1287                struct vfio_device_ioeventfd ioeventfd;
1288                int count;
1289
1290                minsz = offsetofend(struct vfio_device_ioeventfd, fd);
1291
1292                if (copy_from_user(&ioeventfd, (void __user *)arg, minsz))
1293                        return -EFAULT;
1294
1295                if (ioeventfd.argsz < minsz)
1296                        return -EINVAL;
1297
1298                if (ioeventfd.flags & ~VFIO_DEVICE_IOEVENTFD_SIZE_MASK)
1299                        return -EINVAL;
1300
1301                count = ioeventfd.flags & VFIO_DEVICE_IOEVENTFD_SIZE_MASK;
1302
1303                if (hweight8(count) != 1 || ioeventfd.fd < -1)
1304                        return -EINVAL;
1305
1306                return vfio_pci_ioeventfd(vdev, ioeventfd.offset,
1307                                          ioeventfd.data, count, ioeventfd.fd);
1308        } else if (cmd == VFIO_DEVICE_FEATURE) {
1309                struct vfio_device_feature feature;
1310                uuid_t uuid;
1311
1312                minsz = offsetofend(struct vfio_device_feature, flags);
1313
1314                if (copy_from_user(&feature, (void __user *)arg, minsz))
1315                        return -EFAULT;
1316
1317                if (feature.argsz < minsz)
1318                        return -EINVAL;
1319
1320                /* Check unknown flags */
1321                if (feature.flags & ~(VFIO_DEVICE_FEATURE_MASK |
1322                                      VFIO_DEVICE_FEATURE_SET |
1323                                      VFIO_DEVICE_FEATURE_GET |
1324                                      VFIO_DEVICE_FEATURE_PROBE))
1325                        return -EINVAL;
1326
1327                /* GET & SET are mutually exclusive except with PROBE */
1328                if (!(feature.flags & VFIO_DEVICE_FEATURE_PROBE) &&
1329                    (feature.flags & VFIO_DEVICE_FEATURE_SET) &&
1330                    (feature.flags & VFIO_DEVICE_FEATURE_GET))
1331                        return -EINVAL;
1332
1333                switch (feature.flags & VFIO_DEVICE_FEATURE_MASK) {
1334                case VFIO_DEVICE_FEATURE_PCI_VF_TOKEN:
1335                        if (!vdev->vf_token)
1336                                return -ENOTTY;
1337
1338                        /*
1339                         * We do not support GET of the VF Token UUID as this
1340                         * could expose the token of the previous device user.
1341                         */
1342                        if (feature.flags & VFIO_DEVICE_FEATURE_GET)
1343                                return -EINVAL;
1344
1345                        if (feature.flags & VFIO_DEVICE_FEATURE_PROBE)
1346                                return 0;
1347
1348                        /* Don't SET unless told to do so */
1349                        if (!(feature.flags & VFIO_DEVICE_FEATURE_SET))
1350                                return -EINVAL;
1351
1352                        if (feature.argsz < minsz + sizeof(uuid))
1353                                return -EINVAL;
1354
1355                        if (copy_from_user(&uuid, (void __user *)(arg + minsz),
1356                                           sizeof(uuid)))
1357                                return -EFAULT;
1358
1359                        mutex_lock(&vdev->vf_token->lock);
1360                        uuid_copy(&vdev->vf_token->uuid, &uuid);
1361                        mutex_unlock(&vdev->vf_token->lock);
1362
1363                        return 0;
1364                default:
1365                        return -ENOTTY;
1366                }
1367        }
1368
1369        return -ENOTTY;
1370}
1371
1372static ssize_t vfio_pci_rw(struct vfio_pci_device *vdev, char __user *buf,
1373                           size_t count, loff_t *ppos, bool iswrite)
1374{
1375        unsigned int index = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
1376
1377        if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1378                return -EINVAL;
1379
1380        switch (index) {
1381        case VFIO_PCI_CONFIG_REGION_INDEX:
1382                return vfio_pci_config_rw(vdev, buf, count, ppos, iswrite);
1383
1384        case VFIO_PCI_ROM_REGION_INDEX:
1385                if (iswrite)
1386                        return -EINVAL;
1387                return vfio_pci_bar_rw(vdev, buf, count, ppos, false);
1388
1389        case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
1390                return vfio_pci_bar_rw(vdev, buf, count, ppos, iswrite);
1391
1392        case VFIO_PCI_VGA_REGION_INDEX:
1393                return vfio_pci_vga_rw(vdev, buf, count, ppos, iswrite);
1394        default:
1395                index -= VFIO_PCI_NUM_REGIONS;
1396                return vdev->region[index].ops->rw(vdev, buf,
1397                                                   count, ppos, iswrite);
1398        }
1399
1400        return -EINVAL;
1401}
1402
1403static ssize_t vfio_pci_read(struct vfio_device *core_vdev, char __user *buf,
1404                             size_t count, loff_t *ppos)
1405{
1406        struct vfio_pci_device *vdev =
1407                container_of(core_vdev, struct vfio_pci_device, vdev);
1408
1409        if (!count)
1410                return 0;
1411
1412        return vfio_pci_rw(vdev, buf, count, ppos, false);
1413}
1414
1415static ssize_t vfio_pci_write(struct vfio_device *core_vdev, const char __user *buf,
1416                              size_t count, loff_t *ppos)
1417{
1418        struct vfio_pci_device *vdev =
1419                container_of(core_vdev, struct vfio_pci_device, vdev);
1420
1421        if (!count)
1422                return 0;
1423
1424        return vfio_pci_rw(vdev, (char __user *)buf, count, ppos, true);
1425}
1426
1427/* Return 1 on zap and vma_lock acquired, 0 on contention (only with @try) */
1428static int vfio_pci_zap_and_vma_lock(struct vfio_pci_device *vdev, bool try)
1429{
1430        struct vfio_pci_mmap_vma *mmap_vma, *tmp;
1431
1432        /*
1433         * Lock ordering:
1434         * vma_lock is nested under mmap_lock for vm_ops callback paths.
1435         * The memory_lock semaphore is used by both code paths calling
1436         * into this function to zap vmas and the vm_ops.fault callback
1437         * to protect the memory enable state of the device.
1438         *
1439         * When zapping vmas we need to maintain the mmap_lock => vma_lock
1440         * ordering, which requires using vma_lock to walk vma_list to
1441         * acquire an mm, then dropping vma_lock to get the mmap_lock and
1442         * reacquiring vma_lock.  This logic is derived from similar
1443         * requirements in uverbs_user_mmap_disassociate().
1444         *
1445         * mmap_lock must always be the top-level lock when it is taken.
1446         * Therefore we can only hold the memory_lock write lock when
1447         * vma_list is empty, as we'd need to take mmap_lock to clear
1448         * entries.  vma_list can only be guaranteed empty when holding
1449         * vma_lock, thus memory_lock is nested under vma_lock.
1450         *
1451         * This enables the vm_ops.fault callback to acquire vma_lock,
1452         * followed by memory_lock read lock, while already holding
1453         * mmap_lock without risk of deadlock.
1454         */
1455        while (1) {
1456                struct mm_struct *mm = NULL;
1457
1458                if (try) {
1459                        if (!mutex_trylock(&vdev->vma_lock))
1460                                return 0;
1461                } else {
1462                        mutex_lock(&vdev->vma_lock);
1463                }
1464                while (!list_empty(&vdev->vma_list)) {
1465                        mmap_vma = list_first_entry(&vdev->vma_list,
1466                                                    struct vfio_pci_mmap_vma,
1467                                                    vma_next);
1468                        mm = mmap_vma->vma->vm_mm;
1469                        if (mmget_not_zero(mm))
1470                                break;
1471
1472                        list_del(&mmap_vma->vma_next);
1473                        kfree(mmap_vma);
1474                        mm = NULL;
1475                }
1476                if (!mm)
1477                        return 1;
1478                mutex_unlock(&vdev->vma_lock);
1479
1480                if (try) {
1481                        if (!mmap_read_trylock(mm)) {
1482                                mmput(mm);
1483                                return 0;
1484                        }
1485                } else {
1486                        mmap_read_lock(mm);
1487                }
1488                if (try) {
1489                        if (!mutex_trylock(&vdev->vma_lock)) {
1490                                mmap_read_unlock(mm);
1491                                mmput(mm);
1492                                return 0;
1493                        }
1494                } else {
1495                        mutex_lock(&vdev->vma_lock);
1496                }
1497                list_for_each_entry_safe(mmap_vma, tmp,
1498                                         &vdev->vma_list, vma_next) {
1499                        struct vm_area_struct *vma = mmap_vma->vma;
1500
1501                        if (vma->vm_mm != mm)
1502                                continue;
1503
1504                        list_del(&mmap_vma->vma_next);
1505                        kfree(mmap_vma);
1506
1507                        zap_vma_ptes(vma, vma->vm_start,
1508                                     vma->vm_end - vma->vm_start);
1509                }
1510                mutex_unlock(&vdev->vma_lock);
1511                mmap_read_unlock(mm);
1512                mmput(mm);
1513        }
1514}
1515
1516void vfio_pci_zap_and_down_write_memory_lock(struct vfio_pci_device *vdev)
1517{
1518        vfio_pci_zap_and_vma_lock(vdev, false);
1519        down_write(&vdev->memory_lock);
1520        mutex_unlock(&vdev->vma_lock);
1521}
1522
1523u16 vfio_pci_memory_lock_and_enable(struct vfio_pci_device *vdev)
1524{
1525        u16 cmd;
1526
1527        down_write(&vdev->memory_lock);
1528        pci_read_config_word(vdev->pdev, PCI_COMMAND, &cmd);
1529        if (!(cmd & PCI_COMMAND_MEMORY))
1530                pci_write_config_word(vdev->pdev, PCI_COMMAND,
1531                                      cmd | PCI_COMMAND_MEMORY);
1532
1533        return cmd;
1534}
1535
1536void vfio_pci_memory_unlock_and_restore(struct vfio_pci_device *vdev, u16 cmd)
1537{
1538        pci_write_config_word(vdev->pdev, PCI_COMMAND, cmd);
1539        up_write(&vdev->memory_lock);
1540}
1541
1542/* Caller holds vma_lock */
1543static int __vfio_pci_add_vma(struct vfio_pci_device *vdev,
1544                              struct vm_area_struct *vma)
1545{
1546        struct vfio_pci_mmap_vma *mmap_vma;
1547
1548        mmap_vma = kmalloc(sizeof(*mmap_vma), GFP_KERNEL);
1549        if (!mmap_vma)
1550                return -ENOMEM;
1551
1552        mmap_vma->vma = vma;
1553        list_add(&mmap_vma->vma_next, &vdev->vma_list);
1554
1555        return 0;
1556}
1557
1558/*
1559 * Zap mmaps on open so that we can fault them in on access and therefore
1560 * our vma_list only tracks mappings accessed since last zap.
1561 */
1562static void vfio_pci_mmap_open(struct vm_area_struct *vma)
1563{
1564        zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1565}
1566
1567static void vfio_pci_mmap_close(struct vm_area_struct *vma)
1568{
1569        struct vfio_pci_device *vdev = vma->vm_private_data;
1570        struct vfio_pci_mmap_vma *mmap_vma;
1571
1572        mutex_lock(&vdev->vma_lock);
1573        list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1574                if (mmap_vma->vma == vma) {
1575                        list_del(&mmap_vma->vma_next);
1576                        kfree(mmap_vma);
1577                        break;
1578                }
1579        }
1580        mutex_unlock(&vdev->vma_lock);
1581}
1582
1583static vm_fault_t vfio_pci_mmap_fault(struct vm_fault *vmf)
1584{
1585        struct vm_area_struct *vma = vmf->vma;
1586        struct vfio_pci_device *vdev = vma->vm_private_data;
1587        struct vfio_pci_mmap_vma *mmap_vma;
1588        vm_fault_t ret = VM_FAULT_NOPAGE;
1589
1590        mutex_lock(&vdev->vma_lock);
1591        down_read(&vdev->memory_lock);
1592
1593        if (!__vfio_pci_memory_enabled(vdev)) {
1594                ret = VM_FAULT_SIGBUS;
1595                goto up_out;
1596        }
1597
1598        /*
1599         * We populate the whole vma on fault, so we need to test whether
1600         * the vma has already been mapped, such as for concurrent faults
1601         * to the same vma.  io_remap_pfn_range() will trigger a BUG_ON if
1602         * we ask it to fill the same range again.
1603         */
1604        list_for_each_entry(mmap_vma, &vdev->vma_list, vma_next) {
1605                if (mmap_vma->vma == vma)
1606                        goto up_out;
1607        }
1608
1609        if (io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
1610                               vma->vm_end - vma->vm_start,
1611                               vma->vm_page_prot)) {
1612                ret = VM_FAULT_SIGBUS;
1613                zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1614                goto up_out;
1615        }
1616
1617        if (__vfio_pci_add_vma(vdev, vma)) {
1618                ret = VM_FAULT_OOM;
1619                zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
1620        }
1621
1622up_out:
1623        up_read(&vdev->memory_lock);
1624        mutex_unlock(&vdev->vma_lock);
1625        return ret;
1626}
1627
1628static const struct vm_operations_struct vfio_pci_mmap_ops = {
1629        .open = vfio_pci_mmap_open,
1630        .close = vfio_pci_mmap_close,
1631        .fault = vfio_pci_mmap_fault,
1632};
1633
1634static int vfio_pci_mmap(struct vfio_device *core_vdev, struct vm_area_struct *vma)
1635{
1636        struct vfio_pci_device *vdev =
1637                container_of(core_vdev, struct vfio_pci_device, vdev);
1638        struct pci_dev *pdev = vdev->pdev;
1639        unsigned int index;
1640        u64 phys_len, req_len, pgoff, req_start;
1641        int ret;
1642
1643        index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
1644
1645        if (index >= VFIO_PCI_NUM_REGIONS + vdev->num_regions)
1646                return -EINVAL;
1647        if (vma->vm_end < vma->vm_start)
1648                return -EINVAL;
1649        if ((vma->vm_flags & VM_SHARED) == 0)
1650                return -EINVAL;
1651        if (index >= VFIO_PCI_NUM_REGIONS) {
1652                int regnum = index - VFIO_PCI_NUM_REGIONS;
1653                struct vfio_pci_region *region = vdev->region + regnum;
1654
1655                if (region->ops && region->ops->mmap &&
1656                    (region->flags & VFIO_REGION_INFO_FLAG_MMAP))
1657                        return region->ops->mmap(vdev, region, vma);
1658                return -EINVAL;
1659        }
1660        if (index >= VFIO_PCI_ROM_REGION_INDEX)
1661                return -EINVAL;
1662        if (!vdev->bar_mmap_supported[index])
1663                return -EINVAL;
1664
1665        phys_len = PAGE_ALIGN(pci_resource_len(pdev, index));
1666        req_len = vma->vm_end - vma->vm_start;
1667        pgoff = vma->vm_pgoff &
1668                ((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
1669        req_start = pgoff << PAGE_SHIFT;
1670
1671        if (req_start + req_len > phys_len)
1672                return -EINVAL;
1673
1674        /*
1675         * Even though we don't make use of the barmap for the mmap,
1676         * we need to request the region and the barmap tracks that.
1677         */
1678        if (!vdev->barmap[index]) {
1679                ret = pci_request_selected_regions(pdev,
1680                                                   1 << index, "vfio-pci");
1681                if (ret)
1682                        return ret;
1683
1684                vdev->barmap[index] = pci_iomap(pdev, index, 0);
1685                if (!vdev->barmap[index]) {
1686                        pci_release_selected_regions(pdev, 1 << index);
1687                        return -ENOMEM;
1688                }
1689        }
1690
1691        vma->vm_private_data = vdev;
1692        vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1693        vma->vm_pgoff = (pci_resource_start(pdev, index) >> PAGE_SHIFT) + pgoff;
1694
1695        /*
1696         * See remap_pfn_range(), called from vfio_pci_fault() but we can't
1697         * change vm_flags within the fault handler.  Set them now.
1698         */
1699        vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1700        vma->vm_ops = &vfio_pci_mmap_ops;
1701
1702        return 0;
1703}
1704
1705static void vfio_pci_request(struct vfio_device *core_vdev, unsigned int count)
1706{
1707        struct vfio_pci_device *vdev =
1708                container_of(core_vdev, struct vfio_pci_device, vdev);
1709        struct pci_dev *pdev = vdev->pdev;
1710
1711        mutex_lock(&vdev->igate);
1712
1713        if (vdev->req_trigger) {
1714                if (!(count % 10))
1715                        pci_notice_ratelimited(pdev,
1716                                "Relaying device request to user (#%u)\n",
1717                                count);
1718                eventfd_signal(vdev->req_trigger, 1);
1719        } else if (count == 0) {
1720                pci_warn(pdev,
1721                        "No device request channel registered, blocked until released by user\n");
1722        }
1723
1724        mutex_unlock(&vdev->igate);
1725}
1726
1727static int vfio_pci_validate_vf_token(struct vfio_pci_device *vdev,
1728                                      bool vf_token, uuid_t *uuid)
1729{
1730        /*
1731         * There's always some degree of trust or collaboration between SR-IOV
1732         * PF and VFs, even if just that the PF hosts the SR-IOV capability and
1733         * can disrupt VFs with a reset, but often the PF has more explicit
1734         * access to deny service to the VF or access data passed through the
1735         * VF.  We therefore require an opt-in via a shared VF token (UUID) to
1736         * represent this trust.  This both prevents that a VF driver might
1737         * assume the PF driver is a trusted, in-kernel driver, and also that
1738         * a PF driver might be replaced with a rogue driver, unknown to in-use
1739         * VF drivers.
1740         *
1741         * Therefore when presented with a VF, if the PF is a vfio device and
1742         * it is bound to the vfio-pci driver, the user needs to provide a VF
1743         * token to access the device, in the form of appending a vf_token to
1744         * the device name, for example:
1745         *
1746         * "0000:04:10.0 vf_token=bd8d9d2b-5a5f-4f5a-a211-f591514ba1f3"
1747         *
1748         * When presented with a PF which has VFs in use, the user must also
1749         * provide the current VF token to prove collaboration with existing
1750         * VF users.  If VFs are not in use, the VF token provided for the PF
1751         * device will act to set the VF token.
1752         *
1753         * If the VF token is provided but unused, an error is generated.
1754         */
1755        if (!vdev->pdev->is_virtfn && !vdev->vf_token && !vf_token)
1756                return 0; /* No VF token provided or required */
1757
1758        if (vdev->pdev->is_virtfn) {
1759                struct vfio_pci_device *pf_vdev = get_pf_vdev(vdev);
1760                bool match;
1761
1762                if (!pf_vdev) {
1763                        if (!vf_token)
1764                                return 0; /* PF is not vfio-pci, no VF token */
1765
1766                        pci_info_ratelimited(vdev->pdev,
1767                                "VF token incorrectly provided, PF not bound to vfio-pci\n");
1768                        return -EINVAL;
1769                }
1770
1771                if (!vf_token) {
1772                        vfio_device_put(&pf_vdev->vdev);
1773                        pci_info_ratelimited(vdev->pdev,
1774                                "VF token required to access device\n");
1775                        return -EACCES;
1776                }
1777
1778                mutex_lock(&pf_vdev->vf_token->lock);
1779                match = uuid_equal(uuid, &pf_vdev->vf_token->uuid);
1780                mutex_unlock(&pf_vdev->vf_token->lock);
1781
1782                vfio_device_put(&pf_vdev->vdev);
1783
1784                if (!match) {
1785                        pci_info_ratelimited(vdev->pdev,
1786                                "Incorrect VF token provided for device\n");
1787                        return -EACCES;
1788                }
1789        } else if (vdev->vf_token) {
1790                mutex_lock(&vdev->vf_token->lock);
1791                if (vdev->vf_token->users) {
1792                        if (!vf_token) {
1793                                mutex_unlock(&vdev->vf_token->lock);
1794                                pci_info_ratelimited(vdev->pdev,
1795                                        "VF token required to access device\n");
1796                                return -EACCES;
1797                        }
1798
1799                        if (!uuid_equal(uuid, &vdev->vf_token->uuid)) {
1800                                mutex_unlock(&vdev->vf_token->lock);
1801                                pci_info_ratelimited(vdev->pdev,
1802                                        "Incorrect VF token provided for device\n");
1803                                return -EACCES;
1804                        }
1805                } else if (vf_token) {
1806                        uuid_copy(&vdev->vf_token->uuid, uuid);
1807                }
1808
1809                mutex_unlock(&vdev->vf_token->lock);
1810        } else if (vf_token) {
1811                pci_info_ratelimited(vdev->pdev,
1812                        "VF token incorrectly provided, not a PF or VF\n");
1813                return -EINVAL;
1814        }
1815
1816        return 0;
1817}
1818
1819#define VF_TOKEN_ARG "vf_token="
1820
1821static int vfio_pci_match(struct vfio_device *core_vdev, char *buf)
1822{
1823        struct vfio_pci_device *vdev =
1824                container_of(core_vdev, struct vfio_pci_device, vdev);
1825        bool vf_token = false;
1826        uuid_t uuid;
1827        int ret;
1828
1829        if (strncmp(pci_name(vdev->pdev), buf, strlen(pci_name(vdev->pdev))))
1830                return 0; /* No match */
1831
1832        if (strlen(buf) > strlen(pci_name(vdev->pdev))) {
1833                buf += strlen(pci_name(vdev->pdev));
1834
1835                if (*buf != ' ')
1836                        return 0; /* No match: non-whitespace after name */
1837
1838                while (*buf) {
1839                        if (*buf == ' ') {
1840                                buf++;
1841                                continue;
1842                        }
1843
1844                        if (!vf_token && !strncmp(buf, VF_TOKEN_ARG,
1845                                                  strlen(VF_TOKEN_ARG))) {
1846                                buf += strlen(VF_TOKEN_ARG);
1847
1848                                if (strlen(buf) < UUID_STRING_LEN)
1849                                        return -EINVAL;
1850
1851                                ret = uuid_parse(buf, &uuid);
1852                                if (ret)
1853                                        return ret;
1854
1855                                vf_token = true;
1856                                buf += UUID_STRING_LEN;
1857                        } else {
1858                                /* Unknown/duplicate option */
1859                                return -EINVAL;
1860                        }
1861                }
1862        }
1863
1864        ret = vfio_pci_validate_vf_token(vdev, vf_token, &uuid);
1865        if (ret)
1866                return ret;
1867
1868        return 1; /* Match */
1869}
1870
1871static const struct vfio_device_ops vfio_pci_ops = {
1872        .name           = "vfio-pci",
1873        .open           = vfio_pci_open,
1874        .release        = vfio_pci_release,
1875        .ioctl          = vfio_pci_ioctl,
1876        .read           = vfio_pci_read,
1877        .write          = vfio_pci_write,
1878        .mmap           = vfio_pci_mmap,
1879        .request        = vfio_pci_request,
1880        .match          = vfio_pci_match,
1881};
1882
1883static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev);
1884static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck);
1885
1886static int vfio_pci_bus_notifier(struct notifier_block *nb,
1887                                 unsigned long action, void *data)
1888{
1889        struct vfio_pci_device *vdev = container_of(nb,
1890                                                    struct vfio_pci_device, nb);
1891        struct device *dev = data;
1892        struct pci_dev *pdev = to_pci_dev(dev);
1893        struct pci_dev *physfn = pci_physfn(pdev);
1894
1895        if (action == BUS_NOTIFY_ADD_DEVICE &&
1896            pdev->is_virtfn && physfn == vdev->pdev) {
1897                pci_info(vdev->pdev, "Captured SR-IOV VF %s driver_override\n",
1898                         pci_name(pdev));
1899                pdev->driver_override = kasprintf(GFP_KERNEL, "%s",
1900                                                  vfio_pci_ops.name);
1901        } else if (action == BUS_NOTIFY_BOUND_DRIVER &&
1902                   pdev->is_virtfn && physfn == vdev->pdev) {
1903                struct pci_driver *drv = pci_dev_driver(pdev);
1904
1905                if (drv && drv != &vfio_pci_driver)
1906                        pci_warn(vdev->pdev,
1907                                 "VF %s bound to driver %s while PF bound to vfio-pci\n",
1908                                 pci_name(pdev), drv->name);
1909        }
1910
1911        return 0;
1912}
1913
1914static int vfio_pci_vf_init(struct vfio_pci_device *vdev)
1915{
1916        struct pci_dev *pdev = vdev->pdev;
1917        int ret;
1918
1919        if (!pdev->is_physfn)
1920                return 0;
1921
1922        vdev->vf_token = kzalloc(sizeof(*vdev->vf_token), GFP_KERNEL);
1923        if (!vdev->vf_token)
1924                return -ENOMEM;
1925
1926        mutex_init(&vdev->vf_token->lock);
1927        uuid_gen(&vdev->vf_token->uuid);
1928
1929        vdev->nb.notifier_call = vfio_pci_bus_notifier;
1930        ret = bus_register_notifier(&pci_bus_type, &vdev->nb);
1931        if (ret) {
1932                kfree(vdev->vf_token);
1933                return ret;
1934        }
1935        return 0;
1936}
1937
1938static void vfio_pci_vf_uninit(struct vfio_pci_device *vdev)
1939{
1940        if (!vdev->vf_token)
1941                return;
1942
1943        bus_unregister_notifier(&pci_bus_type, &vdev->nb);
1944        WARN_ON(vdev->vf_token->users);
1945        mutex_destroy(&vdev->vf_token->lock);
1946        kfree(vdev->vf_token);
1947}
1948
1949static int vfio_pci_vga_init(struct vfio_pci_device *vdev)
1950{
1951        struct pci_dev *pdev = vdev->pdev;
1952        int ret;
1953
1954        if (!vfio_pci_is_vga(pdev))
1955                return 0;
1956
1957        ret = vga_client_register(pdev, vdev, NULL, vfio_pci_set_vga_decode);
1958        if (ret)
1959                return ret;
1960        vga_set_legacy_decoding(pdev, vfio_pci_set_vga_decode(vdev, false));
1961        return 0;
1962}
1963
1964static void vfio_pci_vga_uninit(struct vfio_pci_device *vdev)
1965{
1966        struct pci_dev *pdev = vdev->pdev;
1967
1968        if (!vfio_pci_is_vga(pdev))
1969                return;
1970        vga_client_register(pdev, NULL, NULL, NULL);
1971        vga_set_legacy_decoding(pdev, VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM |
1972                                              VGA_RSRC_LEGACY_IO |
1973                                              VGA_RSRC_LEGACY_MEM);
1974}
1975
1976static int vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1977{
1978        struct vfio_pci_device *vdev;
1979        struct iommu_group *group;
1980        int ret;
1981
1982        if (vfio_pci_is_denylisted(pdev))
1983                return -EINVAL;
1984
1985        if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1986                return -EINVAL;
1987
1988        /*
1989         * Prevent binding to PFs with VFs enabled, the VFs might be in use
1990         * by the host or other users.  We cannot capture the VFs if they
1991         * already exist, nor can we track VF users.  Disabling SR-IOV here
1992         * would initiate removing the VFs, which would unbind the driver,
1993         * which is prone to blocking if that VF is also in use by vfio-pci.
1994         * Just reject these PFs and let the user sort it out.
1995         */
1996        if (pci_num_vf(pdev)) {
1997                pci_warn(pdev, "Cannot bind to PF with SR-IOV enabled\n");
1998                return -EBUSY;
1999        }
2000
2001        group = vfio_iommu_group_get(&pdev->dev);
2002        if (!group)
2003                return -EINVAL;
2004
2005        vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
2006        if (!vdev) {
2007                ret = -ENOMEM;
2008                goto out_group_put;
2009        }
2010
2011        vfio_init_group_dev(&vdev->vdev, &pdev->dev, &vfio_pci_ops);
2012        vdev->pdev = pdev;
2013        vdev->irq_type = VFIO_PCI_NUM_IRQS;
2014        mutex_init(&vdev->igate);
2015        spin_lock_init(&vdev->irqlock);
2016        mutex_init(&vdev->ioeventfds_lock);
2017        INIT_LIST_HEAD(&vdev->dummy_resources_list);
2018        INIT_LIST_HEAD(&vdev->ioeventfds_list);
2019        mutex_init(&vdev->vma_lock);
2020        INIT_LIST_HEAD(&vdev->vma_list);
2021        init_rwsem(&vdev->memory_lock);
2022
2023        ret = vfio_pci_reflck_attach(vdev);
2024        if (ret)
2025                goto out_free;
2026        ret = vfio_pci_vf_init(vdev);
2027        if (ret)
2028                goto out_reflck;
2029        ret = vfio_pci_vga_init(vdev);
2030        if (ret)
2031                goto out_vf;
2032
2033        vfio_pci_probe_power_state(vdev);
2034
2035        if (!disable_idle_d3) {
2036                /*
2037                 * pci-core sets the device power state to an unknown value at
2038                 * bootup and after being removed from a driver.  The only
2039                 * transition it allows from this unknown state is to D0, which
2040                 * typically happens when a driver calls pci_enable_device().
2041                 * We're not ready to enable the device yet, but we do want to
2042                 * be able to get to D3.  Therefore first do a D0 transition
2043                 * before going to D3.
2044                 */
2045                vfio_pci_set_power_state(vdev, PCI_D0);
2046                vfio_pci_set_power_state(vdev, PCI_D3hot);
2047        }
2048
2049        ret = vfio_register_group_dev(&vdev->vdev);
2050        if (ret)
2051                goto out_power;
2052        dev_set_drvdata(&pdev->dev, vdev);
2053        return 0;
2054
2055out_power:
2056        if (!disable_idle_d3)
2057                vfio_pci_set_power_state(vdev, PCI_D0);
2058out_vf:
2059        vfio_pci_vf_uninit(vdev);
2060out_reflck:
2061        vfio_pci_reflck_put(vdev->reflck);
2062out_free:
2063        kfree(vdev->pm_save);
2064        kfree(vdev);
2065out_group_put:
2066        vfio_iommu_group_put(group, &pdev->dev);
2067        return ret;
2068}
2069
2070static void vfio_pci_remove(struct pci_dev *pdev)
2071{
2072        struct vfio_pci_device *vdev = dev_get_drvdata(&pdev->dev);
2073
2074        pci_disable_sriov(pdev);
2075
2076        vfio_unregister_group_dev(&vdev->vdev);
2077
2078        vfio_pci_vf_uninit(vdev);
2079        vfio_pci_reflck_put(vdev->reflck);
2080        vfio_pci_vga_uninit(vdev);
2081
2082        vfio_iommu_group_put(pdev->dev.iommu_group, &pdev->dev);
2083
2084        if (!disable_idle_d3)
2085                vfio_pci_set_power_state(vdev, PCI_D0);
2086
2087        mutex_destroy(&vdev->ioeventfds_lock);
2088        kfree(vdev->region);
2089        kfree(vdev->pm_save);
2090        kfree(vdev);
2091}
2092
2093static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
2094                                                  pci_channel_state_t state)
2095{
2096        struct vfio_pci_device *vdev;
2097        struct vfio_device *device;
2098
2099        device = vfio_device_get_from_dev(&pdev->dev);
2100        if (device == NULL)
2101                return PCI_ERS_RESULT_DISCONNECT;
2102
2103        vdev = container_of(device, struct vfio_pci_device, vdev);
2104
2105        mutex_lock(&vdev->igate);
2106
2107        if (vdev->err_trigger)
2108                eventfd_signal(vdev->err_trigger, 1);
2109
2110        mutex_unlock(&vdev->igate);
2111
2112        vfio_device_put(device);
2113
2114        return PCI_ERS_RESULT_CAN_RECOVER;
2115}
2116
2117static int vfio_pci_sriov_configure(struct pci_dev *pdev, int nr_virtfn)
2118{
2119        struct vfio_device *device;
2120        int ret = 0;
2121
2122        might_sleep();
2123
2124        if (!enable_sriov)
2125                return -ENOENT;
2126
2127        device = vfio_device_get_from_dev(&pdev->dev);
2128        if (!device)
2129                return -ENODEV;
2130
2131        if (nr_virtfn == 0)
2132                pci_disable_sriov(pdev);
2133        else
2134                ret = pci_enable_sriov(pdev, nr_virtfn);
2135
2136        vfio_device_put(device);
2137
2138        return ret < 0 ? ret : nr_virtfn;
2139}
2140
2141static const struct pci_error_handlers vfio_err_handlers = {
2142        .error_detected = vfio_pci_aer_err_detected,
2143};
2144
2145static struct pci_driver vfio_pci_driver = {
2146        .name                   = "vfio-pci",
2147        .id_table               = NULL, /* only dynamic ids */
2148        .probe                  = vfio_pci_probe,
2149        .remove                 = vfio_pci_remove,
2150        .sriov_configure        = vfio_pci_sriov_configure,
2151        .err_handler            = &vfio_err_handlers,
2152};
2153
2154static DEFINE_MUTEX(reflck_lock);
2155
2156static struct vfio_pci_reflck *vfio_pci_reflck_alloc(void)
2157{
2158        struct vfio_pci_reflck *reflck;
2159
2160        reflck = kzalloc(sizeof(*reflck), GFP_KERNEL);
2161        if (!reflck)
2162                return ERR_PTR(-ENOMEM);
2163
2164        kref_init(&reflck->kref);
2165        mutex_init(&reflck->lock);
2166
2167        return reflck;
2168}
2169
2170static void vfio_pci_reflck_get(struct vfio_pci_reflck *reflck)
2171{
2172        kref_get(&reflck->kref);
2173}
2174
2175static int vfio_pci_reflck_find(struct pci_dev *pdev, void *data)
2176{
2177        struct vfio_pci_reflck **preflck = data;
2178        struct vfio_device *device;
2179        struct vfio_pci_device *vdev;
2180
2181        device = vfio_device_get_from_dev(&pdev->dev);
2182        if (!device)
2183                return 0;
2184
2185        if (pci_dev_driver(pdev) != &vfio_pci_driver) {
2186                vfio_device_put(device);
2187                return 0;
2188        }
2189
2190        vdev = container_of(device, struct vfio_pci_device, vdev);
2191
2192        if (vdev->reflck) {
2193                vfio_pci_reflck_get(vdev->reflck);
2194                *preflck = vdev->reflck;
2195                vfio_device_put(device);
2196                return 1;
2197        }
2198
2199        vfio_device_put(device);
2200        return 0;
2201}
2202
2203static int vfio_pci_reflck_attach(struct vfio_pci_device *vdev)
2204{
2205        bool slot = !pci_probe_reset_slot(vdev->pdev->slot);
2206
2207        mutex_lock(&reflck_lock);
2208
2209        if (pci_is_root_bus(vdev->pdev->bus) ||
2210            vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_reflck_find,
2211                                          &vdev->reflck, slot) <= 0)
2212                vdev->reflck = vfio_pci_reflck_alloc();
2213
2214        mutex_unlock(&reflck_lock);
2215
2216        return PTR_ERR_OR_ZERO(vdev->reflck);
2217}
2218
2219static void vfio_pci_reflck_release(struct kref *kref)
2220{
2221        struct vfio_pci_reflck *reflck = container_of(kref,
2222                                                      struct vfio_pci_reflck,
2223                                                      kref);
2224
2225        kfree(reflck);
2226        mutex_unlock(&reflck_lock);
2227}
2228
2229static void vfio_pci_reflck_put(struct vfio_pci_reflck *reflck)
2230{
2231        kref_put_mutex(&reflck->kref, vfio_pci_reflck_release, &reflck_lock);
2232}
2233
2234static int vfio_pci_get_unused_devs(struct pci_dev *pdev, void *data)
2235{
2236        struct vfio_devices *devs = data;
2237        struct vfio_device *device;
2238        struct vfio_pci_device *vdev;
2239
2240        if (devs->cur_index == devs->max_index)
2241                return -ENOSPC;
2242
2243        device = vfio_device_get_from_dev(&pdev->dev);
2244        if (!device)
2245                return -EINVAL;
2246
2247        if (pci_dev_driver(pdev) != &vfio_pci_driver) {
2248                vfio_device_put(device);
2249                return -EBUSY;
2250        }
2251
2252        vdev = container_of(device, struct vfio_pci_device, vdev);
2253
2254        /* Fault if the device is not unused */
2255        if (vdev->refcnt) {
2256                vfio_device_put(device);
2257                return -EBUSY;
2258        }
2259
2260        devs->devices[devs->cur_index++] = vdev;
2261        return 0;
2262}
2263
2264static int vfio_pci_try_zap_and_vma_lock_cb(struct pci_dev *pdev, void *data)
2265{
2266        struct vfio_devices *devs = data;
2267        struct vfio_device *device;
2268        struct vfio_pci_device *vdev;
2269
2270        if (devs->cur_index == devs->max_index)
2271                return -ENOSPC;
2272
2273        device = vfio_device_get_from_dev(&pdev->dev);
2274        if (!device)
2275                return -EINVAL;
2276
2277        if (pci_dev_driver(pdev) != &vfio_pci_driver) {
2278                vfio_device_put(device);
2279                return -EBUSY;
2280        }
2281
2282        vdev = container_of(device, struct vfio_pci_device, vdev);
2283
2284        /*
2285         * Locking multiple devices is prone to deadlock, runaway and
2286         * unwind if we hit contention.
2287         */
2288        if (!vfio_pci_zap_and_vma_lock(vdev, true)) {
2289                vfio_device_put(device);
2290                return -EBUSY;
2291        }
2292
2293        devs->devices[devs->cur_index++] = vdev;
2294        return 0;
2295}
2296
2297/*
2298 * If a bus or slot reset is available for the provided device and:
2299 *  - All of the devices affected by that bus or slot reset are unused
2300 *    (!refcnt)
2301 *  - At least one of the affected devices is marked dirty via
2302 *    needs_reset (such as by lack of FLR support)
2303 * Then attempt to perform that bus or slot reset.  Callers are required
2304 * to hold vdev->reflck->lock, protecting the bus/slot reset group from
2305 * concurrent opens.  A vfio_device reference is acquired for each device
2306 * to prevent unbinds during the reset operation.
2307 *
2308 * NB: vfio-core considers a group to be viable even if some devices are
2309 * bound to drivers like pci-stub or pcieport.  Here we require all devices
2310 * to be bound to vfio_pci since that's the only way we can be sure they
2311 * stay put.
2312 */
2313static void vfio_pci_try_bus_reset(struct vfio_pci_device *vdev)
2314{
2315        struct vfio_devices devs = { .cur_index = 0 };
2316        int i = 0, ret = -EINVAL;
2317        bool slot = false;
2318        struct vfio_pci_device *tmp;
2319
2320        if (!pci_probe_reset_slot(vdev->pdev->slot))
2321                slot = true;
2322        else if (pci_probe_reset_bus(vdev->pdev->bus))
2323                return;
2324
2325        if (vfio_pci_for_each_slot_or_bus(vdev->pdev, vfio_pci_count_devs,
2326                                          &i, slot) || !i)
2327                return;
2328
2329        devs.max_index = i;
2330        devs.devices = kcalloc(i, sizeof(struct vfio_device *), GFP_KERNEL);
2331        if (!devs.devices)
2332                return;
2333
2334        if (vfio_pci_for_each_slot_or_bus(vdev->pdev,
2335                                          vfio_pci_get_unused_devs,
2336                                          &devs, slot))
2337                goto put_devs;
2338
2339        /* Does at least one need a reset? */
2340        for (i = 0; i < devs.cur_index; i++) {
2341                tmp = devs.devices[i];
2342                if (tmp->needs_reset) {
2343                        ret = pci_reset_bus(vdev->pdev);
2344                        break;
2345                }
2346        }
2347
2348put_devs:
2349        for (i = 0; i < devs.cur_index; i++) {
2350                tmp = devs.devices[i];
2351
2352                /*
2353                 * If reset was successful, affected devices no longer need
2354                 * a reset and we should return all the collateral devices
2355                 * to low power.  If not successful, we either didn't reset
2356                 * the bus or timed out waiting for it, so let's not touch
2357                 * the power state.
2358                 */
2359                if (!ret) {
2360                        tmp->needs_reset = false;
2361
2362                        if (tmp != vdev && !disable_idle_d3)
2363                                vfio_pci_set_power_state(tmp, PCI_D3hot);
2364                }
2365
2366                vfio_device_put(&tmp->vdev);
2367        }
2368
2369        kfree(devs.devices);
2370}
2371
2372static void __exit vfio_pci_cleanup(void)
2373{
2374        pci_unregister_driver(&vfio_pci_driver);
2375        vfio_pci_uninit_perm_bits();
2376}
2377
2378static void __init vfio_pci_fill_ids(void)
2379{
2380        char *p, *id;
2381        int rc;
2382
2383        /* no ids passed actually */
2384        if (ids[0] == '\0')
2385                return;
2386
2387        /* add ids specified in the module parameter */
2388        p = ids;
2389        while ((id = strsep(&p, ","))) {
2390                unsigned int vendor, device, subvendor = PCI_ANY_ID,
2391                        subdevice = PCI_ANY_ID, class = 0, class_mask = 0;
2392                int fields;
2393
2394                if (!strlen(id))
2395                        continue;
2396
2397                fields = sscanf(id, "%x:%x:%x:%x:%x:%x",
2398                                &vendor, &device, &subvendor, &subdevice,
2399                                &class, &class_mask);
2400
2401                if (fields < 2) {
2402                        pr_warn("invalid id string \"%s\"\n", id);
2403                        continue;
2404                }
2405
2406                rc = pci_add_dynid(&vfio_pci_driver, vendor, device,
2407                                   subvendor, subdevice, class, class_mask, 0);
2408                if (rc)
2409                        pr_warn("failed to add dynamic id [%04x:%04x[%04x:%04x]] class %#08x/%08x (%d)\n",
2410                                vendor, device, subvendor, subdevice,
2411                                class, class_mask, rc);
2412                else
2413                        pr_info("add [%04x:%04x[%04x:%04x]] class %#08x/%08x\n",
2414                                vendor, device, subvendor, subdevice,
2415                                class, class_mask);
2416        }
2417}
2418
2419static int __init vfio_pci_init(void)
2420{
2421        int ret;
2422
2423        /* Allocate shared config space permission data used by all devices */
2424        ret = vfio_pci_init_perm_bits();
2425        if (ret)
2426                return ret;
2427
2428        /* Register and scan for devices */
2429        ret = pci_register_driver(&vfio_pci_driver);
2430        if (ret)
2431                goto out_driver;
2432
2433        vfio_pci_fill_ids();
2434
2435        if (disable_denylist)
2436                pr_warn("device denylist disabled.\n");
2437
2438        return 0;
2439
2440out_driver:
2441        vfio_pci_uninit_perm_bits();
2442        return ret;
2443}
2444
2445module_init(vfio_pci_init);
2446module_exit(vfio_pci_cleanup);
2447
2448MODULE_VERSION(DRIVER_VERSION);
2449MODULE_LICENSE("GPL v2");
2450MODULE_AUTHOR(DRIVER_AUTHOR);
2451MODULE_DESCRIPTION(DRIVER_DESC);
2452